Callbacks¶

Run
Metric

Run ¶

BatchOverfitCallback ¶

class catalyst.callbacks.batch_overfit.BatchOverfitCallback(**kwargs)[source]¶

Bases: catalyst.core.callback.Callback

Callback to overfit loaders with specified number of batches. By default we use 1 batch for loader.

Parameters: kwargs – loader names and their number of batches to overfit.

For example, if you have train, train_additional, valid and valid_additional loaders and wan’t to overfit train on first 1 batch, train_additional on first 2 batches, valid - on first 20% of batches and valid_additional - on 50% batches:

from catalyst.dl import (
    SupervisedRunner, BatchOverfitCallback,
)
runner = SupervisedRunner()
runner.train(
    ...
    loaders={
        "train": ...,
        "train_additional": ...,
        "valid": ...,
        "valid_additional":...
    }
    ...
    callbacks=[
        ...
        BatchOverfitCallback(
            train_additional=2,
            valid=0.2,
            valid_additional=0.5
        ),
        ...
    ]
    ...
)

Minimal working example

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# data
num_samples, num_features = int(1e4), int(1e1)
X, y = torch.rand(num_samples, num_features), torch.rand(num_samples)
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, 1)
criterion = torch.nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [3, 6])

# model training
runner = dl.SupervisedRunner()
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    logdir="./logdir",
    num_epochs=8,
    verbose=True,
    callbacks=[dl.BatchOverfitCallback(train=10, valid=0.5)]
)

BatchTransformCallback ¶

class catalyst.callbacks.batch_transform.BatchTransformCallback(transform: Union[Callable, str], scope: str, input_key: Union[List[str], str] = None, output_key: Union[List[str], str] = None, transform_kwargs: Dict[str, Any] = None)[source]¶

Bases: catalyst.core.callback.Callback

Preprocess your batch with specified function.

Parameters

transform (Callable, str) – Function to apply. If string will get function from registry.
scope (str) – "on_batch_end" (post-processing model output) or "on_batch_start" (pre-processing model input).
input_key (Union[List[str], str], optional) – Keys in batch dict to apply function. Defaults to None.
output_key (Union[List[str], str], optional) – Keys for output. If None then will apply function inplace to keys_to_apply. Defaults to None.
transform_kwargs (Dict[str, Any]) – Kwargs for transform.

Raises

TypeError – When keys is not str or a list. When scope is not in ["on_batch_end", "on_batch_start"].

Examples

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_users, num_features, num_items = int(1e4), int(1e1), 10
X = torch.rand(num_users, num_features)
y = (torch.rand(num_users, num_items) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_items)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = SupervisedRunner()
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        dl.BatchTransformCallback(
            input_key="logits", output_key="scores", transform="F.sigmoid",
        ),
        dl.CriterionCallback(
            input_key="logits", target_key="targets", metric_key="loss"
        ),
# uncomment for extra metrics:
#       dl.AUCCallback(
#           input_key="scores", target_key="targets"
#       ),
#       dl.HitrateCallback(
#           input_key="scores", target_key="targets", topk_args=(1, 3, 5)
#       ),
#       dl.MRRCallback(
#           input_key="scores", target_key="targets", topk_args=(1, 3, 5)
#       ),
#       dl.MAPCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
#       dl.NDCGCallback(
#           input_key="scores", target_key="targets", topk_args=(1, 3, 5)
#       ),
        dl.OptimizerCallback(metric_key="loss"),
        dl.SchedulerCallback(),
        dl.CheckpointCallback(
            logdir="./logs", loader_key="valid", metric_key="map01", minimize=False
        ),
    ]
)

class CustomRunner(dl.Runner):

    def handle_batch(self, batch):
        logits = self.model(
            batch["features"].view(batch["features"].size(0), -1)
        )

        loss = F.cross_entropy(logits, batch["targets"])
        accuracy01, accuracy03 = metrics.accuracy(
            logits, batch["targets"], topk=(1, 3)
        )
        self.batch_metrics.update(
            {"loss": loss, "accuracy01": accuracy01, "accuracy03": accuracy03}
        )

        if self.is_train_loader:
            loss.backward()
            self.optimizer.step()
            self.optimizer.zero_grad()


class MnistDataset(torch.utils.data.Dataset):
    def __init__(self, dataset):
        self.dataset = dataset

    def __getitem__(self, item):
        return {"features": self.dataset[item][0], "targets": self.dataset[item][1]}

    def __len__(self):
        return len(self.dataset)

model = torch.nn.Linear(28 * 28, 10)
optimizer = torch.optim.Adam(model.parameters(), lr=0.02)

loaders = {
    "train": DataLoader(
        MnistDataset(
            MNIST(os.getcwd(), train=False, download=True, transform=ToTensor())
        ),
        batch_size=32,
    ),
    "valid": DataLoader(
        MnistDataset(
            MNIST(os.getcwd(), train=False, download=True, transform=ToTensor())
        ),
        batch_size=32,
    ),
}

transrorms = [
    augmentation.RandomAffine(degrees=(-15, 20), scale=(0.75, 1.25)),
]

runner = CustomRunner()

# model training
runner.train(
    model=model,
    optimizer=optimizer,
    loaders=loaders,
    logdir="./logs",
    num_epochs=5,
    verbose=False,
    load_best_on_end=True,
    check=True,
    callbacks=[
        BatchTransformCallback(
            transform=transrorms, scope="on_batch_start", input_key="features"
        )
    ],
)

...
callbacks:
    transform:
        _target_: BatchTransformCallback
        transform: catalyst.ToTensor
        scope: on_batch_start
        input_key: features

CheckpointCallback ¶

class catalyst.callbacks.checkpoint.CheckpointCallback(logdir: str = None, loader_key: str = None, metric_key: str = None, minimize: bool = None, min_delta: float = 1e-06, save_n_best: int = 1, load_on_stage_start: Union[str, Dict[str, str]] = None, load_on_stage_end: Union[str, Dict[str, str]] = None, metrics_filename: str = '_metrics.json', mode: str = 'all', use_logdir_postfix: bool = False, use_runner_logdir: bool = False)[source]¶

Bases: catalyst.callbacks.checkpoint.ICheckpointCallback

Checkpoint callback to save/restore your model/criterion/optimizer/scheduler.

Parameters

logdir – directory to store chekpoints
loader_key – loader key for best model selection (based on metric score over the dataset)
metric_key – metric key for best model selection (based on metric score over the dataset)
minimize – boolean flag to minimize the required metric
min_delta – minimal delta for metric improve
save_n_best – number of best checkpoint to keep, if 0 then store only last state of model and load_on_stage_end should be one of last or last_full.
load_on_stage_start (str or Dict[str, str]) –
load specified state/model at stage start.

If passed string then will be performed initialization from specified state (best/best_full/last/last_full) or checkpoint file.

If passed dict then will be performed initialization only for specified parts - model, criterion, optimizer, scheduler.

Example
```
>>> # possible checkpoints to use:
>>> #   "best"/"best_full"/"last"/"last_full"
>>> #   or path to specific checkpoint
>>> to_load = {
>>>    "model": "path/to/checkpoint.pth",
>>>    "criterion": "best",
>>>    "optimizer": "last_full",
>>>    "scheduler": "best_full",
>>> }
>>> CheckpointCallback(load_on_stage_start=to_load)
```
All other keys instead of "model", "criterion", "optimizer" and "scheduler" will be ignored.

If None or an empty dict (or dict without mentioned above keys) then no action is required at stage start and:
- Config API - will be used best state of model
- Notebook API - no action will be performed (will be used the last state)
NOTE: Loading will be performed on all stages except first.

NOTE: Criterion, optimizer and scheduler are optional keys and should be loaded from full checkpoint.

Model state can be loaded from any checkpoint.

When dict contains keys for model and some other part (for example {"model": "last", "optimizer": "last"}) and they match in prefix ("best" and "best_full") then will be loaded full checkpoint because it contains required states.
load_on_stage_end (str or Dict[str, str]) –
load specified state/model at stage end.

If passed string then will be performed initialization from specified state (best/best_full/last/last_full) or checkpoint file.

If passed dict then will be performed initialization only for specified parts - model, criterion, optimizer, scheduler. Logic for dict is the same as for load_on_stage_start.

If None then no action is required at stage end and will be used the last runner.

NOTE: Loading will be performed always at stage end.
metrics_filename – filename to save metrics in checkpoint folder. Must ends on .json or .yml
mode – checkpoining mode, could be all, full, model
use_logdir_postfix – boolean flag to use extra prefix checkpoints for logdir
use_runner_logdir – boolean flag to use runner._logdir as logdir

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_users, num_features, num_items = int(1e4), int(1e1), 10
X = torch.rand(num_users, num_features)
y = (torch.rand(num_users, num_items) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_items)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        dl.BatchTransformCallback(
            transform=torch.sigmoid,
            scope="on_batch_end",
            input_key="logits",
            output_key="scores"
        ),
        dl.CriterionCallback(
            input_key="logits", target_key="targets", metric_key="loss"
        ),
        dl.AUCCallback(input_key="scores", target_key="targets"),
        dl.HitrateCallback(
            input_key="scores", target_key="targets", topk_args=(1, 3, 5)
        ),
        dl.MRRCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.MAPCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.NDCGCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.OptimizerCallback(metric_key="loss"),
        dl.SchedulerCallback(),
        dl.CheckpointCallback(
            logdir="./logs", loader_key="valid", metric_key="loss", minimize=True
        ),
    ]
)

Note

Please follow the minimal examples sections for more use cases.

ControlFlowCallback ¶

class catalyst.callbacks.control_flow.ControlFlowCallback(base_callback: catalyst.core.callback.Callback, epochs: Union[int, Sequence[int]] = None, ignore_epochs: Union[int, Sequence[int]] = None, loaders: Union[str, Sequence[str], Mapping[str, Union[int, Sequence[int]]]] = None, ignore_loaders: Union[str, Sequence[str], Mapping[str, Union[int, Sequence[int]]]] = None, filter_fn: Union[str, Callable[[str, int, str], bool]] = None, use_global_epochs: bool = False)[source]¶

Bases: catalyst.core.callback.CallbackWrapper

Enable/disable callback execution on different stages, loaders and epochs.

Parameters

base_callback – callback to wrap
epochs –
epochs where need to enable callback, on other epochs callback will be disabled.

If passed int/float then callback will be enabled with period specified as epochs value (epochs expression epoch_number % epochs == 0) and disabled on other epochs.

If passed list of epochs then will be executed callback on specified epochs.

Default value is None.
ignore_epochs: –
epochs where: need to disable callback, on other epochs callback will be enabled.

If passed int/float then callback will be disabled with period specified as epochs value (epochs expression epoch_number % epochs != 0) and enabled on other epochs.

If passed list of epochs then will be disabled callback on specified epochs.

Default value is None.
loaders (str/Sequence[str]/Mapping[str, int/Sequence[str]]) –
loaders where should be enabled callback, on other loaders callback will be disabled.

If passed string object then will be disabled callback for loader with specified name.

If passed list/tuple of strings then will be disabled callback for loaders with specified names.

If passed dictionary where key is a string and values int or list of integers then callback will be disabled on epochs (dictionary value) for specified loader (dictionary key).

Default value is None.
ignore_loaders (str/Sequence[str]/Mapping[str, int/Sequence[str]]) –
loader names where should be disabled callback, on other loaders callback will be enabled.

If passed string object then will be disabled callback for loader with specified name.

If passed list/tuple of strings then will be disabled callback for loaders with specified names.

If passed dictionary where key is a string and values int or list of integers then callback will be disabled on epochs (dictionary value) for specified loader (dictionary key).

Default value is None.
filter_fn (str or Callable[[str, int, str], bool]) –
function to use instead of loaders or epochs arguments.

If the object passed to a filter_fn is a string then it will be interpreted as python code. Expected lambda function with three arguments stage name (str), epoch number (int), loader name (str) and this function should return True if callback should be enabled on some condition.

If passed callable object then it should accept three arguments - stage name (str), epoch number (int), loader name (str) and should return True if callback should be enabled on some condition othervise should return False.

Default value is None.

Examples:
```
# enable callback on all loaders
# exept "train" loader every 2 epochs
ControlFlowCallback(
    ...
    filter_fn=lambda s, e, l: l != "train" and e % 2 == 0
    ...
)
# or with string equivalent
ControlFlowCallback(
    ...
    filter_fn="lambda s, e, l: l != 'train' and e % 2 == 0"
    ...
)
```
use_global_epochs – if True then will be used global epochs instead of epochs in a stage, the default value is False

Note

Please run experiment with check option to check if everything works as expected with this callback.

For example, if you don’t want to compute loss on a validation you can ignore CriterionCallback, for notebook API need to wrap callback:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst.dl import (
    SupervisedRunner, AccuracyCallback,
    CriterionCallback, ControlFlowCallback,
)

num_samples, num_features = 10_000, 10
n_classes = 10
X = torch.rand(num_samples, num_features)
y = torch.randint(0, n_classes, [num_samples])
loader = DataLoader(TensorDataset(X, y), batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

model = torch.nn.Linear(num_features, n_classes)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [3, 6])

runner = SupervisedRunner()
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    logdir="./logdir",
    num_epochs=5,
    verbose=False,
    valid_metric="accuracy03",
    minimize_metric=False,
    callbacks=[
        AccuracyCallback(
            accuracy_args=[1, 3, 5]
        ),
        ControlFlowCallback(
            base_callback=CriterionCallback(),
            ignore_loaders="valid"  # or loaders="train"
        )
    ]
)

In config API need to use _wrapper argument:

callbacks_params:
  ...
  loss:
    _wrapper:
       callback: ControlFlowCallback
       ignore_loaders: valid
    callback: CriterionCallback
  ...

CriterionCallback ¶

class catalyst.callbacks.criterion.CriterionCallback(input_key: str, target_key: str, metric_key: str, criterion_key: str = None, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metrics.functional_metric.FunctionalMetricCallback, catalyst.callbacks.criterion.ICriterionCallback

Criterion callback, abstraction over criterion step.

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
metric_key – key to store computed metric in runner.batch_metrics dictionary
criterion_key – A key to take a criterion in case there are several of them, and they are in a dictionary format.

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_users, num_features, num_items = int(1e4), int(1e1), 10
X = torch.rand(num_users, num_features)
y = (torch.rand(num_users, num_items) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_items)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        dl.BatchTransformCallback(
            transform=torch.sigmoid,
            scope="on_batch_end",
            input_key="logits",
            output_key="scores"
        ),
        dl.CriterionCallback(
            input_key="logits", target_key="targets", metric_key="loss"
        ),
        dl.AUCCallback(input_key="scores", target_key="targets"),
        dl.HitrateCallback(
            input_key="scores", target_key="targets", topk_args=(1, 3, 5)
        ),
        dl.MRRCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.MAPCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.NDCGCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.OptimizerCallback(metric_key="loss"),
        dl.SchedulerCallback(),
        dl.CheckpointCallback(
            logdir="./logs", loader_key="valid", metric_key="loss", minimize=True
        ),
    ]
)

Note

Please follow the minimal examples sections for more use cases.

Metric – BatchMetricCallback ¶

class catalyst.callbacks.metric.BatchMetricCallback(metric: catalyst.metrics._metric.ICallbackBatchMetric, input_key: Union[str, Iterable[str], Dict[str, str]], target_key: Union[str, Iterable[str], Dict[str, str]], log_on_batch: bool = True)[source]¶

Bases: catalyst.callbacks.metric.MetricCallback

BatchMetricCallback implements batch-based metrics update and computation over loader

Parameters

metric – metric to calculate in callback
input_key – keys of tensors that should be used as inputs in metric calculation
target_key – keys of tensors that should be used as targets in metric calculation
log_on_batch – boolean flag to log computed metrics every batch

Metric – LoaderMetricCallback ¶

class catalyst.callbacks.metric.LoaderMetricCallback(metric: catalyst.metrics._metric.ICallbackLoaderMetric, input_key: Union[str, Iterable[str], Dict[str, str]], target_key: Union[str, Iterable[str], Dict[str, str]])[source]¶

Bases: catalyst.callbacks.metric.MetricCallback

LoaderMetricCallback implements loader-based metrics update and computation over loader

Parameters

metric – metric to calculate in callback
input_key – keys of tensors that should be used as inputs in metric calculation
target_key – keys of tensors that should be used as targets in metric calculation

Metric – MetricAggregationCallback ¶

class catalyst.callbacks.metric_aggregation.MetricAggregationCallback(metric_key: str, metrics: Union[str, List[str], Dict[str, float]] = None, mode: Union[str, Callable] = 'mean', scope: str = 'batch', multiplier: float = 1.0)[source]¶

Bases: catalyst.core.callback.Callback

A callback to aggregate several metrics in one value.

Parameters

metric_key – new key for aggregated metric.
metrics (Union[str, List[str], Dict[str, float]]) – If not None, it aggregates only the values from the metric by these keys. for weighted_sum aggregation it must be a Dict[str, float].
mode – function for aggregation. Must be either sum, mean or weighted_sum or user’s function to aggregate metrics. This function must get dict of metrics and runner and return aggregated metric. It can be useful for complicated fine tuning with different losses that depends on epochs and loader or something also
scope – type of metric. Must be either batch or loader
multiplier – scale factor for the aggregated metric.

Python example - loss is a weighted sum of cross entropy loss and binary cross entropy loss:

import torch
from torch.utils.data import DataLoader, TensorDataset

from catalyst import dl

# data
num_samples, num_features, num_classes = int(1e4), int(1e1), 4
X = torch.rand(num_samples, num_features)
y = (torch.rand(num_samples,) * num_classes).to(torch.int64)
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_classes)
criterion = {"ce": torch.nn.CrossEntropyLoss(), "bce": torch.nn.BCEWithLogitsLoss()}
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# runner
class CustomRunner(dl.Runner):
    def handle_batch(self, batch):
        x, y = batch
        logits = self.model(x)
        num_classes = logits.shape[-1]
        targets_onehot = torch.nn.functional.one_hot(y, num_classes=num_classes)
        self.batch = {
            "features": x,
            "logits": logits,
            "targets": y,
            "targets_onehot": targets_onehot.float(),
        }


# training
runner = CustomRunner()
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    logdir="./logdir",
    valid_loader="valid",
    valid_metric="loss",
    minimize_valid_metric=True,
    num_epochs=3,
    callbacks=[
        dl.AccuracyCallback(
            input_key="logits", target_key="targets", num_classes=num_classes
        ),
        dl.CriterionCallback(
            input_key="logits",
            target_key="targets",
            metric_key="loss_ce",
            criterion_key="ce",
        ),
        dl.CriterionCallback(
            input_key="logits",
            target_key="targets_onehot",
            metric_key="loss_bce",
            criterion_key="bce",
        ),
        # loss aggregation
        dl.MetricAggregationCallback(
            metric_key="loss",
            metrics={"loss_ce": 0.6, "loss_bce": 0.4},
            mode="weighted_sum",
        ),
        dl.OptimizerCallback(metric_key="loss"),
    ],
)

Misc – CheckRunCallback ¶

class catalyst.callbacks.misc.CheckRunCallback(num_batch_steps: int = 3, num_epoch_steps: int = 3)[source]¶

Bases: catalyst.core.callback.Callback

Executes only a pipeline part from the run.

Parameters

num_batch_steps – number of batches to iterate in epoch
num_epoch_steps – number of epoch to perform in a stage

Minimal working example (Notebook API):

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# data
num_samples, num_features = int(1e4), int(1e1)
X, y = torch.rand(num_samples, num_features), torch.rand(num_samples)
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, 1)
criterion = torch.nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [3, 6])

# model training
runner = dl.SupervisedRunner()
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    logdir="./logdir",
    num_epochs=8,
    verbose=True,
    callbacks=[
        dl.CheckRunCallback(num_batch_steps=3, num_epoch_steps=3)
    ]
)

Misc – EarlyStoppingCallback ¶

class catalyst.callbacks.misc.EarlyStoppingCallback(patience: int, loader_key: str, metric_key: str, minimize: bool, min_delta: float = 1e-06)[source]¶

Bases: catalyst.callbacks.misc.IEpochMetricHandlerCallback

Stage early stop based on metric.

Parameters

patience – number of epochs with no improvement after which training will be stopped.
loader_key – loader key for early stopping (based on metric score over the dataset)
metric_key – metric key for early stopping (based on metric score over the dataset)
minimize – if True then expected that metric should decrease and early stopping will be performed only when metric stops decreasing. If False then expected that metric should increase. Default value True.
min_delta – minimum change in the monitored metric to qualify as an improvement, i.e. an absolute change of less than min_delta, will count as no improvement, default value is 1e-6.

handle_score_is_better(runner: catalyst.core.runner.IRunner)[source]¶: Event handler.

handle_score_is_not_better(runner: catalyst.core.runner.IRunner)[source]¶: Event handler.

Misc – TimerCallback ¶

class catalyst.callbacks.misc.TimerCallback[source]¶

Bases: catalyst.core.callback.Callback

Logs pipeline execution time.

Misc – TqdmCallback ¶

class catalyst.callbacks.misc.TqdmCallback[source]¶

Bases: catalyst.core.callback.Callback

Logs the params into tqdm console.

on_exception(runner: catalyst.core.runner.IRunner)[source]¶: Called if an Exception was raised.

MixUp Callback ¶

class catalyst.callbacks.mixup.MixupCallback(keys: Union[str, List[str]], alpha=0.2, mode='replace', on_train_only=True, **kwargs)[source]¶

Bases: catalyst.core.callback.Callback

Callback to do mixup augmentation. More details about mixin can be found in the paper mixup: Beyond Empirical Risk Minimization: https://arxiv.org/abs/1710.09412 .

Examples:

from typing import Any, Dict
import os

import numpy as np
import torch
from torch import nn
from torch.utils.data import DataLoader

from catalyst import dl
from catalyst.callbacks import MixupCallback
from catalyst.contrib.datasets import MNIST
from catalyst.data.transforms import ToTensor


class SimpleNet(nn.Module):
    def __init__(self, in_channels, in_hw, out_features):
        super().__init__()
        self.encoder = nn.Sequential(nn.Conv2d(in_channels,
                                               in_channels, 3, 1, 1), nn.Tanh())
        self.clf = nn.Linear(in_channels * in_hw * in_hw, out_features)

    def forward(self, x):
        features = self.encoder(x)
        features = features.view(features.size(0), -1)
        logits = self.clf(features)
        return logits


class SimpleDataset(torch.utils.data.Dataset):
    def __init__(self, train: bool = False):
        self.mnist = MNIST(os.getcwd(), train=train, download=True, transform=ToTensor())

    def __len__(self) -> int:
        return len(self.mnist)

    def __getitem__(self, idx: int) -> Dict[str, Any]:
        x, y = self.mnist.__getitem__(idx)
        y_one_hot = np.zeros(10)
        y_one_hot[y] = 1
        return {"image": x,
                "clf_targets": y,
                "clf_targets_one_hot": torch.Tensor(y_one_hot)}


model = SimpleNet(1, 28, 10)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.02)

loaders = {
    "train": DataLoader(SimpleDataset(train=True), batch_size=32),
    "valid": DataLoader(SimpleDataset(train=False), batch_size=32),
}


class CustomRunner(dl.Runner):
    def handle_batch(self, batch):
        image = batch["image"]
        clf_logits = self.model(image)
        self.batch["clf_logits"] = clf_logits


runner = CustomRunner()
runner.train(
    loaders=loaders,
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    logdir="./logdir14",
    num_epochs=2,
    verbose=True,
    valid_loader="valid",
    valid_metric="loss",
    minimize_valid_metric=True,
    callbacks={
        "mixup": MixupCallback(keys=["image", "clf_targets_one_hot"]),
        "criterion": dl.CriterionCallback(
            metric_key="loss", input_key="clf_logits", target_key="clf_targets_one_hot",
        ),
        "optimizer": dl.OptimizerCallback(metric_key="loss"),
        "classification": dl.ControlFlowCallback(
            dl.PrecisionRecallF1SupportCallback(
                input_key="clf_logits", target_key="clf_targets", num_classes=10,
            ),
            ignore_loaders="train",
        ),
    },
)

OnnxCallback ¶

class catalyst.callbacks.onnx.OnnxCallback(input_key: str, logdir: Union[str, pathlib.Path] = None, filename: str = 'onnx.py', method_name: str = 'forward', input_names: Iterable = None, output_names: List[str] = None, dynamic_axes: Union[Dict[str, int], Dict[str, Dict[str, int]]] = None, opset_version: int = 9, do_constant_folding: bool = False, verbose: bool = False)[source]¶

Bases: catalyst.core.callback.Callback

Callback for converting model to onnx runtime.

Parameters

input_key – input key from runner.batch to use for onnx export
logdir – path to folder for saving
filename – filename
method_name (str, optional) – Forward pass method to be converted. Defaults to “forward”.
input_names (Iterable, optional) – name of inputs in graph. Defaults to None.
output_names (List[str], optional) – name of outputs in graph. Defaults to None.
dynamic_axes (Union[Dict[str, int], Dict[str, Dict[str, int]]], optional) – axes with dynamic shapes. Defaults to None.
opset_version (int, optional) – Defaults to 9.
do_constant_folding (bool, optional) – If True, the constant-folding optimization is applied to the model during export. Defaults to False.
verbose (bool, default False) – if specified, we will print out a debug description of the trace being exported.

Example

import os

import torch
from torch import nn
from torch.utils.data import DataLoader

from catalyst import dl
from catalyst.data import ToTensor
from catalyst.contrib.datasets import MNIST
from catalyst.contrib.nn.modules import Flatten

loaders = {
    "train": DataLoader(
        MNIST(
            os.getcwd(), train=False, download=True, transform=ToTensor()
        ),
        batch_size=32,
    ),
    "valid": DataLoader(
        MNIST(
            os.getcwd(), train=False, download=True, transform=ToTensor()
        ),
        batch_size=32,
    ),
}

model = nn.Sequential(Flatten(), nn.Linear(784, 512), nn.ReLU(), nn.Linear(512, 10))
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-2)
runner = dl.SupervisedRunner()
runner.train(
    model=model,
    callbacks=[dl.OnnxCallback(input_key="features", logdir="./logs")],
    loaders=loaders,
    criterion=criterion,
    optimizer=optimizer,
    num_epochs=1,
    logdir="./logs",
)

OptimizerCallback ¶

class catalyst.callbacks.optimizer.OptimizerCallback(metric_key: str, model_key: str = None, optimizer_key: str = None, accumulation_steps: int = 1, grad_clip_fn: Union[str, Callable] = None, grad_clip_params: Dict = None)[source]¶

Bases: catalyst.callbacks.optimizer.IOptimizerCallback

Optimizer callback, abstraction over optimizer step.

Parameters

metric_key – a key to get loss from runner.batch_metrics
model_key – a key to select a model from runner.model in case there are several of them and they are in a dictionary format.
optimizer_key – a key to select a optimizer from runner.optimizer in case there are several of them and they are in a dictionary format.
accumulation_steps – number of steps before model.zero_grad()
grad_clip_fn – callable gradient cliping function or it’s name or
grad_clip_params – key-value parameters for grad_clip_fn

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_users, num_features, num_items = int(1e4), int(1e1), 10
X = torch.rand(num_users, num_features)
y = (torch.rand(num_users, num_items) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_items)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        dl.BatchTransformCallback(
            transform=torch.sigmoid,
            scope="on_batch_end",
            input_key="logits",
            output_key="scores"
        ),
        dl.CriterionCallback(
            input_key="logits", target_key="targets", metric_key="loss"
        ),
        dl.AUCCallback(input_key="scores", target_key="targets"),
        dl.HitrateCallback(
            input_key="scores", target_key="targets", topk_args=(1, 3, 5)
        ),
        dl.MRRCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.MAPCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.NDCGCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.OptimizerCallback(metric_key="loss"),
        dl.SchedulerCallback(),
        dl.CheckpointCallback(
            logdir="./logs", loader_key="valid", metric_key="loss", minimize=True
        ),
    ]
)

Note

Please follow the minimal examples sections for more use cases.

OptunaPruningCallback ¶

class catalyst.callbacks.optuna.OptunaPruningCallback(loader_key: str, metric_key: str, minimize: bool, min_delta: float = 1e-06, trial: Optional[optuna.trial._trial.Trial] = None)[source]¶

Bases: catalyst.core.callback.Callback

Optuna callback for pruning unpromising runs. This callback can be used for early stopping (pruning) unpromising runs.

Parameters

loader_key – loader key for best model selection (based on metric score over the dataset)
metric_key – metric key for best model selection (based on metric score over the dataset)
minimize – boolean flag to minimize the required metric
min_delta – minimal delta for metric improve
trial – Optuna.Trial for experiment.

import optuna

from catalyst.dl import SupervisedRunner, OptunaPruningCallback

# some python code ...

def objective(trial: optuna.Trial):
    # standard optuna code for model and/or optimizer suggestion ...
    runner = SupervisedRunner()
    runner.train(
        model=model,
        loaders=loaders,
        criterion=criterion,
        optimizer=optimizer,
        callbacks=[
            OptunaPruningCallback(trial)
            # some other callbacks ...
        ],
        num_epochs=num_epochs,
    )
    return runner.best_valid_metrics[runner.valid_metric]

study = optuna.create_study()
study.optimize(objective, n_trials=100, timeout=600)

Config API is supported through catalyst-dl tune command.

PeriodicLoaderCallback ¶

class catalyst.callbacks.periodic_loader.PeriodicLoaderCallback(valid_loader_key: str, valid_metric_key: str, minimize: bool = True, **kwargs)[source]¶

Bases: catalyst.core.callback.Callback

Callback for runing loaders with specified period. To disable loader use 0 as period (if specified 0 for validation loader then will be raised an error).

Parameters: kwargs – loader names and their run periods.

For example, if you have train, train_additional, valid and valid_additional loaders and wan’t to use train_additional every 2 epochs, valid - every 3 epochs and valid_additional - every 5 epochs:

from catalyst.dl import (
    SupervisedRunner, PeriodicLoaderCallback,
)
runner = SupervisedRunner()
runner.train(
    ...
    loaders={
        "train": ...,
        "train_additional": ...,
        "valid": ...,
        "valid_additional":...
    }
    ...
    callbacks=[
        ...
        PeriodicLoaderCallback(
            train_additional=2,
            valid=3,
            valid_additional=5
        ),
        ...
    ]
    ...
)

PruningCallback ¶

class catalyst.callbacks.pruning.PruningCallback(pruning_fn: Union[Callable, str], amount: Union[int, float], keys_to_prune: Optional[List[str]] = None, prune_on_epoch_end: Optional[bool] = False, prune_on_stage_end: Optional[bool] = True, remove_reparametrization_on_stage_end: Optional[bool] = True, layers_to_prune: Optional[List[str]] = None, dim: Optional[int] = None, l_norm: Optional[int] = None)[source]¶

Bases: catalyst.core.callback.Callback

This callback prunes network parameters during and/or after training.

Parameters

pruning_fn – function from torch.nn.utils.prune module or your based on BasePruningMethod. Can be string e.g. “l1_unstructured”. See pytorch docs for more details.
amount – quantity of parameters to prune. If float, should be between 0.0 and 1.0 and represent the fraction of parameters to prune. If int, it represents the absolute number of parameters to prune.
keys_to_prune – list of strings. Determines which tensor in modules will be pruned.
prune_on_epoch_end – bool flag determines call or not to call pruning_fn on epoch end.
prune_on_stage_end – bool flag determines call or not to call pruning_fn on stage end.
remove_reparametrization_on_stage_end – if True then all reparametrization pre-hooks and tensors with mask will be removed on stage end.
layers_to_prune – list of strings - module names to be pruned. If None provided then will try to prune every module in model.
dim – if you are using structured pruning method you need to specify dimension.
l_norm – if you are using ln_structured you need to specify l_norm.

ProfilerCallback ¶

class catalyst.callbacks.profiler.ProfilerCallback(loader_key: str = None, epoch: int = 1, num_batches: int = None, profiler_kwargs: Dict[str, Any] = None, tensorboard_path: str = None, export_chrome_trace_path: str = None, export_stacks_kwargs: Dict[str, Any] = None)[source]¶

Bases: catalyst.core.callback.Callback

Profile specified epoch or some fixed number of batches.

Parameters

loader_key – name of the loader to use for profiling. If None then will be used first loader from experiment.
epoch – epoch number to use for profiling.
num_batches – number of batches to use in epoch to do a profiling. If None then will be used all batches in loader.
profiler_kwargs – arguments to pass to a profiler. To get more info about possible arguments please use PyTorch profiler docs.
tensorboard_path – path where should be stored logs for tensorboard. If None then will be ignored.
export_chrome_trace_path – path to export chrome trace. If None then will be ignored exporting chrome trace to a file.
export_stacks_kwargs –
arguments to pass to a profiler.export_stacks method. If None then triggering profiler.export_stacks will be avoided.

Example of using FlameGraph tool:
```
git clone https://github.com/brendangregg/FlameGraph
cd FlameGraph
./flamegraph.pl –title “CPU time” –countname “us.” profiler.stacks > perf_viz.svg
```

Note

Export to tensorboard and chrome trace mutually exclusive and specifying both of them will raise an error.

Example

import os

import torch
from torch import nn
from torch.utils.data import DataLoader

from catalyst import dl
from catalyst.data import ToTensor
from catalyst.contrib.datasets import MNIST
from catalyst.contrib.nn.modules import Flatten

loaders = {
    "train": DataLoader(
        MNIST(os.getcwd(), train=False, download=True, transform=ToTensor()),
        batch_size=32,
    ),
    "valid": DataLoader(
        MNIST(os.getcwd(), train=False, download=True, transform=ToTensor()),
        batch_size=32,
    ),
}

model = nn.Sequential(Flatten(), nn.Linear(784, 512), nn.ReLU(), nn.Linear(512, 10))
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-2)
runner = dl.SupervisedRunner()
runner.train(
    model=model,
    callbacks=[dl.ProfilerCallback(
        loader_key="train", epoch=3,
        profiler_kwargs=dict(
            activities=[
                torch.profiler.ProfilerActivity.CPU,
                torch.profiler.ProfilerActivity.CUDA,
            ],
            on_trace_ready=torch.profiler.tensorboard_trace_handler(
                "./logs/tb_profile"
            ),
            with_stack=True,
            with_flops=True,
        )
    )],
    loaders=loaders,
    criterion=criterion,
    optimizer=optimizer,
    num_epochs=5,
    logdir="./logs",
)

QuantizationCallback ¶

class catalyst.callbacks.quantization.QuantizationCallback(logdir: Union[str, pathlib.Path] = None, filename: str = 'quantized.pth', qconfig_spec: Dict = None, dtype: Union[str, torch.dtype, None] = 'qint8')[source]¶

Bases: catalyst.core.callback.Callback

Callback for model quantiztion.

Parameters

logdir – path to folder for saving
filename – filename
qconfig_spec (Dict, optional) – quantization config in PyTorch format. Defaults to None.
dtype (Union[str, Optional[torch.dtype]], optional) – Type of weights after quantization. Defaults to “qint8”.

Example

import os

import torch
from torch import nn
from torch.utils.data import DataLoader

from catalyst import dl
from catalyst.data import ToTensor
from catalyst.contrib.datasets import MNIST
from catalyst.contrib.nn.modules import Flatten

loaders = {
    "train": DataLoader(
        MNIST(os.getcwd(),
        train=False,
        download=True,
        transform=ToTensor()),
        batch_size=32,
    ),
    "valid": DataLoader(
        MNIST(os.getcwd(),
        train=False,
        download=True,
        transform=ToTensor()),
        batch_size=32,
    ),
}

model = nn.Sequential(Flatten(), nn.Linear(784, 512), nn.ReLU(), nn.Linear(512, 10))
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-2)
runner = dl.SupervisedRunner()
runner.train(
    model=model,
    callbacks=[dl.QuantizationCallback(logdir="./logs")],
    loaders=loaders,
    criterion=criterion,
    optimizer=optimizer,
    num_epochs=1,
    logdir="./logs",
)

Scheduler – SchedulerCallback ¶

class catalyst.callbacks.scheduler.SchedulerCallback(scheduler_key: str = None, mode: str = None, loader_key: str = None, metric_key: str = None)[source]¶

Bases: catalyst.callbacks.scheduler.ISchedulerCallback

Scheduler callback, abstraction over scheduler step.

Parameters

scheduler_key – scheduler name, if None, default is None.
mode – scheduler mode, should be one of "epoch" or "batch", default is None. If None and object is instance of BatchScheduler or OneCycleLRWithWarmup then will be used "batch" otherwise - "epoch".
loader_key – loader name to look after for ReduceLROnPlateau scheduler
metric_key – metric name to forward to scheduler object, if None then will be used main metric specified in experiment.

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_users, num_features, num_items = int(1e4), int(1e1), 10
X = torch.rand(num_users, num_features)
y = (torch.rand(num_users, num_items) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_items)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        dl.BatchTransformCallback(
            transform=torch.sigmoid,
            scope="on_batch_end",
            input_key="logits",
            output_key="scores"
        ),
        dl.CriterionCallback(
            input_key="logits", target_key="targets", metric_key="loss"
        ),
        dl.AUCCallback(input_key="scores", target_key="targets"),
        dl.HitrateCallback(
            input_key="scores", target_key="targets", topk_args=(1, 3, 5)
        ),
        dl.MRRCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.MAPCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.NDCGCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.OptimizerCallback(metric_key="loss"),
        dl.SchedulerCallback(),
        dl.CheckpointCallback(
            logdir="./logs", loader_key="valid", metric_key="loss", minimize=True
        ),
    ]
)

Note

Please follow the minimal examples sections for more use cases.

make_batch_step(runner: IRunner) → None[source]¶

Perform scheduler step and update batch metrics

Parameters: runner – current runner

make_epoch_step(runner: IRunner) → None[source]¶

Perform scheduler step and update epoch metrics

Parameters: runner – current runner

Scheduler – LRFinder ¶

class catalyst.callbacks.scheduler.LRFinder(final_lr: float, scale: str = 'log', num_steps: Optional[int] = None, optimizer_key: str = None)[source]¶

Bases: catalyst.callbacks.scheduler.ILRUpdater

Helps you find an optimal learning rate for a model, as per suggestion of Cyclical Learning Rates for Training Neural Networks paper. Learning rate is increased in linear or log scale, depending on user input.

See How Do You Find A Good Learning Rate article for details.

calc_lr()[source]¶

Calculates learning rate.

Returns: learning rate.

calc_momentum()[source]¶: Calculates new momentum.

Tracing ¶

class catalyst.callbacks.tracing.TracingCallback(input_key: Union[str, List[str]], logdir: Union[str, pathlib.Path] = None, filename: str = 'traced_model.pth', method_name: str = 'forward')[source]¶

Bases: catalyst.core.callback.Callback

Callback for model tracing.

Parameters

input_key – input key from runner.batch to use for model tracing
logdir – path to folder for saving
filename – filename
method_name – Model’s method name that will be used as entrypoint during tracing

Example

import os

import torch
from torch import nn
from torch.utils.data import DataLoader

from catalyst import dl
from catalyst.data import ToTensor
from catalyst.contrib.datasets import MNIST
from catalyst.contrib.nn.modules import Flatten

loaders = {
    "train": DataLoader(
        MNIST(
            os.getcwd(), train=False, download=True, transform=ToTensor()
        ),
        batch_size=32,
    ),
    "valid": DataLoader(
        MNIST(
            os.getcwd(), train=False, download=True, transform=ToTensor()
        ),
        batch_size=32,
    ),
}

model = nn.Sequential(
    Flatten(), nn.Linear(784, 512), nn.ReLU(), nn.Linear(512, 10)
)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-2)
runner = dl.SupervisedRunner()
runner.train(
    model=model,
    callbacks=[dl.TracingCallback(input_key="features", logdir="./logs")],
    loaders=loaders,
    criterion=criterion,
    optimizer=optimizer,
    num_epochs=1,
    logdir="./logs",
)

Metric ¶

Accuracy - AccuracyCallback ¶

class catalyst.callbacks.metrics.accuracy.AccuracyCallback(input_key: str, target_key: str, topk_args: List[int] = None, num_classes: int = None, log_on_batch: bool = True, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.BatchMetricCallback

Accuracy metric callback. Computes multiclass accuracy@topk for the specified values of topk.

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
topk_args – specifies which accuracy@K to log: [1] - accuracy [1, 3] - accuracy at 1 and 3 [1, 3, 5] - accuracy at 1, 3 and 5
num_classes – number of classes to calculate topk_args if accuracy_args is None
log_on_batch – boolean flag to log computed metrics every batch
prefix – metric prefix
suffix – metric suffix

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_samples, num_features, num_classes = int(1e4), int(1e1), 4
X = torch.rand(num_samples, num_features)
y = (torch.rand(num_samples,) * num_classes).to(torch.int64)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_classes)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    logdir="./logdir",
    num_epochs=3,
    valid_loader="valid",
    valid_metric="accuracy03",
    minimize_valid_metric=False,
    verbose=True,
    callbacks=[
        dl.AccuracyCallback(
            input_key="logits", target_key="targets", num_classes=num_classes
        ),
        dl.PrecisionRecallF1SupportCallback(
            input_key="logits", target_key="targets", num_classes=num_classes
        ),
        dl.AUCCallback(input_key="logits", target_key="targets"),
    ],
)

Note

Please follow the minimal examples sections for more use cases.

Accuracy - MultilabelAccuracyCallback ¶

class catalyst.callbacks.metrics.accuracy.MultilabelAccuracyCallback(input_key: str, target_key: str, threshold: Union[float, torch.Tensor] = 0.5, log_on_batch: bool = True, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.BatchMetricCallback

Multilabel accuracy metric callback. Computes multilabel accuracy@topk for the specified values of topk.

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
threshold – thresholds for model scores
log_on_batch – boolean flag to log computed metrics every batch
prefix – metric prefix
suffix – metric suffix

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_samples, num_features, num_classes = int(1e4), int(1e1), 4
X = torch.rand(num_samples, num_features)
y = (torch.rand(num_samples, num_classes) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_classes)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    logdir="./logdir",
    num_epochs=3,
    valid_loader="valid",
    valid_metric="accuracy",
    minimize_valid_metric=False,
    verbose=True,
    callbacks=[
        dl.AUCCallback(input_key="logits", target_key="targets"),
        dl.MultilabelAccuracyCallback(
            input_key="logits", target_key="targets", threshold=0.5
        )
    ]

Note

Please follow the minimal examples sections for more use cases.

AUCCallback ¶

class catalyst.callbacks.metrics.auc.AUCCallback(input_key: str, target_key: str, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.LoaderMetricCallback

ROC-AUC metric callback.

Parameters

input_key – input key to use for auc calculation, specifies our y_true.
target_key – output key to use for auc calculation, specifies our y_pred.
prefix – metric prefix
suffix – metric suffix

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_samples, num_features, num_classes = int(1e4), int(1e1), 4
X = torch.rand(num_samples, num_features)
y = (torch.rand(num_samples,) * num_classes).to(torch.int64)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_classes)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    logdir="./logdir",
    num_epochs=3,
    valid_loader="valid",
    valid_metric="accuracy03",
    minimize_valid_metric=False,
    verbose=True,
    callbacks=[
        dl.AccuracyCallback(
            input_key="logits", target_key="targets", num_classes=num_classes
        ),
        dl.PrecisionRecallF1SupportCallback(
            input_key="logits", target_key="targets", num_classes=num_classes
        ),
        dl.AUCCallback(input_key="logits", target_key="targets"),
    ],
)

Note

Please follow the minimal examples sections for more use cases.

Classification – PrecisionRecallF1SupportCallback ¶

class catalyst.callbacks.metrics.classification.PrecisionRecallF1SupportCallback(input_key: str, target_key: str, num_classes: int, zero_division: int = 0, log_on_batch: bool = True, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.BatchMetricCallback

Multiclass PrecisionRecallF1Support metric callback.

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
num_classes – number of classes
zero_division – value to set in case of zero division during metrics (precision, recall) computation; should be one of 0 or 1
log_on_batch – boolean flag to log computed metrics every batch
prefix – metric prefix
suffix – metric suffix

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_samples, num_features, num_classes = int(1e4), int(1e1), 4
X = torch.rand(num_samples, num_features)
y = (torch.rand(num_samples,) * num_classes).to(torch.int64)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_classes)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    logdir="./logdir",
    num_epochs=3,
    valid_loader="valid",
    valid_metric="accuracy03",
    minimize_valid_metric=False,
    verbose=True,
    callbacks=[
        dl.AccuracyCallback(
            input_key="logits", target_key="targets", num_classes=num_classes
        ),
        dl.PrecisionRecallF1SupportCallback(
            input_key="logits", target_key="targets", num_classes=num_classes
        ),
        dl.AUCCallback(input_key="logits", target_key="targets"),
    ],
)

Note

Please follow the minimal examples sections for more use cases.

Classification – MultilabelPrecisionRecallF1SupportCallback ¶

class catalyst.callbacks.metrics.classification.MultilabelPrecisionRecallF1SupportCallback(input_key: str, target_key: str, num_classes: int, zero_division: int = 0, log_on_batch: bool = True, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.BatchMetricCallback

Multilabel PrecisionRecallF1Support metric callback.

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
num_classes – number of classes
zero_division – value to set in case of zero division during metrics (precision, recall) computation; should be one of 0 or 1
log_on_batch – boolean flag to log computed metrics every batch
prefix – metric prefix
suffix – metric suffix

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_samples, num_features, num_classes = int(1e4), int(1e1), 4
X = torch.rand(num_samples, num_features)
y = (torch.rand(num_samples, num_classes) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_classes)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    logdir="./logdir",
    num_epochs=3,
    valid_loader="valid",
    valid_metric="accuracy",
    minimize_valid_metric=False,
    verbose=True,
    callbacks=[
        dl.BatchTransformCallback(
            transform=torch.sigmoid,
            scope="on_batch_end",
            input_key="logits",
            output_key="scores"
        ),
        dl.AUCCallback(input_key="scores", target_key="targets"),
        dl.MultilabelAccuracyCallback(
            input_key="scores", target_key="targets", threshold=0.5
        ),
        dl.MultilabelPrecisionRecallF1SupportCallback(
            input_key="scores", target_key="targets", num_classes=num_classes
        ),
    ]
)

Note

Please follow the minimal examples sections for more use cases.

CMCScoreCallback ¶

class catalyst.callbacks.metrics.cmc_score.CMCScoreCallback(embeddings_key: str, labels_key: str, is_query_key: str, topk_args: List[int] = None, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.LoaderMetricCallback

Cumulative Matching Characteristics callback.

This callback was designed to count cumulative matching characteristics. If current object is from query your dataset should output True in is_query_key and false if current object is from gallery. You can see QueryGalleryDataset in catalyst.contrib.datasets.metric_learning for more information. On batch end callback accumulate all embeddings

Parameters

embeddings_key – embeddings key in output dict
labels_key – labels key in output dict
is_query_key – bool key True if current object is from query
topk_args – specifies which cmc@K to log. [1] - cmc@1 [1, 3] - cmc@1 and cmc@3 [1, 3, 5] - cmc@1, cmc@3 and cmc@5
prefix – metric prefix
suffix – metric suffix

Note

You should use it with ControlFlowCallback and add all query/gallery sets to loaders. Loaders should contain “is_query” and “label” key.

Examples:

import os
from torch.optim import Adam
from torch.utils.data import DataLoader
from catalyst import data, dl
from catalyst.contrib import datasets, models, nn
from catalyst.data.transforms import Compose, Normalize, ToTensor


# 1. train and valid loaders
transforms = Compose([ToTensor(), Normalize((0.1307,), (0.3081,))])

train_dataset = datasets.MnistMLDataset(
    root=os.getcwd(), download=True, transform=transforms
    )
sampler = data.BalanceBatchSampler(labels=train_dataset.get_labels(), p=5, k=10)
train_loader = DataLoader(
    dataset=train_dataset, sampler=sampler, batch_size=sampler.batch_size
    )

valid_dataset = datasets.MnistQGDataset(
    root=os.getcwd(), transform=transforms, gallery_fraq=0.2
)
valid_loader = DataLoader(dataset=valid_dataset, batch_size=1024)

# 2. model and optimizer
model = models.MnistSimpleNet(out_features=16)
optimizer = Adam(model.parameters(), lr=0.001)

# 3. criterion with triplets sampling
sampler_inbatch = data.HardTripletsSampler(norm_required=False)
criterion = nn.TripletMarginLossWithSampler(margin=0.5, sampler_inbatch=sampler_inbatch)

# 4. training with catalyst Runner
class CustomRunner(dl.SupervisedRunner):
    def handle_batch(self, batch) -> None:
        if self.is_train_loader:
            images, targets = batch["features"].float(), batch["targets"].long()
            features = self.model(images)
            self.batch = {"embeddings": features, "targets": targets,}
        else:
            images, targets, is_query =                         batch["features"].float(),                         batch["targets"].long(),                         batch["is_query"].bool()
            features = self.model(images)
            self.batch = {
                "embeddings": features, "targets": targets, "is_query": is_query
            }

callbacks = [
    dl.ControlFlowCallback(
        dl.CriterionCallback(
            input_key="embeddings", target_key="targets", metric_key="loss"
        ),
        loaders="train",
    ),
    dl.ControlFlowCallback(
        dl.CMCScoreCallback(
            embeddings_key="embeddings",
            labels_key="targets",
            is_query_key="is_query",
            topk_args=[1],
        ),
        loaders="valid",
    ),
    dl.PeriodicLoaderCallback(
        valid_loader_key="valid", valid_metric_key="cmc01", minimize=False, valid=2
    ),
]

runner = CustomRunner(input_key="features", output_key="embeddings")
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    callbacks=callbacks,
    loaders={"train": train_loader, "valid": valid_loader},
    verbose=False,
    logdir="./logs",
    valid_loader="valid",
    valid_metric="cmc01",
    minimize_valid_metric=False,
    num_epochs=10,
)

Note

Please follow the minimal examples sections for more use cases.

ReidCMCScoreCallback ¶

class catalyst.callbacks.metrics.cmc_score.ReidCMCScoreCallback(embeddings_key: str, pids_key: str, cids_key: str, is_query_key: str, topk_args: List[int] = None, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.LoaderMetricCallback

Cumulative Matching Characteristics callback for reID case. More information about cmc-based callbacks in CMCScoreCallback’s docs.

Parameters

embeddings_key – embeddings key in output dict
pids_key – pids key in output dict
cids_key – cids key in output dict
is_query_key – bool key True if current object is from query
topk_args – specifies which cmc@K to log. [1] - cmc@1 [1, 3] - cmc@1 and cmc@3 [1, 3, 5] - cmc@1, cmc@3 and cmc@5
prefix – metric prefix
suffix – metric suffix

ConfusionMatrixCallback ¶

class catalyst.callbacks.metrics.confusion_matrix.ConfusionMatrixCallback(input_key: str, target_key: str, prefix: str = None, class_names: List[str] = None, num_classes: int = None, normalized: bool = False, plot_params: Dict = None)[source]¶

Bases: catalyst.core.callback.Callback

Callback to plot your confusion matrix to the loggers.

Parameters

input_key – key to use from runner.batch, specifies our y_pred
target_key – key to use from runner.batch, specifies our y_true
prefix – plot name for monitoring tools
class_names – list with class names
num_classes – number of classes
normalized – boolean flag for confusion matrix normalization
plot_params – extra params for plt.figure rendering

Note

catalyst[ml] required for this callback

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_samples, num_features, num_classes = int(1e4), int(1e1), 4
X = torch.rand(num_samples, num_features)
y = (torch.rand(num_samples,) * num_classes).to(torch.int64)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_classes)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    logdir="./logdir",
    num_epochs=3,
    valid_loader="valid",
    valid_metric="accuracy03",
    minimize_valid_metric=False,
    verbose=True,
    callbacks=[
        dl.AccuracyCallback(
            input_key="logits", target_key="targets", num_classes=num_classes
        ),
        dl.PrecisionRecallF1SupportCallback(
            input_key="logits", target_key="targets", num_classes=num_classes
        ),
        dl.AUCCallback(input_key="logits", target_key="targets"),
        dl.ConfusionMatrixCallback(
            input_key="logits", target_key="targets", num_classes=num_classes
        ),
    ],
)

Note

Please follow the minimal examples sections for more use cases.

FunctionalMetricCallback ¶

class catalyst.callbacks.metrics.functional_metric.FunctionalMetricCallback(input_key: Union[str, Iterable[str], Dict[str, str]], target_key: Union[str, Iterable[str], Dict[str, str]], metric_fn: Callable, metric_key: str, compute_on_call: bool = True, log_on_batch: bool = True, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.FunctionalBatchMetricCallback

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
metric_fn – metric function, that get outputs, targets and return score as torch.Tensor
metric_key – key to store computed metric in runner.batch_metrics dictionary
compute_on_call – Computes and returns metric value during metric call. Used for per-batch logging. default: True
log_on_batch – boolean flag to log computed metrics every batch
prefix – metric prefix
suffix – metric suffix

RecSys – HitrateCallback ¶

class catalyst.callbacks.metrics.recsys.HitrateCallback(input_key: str, target_key: str, topk_args: List[int] = None, log_on_batch: bool = True, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.BatchMetricCallback

Hitrate metric callback. Computes HR@topk for the specified values of topk.

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
topk_args – specifies which HR@K to log: [1] - HR [1, 3] - HR at 1 and 3 [1, 3, 5] - HR at 1, 3 and 5
log_on_batch – boolean flag to log computed metrics every batch
prefix – metric prefix
suffix – metric suffix

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_users, num_features, num_items = int(1e4), int(1e1), 10
X = torch.rand(num_users, num_features)
y = (torch.rand(num_users, num_items) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_items)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        dl.BatchTransformCallback(
            transform=torch.sigmoid,
            scope="on_batch_end",
            input_key="logits",
            output_key="scores"
        ),
        dl.CriterionCallback(
            input_key="logits", target_key="targets", metric_key="loss"
        ),
        dl.AUCCallback(input_key="scores", target_key="targets"),
        dl.HitrateCallback(
            input_key="scores", target_key="targets", topk_args=(1, 3, 5)
        ),
        dl.MRRCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.MAPCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.NDCGCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.OptimizerCallback(metric_key="loss"),
        dl.SchedulerCallback(),
        dl.CheckpointCallback(
            logdir="./logs", loader_key="valid", metric_key="loss", minimize=True
        ),
    ]
)

Note

Please follow the minimal examples sections for more use cases.

RecSys – MAPCallback ¶

class catalyst.callbacks.metrics.recsys.MAPCallback(input_key: str, target_key: str, topk_args: List[int] = None, log_on_batch: bool = True, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.BatchMetricCallback

MAP metric callback. Computes MAP@topk for the specified values of topk.

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
prefix – key for the metric’s name
topk_args – specifies which MAP@K to log: [1] - MAP [1, 3] - MAP at 1 and 3 [1, 3, 5] - MAP at 1, 3 and 5
log_on_batch – boolean flag to log computed metrics every batch
prefix – metric prefix
suffix – metric suffix

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_users, num_features, num_items = int(1e4), int(1e1), 10
X = torch.rand(num_users, num_features)
y = (torch.rand(num_users, num_items) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_items)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        dl.BatchTransformCallback(
            transform=torch.sigmoid,
            scope="on_batch_end",
            input_key="logits",
            output_key="scores"
        ),
        dl.CriterionCallback(
            input_key="logits", target_key="targets", metric_key="loss"
        ),
        dl.AUCCallback(input_key="scores", target_key="targets"),
        dl.HitrateCallback(
            input_key="scores", target_key="targets", topk_args=(1, 3, 5)
        ),
        dl.MRRCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.MAPCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.NDCGCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.OptimizerCallback(metric_key="loss"),
        dl.SchedulerCallback(),
        dl.CheckpointCallback(
            logdir="./logs", loader_key="valid", metric_key="loss", minimize=True
        ),
    ]
)

Note

Please follow the minimal examples sections for more use cases.

RecSys – MRRCallback ¶

class catalyst.callbacks.metrics.recsys.MRRCallback(input_key: str, target_key: str, topk_args: List[int] = None, log_on_batch: bool = True, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.BatchMetricCallback

MRR metric callback. Computes MRR@topk for the specified values of topk.

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
prefix – key for the metric’s name
topk_args – specifies which MRR@K to log: [1] - MRR [1, 3] - MRR at 1 and 3 [1, 3, 5] - MRR at 1, 3 and 5
log_on_batch – boolean flag to log computed metrics every batch
prefix – metric prefix
suffix – metric suffix

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_users, num_features, num_items = int(1e4), int(1e1), 10
X = torch.rand(num_users, num_features)
y = (torch.rand(num_users, num_items) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_items)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        dl.BatchTransformCallback(
            transform=torch.sigmoid,
            scope="on_batch_end",
            input_key="logits",
            output_key="scores"
        ),
        dl.CriterionCallback(
            input_key="logits", target_key="targets", metric_key="loss"
        ),
        dl.AUCCallback(input_key="scores", target_key="targets"),
        dl.HitrateCallback(
            input_key="scores", target_key="targets", topk_args=(1, 3, 5)
        ),
        dl.MRRCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.MAPCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.NDCGCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.OptimizerCallback(metric_key="loss"),
        dl.SchedulerCallback(),
        dl.CheckpointCallback(
            logdir="./logs", loader_key="valid", metric_key="loss", minimize=True
        ),
    ]
)

Note

Please follow the minimal examples sections for more use cases.

RecSys – NDCGCallback ¶

class catalyst.callbacks.metrics.recsys.NDCGCallback(input_key: str, target_key: str, topk_args: List[int] = None, log_on_batch: bool = True, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.BatchMetricCallback

NDCG metric callback. Computes NDCG@topk for the specified values of topk.

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
prefix – key for the metric’s name
topk_args – specifies which NDCG@K to log: [1] - NDCG [1, 3] - NDCG at 1 and 3 [1, 3, 5] - NDCG at 1, 3 and 5
log_on_batch – boolean flag to log computed metrics every batch
prefix – metric prefix
suffix – metric suffix

Examples:

import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl

# sample data
num_users, num_features, num_items = int(1e4), int(1e1), 10
X = torch.rand(num_users, num_features)
y = (torch.rand(num_users, num_items) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_items)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        dl.BatchTransformCallback(
            transform=torch.sigmoid,
            scope="on_batch_end",
            input_key="logits",
            output_key="scores"
        ),
        dl.CriterionCallback(
            input_key="logits", target_key="targets", metric_key="loss"
        ),
        dl.AUCCallback(input_key="scores", target_key="targets"),
        dl.HitrateCallback(
            input_key="scores", target_key="targets", topk_args=(1, 3, 5)
        ),
        dl.MRRCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.MAPCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.NDCGCallback(input_key="scores", target_key="targets", topk_args=(1, 3, 5)),
        dl.OptimizerCallback(metric_key="loss"),
        dl.SchedulerCallback(),
        dl.CheckpointCallback(
            logdir="./logs", loader_key="valid", metric_key="loss", minimize=True
        ),
    ]
)

Note

Please follow the minimal examples sections for more use cases.

Segmentation – DiceCallback ¶

class catalyst.callbacks.metrics.segmentation.DiceCallback(input_key: str, target_key: str, class_dim: int = 1, weights: Optional[List[float]] = None, class_names: Optional[List[str]] = None, threshold: Optional[float] = None, log_on_batch: bool = True, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.BatchMetricCallback

Dice metric callback.

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
class_dim – indicates class dimension (K) for outputs and targets tensors (default = 1)
weights – class weights
class_names – class names
threshold – threshold for outputs binarization
log_on_batch – boolean flag to log computed metrics every batch
prefix – metric prefix
suffix – metric suffix

Examples:

import os
import torch
from torch import nn
from torch.utils.data import DataLoader
from catalyst import dl
from catalyst.data import ToTensor
from catalyst.contrib.datasets import MNIST
from catalyst.contrib.nn import IoULoss


model = nn.Sequential(
    nn.Conv2d(1, 1, 3, 1, 1), nn.ReLU(),
    nn.Conv2d(1, 1, 3, 1, 1), nn.Sigmoid(),
)
criterion = IoULoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.02)

loaders = {
    "train": DataLoader(
        MNIST(os.getcwd(), train=True, download=True, transform=ToTensor()),
        batch_size=32
    ),
    "valid": DataLoader(
        MNIST(os.getcwd(), train=False, download=True, transform=ToTensor()),
        batch_size=32
    ),
}

class CustomRunner(dl.SupervisedRunner):
    def handle_batch(self, batch):
        x = batch[self._input_key]
        x_noise = (x + torch.rand_like(x)).clamp_(0, 1)
        x_ = self.model(x_noise)
        self.batch = {self._input_key: x, self._output_key: x_, self._target_key: x}

runner = CustomRunner(
    input_key="features", output_key="scores", target_key="targets", loss_key="loss"
)
# model training
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    loaders=loaders,
    num_epochs=1,
    callbacks=[
        dl.IOUCallback(input_key="scores", target_key="targets"),
        dl.DiceCallback(input_key="scores", target_key="targets"),
        dl.TrevskyCallback(input_key="scores", target_key="targets", alpha=0.2),
    ],
    logdir="./logdir",
    valid_loader="valid",
    valid_metric="loss",
    minimize_valid_metric=True,
    verbose=True,
)

Note

Please follow the minimal examples sections for more use cases.

Segmentation – IOUCallback ¶

class catalyst.callbacks.metrics.segmentation.IOUCallback(input_key: str, target_key: str, class_dim: int = 1, weights: Optional[List[float]] = None, class_names: Optional[List[str]] = None, threshold: Optional[float] = None, log_on_batch: bool = True, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.BatchMetricCallback

IOU metric callback.

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
class_dim – indicates class dimension (K) for outputs and targets tensors (default = 1)
weights – class weights
class_names – class names
threshold – threshold for outputs binarization
log_on_batch – boolean flag to log computed metrics every batch
prefix – metric prefix
suffix – metric suffix

Examples:

import os
import torch
from torch import nn
from torch.utils.data import DataLoader
from catalyst import dl
from catalyst.data import ToTensor
from catalyst.contrib.datasets import MNIST
from catalyst.contrib.nn import IoULoss


model = nn.Sequential(
    nn.Conv2d(1, 1, 3, 1, 1), nn.ReLU(),
    nn.Conv2d(1, 1, 3, 1, 1), nn.Sigmoid(),
)
criterion = IoULoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.02)

loaders = {
    "train": DataLoader(
        MNIST(os.getcwd(), train=True, download=True, transform=ToTensor()),
        batch_size=32
    ),
    "valid": DataLoader(
        MNIST(os.getcwd(), train=False, download=True, transform=ToTensor()),
        batch_size=32
    ),
}

class CustomRunner(dl.SupervisedRunner):
    def handle_batch(self, batch):
        x = batch[self._input_key]
        x_noise = (x + torch.rand_like(x)).clamp_(0, 1)
        x_ = self.model(x_noise)
        self.batch = {self._input_key: x, self._output_key: x_, self._target_key: x}

runner = CustomRunner(
    input_key="features", output_key="scores", target_key="targets", loss_key="loss"
)
# model training
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    loaders=loaders,
    num_epochs=1,
    callbacks=[
        dl.IOUCallback(input_key="scores", target_key="targets"),
        dl.DiceCallback(input_key="scores", target_key="targets"),
        dl.TrevskyCallback(input_key="scores", target_key="targets", alpha=0.2),
    ],
    logdir="./logdir",
    valid_loader="valid",
    valid_metric="loss",
    minimize_valid_metric=True,
    verbose=True,
)

Note

Please follow the minimal examples sections for more use cases.

Segmentation – TrevskyCallback ¶

class catalyst.callbacks.metrics.segmentation.TrevskyCallback(input_key: str, target_key: str, alpha: float, beta: Optional[float] = None, class_dim: int = 1, weights: Optional[List[float]] = None, class_names: Optional[List[str]] = None, threshold: Optional[float] = None, log_on_batch: bool = True, prefix: str = None, suffix: str = None)[source]¶

Bases: catalyst.callbacks.metric.BatchMetricCallback

Trevsky metric callback.

Parameters

input_key – input key to use for metric calculation, specifies our y_pred
target_key – output key to use for metric calculation, specifies our y_true
alpha – false negative coefficient, bigger alpha bigger penalty for false negative. if beta is None, alpha must be in (0, 1)
beta – false positive coefficient, bigger alpha bigger penalty for false positive. Must be in (0, 1), if None beta = (1 - alpha)
class_dim – indicates class dimension (K) for outputs and targets tensors (default = 1)
weights – class weights
class_names – class names
threshold – threshold for outputs binarization
log_on_batch – boolean flag to log computed metrics every batch
prefix – metric prefix
suffix – metric suffix

Examples:

import os
import torch
from torch import nn
from torch.utils.data import DataLoader
from catalyst import dl
from catalyst.data import ToTensor
from catalyst.contrib.datasets import MNIST
from catalyst.contrib.nn import IoULoss


model = nn.Sequential(
    nn.Conv2d(1, 1, 3, 1, 1), nn.ReLU(),
    nn.Conv2d(1, 1, 3, 1, 1), nn.Sigmoid(),
)
criterion = IoULoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.02)

loaders = {
    "train": DataLoader(
        MNIST(os.getcwd(), train=True, download=True, transform=ToTensor()),
        batch_size=32
    ),
    "valid": DataLoader(
        MNIST(os.getcwd(), train=False, download=True, transform=ToTensor()),
        batch_size=32
    ),
}

class CustomRunner(dl.SupervisedRunner):
    def handle_batch(self, batch):
        x = batch[self._input_key]
        x_noise = (x + torch.rand_like(x)).clamp_(0, 1)
        x_ = self.model(x_noise)
        self.batch = {self._input_key: x, self._output_key: x_, self._target_key: x}

runner = CustomRunner(
    input_key="features", output_key="scores", target_key="targets", loss_key="loss"
)
# model training
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    loaders=loaders,
    num_epochs=1,
    callbacks=[
        dl.IOUCallback(input_key="scores", target_key="targets"),
        dl.DiceCallback(input_key="scores", target_key="targets"),
        dl.TrevskyCallback(input_key="scores", target_key="targets", alpha=0.2),
    ],
    logdir="./logdir",
    valid_loader="valid",
    valid_metric="loss",
    minimize_valid_metric=True,
    verbose=True,
)

Note

Please follow the minimal examples sections for more use cases.

SklearnBatchCallback ¶

class catalyst.callbacks.metrics.scikit_learn.SklearnBatchCallback(keys: Mapping[str, Any], metric_fn: Union[Callable, str], metric_key: str, log_on_batch: bool = True, **metric_kwargs)[source]¶

Bases: catalyst.callbacks.metric.FunctionalBatchMetricCallback

SklearnBatchCallback implements an integration of batch-based Sklearn metrics

Parameters

keys – a dictionary containing: a mapping between metric_fn arguments and keys in runner.batch other arguments needed for metric_fn
metric_fn – metric function that gets outputs, targets, and other arguments given in keys and returns score
metric_key – key to store computed metric in runner.batch_metrics dictionary
log_on_batch – boolean flag to log computed metrics every batch
metric_kwargs – additional parameters for metric_fn

Note

catalyst[ml] required for this callback

Examples:

import sklearn
import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl
from functools import partial

# sample data
num_samples, num_features, num_classes = int(1e4), int(1e1), 4
X = torch.rand(num_samples, num_features)
y = (torch.rand(num_samples, num_classes) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_classes)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        dl.BatchTransformCallback(
            input_key="targets",
            output_key="labels",
            transform=partial(torch.argmax, dim=1),
            scope="on_batch_end",
        ),
        dl.BatchTransformCallback(
            input_key="logits",
            output_key="scores",
            transform=partial(torch.softmax, dim=1),
            scope="on_batch_end",
        ),
        dl.BatchTransformCallback(
            input_key="scores",
            output_key="preds",
            transform=partial(torch.argmax, dim=1),
            scope="on_batch_end",
        ),
        dl.MultilabelAccuracyCallback(
            input_key="logits", target_key="targets", threshold=0.5
        ),
        dl.SklearnBatchCallback(
            keys={"y_pred": "preds", "y_true": "labels"},
            metric_fn="f1_score",
            metric_key="sk_f1",
            average="macro",
            zero_division=1,
        )
    ]
)

Note

Please follow the minimal examples sections for more use cases.

SklearnLoaderCallback ¶

class catalyst.callbacks.metrics.scikit_learn.SklearnLoaderCallback(keys: Mapping[str, Any], metric_fn: Union[Callable, str], metric_key: str, **metric_kwargs)[source]¶

Bases: catalyst.callbacks.metric.LoaderMetricCallback

SklearnLoaderCallback implements an integration of loader-based Sklearn metrics

Parameters

keys – a mapping between metric_fn arguments and keys in runner.batch
metric_fn – metric function that gets outputs, targets, and other arguments given in keys and returns score
metric_key – key to store computed metric in runner.batch_metrics dictionary
metric_kwargs – additional parameters for metric_fn

Note

catalyst[ml] required for this callback

Examples:

import sklearn
import torch
from torch.utils.data import DataLoader, TensorDataset
from catalyst import dl
from functools import partial

# sample data
num_samples, num_features, num_classes = int(1e4), int(1e1), 4
X = torch.rand(num_samples, num_features)
y = (torch.rand(num_samples, num_classes) > 0.5).to(torch.float32)

# pytorch loaders
dataset = TensorDataset(X, y)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

# model, criterion, optimizer, scheduler
model = torch.nn.Linear(num_features, num_classes)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters())
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [2])

# model training
runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        dl.BatchTransformCallback(
            input_key="targets",
            output_key="labels",
            transform=partial(torch.argmax, dim=1),
            scope="on_batch_end",
        ),
        dl.BatchTransformCallback(
            input_key="logits",
            output_key="scores",
            transform=partial(torch.softmax, dim=1),
            scope="on_batch_end",
        ),
        dl.BatchTransformCallback(
            input_key="scores",
            output_key="preds",
            transform=partial(torch.argmax, dim=1),
            scope="on_batch_end",
        ),
        dl.MultilabelAccuracyCallback(
            input_key="logits", target_key="targets", threshold=0.5
        ),
        dl.SklearnLoaderCallback(
            keys={"y_score": "scores", "y_true": "labels"},
            metric_fn="roc_auc_score",
            metric_key="roc_auc_score",
            average="macro",
            multi_class="ovo"
        )
    ]
)

Note

Please follow the minimal examples sections for more use cases.

SklearnModelCallback ¶

class catalyst.callbacks.sklearn_model.SklearnModelCallback(feature_key: str, target_key: Optional[str], train_loader: str, valid_loaders: Union[str, List[str]], model_fn: Union[Callable, str], predict_method: str = 'predict', predict_key: str = 'sklearn_predict', **model_kwargs)[source]¶

Bases: catalyst.core.callback.Callback

Callback to train a classifier on the train loader and to give predictions on the valid loader.

Parameters

feature_key – keys of tensors that should be used as features in the classifier calculations
target_key – keys of tensors that should be used as targets in the classifier calculations
train_loader – train loader name
valid_loaders – valid loaders where model should be predicted
model_fn – fabric to produce objects with .fit and predict method
predict_method – predict method name for the classifier
predict_key – key to store computed classifier predicts in runner.batch dictionary
model_kwargs – additional parameters for model_fn

Note

catalyst[ml] required for this callback

Examples

import os

from sklearn.linear_model import LogisticRegression
from torch.optim import Adam
from torch.utils.data import DataLoader

from catalyst import data, dl
from catalyst.contrib import datasets, models, nn
from catalyst.data.transforms import Compose, Normalize, ToTensor

# 1. train and valid loaders
transforms = Compose([ToTensor(), Normalize((0.1307,), (0.3081,))])

train_dataset = datasets.MnistMLDataset(
    root=os.getcwd(),
    download=True,
    transform=transforms
)
sampler = data.BalanceBatchSampler(labels=train_dataset.get_labels(), p=5, k=10)
train_loader = DataLoader(
    dataset=train_dataset,
    sampler=sampler,
    batch_size=sampler.batch_size)

valid_dataset = datasets.MNIST(root=os.getcwd(), transform=transforms, train=False)
valid_loader = DataLoader(dataset=valid_dataset, batch_size=1024)

# 2. model and optimizer
model = models.MnistSimpleNet(out_features=16)
optimizer = Adam(model.parameters(), lr=0.001)

# 3. criterion with triplets sampling
sampler_inbatch = data.HardTripletsSampler(norm_required=False)
criterion = nn.TripletMarginLossWithSampler(
    margin=0.5,
    sampler_inbatch=sampler_inbatch
)

# 4. training with catalyst Runner
class CustomRunner(dl.SupervisedRunner):
    def handle_batch(self, batch) -> None:
        images, targets = batch["features"].float(), batch["targets"].long()
        features = self.model(images)
        self.batch = {
            "embeddings": features,
            "targets": targets,
        }

callbacks = [
    dl.ControlFlowCallback(
        dl.CriterionCallback(
            input_key="embeddings",
            target_key="targets",
            metric_key="loss"),
        loaders="train",
    ),
    dl.SklearnModelCallback(
        feature_key="embeddings",
        target_key="targets",
        train_loader="train",
        valid_loaders="valid",
        model_fn=LogisticRegression,
        predict_method="predict_proba",
        predict_key="sklearn_predict"
    ),
    dl.ControlFlowCallback(
        dl.AccuracyCallback(
            target_key="targets", input_key="sklearn_predict", topk_args=(1, 3)
        ),
        loaders="valid"
    )
]

runner = CustomRunner(input_key="features", output_key="embeddings")
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    callbacks=callbacks,
    loaders={"train": train_loader, "valid": valid_loader},
    verbose=False,
    logdir="./logs",
    valid_loader="valid",
    valid_metric="accuracy",
    minimize_valid_metric=False,
    num_epochs=100,
)

import os

from torch.optim import Adam
from torch.utils.data import DataLoader

from catalyst import data, dl
from catalyst.contrib import datasets, models, nn
from catalyst.data.transforms import Compose, Normalize, ToTensor

# 1. train and valid loaders
transforms = Compose([ToTensor(), Normalize((0.1307,), (0.3081,))])

train_dataset = datasets.MnistMLDataset(
    root=os.getcwd(),
    download=True,
    transform=transforms
)
sampler = data.BalanceBatchSampler(labels=train_dataset.get_labels(), p=5, k=10)
train_loader = DataLoader(
    dataset=train_dataset,
    sampler=sampler,
    batch_size=sampler.batch_size)

valid_dataset = datasets.MNIST(root=os.getcwd(), transform=transforms, train=False)
valid_loader = DataLoader(dataset=valid_dataset, batch_size=1024)

# 2. model and optimizer
model = models.MnistSimpleNet(out_features=16)
optimizer = Adam(model.parameters(), lr=0.001)

# 3. criterion with triplets sampling
sampler_inbatch = data.HardTripletsSampler(norm_required=False)
criterion = nn.TripletMarginLossWithSampler(
    margin=0.5,
    sampler_inbatch=sampler_inbatch
)

# 4. training with catalyst Runner
class CustomRunner(dl.SupervisedRunner):
    def handle_batch(self, batch) -> None:
        images, targets = batch["features"].float(), batch["targets"].long()
        features = self.model(images)
        self.batch = {
            "embeddings": features,
            "targets": targets,
        }

callbacks = [
    dl.ControlFlowCallback(
        dl.CriterionCallback(
            input_key="embeddings",
            target_key="targets",
            metric_key="loss"),
        loaders="train",
    ),
    dl.SklearnModelCallback(
        feature_key="embeddings",
        target_key="targets",
        train_loader="train",
        valid_loaders="valid",
        model_fn="linear_model.LogisticRegression",
        predict_method="predict_proba",
        predict_key="sklearn_predict"
    ),
    dl.ControlFlowCallback(
        dl.AccuracyCallback(
            target_key="targets", input_key="sklearn_predict", topk_args=(1, 3)
        ),
        loaders="valid"
    )
]

runner = CustomRunner(input_key="features", output_key="embeddings")
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    callbacks=callbacks,
    loaders={"train": train_loader, "valid": valid_loader},
    verbose=False,
    logdir="./logs",
    valid_loader="valid",
    valid_metric="accuracy",
    minimize_valid_metric=False,
    num_epochs=100,
)