python refactoring & project restructuration (#223)

* py3 super

* contrib

* data

* dl WIP

* rl

* dl new structure

* rl utils

* codestyle

* tests

* codestyle

* working notebooks

* update

* now dl works

* rl should work

* init update

* docs update

* doc fix

* modules fix

* no registry docs

* tests update

* imread update

* rl core refactoring

* rl algorithms refactored

* docs update

* example fixes

* tests update

* tests check

* get_loader -> dl

* import fix

.travis.yml

@@ -54,7 +54,7 @@ jobs:
         - pip install tifffile
         - wget -P ./data/ https://www.dropbox.com/s/0rvuae4mj6jn922/isbi.tar.gz
         - tar -xf ./data/isbi.tar.gz -C ./data/
-        - for f in examples/_tests_scripts/*.py; do PYTHONPATH=./catalyst:${PYTHONPATH} python "$f"; done
+        - (set -e; for f in examples/_tests_scripts/*.py; do PYTHONPATH=./catalyst:${PYTHONPATH} python "$f"; done)
         - PYTHONPATH=./examples:./catalyst:${PYTHONPATH}
           python catalyst/dl/scripts/run.py
           --expdir=./examples/_tests_mnist_stages

catalyst/__version__.py

@@ -1 +1 @@
-__version__ = "19.06.2"
+__version__ = "19.06.3"

catalyst/contrib/criterion/__init__.py

@@ -1,11 +1,13 @@
 # flake8: noqa
+
 from torch.nn.modules.loss import *
-from .ce import *
-from .center import *
-from .contrastive import *
-from .dice import *
-from .focal import *
-from .huber import *
-from .iou import *
-from .lovasz import *
-from .wing import *
+from .ce import NaiveCrossEntropyLoss
+from .center import CenterLoss
+from .contrastive import ContrastiveDistanceLoss, ContrastiveEmbeddingLoss
+from .dice import DiceLoss, BCEDiceLoss
+from .focal import FocalLossBinary, FocalLossMultiClass
+from .huber import HuberLoss
+from .iou import IoULoss, BCEIoULoss
+from .lovasz import LovaszLossBinary, LovaszLossMultiClass, \
+    LovaszLossMultiLabel
+from .wing import WingLoss

catalyst/contrib/criterion/dice.py

@@ -1,7 +1,7 @@
 from functools import partial
 
 import torch.nn as nn
-from catalyst.dl import metrics
+from catalyst.dl.utils import criterion
 
 
 class DiceLoss(nn.Module):
@@ -11,10 +11,10 @@ def __init__(
         threshold: float = None,
         activation: str = "Sigmoid"
     ):
-        super(DiceLoss, self).__init__()
+        super().__init__()
 
         self.loss_fn = partial(
-            metrics.dice,
+            criterion.dice,
             eps=eps,
             threshold=threshold,
             activation=activation)

catalyst/contrib/criterion/focal.py

@@ -1,7 +1,7 @@
 from functools import partial
 
 from torch.nn.modules.loss import _Loss
-from catalyst.dl.losses import sigmoid_focal_loss, reduced_focal_loss
+from catalyst.dl.utils import criterion
 
 
 class FocalLossBinary(_Loss):
@@ -22,13 +22,13 @@ def __init__(
 
         if reduced:
             self.loss_fn = partial(
-                reduced_focal_loss,
+                criterion.reduced_focal_loss,
                 gamma=gamma,
                 threshold=threshold,
                 reduction=reduction)
         else:
             self.loss_fn = partial(
-                sigmoid_focal_loss,
+                criterion.sigmoid_focal_loss,
                 gamma=gamma,
                 alpha=alpha,
                 reduction=reduction)
@@ -87,32 +87,33 @@ def forward(self, logits, targets):
         return loss
 
 
-class FocalLossMultiLabel(_Loss):
-    """
-    Compute focal loss for multi-label problem.
-    Ignores targets having -1 label
-    """
-
-    def forward(self, logits, targets):
-        """
-        Args:
-            logits: [bs; num_classes]
-            targets: [bs; num_classes]
-        """
-        num_classes = logits.size(1)
-        loss = 0
-
-        for cls in range(num_classes):
-            # Filter anchors with -1 label from loss computation
-            if cls == self.ignore:
-                continue
-
-            cls_label_target = targets[..., cls].long()
-            cls_label_input = logits[..., cls]
-
-            loss += self.loss_fn(cls_label_input, cls_label_target)
-
-        return loss
-
-
-__all__ = ["FocalLossBinary", "FocalLossMultiClass", "FocalLossMultiLabel"]
+# @TODO: check
+# class FocalLossMultiLabel(_Loss):
+#     """
+#     Compute focal loss for multi-label problem.
+#     Ignores targets having -1 label
+#     """
+#
+#     def forward(self, logits, targets):
+#         """
+#         Args:
+#             logits: [bs; num_classes]
+#             targets: [bs; num_classes]
+#         """
+#         num_classes = logits.size(1)
+#         loss = 0
+#
+#         for cls in range(num_classes):
+#             # Filter anchors with -1 label from loss computation
+#             if cls == self.ignore:
+#                 continue
+#
+#             cls_label_target = targets[..., cls].long()
+#             cls_label_input = logits[..., cls]
+#
+#             loss += self.loss_fn(cls_label_input, cls_label_target)
+#
+#         return loss
+
+
+__all__ = ["FocalLossBinary", "FocalLossMultiClass"]

catalyst/contrib/criterion/huber.py

@@ -4,7 +4,7 @@
 
 class HuberLoss(nn.Module):
     def __init__(self, clip_delta=1.0, reduction="elementwise_mean"):
-        super(HuberLoss, self).__init__()
+        super().__init__()
         self.clip_delta = clip_delta
         self.reduction = reduction or "none"
 

catalyst/contrib/criterion/iou.py

@@ -1,6 +1,6 @@
 from functools import partial
 import torch.nn as nn
-from catalyst.dl import metrics
+from catalyst.dl.utils import criterion
 
 
 class IoULoss(nn.Module):
@@ -21,7 +21,7 @@ def __init__(
     ):
         super().__init__()
         self.metric_fn = partial(
-            metrics.iou,
+            criterion.iou,
             eps=eps,
             threshold=threshold,
             activation=activation)

catalyst/contrib/criterion/wing.py

@@ -1,7 +1,41 @@
 from functools import partial
+import math
 
+import torch
 import torch.nn as nn
-from catalyst.dl import losses
+
+
+def wing_loss(
+    outputs: torch.Tensor,
+    targets: torch.Tensor,
+    width: int = 5,
+    curvature: float = 0.5,
+    reduction: str = "mean"
+):
+    """
+    https://arxiv.org/pdf/1711.06753.pdf
+
+    Source https://github.com/BloodAxe/pytorch-toolbelt
+    See :class:`~pytorch_toolbelt.losses` for details.
+    """
+    diff_abs = (targets - outputs).abs()
+    loss = diff_abs.clone()
+
+    idx_smaller = diff_abs < width
+    idx_bigger = diff_abs >= width
+
+    loss[idx_smaller] = \
+        width * torch.log(1 + diff_abs[idx_smaller] / curvature)
+
+    C = width - width * math.log(1 + width / curvature)
+    loss[idx_bigger] = loss[idx_bigger] - C
+
+    if reduction == "sum":
+        loss = loss.sum()
+    if reduction == "mean":
+        loss = loss.mean()
+
+    return loss
 
 
 class WingLoss(nn.Module):
@@ -13,7 +47,7 @@ def __init__(
     ):
         super().__init__()
         self.loss_fn = partial(
-            losses.wing_loss,
+            wing_loss,
             width=width,
             curvature=curvature,
             reduction=reduction)

catalyst/contrib/models/__init__.py

@@ -1,5 +1,2 @@
 # flake8: noqa
 from .sequential import *
-# from .encoder import *
-# from .classification import *
-# from .segmentation import *

catalyst/contrib/models/classification/mobilenetv2.py

@@ -54,7 +54,7 @@ def conv_1x1_bn(inp, oup):
 
 class InvertedResidual(nn.Module):
     def __init__(self, inp, oup, stride, expand_ratio):
-        super(InvertedResidual, self).__init__()
+        super().__init__()
         assert stride in [1, 2]
 
         hidden_dim = round(inp * expand_ratio)

catalyst/contrib/models/classification/mobilenetv3.py

@@ -31,7 +31,7 @@ def _make_divisible(v, divisor, min_value=None):
 
 class h_sigmoid(nn.Module):
     def __init__(self, inplace=True):
-        super(h_sigmoid, self).__init__()
+        super().__init__()
         self.relu = nn.ReLU6(inplace=inplace)
 
     def forward(self, x):
@@ -40,7 +40,7 @@ def forward(self, x):
 
 class h_swish(nn.Module):
     def __init__(self, inplace=True):
-        super(h_swish, self).__init__()
+        super().__init__()
         self.sigmoid = h_sigmoid(inplace=inplace)
 
     def forward(self, x):
@@ -49,7 +49,7 @@ def forward(self, x):
 
 class SELayer(nn.Module):
     def __init__(self, channel, reduction=4):
-        super(SELayer, self).__init__()
+        super().__init__()
         self.avg_pool = nn.AdaptiveAvgPool2d(1)
         self.fc = nn.Sequential(
             nn.Linear(channel, channel // reduction),

catalyst/contrib/models/encoder/__init__.py

@@ -1,4 +1,4 @@
 from .resnet import ResnetEncoder
-from .mobilenet import MobileNetEncoder
+from .mobilenetv2 import MobileNetV2Encoder
 
-__all__ = ["ResnetEncoder", "MobileNetEncoder"]
+__all__ = ["ResnetEncoder", "MobileNetV2Encoder"]

catalyst/contrib/models/encoder/mobilenet.py → catalyst/contrib/models/encoder/mobilenetv2.py

@@ -6,7 +6,7 @@
 from ..classification import MobileNetV2
 
 
-class MobileNetEncoder(nn.Module):
+class MobileNetV2Encoder(nn.Module):
     def __init__(
         self,
         input_size=224,

catalyst/contrib/models/segmentation/__init__.py

@@ -14,3 +14,10 @@
 from .linknet import *
 from .psp import *
 from .unet import *
+
+__all__ = [
+    "UnetMetaSpec", "UnetSpec", "ResnetUnetSpec",
+    "Unet", "Linknet", "FPNUnet", "PSPnet",
+    "ResnetUnet", "ResnetLinknet", "ResnetFPNUnet", "ResnetPSPnet",
+    "MobileUnet", "ResNetUnet", "ResNetLinknet"
+]

catalyst/contrib/models/segmentation/core.py

@@ -9,7 +9,7 @@
 from .head import HeadSpec
 
 
-class UnetSpec(nn.Module):
+class UnetMetaSpec(nn.Module):
     def __init__(
         self,
         encoder: EncoderSpec,
@@ -31,7 +31,7 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
         return output
 
 
-class _UnetSpec(UnetSpec):
+class UnetSpec(UnetMetaSpec):
     def __init__(
         self,
         num_classes: int = 1,
@@ -77,7 +77,7 @@ def _get_components(
         raise NotImplementedError()
 
 
-class _ResnetUnetSpec(UnetSpec):
+class ResnetUnetSpec(UnetMetaSpec):
     def __init__(
         self,
         num_classes: int = 1,

catalyst/contrib/models/segmentation/fpn.py

@@ -5,10 +5,10 @@
 from .bridge import UnetBridge
 from .decoder import FPNDecoder
 from .head import FPNHead
-from .core import _UnetSpec, _ResnetUnetSpec
+from .core import UnetSpec, ResnetUnetSpec
 
 
-class FPNUnet(_UnetSpec):
+class FPNUnet(UnetSpec):
 
     def _get_components(
         self,
@@ -42,7 +42,7 @@ def _get_components(
         return encoder, bridge, decoder, head
 
 
-class ResnetFPNUnet(_ResnetUnetSpec):
+class ResnetFPNUnet(ResnetUnetSpec):
 
     def _get_components(
         self,

catalyst/contrib/models/segmentation/linknet.py

@@ -7,10 +7,10 @@
 from .bridge import UnetBridge
 from .decoder import UNetDecoder
 from .head import UnetHead
-from .core import _UnetSpec, _ResnetUnetSpec
+from .core import UnetSpec, ResnetUnetSpec
 
 
-class Linknet(_UnetSpec):
+class Linknet(UnetSpec):
 
     def _get_components(
         self,
@@ -43,7 +43,7 @@ def _get_components(
         return encoder, bridge, decoder, head
 
 
-class ResnetLinknet(_ResnetUnetSpec):
+class ResnetLinknet(ResnetUnetSpec):
 
     def _get_components(
         self,

catalyst/contrib/models/segmentation/models/__init__.py

@@ -2,3 +2,7 @@
 from .mobileunet import *
 from .resnetunet import *
 from .resnetlinknet import *
+
+__all__ = [
+    "MobileUnet", "ResNetUnet", "ResNetLinknet"
+]

catalyst/contrib/models/segmentation/models/resnetunet.py

@@ -81,25 +81,8 @@ def forward(self, x):
 class ResNetUnet(nn.Module):
     """
     U-Net inspired encoder-decoder architecture for semantic segmentation,
-        with a ResNet encoder as proposed by Alexander Buslaev.
-
+    with a ResNet encoder as proposed by Alexander Buslaev.
     Also known as AlbuNet
-
-    See:
-    - https://arxiv.org/abs/1505.04597 -
-        U-Net: Convolutional Networks for Biomedical Image Segmentation
-    - https://arxiv.org/abs/1411.4038  -
-        Fully Convolutional Networks for Semantic Segmentation
-    - https://arxiv.org/abs/1512.03385 -
-        Deep Residual Learning for Image Recognition
-    - https://arxiv.org/abs/1801.05746 -
-        TernausNet: U-Net with VGG11
-        Encoder Pre-Trained on ImageNet for Image Segmentation
-    - https://arxiv.org/abs/1806.00844 -
-        TernausNetV2: Fully Convolutional Network for Instance Segmentation
-
-    based on https://github.com/mapbox/robosat/blob/master/robosat/unet.py
-
     """
     def __init__(
         self,