EarlyStoppingCallback considers first epoch as bad

[alyaxey]

🐛 Bug Report

EarlyStoppingCallback considers first epoch as bad. This can lead for example to always stopping after first epoch if patience=1.

How To Reproduce

You can train a model with early stopping and patience=1 and see that it always stops after first epoch. Or you can use the unit test below that I added to pull request.

Code sample

from unittest.mock import MagicMock, PropertyMock

from catalyst.core import EarlyStoppingCallback


def test_patience1():
    """@TODO: Docs. Contribution is welcome."""
    early_stop = EarlyStoppingCallback(1)
    runner = MagicMock()
    type(runner).stage_name = PropertyMock(return_value="training")
    type(runner).valid_metrics = PropertyMock(return_value={"loss": 0.001})
    stop_mock = PropertyMock(return_value=False)
    type(runner).need_early_stop = stop_mock

    early_stop.on_epoch_end(runner)

    assert stop_mock.mock_calls == []

Expected behavior

Training doesn't stop after first epoch. And the unit test passes.

Environment

Catalyst version: 20.06
PyTorch version: 1.5.1
Is debug build: No
CUDA used to build PyTorch: None
TensorFlow version: N/A
TensorBoard version: 2.2.2

OS: Mac OSX 10.15.5
GCC version: Could not collect
CMake version: version 3.8.0

Python version: 3.7
Is CUDA available: No
CUDA runtime version: No CUDA
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA

Versions of relevant libraries:
[pip3] catalyst-codestyle==20.4
[pip3] catalyst-sphinx-theme==1.1.1
[pip3] efficientnet-pytorch==0.6.3
[pip3] numpy==1.18.5
[pip3] segmentation-models-pytorch==0.1.0
[pip3] tensorboard==2.2.2
[pip3] tensorboard-plugin-wit==1.6.0.post3
[pip3] tensorboardX==2.0
[pip3] torch==1.5.1
[pip3] torchvision==0.6.1
[conda] catalyst-codestyle        20.4                      <pip>
[conda] catalyst-sphinx-theme     1.1.1                     <pip>
[conda] efficientnet-pytorch      0.6.3                     <pip>
[conda] numpy                     1.18.5                    <pip>
[conda] segmentation-models-pytorch 0.1.0                     <pip>
[conda] tensorboard               2.2.2                     <pip>
[conda] tensorboard-plugin-wit    1.6.0.post3               <pip>
[conda] tensorboardX              2.0                       <pip>
[conda] torch                     1.5.1                     <pip>
[conda] torchvision               0.6.1                     <pip>

[github-actions]

Hi! Thank you for your contribution! Great first issue!