alxmamaev/catalyst

Accelerated Deep Learning R&D

Catalyst is a PyTorch framework for Deep Learning Research and Development. Catalyst focuses on reproducibility, rapid experimentation, and codebase reuse so you can create something new rather than write yet another train loop.
Break the cycle, use Catalyst!

Read more about our vision in the Project Manifest. Catalyst is a part of the PyTorch Ecosystem. Catalyst Ecosystem consists of:

Alchemy - experiments logging & visualization
Catalyst - accelerated deep learning R&D
Reaction - convenient deep learning model serving

Catalyst at AI Landscape

Getting started

pip install -U catalyst

import os
from torch import nn, optim
from torch.utils.data import DataLoader
from catalyst import dl, utils
from catalyst.data.transforms import ToTensor
from catalyst.contrib.datasets import MNIST

model = nn.Sequential(nn.Flatten(), nn.Linear(28 * 28, 10))
criterion = nn.CrossEntropyLoss()
optimizer = 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
    ),
}

runner = dl.SupervisedRunner(
    input_key="features", output_key="logits", target_key="targets", loss_key="loss"
)
# model training
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    loaders=loaders,
    num_epochs=1,
    callbacks=[
        dl.AccuracyCallback(input_key="logits", target_key="targets", topk_args=(1, 3, 5)),
        dl.PrecisionRecallF1SupportCallback(
            input_key="logits", target_key="targets", num_classes=10
        ),
        dl.AUCCallback(input_key="logits", target_key="targets"),
        # catalyst[ml] required ``pip install catalyst[ml]``
        # dl.ConfusionMatrixCallback(input_key="logits", target_key="targets", num_classes=10),
    ],
    logdir="./logs",
    valid_loader="valid",
    valid_metric="loss",
    minimize_valid_metric=True,
    verbose=True,
    load_best_on_end=True,
)
# model inference
for prediction in runner.predict_loader(loader=loaders["valid"]):
    assert prediction["logits"].detach().cpu().numpy().shape[-1] == 10

features_batch = next(iter(loaders["valid"]))[0]
# model stochastic weight averaging
model.load_state_dict(utils.get_averaged_weights_by_path_mask(logdir="./logs", path_mask="*.pth"))
# model tracing
utils.trace_model(model=runner.model, batch=features_batch)
# model quantization
utils.quantize_model(model=runner.model)
# model pruning
utils.prune_model(model=runner.model, pruning_fn="l1_unstructured", amount=0.8)
# onnx export
utils.onnx_export(model=runner.model, batch=features_batch, file="./logs/mnist.onnx", verbose=True)

Step-by-step Guide

Start with Catalyst 101 — Accelerated PyTorch introduction.
Check the minimal examples.
Try notebook tutorials with Google Colab.
Read the blog posts with use-cases and guides.
Learn machine learning with our "Deep Learning with Catalyst" course.
If you would like to contribute to the project, follow our contribution guidelines.
If you want to support the project, feel free to donate on patreon page or write us with your proposals.
And finally, join our slack if you want to contribute or chat about the project.

Overview
Catalyst
- Documentation
- Notebooks
- Blog Posts
- Talks
- Projects
Community

Overview

Catalyst helps you implement compact but full-featured Deep Learning pipelines with just a few lines of code. You get a training loop with metrics, early-stopping, model checkpointing, and other features without the boilerplate.

Installation

Generic installation:

pip install -U catalyst

Specialized versions, extra requirements might apply

pip install catalyst[ml]         # installs ML-based Catalyst
pip install catalyst[cv]         # installs CV-based Catalyst
# master version installation
pip install git+https://github.com/catalyst-team/catalyst@master --upgrade

Catalyst is compatible with: Python 3.6+. PyTorch 1.3+.
Tested on Ubuntu 16.04/18.04/20.04, macOS 10.15, Windows 10, and Windows Subsystem for Linux.

Minimal Examples

MNIST-based examples.

ML - linear regression

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",
    valid_loader="valid",
    valid_metric="loss",
    minimize_valid_metric=True,
    num_epochs=8,
    verbose=True,
)

ML - multiclass classification

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),
        # uncomment for extra metrics:
        # dl.PrecisionRecallF1SupportCallback(
        #     input_key="logits", target_key="targets", num_classes=num_classes
        # ),
        # dl.AUCCallback(input_key="logits", target_key="targets"),
        # catalyst[ml] required
        # dl.ConfusionMatrixCallback(
        #     input_key="logits", target_key="targets", num_classes=num_classes
        # ), 
    ],
)

ML - multilabel classification

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)
    ]
)

ML - multihead classification

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

# sample data
num_samples, num_features, num_classes1, num_classes2 = int(1e4), int(1e1), 4, 10
X = torch.rand(num_samples, num_features)
y1 = (torch.rand(num_samples,) * num_classes1).to(torch.int64)
y2 = (torch.rand(num_samples,) * num_classes2).to(torch.int64)

# pytorch loaders
dataset = TensorDataset(X, y1, y2)
loader = DataLoader(dataset, batch_size=32, num_workers=1)
loaders = {"train": loader, "valid": loader}

class CustomModule(nn.Module):
    def __init__(self, in_features: int, out_features1: int, out_features2: int):
        super().__init__()
        self.shared = nn.Linear(in_features, 128)
        self.head1 = nn.Linear(128, out_features1)
        self.head2 = nn.Linear(128, out_features2)
    
    def forward(self, x):
        x = self.shared(x)
        y1 = self.head1(x)
        y2 = self.head2(x)
        return y1, y2

# model, criterion, optimizer, scheduler
model = CustomModule(num_features, num_classes1, num_classes2)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters())
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, [2])

class CustomRunner(dl.Runner):
    def handle_batch(self, batch):
        x, y1, y2 = batch
        y1_hat, y2_hat = self.model(x)
        self.batch = {
            "features": x,
            "logits1": y1_hat,
            "logits2": y2_hat,
            "targets1": y1,
            "targets2": y2,
        }

# model training
runner = CustomRunner()
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        dl.CriterionCallback(metric_key="loss1", input_key="logits1", target_key="targets1"),
        dl.CriterionCallback(metric_key="loss2", input_key="logits2", target_key="targets2"),
        dl.MetricAggregationCallback(prefix="loss", metrics=["loss1", "loss2"], mode="mean"),
        dl.OptimizerCallback(metric_key="loss"),
        dl.SchedulerCallback(),
        dl.AccuracyCallback(
            input_key="logits1", target_key="targets1", num_classes=num_classes1, prefix="one_"
        ),
        dl.AccuracyCallback(
            input_key="logits2", target_key="targets2", num_classes=num_classes2, prefix="two_"
        ),
        # catalyst[ml] required
        # dl.ConfusionMatrixCallback(
        #     input_key="logits1", target_key="targets1", num_classes=num_classes1, prefix="one_cm"
        # ), 
        # dl.ConfusionMatrixCallback(
        #     input_key="logits2", target_key="targets2", num_classes=num_classes2, prefix="two_cm"
        # ),
        dl.CheckpointCallback(
            logdir="./logs/one", 
            loader_key="valid", metric_key="one_accuracy", minimize=False, save_n_best=1
        ),
        dl.CheckpointCallback(
            logdir="./logs/two", 
            loader_key="valid", metric_key="two_accuracy03", minimize=False, save_n_best=3
        ),
    ],
    loggers={"console": dl.ConsoleLogger(), "tb": dl.TensorboardLogger("./logs/tb")},
)

ML – RecSys

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])

class CustomRunner(dl.Runner):
    def handle_batch(self, batch):
        x, y = batch
        logits = self.model(x)
        self.batch = {
            "features": x, "logits": logits, "scores": torch.sigmoid(logits), "targets": y
        }

# model training
runner = CustomRunner()
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    loaders=loaders,
    num_epochs=3,
    verbose=True,
    callbacks=[
        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
        ),
    ]
)

CV - MNIST classification

import os
from torch import nn, optim
from torch.utils.data import DataLoader
from catalyst import dl
from catalyst.data.transforms import ToTensor
from catalyst.contrib.datasets import MNIST

model = nn.Sequential(nn.Flatten(), nn.Linear(28 * 28, 10))
criterion = nn.CrossEntropyLoss()
optimizer = 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
    ),
}

runner = dl.SupervisedRunner()
# model training
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    loaders=loaders,
    num_epochs=1,
    logdir="./logs",
    valid_loader="valid",
    valid_metric="loss",
    minimize_valid_metric=True,
    verbose=True,
# uncomment for extra metrics:
#     callbacks=[
#         dl.AccuracyCallback(input_key="logits", target_key="targets", num_classes=10),
#         dl.PrecisionRecallF1SupportCallback(
#             input_key="logits", target_key="targets", num_classes=10
#         ),
#         dl.AUCCallback(input_key="logits", target_key="targets"),
#         # catalyst[ml] required ``pip install catalyst[ml]``
#         dl.ConfusionMatrixCallback(
#             input_key="logits", target_key="targets", num_classes=num_classes
#         ), 
#     ]
)

CV - MNIST segmentation

import os
import torch
from torch import nn
from torch.utils.data import DataLoader
from catalyst import dl
from catalyst.data.transforms 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,
)

CV - MNIST model distillation

import os
import torch
from torch import nn, optim
from torch.nn import functional as F
from torch.utils.data import DataLoader
from catalyst import dl
from catalyst.data.transforms import ToTensor
from catalyst.contrib.datasets import MNIST

# [!] teacher model should be already pretrained
teacher = nn.Sequential(nn.Flatten(), nn.Linear(28 * 28, 10))
student = nn.Sequential(nn.Flatten(), nn.Linear(28 * 28, 10))
criterion = {"cls": nn.CrossEntropyLoss(), "kl": nn.KLDivLoss(reduction="batchmean")}
optimizer = optim.Adam(student.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 DistilRunner(dl.Runner):
    def handle_batch(self, batch):
        x, y = batch

        teacher.eval()  # let's manually set teacher model to eval mode
        with torch.no_grad():
            t_logits = self.model["teacher"](x)

        s_logits = self.model["student"](x)
        self.batch = {
            "t_logits": t_logits, "s_logits": s_logits, "targets": y,
            "s_logprobs": F.log_softmax(s_logits, dim=-1), "t_probs": F.softmax(t_logits, dim=-1)
        }

runner = DistilRunner()
callbacks = [
    dl.AccuracyCallback(
        input_key="t_logits", target_key="targets", num_classes=2, prefix="teacher_"
    ),
    dl.AccuracyCallback(
        input_key="s_logits", target_key="targets", num_classes=2, prefix="student_"
    ),
    dl.CriterionCallback(
        input_key="s_logits", target_key="targets", metric_key="cls_loss", criterion_key="cls"
    ),
    dl.CriterionCallback(
        input_key="s_logprobs", target_key="t_probs", metric_key="kl_div_loss", criterion_key="kl"
    ),
    dl.MetricAggregationCallback(prefix="loss", metrics=["kl_div_loss", "cls_loss"], mode="mean"),
    dl.OptimizerCallback(metric_key="loss", model_key="student"),
    dl.CheckpointCallback(
        logdir="./logs", loader_key="valid", metric_key="loss", minimize=True, save_n_best=3
    ),
]
# model training
runner.train(
    model={"teacher": teacher, "student": student},
    criterion=criterion,
    optimizer=optimizer,
    loaders=loaders,
    num_epochs=1,
    logdir="./logs",
    verbose=True,
    callbacks=callbacks,
)

CV - MNIST metric learning

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,
)

CV - MNIST GAN

import os
import torch
from torch import nn
from torch.utils.data import DataLoader
from catalyst import dl
from catalyst.contrib.datasets import MNIST
from catalyst.contrib.nn.modules import Flatten, GlobalMaxPool2d, Lambda
from catalyst.data.transforms import ToTensor

latent_dim = 128
generator = nn.Sequential(
    # We want to generate 128 coefficients to reshape into a 7x7x128 map
    nn.Linear(128, 128 * 7 * 7),
    nn.LeakyReLU(0.2, inplace=True),
    Lambda(lambda x: x.view(x.size(0), 128, 7, 7)),
    nn.ConvTranspose2d(128, 128, (4, 4), stride=(2, 2), padding=1),
    nn.LeakyReLU(0.2, inplace=True),
    nn.ConvTranspose2d(128, 128, (4, 4), stride=(2, 2), padding=1),
    nn.LeakyReLU(0.2, inplace=True),
    nn.Conv2d(128, 1, (7, 7), padding=3),
    nn.Sigmoid(),
)
discriminator = nn.Sequential(
    nn.Conv2d(1, 64, (3, 3), stride=(2, 2), padding=1),
    nn.LeakyReLU(0.2, inplace=True),
    nn.Conv2d(64, 128, (3, 3), stride=(2, 2), padding=1),
    nn.LeakyReLU(0.2, inplace=True),
    GlobalMaxPool2d(),
    Flatten(),
    nn.Linear(128, 1),
)

model = {"generator": generator, "discriminator": discriminator}
criterion = {"generator": nn.BCEWithLogitsLoss(), "discriminator": nn.BCEWithLogitsLoss()}
optimizer = {
    "generator": torch.optim.Adam(generator.parameters(), lr=0.0003, betas=(0.5, 0.999)),
    "discriminator": torch.optim.Adam(discriminator.parameters(), lr=0.0003, betas=(0.5, 0.999)),
}
loaders = {
    "train": DataLoader(
        MNIST(os.getcwd(), train=False, download=True, transform=ToTensor()), batch_size=32
    )
}

class CustomRunner(dl.Runner):
    def predict_batch(self, batch):
        batch_size = 1
        # Sample random points in the latent space
        random_latent_vectors = torch.randn(batch_size, latent_dim).to(self.device)
        # Decode them to fake images
        generated_images = self.model["generator"](random_latent_vectors).detach()
        return generated_images

    def handle_batch(self, batch):
        real_images, _ = batch
        batch_size = real_images.shape[0]

        # Sample random points in the latent space
        random_latent_vectors = torch.randn(batch_size, latent_dim).to(self.device)

        # Decode them to fake images
        generated_images = self.model["generator"](random_latent_vectors).detach()
        # Combine them with real images
        combined_images = torch.cat([generated_images, real_images])

        # Assemble labels discriminating real from fake images
        labels = \
            torch.cat([torch.ones((batch_size, 1)), torch.zeros((batch_size, 1))]).to(self.device)
        # Add random noise to the labels - important trick!
        labels += 0.05 * torch.rand(labels.shape).to(self.device)

        # Discriminator forward
        combined_predictions = self.model["discriminator"](combined_images)

        # Sample random points in the latent space
        random_latent_vectors = torch.randn(batch_size, latent_dim).to(self.device)
        # Assemble labels that say "all real images"
        misleading_labels = torch.zeros((batch_size, 1)).to(self.device)

        # Generator forward
        generated_images = self.model["generator"](random_latent_vectors)
        generated_predictions = self.model["discriminator"](generated_images)

        self.batch = {
            "combined_predictions": combined_predictions,
            "labels": labels,
            "generated_predictions": generated_predictions,
            "misleading_labels": misleading_labels,
        }


runner = CustomRunner()
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    loaders=loaders,
    callbacks=[
        dl.CriterionCallback(
            input_key="combined_predictions",
            target_key="labels",
            metric_key="loss_discriminator",
            criterion_key="discriminator",
        ),
        dl.CriterionCallback(
            input_key="generated_predictions",
            target_key="misleading_labels",
            metric_key="loss_generator",
            criterion_key="generator",
        ),
        dl.OptimizerCallback(
            model_key="generator", 
            optimizer_key="generator", 
            metric_key="loss_generator"
        ),
        dl.OptimizerCallback(
            model_key="discriminator", 
            optimizer_key="discriminator", 
            metric_key="loss_discriminator"
        ),
    ],
    valid_loader="train",
    valid_metric="loss_generator",
    minimize_valid_metric=True,
    num_epochs=20,
    verbose=True,
    logdir="./logs_gan",
)

# visualization (matplotlib required):
# import matplotlib.pyplot as plt
# %matplotlib inline
# plt.imshow(runner.predict_batch(None)[0, 0].cpu().numpy())

CV - MNIST VAE

import os
import torch
from torch import nn, optim
from torch.nn import functional as F
from torch.utils.data import DataLoader
from catalyst import dl, metrics
from catalyst.contrib.datasets import MNIST
from catalyst.data.transforms import ToTensor

LOG_SCALE_MAX = 2
LOG_SCALE_MIN = -10

def normal_sample(loc, log_scale):
    scale = torch.exp(0.5 * log_scale)
    return loc + scale * torch.randn_like(scale)

class VAE(nn.Module):
    def __init__(self, in_features, hid_features):
        super().__init__()
        self.hid_features = hid_features
        self.encoder = nn.Linear(in_features, hid_features * 2)
        self.decoder = nn.Sequential(nn.Linear(hid_features, in_features), nn.Sigmoid())

    def forward(self, x, deterministic=False):
        z = self.encoder(x)
        bs, z_dim = z.shape

        loc, log_scale = z[:, : z_dim // 2], z[:, z_dim // 2 :]
        log_scale = torch.clamp(log_scale, LOG_SCALE_MIN, LOG_SCALE_MAX)

        z_ = loc if deterministic else normal_sample(loc, log_scale)
        z_ = z_.view(bs, -1)
        x_ = self.decoder(z_)

        return x_, loc, log_scale

class CustomRunner(dl.IRunner):
    def __init__(self, logdir, device):
        super().__init__()
        self._logdir = logdir
        self._device = device

    def get_engine(self):
        return dl.DeviceEngine(self._device)

    def get_loggers(self):
        return {
            "console": dl.ConsoleLogger(),
            "csv": dl.CSVLogger(logdir=self._logdir),
            "tensorboard": dl.TensorboardLogger(logdir=self._logdir),
        }

    @property
    def stages(self):
        return ["train"]

    def get_stage_len(self, stage: str) -> int:
        return 3

    def get_loaders(self, stage: str):
        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
            ),
        }
        return loaders

    def get_model(self, stage: str):
        model = self.model if self.model is not None else VAE(28 * 28, 64)
        return model

    def get_optimizer(self, stage: str, model):
        return optim.Adam(model.parameters(), lr=0.02)

    def get_callbacks(self, stage: str):
        return {
            "optimizer": dl.OptimizerCallback(metric_key="loss"),
            "checkpoint": dl.CheckpointCallback(
                self._logdir, loader_key="valid", metric_key="loss", minimize=True
            ),
        }

    def on_loader_start(self, runner):
        super().on_loader_start(runner)
        self.meters = {
            key: metrics.AdditiveValueMetric(compute_on_call=False)
            for key in ["loss_ae", "loss_kld", "loss"]
        }

    def handle_batch(self, batch):
        x, _ = batch
        x = x.view(x.size(0), -1)
        x_, loc, log_scale = self.model(x, deterministic=not self.is_train_loader)

        loss_ae = F.mse_loss(x_, x)
        loss_kld = (-0.5 * torch.sum(1 + log_scale - loc.pow(2) - log_scale.exp(), dim=1)).mean()
        loss = loss_ae + loss_kld * 0.01

        self.batch_metrics = {"loss_ae": loss_ae, "loss_kld": loss_kld, "loss": loss}
        for key in ["loss_ae", "loss_kld", "loss"]:
            self.meters[key].update(self.batch_metrics[key].item(), self.batch_size)

    def on_loader_end(self, runner):
        for key in ["loss_ae", "loss_kld", "loss"]:
            self.loader_metrics[key] = self.meters[key].compute()[0]
        super().on_loader_end(runner)
    
    def predict_batch(self, batch):
        random_latent_vectors = torch.randn(1, self.model.hid_features).to(self.device)
        generated_images = self.model.decoder(random_latent_vectors).detach()
        return generated_images

runner = CustomRunner("./logs", "cpu")
runner.run()
# visualization (matplotlib required):
# import matplotlib.pyplot as plt
# %matplotlib inline
# plt.imshow(runner.predict_batch(None)[0].cpu().numpy().reshape(28, 28))

CV - MNIST multistage finetuning

import os
from torch import nn, optim
from torch.utils.data import DataLoader
from catalyst import dl, utils
from catalyst.contrib.datasets import MNIST
from catalyst.data.transforms import ToTensor


class CustomRunner(dl.IRunner):
    def __init__(self, logdir, device):
        super().__init__()
        self._logdir = logdir
        self._device = device

    def get_engine(self):
        return dl.DeviceEngine(self._device)

    def get_loggers(self):
        return {
            "console": dl.ConsoleLogger(),
            "csv": dl.CSVLogger(logdir=self._logdir),
            "tensorboard": dl.TensorboardLogger(logdir=self._logdir),
        }

    @property
    def stages(self):
        return ["train_freezed", "train_unfreezed"]

    def get_stage_len(self, stage: str) -> int:
        return 3

    def get_loaders(self, stage: str):
        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
            ),
        }
        return loaders

    def get_model(self, stage: str):
        model = (
            self.model
            if self.model is not None
            else nn.Sequential(nn.Flatten(), nn.Linear(784, 128), nn.ReLU(), nn.Linear(128, 10))
        )
        if stage == "train_freezed":
            # freeze layer
            utils.set_requires_grad(model[1], False)
        else:
            utils.set_requires_grad(model, True)
        return model

    def get_criterion(self, stage: str):
        return nn.CrossEntropyLoss()

    def get_optimizer(self, stage: str, model):
        if stage == "train_freezed":
            return optim.Adam(model.parameters(), lr=1e-3)
        else:
            return optim.SGD(model.parameters(), lr=1e-1)

    def get_scheduler(self, stage: str, optimizer):
        return None

    def get_callbacks(self, stage: str):
        return {
            "criterion": dl.CriterionCallback(
                metric_key="loss", input_key="logits", target_key="targets"
            ),
            "optimizer": dl.OptimizerCallback(metric_key="loss"),
            # "scheduler": dl.SchedulerCallback(loader_key="valid", metric_key="loss"),
            # "accuracy": dl.AccuracyCallback(
            #     input_key="logits", target_key="targets", topk_args=(1, 3, 5)
            # ),
            # "classification": dl.PrecisionRecallF1SupportCallback(
            #     input_key="logits", target_key="targets", num_classes=10
            # ),
            # "confusion_matrix": dl.ConfusionMatrixCallback(
            #     input_key="logits", target_key="targets", num_classes=10
            # ),
            "checkpoint": dl.CheckpointCallback(
                self._logdir, loader_key="valid", metric_key="loss", minimize=True, save_n_best=3
            ),
        }

    def handle_batch(self, batch):
        x, y = batch
        logits = self.model(x)

        self.batch = {
            "features": x,
            "targets": y,
            "logits": logits,
        }

runner = CustomRunner("./logs", "cpu")
runner.run()

CV - MNIST multistage finetuning (distributed)

import os
from torch import nn, optim
from torch.utils.data import DataLoader
from catalyst import dl, utils
from catalyst.contrib.datasets import MNIST
from catalyst.data.transforms import ToTensor


class CustomRunner(dl.IRunner):
    def __init__(self, logdir):
        super().__init__()
        self._logdir = logdir

    def get_engine(self):
        return dl.DistributedDataParallelEngine()

    def get_loggers(self):
        return {
            "console": dl.ConsoleLogger(),
            "csv": dl.CSVLogger(logdir=self._logdir),
            "tensorboard": dl.TensorboardLogger(logdir=self._logdir),
        }

    @property
    def stages(self):
        return ["train_freezed", "train_unfreezed"]

    def get_stage_len(self, stage: str) -> int:
        return 3

    def get_loaders(self, stage: str):
        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
            ),
        }
        return loaders

    def get_model(self, stage: str):
        model = nn.Sequential(nn.Flatten(), nn.Linear(784, 128), nn.ReLU(), nn.Linear(128, 10))
        if stage == "train_freezed":  # freeze layer
            utils.set_requires_grad(model[1], False)
        else:
            utils.set_requires_grad(model, True)
        return model

    def get_criterion(self, stage: str):
        return nn.CrossEntropyLoss()

    def get_optimizer(self, stage: str, model):
        if stage == "train_freezed":
            return optim.Adam(model.parameters(), lr=1e-3)
        else:
            return optim.SGD(model.parameters(), lr=1e-1)

    def get_callbacks(self, stage: str):
        return {
            "criterion": dl.CriterionCallback(
                metric_key="loss", input_key="logits", target_key="targets"
            ),
            "optimizer": dl.OptimizerCallback(metric_key="loss"),
            "accuracy": dl.AccuracyCallback(
                input_key="logits", target_key="targets", topk_args=(1, 3, 5)
            ),
            "classification": dl.PrecisionRecallF1SupportCallback(
                input_key="logits", target_key="targets", num_classes=10
            ),
            # catalyst[ml] required
            "confusion_matrix": dl.ConfusionMatrixCallback(
                input_key="logits", target_key="targets", num_classes=10
            ),
            "checkpoint": dl.CheckpointCallback(
                self._logdir,
                loader_key="valid",
                metric_key="loss",
                minimize=True,
                save_n_best=3,
                load_on_stage_start={
                    "model": "best",
                    "global_epoch_step": "last",
                    "global_batch_step": "last",
                    "global_sample_step": "last",
                },
            ),
            "verbose": dl.TqdmCallback(),
        }

    def handle_batch(self, batch):
        x, y = batch
        logits = self.model(x)

        self.batch = {
            "features": x,
            "targets": y,
            "logits": logits,
        }


if __name__ == "__main__":
    runner = CustomRunner("./logs")
    runner.run()

AutoML - hyperparameters optimization with Optuna

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

def objective(trial):
    lr = trial.suggest_loguniform("lr", 1e-3, 1e-1)
    num_hidden = int(trial.suggest_loguniform("num_hidden", 32, 128))

    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
        ),
    }
    model = nn.Sequential(
        nn.Flatten(), nn.Linear(784, num_hidden), nn.ReLU(), nn.Linear(num_hidden, 10)
    )
    optimizer = torch.optim.Adam(model.parameters(), lr=lr)
    criterion = nn.CrossEntropyLoss()

    runner = dl.SupervisedRunner(input_key="features", output_key="logits", target_key="targets")
    runner.train(
        model=model,
        criterion=criterion,
        optimizer=optimizer,
        loaders=loaders,
        callbacks={
            "accuracy": dl.AccuracyCallback(
                input_key="logits", target_key="targets", num_classes=10
            ),
            "optuna": dl.OptunaPruningCallback(
                loader_key="valid", metric_key="accuracy01", minimize=False, trial=trial
            ),
        },
        num_epochs=3,
    )
    score = runner.callbacks["optuna"].best_score
    return score

study = optuna.create_study(
    direction="maximize",
    pruner=optuna.pruners.MedianPruner(
        n_startup_trials=1, n_warmup_steps=0, interval_steps=1
    ),
)
study.optimize(objective, n_trials=3, timeout=300)
print(study.best_value, study.best_params)

Minimal example for Config/Hydra API

Features

Universal train/inference loop.
Configuration files for model and data hyperparameters.
Reproducibility – all source code and environment variables are saved.
Callbacks – reusable train/inference pipeline parts with easy customization.
Training stages support.
Deep Learning best practices: SWA, AdamW, Ranger optimizer, OneCycle, and more.
Workflow best practices: fp16 support, distributed training, slurm support.

Tests

All Catalyst code, features, and pipelines are fully tested. We also have our own catalyst-codestyle.

During testing, we train a variety of different models: image classification, image segmentation, text classification, GANs, and much more. We then compare their convergence metrics in order to verify the correctness of the training procedure and its reproducibility.

As a result, Catalyst provides fully tested and reproducible best practices for your deep learning research and development.

Catalyst

Documentation

Notebooks

Customizing what happens in train
Demo with minimal examples for ML, CV, NLP, GANs and RecSys
Metric Learning Tutorial
Catalyst with Google TPUs

Notable Blog Posts

Talks

Catalyst-team YouTube channel
Catalyst.RL – reproducible RL research framework at Stachka
Catalyst.DL – reproducible DL research framework (rus) and slides (eng) at RIF
Catalyst.DL – reproducible DL research framework (rus) and slides (eng) at AI-Journey
Catalyst.DL – fast & reproducible DL at Datastart
Catalyst.RL - NeurIPS 2019: Learn to Move - Walk Around and slides (eng) at RL reading group Meetup
Catalyst – accelerated DL & RL (rus) and slides (eng) at Facebook Developer Circle: Moscow | ML & AI Meetup
Catalyst.RL - Learn to Move - Walk Around 2nd place solution at NeurIPS competition track
Open Source ML 2019 edition at Datafest.elka

Projects

Examples, Notebooks, and Starter Kits

CamVid Segmentation Example - Example of semantic segmentation for CamVid dataset
Notebook API tutorial for segmentation in Understanding Clouds from Satellite Images Competition
Catalyst.RL - NeurIPS 2019: Learn to Move - Walk Around – starter kit
Catalyst.RL - NeurIPS 2019: Animal-AI Olympics - starter kit
Inria Segmentation Example - An example of training segmentation model for Inria Sattelite Segmentation Challenge
iglovikov_segmentation - Semantic segmentation pipeline using Catalyst

Competitions

Kaggle Quick, Draw! Doodle Recognition Challenge - 11th place
Catalyst.RL - NeurIPS 2018: AI for Prosthetics Challenge – 3rd place
Kaggle Google Landmark 2019 - 30th place
iMet Collection 2019 - FGVC6 - 24th place
ID R&D Anti-spoofing Challenge - 14th place
NeurIPS 2019: Recursion Cellular Image Classification - 4th place
MICCAI 2019: Automatic Structure Segmentation for Radiotherapy Planning Challenge 2019
- 3rd place solution for Task 3: Organ-at-risk segmentation from chest CT scans
- and 4th place solution for Task 4: Gross Target Volume segmentation of lung cancer
Kaggle Seversteal steel detection - 5th place
RSNA Intracranial Hemorrhage Detection - 5th place
APTOS 2019 Blindness Detection – 7th place
Catalyst.RL - NeurIPS 2019: Learn to Move - Walk Around – 2nd place
xView2 Damage Assessment Challenge - 3rd place

Research Papers

Toolkits

Catalyst.RL – A Distributed Framework for Reproducible RL Research by Scitator
Catalyst.Classification - Comprehensive classification pipeline with Pseudo-Labeling by Bagxi and Pdanilov
Catalyst.Segmentation - Segmentation pipelines - binary, semantic and instance, by Bagxi
Catalyst.Detection - Anchor-free detection pipeline by Avi2011class and TezRomacH
Catalyst.GAN - Reproducible GANs pipelines by Asmekal
Catalyst.Neuro - Brain image analysis project, in collaboration with TReNDS Center
MLComp – Distributed DAG framework for machine learning with UI by Lightforever
Pytorch toolbelt - PyTorch extensions for fast R&D prototyping and Kaggle farming by BloodAxe
Helper functions - An assorted collection of helper functions by Ternaus
BERT Distillation with Catalyst by elephantmipt

See other projects at the GitHub dependency graph.

If your project implements a paper, a notable use-case/tutorial, or a Kaggle competition solution, or if your code simply presents interesting results and uses Catalyst, we would be happy to add your project to the list above!
Do not hesitate to send us a PR with a brief description of the project similar to the above.

Community

Contribution Guide

We appreciate all contributions. If you are planning to contribute back bug-fixes, there is no need to run that by us; just send a PR. If you plan to contribute new features, new utility functions, or extensions, please open an issue first and discuss it with us.

Please see the Contribution Guide for more information.
By participating in this project, you agree to abide by its Code of Conduct.

User Feedback

We've created feedback@catalyst-team.com as an additional channel for user feedback.

If you like the project and want to thanks us, this the right place.
If you would like to start a collaboration between your team and Catalyst team to improve Deep Learning R&D, you are always welcome.
If you just don't like Github Issues and this prefer email, feel free to email us.
Finally, if you do not like something, please, share it with us and we can see how to improve it.

We appreciate any type of feedback. Thank you!

Acknowledgments

Since the beginning of the Сatalyst development, a lot of people have influenced it in a lot of different ways.

Catalyst.Team

Catalyst.Contributors

Trusted by

Awecom
Researchers at the Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Deep Learning School
Researchers at Emory University
Evil Martians
Researchers at the Georgia Institute of Technology
Researchers at Georgia State University
Helios
HPCD Lab
iFarm
Kinoplan
Researchers at the Moscow Institute of Physics and Technology
Neuromation
Poteha Labs
Provectus
Researchers at the Skolkovo Institute of Science and Technology
SoftConstruct
Researchers at Tinkoff
Researchers at Yandex.Research

Citation

Please use this bibtex if you want to cite this repository in your publications:

@misc{catalyst,
    author = {Kolesnikov, Sergey},
    title = {Accelerated deep learning R&D},
    year = {2018},
    publisher = {GitHub},
    journal = {GitHub repository},
    howpublished = {\url{https://github.com/catalyst-team/catalyst}},
}

alxmamaev / catalyst