catalyst 21.3

Catalyst. PyTorch framework for DL research and development.

Accelerated deep learning R&D

PyTorch framework for Deep Learning research and development. It focuses on reproducibility, rapid experimentation, and codebase reuse so you can create something new rather than write another regular train loop.
Break the cycle - use the Catalyst!

Project manifest. Part of PyTorch Ecosystem. Part of Catalyst Ecosystem:

• Alchemy - experiments logging & visualization
• Catalyst - accelerated deep learning R&D
• Reaction - convenient deep learning models 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"),
#         dl.ConfusionMatrixCallback(input_key="logits", target_key="targets", num_classes=10),  # catalyst[ml] required
    ],
    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

1. Start with Catalyst 101 — Accelerated PyTorch introduction.
2. Check minimal examples.
3. Try notebook tutorials with Google Colab.
4. Read blogposts with use-cases and guides.
5. Learn machine learning with our "Deep Learning with Catalyst" course.
6. If you would like to contribute to the project, follow our contribution guidelines.
7. If you want to support the project, feel free to donate on patreon page or write us with your proposals.
8. And do not forget to join our slack for collaboration.

Table of Contents

• Overview
  • Installation
  • Minimal examples
  • Features
  • Tests
• Catalyst
  • Documentation
  • Notebooks
  • Blogposts
  • Talks
  • Projects
• Community
  • Contribution guide
  • User feedback
  • Acknowledgments
  • Trusted by
  • Supported by
  • Citation

Overview

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

Installation

Common installation:

pip install -U catalyst

Specific versions with additional requirements

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-driven development.

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,
# uncomment for extra metrics:
    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), # catalyst[ml] required
    ],
)

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_"),
        dl.ConfusionMatrixCallback(input_key="logits1", target_key="targets1", num_classes=num_classes1, prefix="one_cm"), # catalyst[ml] required
        dl.ConfusionMatrixCallback(input_key="logits2", target_key="targets2", num_classes=num_classes2, prefix="two_cm"), # catalyst[ml] required
        dl.CheckpointCallback("./logs/one", loader_key="valid", metric_key="one_accuracy", minimize=False, save_n_best=1),
        dl.CheckpointCallback("./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"),
#         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"),
#         dl.ConfusionMatrixCallback(input_key="logits", target_key="targets", num_classes=num_classes), # catalyst[ml] required
#     ]
)

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 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, epoch: int = None):
        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, epoch: int = None):
        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, epoch: int = None):
        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),
            "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)

Features

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

Tests

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

In fact, we train a number of different models for various of tasks - image classification, image segmentation, text classification, GANs training and much more. During the tests, we 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.

Catalyst

Documentation

• master
• 20.12
• 20.11
• 20.10
• 20.09
• 20.08.2
• 20.07 - dev blog: 20.07 release
• 20.06
• 20.05, 20.05.1
• 20.04, 20.04.1, 20.04.2

Notebooks

• Customizing what happens in train
• Demo with minimal examples for ML, CV, NLP, GANs and RecSys
• Metric Learning tutorial
• Catalyst with Google TPU
• Detailed classification tutorial
• Advanced segmentation tutorial

Blogposts

• Catalyst 101 — Accelerated PyTorch
• Catalyst 102 — Core Trinity
• BERT Distillation with Catalyst
• Metric Learning with Catalyst
• Pruning with Catalyst
• Distributed training best practices
• Addressing the Cocktail Party Problem using PyTorch
• Beyond fashion: Deep Learning with Catalyst (Config API)
• Tutorial from Notebook API to Config API (RU)

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 solution
• Catalyst.RL - NeurIPS 2018: AI for Prosthetics Challenge – 3rd place solution
• Kaggle Google Landmark 2019 - 30th place solution
• iMet Collection 2019 - FGVC6 - 24th place solution
• ID R&D Anti-spoofing Challenge - 14th place solution
• NeurIPS 2019: Recursion Cellular Image Classification - 4th place solution
• 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 solution
• RSNA Intracranial Hemorrhage Detection - 5th place solution
• APTOS 2019 Blindness Detection – 7th place solution
• Catalyst.RL - NeurIPS 2019: Learn to Move - Walk Around – 2nd place solution
• xView2 Damage Assessment Challenge - 3rd place solution

Research papers

• Hierarchical attention for sentiment classification with visualization
• Pediatric bone age assessment
• Implementation of paper "Tell Me Where to Look: Guided Attention Inference Network"
• Implementation of paper "Filter Response Normalization Layer: Eliminating Batch Dependence in the Training of Deep Neural Networks"
• Implementation of paper "Utterance-level Aggregation For Speaker Recognition In The Wild"
• Implementation of paper "Looking to Listen at the Cocktail Party: A Speaker-Independent Audio-Visual Model for Speech Separation"
• Implementation of paper "ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks"

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 unstructured set of helper functions by Ternaus
• BERT Distillation with Catalyst by elephantmipt

See other projects at "Used by".

If your project implements a paper, represents use-case/tutorial, Kaggle competition's code or just your code presents interesting results and uses Catalyst... we would be happy to add your project to the list above, so please send a PR with a brief description of the project.

Community

Contribution guide

We appreciate all contributions. If you are planning to contribute back bug-fixes, please do so without any further discussion. If you plan to contribute new features, utility functions or extensions, please first open an issue and discuss the feature 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 have created [e-mail] for "user feedback".

• If you like the project and want to say thanks, this the right place.
• If you would like to start a collaboration between your team and Catalyst team to do better Deep Learning R&D - you are always welcome.
• If you just don't like Github issues and this ways suits you better - 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 development of the Сatalyst, a lot of people have influenced it in a lot of different ways.

Catalyst.Team

• Eugene Kachan (bagxi)
• Dmytro Doroshenko (ditwoo)
• Sergey Kolesnikov (scitator)

Catalyst - Metric Learning Team

• Julia Shenshina (julia-shenshina)
• Nikita Balagansky (elephantmipt)
• Aleksey Shabanov (AlekseySh)

Catalyst.Contributors

• Aleksey Grinchuk (alexgrinch)
• Alex Gaziev (gazay)
• Andrey Zharkov (asmekal)
• Artem Zolkin (arquestro)
• David Kuryakin (dkuryakin)
• Evgeny Semyonov (lightforever)
• Eugene Khvedchenya (bloodaxe)
• Roman Tezikov (TezRomacH)
• Valentin Khrulkov (khrulkovv)
• Vsevolod Poletaev (hexfaker)
• Yury Kashnitsky (yorko)

Catalyst.Friends

• Ivan Stepanenko
• Nguyen Xuan Bac (ngxbac)
• Vladimir Iglovikov (ternaus)

Trusted by

• Awecom
• Researchers@Center for Translational Research in Neuroimaging and Data Science (TReNDS)
• Deep Learning School
• Researchers@Emory University
• Evil Martians
• Researchers@Georgia Institute of Technology
• Researchers@Georgia State University
• Helios
• HPCD Lab
• iFarm
• Kinoplan
• Researchers@Moscow Institute of Physics and Technology
• Neuromation
• Poteha Labs
• Provectus
• Researchers@Skolkovo Institute of Science and Technology
• SoftConstruct
• Researchers@Tinkoff
• [e-mail]

Supported by

• HostKey
• Moscow Institute of Physics and Technology

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}},
}