Multiple input and output keys

Catalyst supports models with multiple input arguments and multiple outputs.

Suppose that we need to train a siamese network. Firstly, need to create a dataset that will yield pairs of images and the same class indicator which later can be used in a contrastive loss.

import cv2
import numpy as np
from torch.utils.data import Dataset

class SiameseDataset(Dataset):
    def __init__(self, images, labels):
        self.images = images
        self.labels = labels

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

    def __getitem__(self, idx):
        original_image = ... # load image using `idx`
        is_same = np.random.uniform() >= 0.5  # use same or opposite class
        if is_same:
            pair_image = ... # load image from the same class and with index != `idx`
        else:
            pair_image = ... # load image from another class
        label = torch.FloatTensor([is_same])
        return {"first": original_image, "second": pair_image, "labels": label}

Do not forget about contrastive loss:

import torch.nn as nn

class ContrastiveLoss(nn.Module):
    def __init__(self, margin=1.0):
        super().__init__()
        self.margin = margin

    def forward(self, l2_distance, labels, **kwargs):
        # ...
        return loss

Suppose you have a model which accepts two tensors - first and second and returns embeddings for input batches and distance between them:

import torch.nn as nn

class SiameseNet(nn.Module):
    def __init__(self, in_features, out_features):
        super().__init__()
        self.layers = nn.Sequential(
            nn.Linear(in_features, in_features * 2),
            nn.ReLU(),
            nn.Linear(in_features * 2, out_features),
        )

    def get_embeddings(self, batch):
        """Generate embeddings for a given batch of images.

        Args:
            batch (torch.Tensor): batch with images,
                expected shapes - [B, C, H, W].

        Returns:
            embeddings (torch.Tensor) for a given batch of images,
                output shapes - [B, out_features].
        """
        return self.layers(batch)


    def forward(self, first, second):
        """Forward pass.

        Args:
            first (torch.Tensor): batch with images,
                expected shapes - [B, C, H, W].
            second (torch.Tensor): batch with images,
                expected shapes - [B, C, H, W].

        Returns:
            embeddings (torch.Tensor) for a first batch of images,
                output shapes - [B, out_features]
            embeddings (torch.Tensor) for a second batch of images,
                output shapes - [B, out_features]
            l2 distance (torch.Tensor) between first and second image embeddings,
                output shapes - [B,]
        """
        first = self.get_embeddings(first)
        second = self.get_embeddings(second)
        difference = torch.sqrt(torch.sum(torch.pow(first - second, 2), 1))
        return first, second, distance

And then for python API:

import torch
from torch import nn, optim
from torch.utils.data import DataLoader
from catalyst import dl

dataset = SiameseDataset(...)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

model = SiameseNet(...)
optimizer = optim.Adam(model.parameters(), lr=1e-3)
criterion = ContrastiveLoss(margin=1.1)

runner = dl.SupervisedRunner(
    input_key=["first", "second"],  # model inputs, should be the same as in forward method
    output_key=["first_emb", "second_emb", "l2_distance"],  # model outputs, part of them will be passed to loss
    target_key=["labels"],  # key from dataset
    loss_key="loss",  # key to use for loss values
)
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    loaders=loaders,
    num_epochs=3,
    callbacks=[
        dl.CriterionCallback(
            input_key="l2_distance", target_key="labels", metric_key="loss"
        ),
    ],
    logdir="./siamese_logs",
    valid_loader="valid",
    valid_metric="loss",
    minimize_valid_metric=True,
    verbose=True,
    load_best_on_end=True,
)