Catalyst 19.05 -> 19.06

Overall

• docs, docker and travis update (https://github.com/catalyst-team/catalyst/pull/205, https://github.com/catalyst-team/catalyst/pull/201, https://github.com/catalyst-team/catalyst/pull/203)
• Removed dependency on TensorFlow, new tensorboard logs parsing mechanism (https://github.com/catalyst-team/catalyst/pull/182)

DL

• FP16 & distributed training support, nvidia-apex support (https://github.com/catalyst-team/catalyst/pull/183, https://github.com/catalyst-team/catalyst/pull/207)
• PyTorch auto-tracing support (https://github.com/catalyst-team/catalyst/pull/189)
• logdir autonaming based on baselogdir, timestamp and config hash (https://github.com/catalyst-team/catalyst/pull/204/commits/a141e4c95dde81aa15d0eb0c6daaaf8648df8e20)
• 1Cycle PyTorch scheduler (https://github.com/catalyst-team/catalyst/pull/162)

RL

• on-policy model-free algorithms support, PPO & REINFORCE implementations (https://github.com/catalyst-team/catalyst/pull/174)
• mongo DB support fix (https://github.com/catalyst-team/catalyst/pull/190)
• updated samplers, no trajectory length limit now (https://github.com/catalyst-team/catalyst/pull/196)

Contrib

• segmentation contrib (https://github.com/catalyst-team/catalyst/pull/204)
• losses and metrics refactoring (https://github.com/catalyst-team/catalyst/pull/191)

Improvements and fixes

• optimizer transferring from previous stage support (https://github.com/catalyst-team/catalyst/pull/181)
• additional checkpoint data support (https://github.com/catalyst-team/catalyst/pull/184)
• --version feature (https://github.com/catalyst-team/catalyst/pull/188)

Breaking changes

• all methods with prepare_*, that have returned something, renamed to get_*;
• Experiment._prepare_logdir renamed to Experiment._get_logdir
• UtilsFactory.prepare_models renamed to UtilsFactory.process_components and supports PyTorch model, criterion, optimizer and scheduler now
• Previous segmentation models were moved to catalyst.contrib.models.segmentation.models: ResNetUnet and ResNetLinknet
• per_gpu_batch_size renamed to per_gpu_scaling and affects on batch_size and num_workers now

Features examples

Segmentation

from catalyst.contrib.models.segmentation import \
    Unet, Linknet, FPNUnet, PSPnet, \
    ResnetUnet, ResnetLinknet, ResnetFPNUnet, ResnetPSPnet

Apex

runner = SupervisedRunner()
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    fp16=True
    ...)

or

runner = SupervisedRunner()
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    fp16={"opt_level": "O1"} # and other apex.initalize kwargs
    ...)

Distributed training bash example

#!/usr/bin/env bash

export MASTER_ADDR="127.0.0.1"
export MASTER_PORT=29500
export WORLD_SIZE=2  # number of gpus

RANK=0 LOCAL_RANK=0 catalyst-dl run --config=config.yml --distributed_params/rank=0:int &  # gpu 0
sleep 5
RANK=1 LOCAL_RANK=1 catalyst-dl run --config=config.yml --distributed_params/rank=1:int &  # gpu 1