Version 0.3.0 (#6)

.github/workflows/publish_docker_image.yml

@@ -13,11 +13,6 @@ jobs:
       - name: Login to Github Container Registry
         run: echo ${{ secrets.CR_PAT }} | docker login ghcr.io -u ${{ github.actor }} --password-stdin
       - name: Build image
-        run: |
-          docker build
-          -f docker/Dockerfile
-          --tag ghcr.io/albumentations-team/autoalbument:${{ github.event.release.tag_name }}
-          --tag ghcr.io/albumentations-team/autoalbument:latest
-          .
+        run: docker build -f docker/Dockerfile --tag ghcr.io/albumentations-team/autoalbument:${{ github.event.release.tag_name }} --tag ghcr.io/albumentations-team/autoalbument:latest .
       - name: Push image
         run: docker push ghcr.io/albumentations-team/autoalbument

autoalbument/__init__.py

@@ -1 +1 @@
-__version__ = "0.2.0"
+__version__ = "0.3.0"

autoalbument/cli/conf/data/default.yaml

@@ -42,6 +42,6 @@ dataloader:
   _target_: torch.utils.data.DataLoader
   batch_size: 128
   shuffle: True
-  num_workers: 4
+  num_workers: 8
   pin_memory: True
   drop_last: True

autoalbument/cli/templates/classification/search.yaml.tmpl

@@ -19,7 +19,7 @@ policy_model:
   # number of sub-policies leads to a more diverse set of augmentations and better performance of a model trained on
   # augmented images. However, an increase in the number of sub-policies leads to the exponential growth of a search
   # space of augmentations, so you need more training data for Policy Model to find good augmentation policies.
-  num_sub_policies: 20
+  num_sub_policies: 50
 
   # Number of chunks in a batch. Faster AutoAugment splits each batch of images into `num_chunks` chunks. Then it
   # applies the same sub-policy with the same parameters to each image in a chunk. This parameter controls the tradeoff
@@ -101,7 +101,7 @@ data:
   dataloader:
     _target_: torch.utils.data.DataLoader
     batch_size: 64
-    num_workers: 4
+    num_workers: 8
 
 optim:
   # Number of epochs to search parameters of augmentations.

autoalbument/cli/templates/classification/search_full.yaml.tmpl

@@ -25,7 +25,7 @@ policy_model:
   # number of sub-policies leads to a more diverse set of augmentations and better performance of a model trained on
   # augmented images. However, an increase in the number of sub-policies leads to the exponential growth of a search
   # space of augmentations, so you need more training data for Policy Model to find good augmentation policies.
-  num_sub_policies: 20
+  num_sub_policies: 50
 
   # Number of chunks in a batch. Faster AutoAugment splits each batch of images into `num_chunks` chunks. Then it
   # applies the same sub-policy with the same parameters to each image in a chunk. This parameter controls the tradeoff
@@ -128,7 +128,7 @@ data:
     _target_: torch.utils.data.DataLoader
     batch_size: 64
     shuffle: True
-    num_workers: 4
+    num_workers: 8
     pin_memory: True
     drop_last: True
 

autoalbument/cli/templates/semantic_segmentation/search.yaml.tmpl

@@ -19,7 +19,7 @@ policy_model:
   # number of sub-policies leads to a more diverse set of augmentations and better performance of a model trained on
   # augmented images. However, an increase in the number of sub-policies leads to the exponential growth of a search
   # space of augmentations, so you need more training data for Policy Model to find good augmentation policies.
-  num_sub_policies: 20
+  num_sub_policies: 25
 
   # Number of chunks in a batch. Faster AutoAugment splits each batch of images into `num_chunks` chunks. Then it
   # applies the same sub-policy with the same parameters to each image in a chunk. This parameter controls the tradeoff
@@ -110,7 +110,7 @@ data:
   dataloader:
     _target_: torch.utils.data.DataLoader
     batch_size: 64
-    num_workers: 4
+    num_workers: 8
 
 optim:
   # Number of epochs to search parameters of augmentations.

autoalbument/cli/templates/semantic_segmentation/search_full.yaml.tmpl

@@ -23,7 +23,7 @@ policy_model:
   # number of sub-policies leads to a more diverse set of augmentations and better performance of a model trained on
   # augmented images. However, an increase in the number of sub-policies leads to the exponential growth of a search
   # space of augmentations, so you need more training data for Policy Model to find good augmentation policies.
-  num_sub_policies: 20
+  num_sub_policies: 25
 
   # Number of chunks in a batch. Faster AutoAugment splits each batch of images into `num_chunks` chunks. Then it
   # applies the same sub-policy with the same parameters to each image in a chunk. This parameter controls the tradeoff
@@ -135,7 +135,7 @@ data:
     _target_: torch.utils.data.DataLoader
     batch_size: 64
     shuffle: True
-    num_workers: 4
+    num_workers: 8
     pin_memory: True
     drop_last: True
 

autoalbument/faster_autoaugment/albumentations_pytorch/functional.py

@@ -1,6 +1,7 @@
 import random
 
 import torch
+import torch.nn.functional as F
 
 from autoalbument.faster_autoaugment.albumentations_pytorch.affine import (
     get_scaling_matrix,
@@ -66,11 +67,10 @@ def scale(img_batch, scale, padding_mode=TorchPadding.REFLECTION):
 def cutout(img_batch, num_holes, hole_size, fill_value=0):
     img_batch = img_batch.clone()
     height, width = img_batch.shape[-2:]
-    for i in range(len(img_batch)):
-        for _n in range(num_holes):
-            y1 = random.randint(0, height - hole_size)
-            x1 = random.randint(0, width - hole_size)
-            y2 = y1 + hole_size
-            x2 = x1 + hole_size
-            img_batch[i, :, y1:y2, x1:x2] = fill_value
+    for _n in range(num_holes):
+        y1 = random.randint(0, height - hole_size)
+        x1 = random.randint(0, width - hole_size)
+        y2 = y1 + hole_size
+        x2 = x1 + hole_size
+        img_batch[:, :, y1:y2, x1:x2] = fill_value
     return img_batch

autoalbument/faster_autoaugment/operations.py

@@ -6,6 +6,7 @@
 import warnings
 
 import albumentations as A
+import cv2
 import torch
 from torch import nn
 from torch.autograd import Function
@@ -292,7 +293,7 @@ def as_transform(self, value, p):
 
 class Cutout(Operation):
     def __init__(self, temperature, value_range=(0.0, 1.0)):
-        super().__init__(temperature, value_range=value_range)
+        super().__init__(temperature, value_range=value_range, ste=True)
         self.register_buffer("saved_image_shape", torch.Tensor([0, 0]).type(torch.int64))
         self.is_image_shape_saved = False
 
@@ -325,7 +326,7 @@ def _as_cutout_transform(self, value, p, image_shape):
         raise NotImplementedError
 
 
-class CutoutFixedNumerOfHoles(Cutout):
+class CutoutFixedNumberOfHoles(Cutout):
     def __init__(self, temperature, num_holes=16):
         super().__init__(temperature)
         self.num_holes = num_holes

autoalbument/faster_autoaugment/policy.py

@@ -14,7 +14,7 @@
 from torch import Tensor, nn
 
 from autoalbument.faster_autoaugment.operations import (
-    CutoutFixedNumerOfHoles,
+    CutoutFixedNumberOfHoles,
     CutoutFixedSize,
     HorizontalFlip,
     RandomBrightness,
@@ -164,7 +164,7 @@ def dda_operations(temperature):
             ShiftX(temperature=temperature),
             ShiftY(temperature=temperature),
             Scale(temperature=temperature),
-            CutoutFixedNumerOfHoles(temperature=temperature),
+            CutoutFixedNumberOfHoles(temperature=temperature),
             CutoutFixedSize(temperature=temperature),
         ]
 

autoalbument/faster_autoaugment/search.py

@@ -13,6 +13,7 @@
 import torch
 from albumentations.pytorch import ToTensorV2
 from hydra.utils import instantiate
+from omegaconf import OmegaConf
 from torch import Tensor, nn
 from torch.utils.tensorboard import SummaryWriter
 from tqdm import tqdm
@@ -51,7 +52,8 @@ def set_seed(self):
 
     def create_tensorboard_writer(self):
         if self.cfg.tensorboard_logs_dir:
-            return SummaryWriter(os.path.join(self.cfg.tensorboard_logs_dir, os.getcwd().replace(os.sep, ".")))
+            filename = os.getcwd().replace(os.sep, ".").lstrip(".")
+            return SummaryWriter(os.path.join(self.cfg.tensorboard_logs_dir, filename))
         return None
 
     def get_policy_state_dict(self):
@@ -83,19 +85,21 @@ def create_metric_tracker(self):
 
     def get_preprocessing_transforms(self):
         preprocessing_config = self.cfg.data.preprocessing
+        if not preprocessing_config:
+            return []
+        preprocessing_config = OmegaConf.to_container(preprocessing_config, resolve=True)
         preprocessing_transforms = []
-        if preprocessing_config:
-            for preprocessing_transform in preprocessing_config:
-                for transform_name, transform_args in preprocessing_transform.items():
-                    transform = A.from_dict(
-                        {
-                            "transform": {
-                                "__class_fullname__": "albumentations.augmentations.transforms." + transform_name,
-                                **transform_args,
-                            }
+        for preprocessing_transform in preprocessing_config:
+            for transform_name, transform_args in preprocessing_transform.items():
+                transform = A.from_dict(
+                    {
+                        "transform": {
+                            "__class_fullname__": "albumentations.augmentations.transforms." + transform_name,
+                            **transform_args,
                         }
-                    )
-                    preprocessing_transforms.append(transform)
+                    }
+                )
+                preprocessing_transforms.append(transform)
         return preprocessing_transforms
 
     def create_preprocessing_transform(self):

docker/Dockerfile

@@ -9,6 +9,9 @@ RUN useradd --create-home --shell /bin/bash --no-log-init autoalbument
 USER autoalbument
 ENV PATH="/home/autoalbument/.local/bin:${PATH}"
 WORKDIR /opt/autoalbument
+COPY ./docker/requirements.txt /opt/autoalbument/docker/requirements.txt
+RUN pip install --no-cache-dir -r /opt/autoalbument/docker/requirements.txt
+
 COPY . .
 RUN pip install --no-cache-dir .
 COPY docker/entrypoint.sh entrypoint.sh

docker/requirements.txt

@@ -0,0 +1,89 @@
+absl-py==0.11.0
+albumentations==0.5.2
+antlr4-python3-runtime==4.8
+backcall==0.2.0
+beautifulsoup4==4.9.3
+cachetools==4.2.0
+certifi==2020.6.20
+cffi==1.14.0
+chardet==3.0.4
+click==7.1.2
+colorama==0.4.4
+cryptography==2.9.2
+cycler==0.10.0
+dataclasses==0.6
+decorator==4.4.2
+dnspython==2.0.0
+efficientnet-pytorch==0.6.3
+filelock==3.0.12
+future==0.18.2
+glob2==0.7
+google-auth==1.24.0
+google-auth-oauthlib==0.4.2
+grpcio==1.34.0
+hydra-core==1.0.4
+idna==2.9
+imageio==2.9.0
+imgaug==0.4.0
+importlib-resources==4.0.0
+ipython==7.18.1
+ipython-genutils==0.2.0
+jedi==0.17.2
+Jinja2==2.11.2
+kiwisolver==1.3.1
+libarchive-c==2.9
+Markdown==3.3.3
+MarkupSafe==1.1.1
+matplotlib==3.3.3
+munch==2.5.0
+networkx==2.5
+numpy==1.19.2
+oauthlib==3.1.0
+olefile==0.46
+omegaconf==2.0.5
+opencv-python==4.4.0.46
+opencv-python-headless==4.4.0.46
+parso==0.7.0
+pexpect==4.8.0
+pickleshare==0.7.5
+Pillow==8.0.0
+pkginfo==1.6.0
+pretrainedmodels==0.7.4
+prompt-toolkit==3.0.8
+protobuf==3.14.0
+psutil==5.7.2
+ptyprocess==0.6.0
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pycosat==0.6.3
+pycparser==2.20
+Pygments==2.7.1
+pyOpenSSL==19.1.0
+pyparsing==2.4.7
+PySocks==1.7.1
+python-dateutil==2.8.1
+python-etcd==0.4.5
+pytz==2020.1
+PyWavelets==1.1.1
+PyYAML==5.3.1
+requests==2.23.0
+requests-oauthlib==1.3.0
+rsa==4.6
+ruamel-yaml==0.15.87
+scikit-image==0.18.1
+scipy==1.5.4
+segmentation-models-pytorch==0.1.3
+Shapely==1.7.1
+six==1.14.0
+soupsieve==2.0.1
+tensorboard==2.4.0
+tensorboard-plugin-wit==1.7.0
+tifffile==2020.12.8
+timm==0.3.2
+torchelastic==0.2.1
+tqdm==4.46.0
+traitlets==5.0.5
+typing-extensions==3.7.4.3
+urllib3==1.25.8
+wcwidth==0.2.5
+Werkzeug==1.0.1

examples/cifar10/dataset.py

@@ -1,7 +1,12 @@
 import torchvision
 
+import cv2
 
-class SearchDataset(torchvision.datasets.CIFAR10):
+cv2.setNumThreads(0)
+cv2.ocl.setUseOpenCL(False)
+
+
+class Cifar10SearchDataset(torchvision.datasets.CIFAR10):
     def __init__(self, root="~/data/cifar10", train=True, download=True, transform=None):
         super().__init__(root=root, train=train, download=download, transform=transform)