Skip to content

Semantic segmentation on the Pascal VOC dataset

The following files are also available on GitHub -

import cv2
import numpy as np
import torchvision


class CityscapesSearchDataset(torchvision.datasets.Cityscapes):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs, target_type="semantic")
        self.semantic_target_type_index = [i for i, t in enumerate(self.target_type) if t == "semantic"][0]
        self.colormap = self._generate_colormap()

    def _generate_colormap(self):
        colormap = {}
        for class_ in self.classes:
            if class_.train_id in (-1, 255):
            colormap[class_.train_id] =
        return colormap

    def _convert_to_segmentation_mask(self, mask):
        height, width = mask.shape[:2]
        segmentation_mask = np.zeros((height, width, len(self.colormap)), dtype=np.float32)
        for label_index, label in self.colormap.items():
            segmentation_mask[:, :, label_index] = (mask == label).astype(float)
        return segmentation_mask

    def __getitem__(self, index):
        image = cv2.imread(self.images[index])
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
        mask = cv2.imread(self.targets[index][self.semantic_target_type_index], cv2.IMREAD_UNCHANGED)

        mask = self._convert_to_segmentation_mask(mask)

        if self.transform is not None:
            transformed = self.transform(image=image, mask=mask)
            image = transformed["image"]
            mask = transformed["mask"]

        return image, mask


# @package _global_

_version: 2  # An internal value that indicates a version of the config schema. This value is used by
# `autoalbument-search` and `autoalbument-migrate` to upgrade the config to the latest version if necessary.
# Please do not change it manually.

task: semantic_segmentation # Deep learning task. Should either be `classification` or `semantic_segmentation`.

  # Settings for Policy Model that searches augmentation policies.

  task_factor: 0.1
  # Multiplier for segmentation loss of a model. Faster AutoAugment uses segmentation loss to prevent augmentations
  # from transforming images of a particular class to another class.

  gp_factor: 10
  # Multiplier for the gradient penalty for WGAN-GP training. 10 is the default value that was proposed in
  # `Improved Training of Wasserstein GANs`.

  temperature: 0.05
  # Temperature for Relaxed Bernoulli distribution. The probability of applying a certain augmentation is sampled from
  # Relaxed Bernoulli distribution (because Bernoulli distribution is not differentiable). With lower values of
  # `temperature` Relaxed Bernoulli distribution behaves like Bernoulli distribution. In the paper, the authors
  # of Faster AutoAugment used 0.05 as a default value for `temperature`.

  num_sub_policies: 25
  # Number of augmentation sub-policies. When an image passes through an augmentation pipeline, Faster AutoAugment
  # randomly chooses one sub-policy and uses augmentations from that sub-policy to transform an input image. A larger
  # 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_chunks: 4
  # 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
  # between the speed of augmentation search and diversity of augmentations. Larger `num_chunks` values will lead to
  # faster searching but less diverse set of augmentations. Note that this parameter is used only in the searching
  # phase. When you train a model with found sub-policies, Albumentations will apply a distinct set of transformations
  # to each image separately.

  operation_count: 4
  # Number of consecutive augmentations in each sub-policy. Faster AutoAugment will sequentially apply `operation_count`
  # augmentations from a sub-policy to an image. Larger values of `operation_count` lead to better performance of
  # a model trained on augmented images. Simultaneously, larger values of `operation_count` affect the speed of search
  # and increase the searching time.

  # Settings for Semantic Segmentation Model that is used for two purposes:
  # 1. As a model that performs semantic segmentation of input images.
  # 2. As a Discriminator for Policy Model.

  _target_: autoalbument.faster_autoaugment.models.SemanticSegmentationModel
  # By default, AutoAlbument uses an instance of `autoalbument.faster_autoaugment.models.SemanticSegmentationModel` as
  # a semantic segmentation model.
  # This model takes four parameters: `num_classes`, `architecture`, `encoder_architecture` and `pretrained`.

  num_classes: 19
  # The number of classes in the dataset. The dataset implementation should return a mask as a NumPy array with
  # the shape [height, width, num_classes]. In a case of binary segmentation you can set `num_classes` to 1.

  architecture: DeepLabV3Plus
  # The architecture of Semantic Segmentation Model. AutoAlbument uses models from
  # Please refer to its documentation to get a list of available
  # models -

  encoder_architecture: resnet50
  # The architecture of encoder in Semantic Segmentation Model. Please refer to Segmentation Models' documentation to
  # get a list of available encoders -

  pretrained: true
  # Either boolean flag or string with that indicates whether the selected encoder architecture should load pretrained
  # weights or use randomly initialized weights.
  # - In the case of boolean flag `true` means using pretrained weights from ImageNet and `false` means using randomly
  #   initialized weights.
  # - In the case of string the value should specify the name of the weights. For the list of available weights please
  #   refer to

    _target_: dataset.CityscapesSearchDataset
    root: ~/data/cityscapes/data
    split: train
  # Class for the PyTorch Dataset and arguments to it. AutoAlbument will create an object of this class using
  # the `instantiate` method from Hydra -
  # Note that the target class value in the `_target_` argument should be located inside PYTHONPATH so Hydra could
  # find it. The directory with the config file is automatically added to PYTHONPATH, so the default value
  # `dataset.SearchDataset` points to the class `SearchDataset` from the `` file. This `` file is
  # located along with the `search.yaml` file in the same directory provided by `--config-dir`.
  # As an alternative, you could provide a path to a Python file with the dataset using the `dataset_file` parameter
  # instead of the `dataset` parameter. The Python file should contain the implementation of a PyTorch dataset for
  # augmentation search. The dataset class should have named `SearchDataset`. The value in `dataset_file` could either
  # be a relative or an absolute path ; in the case of a relative path, the path should be relative to this config
  # file's location.
  # - Example of a relative path:
  # dataset_file:
  # - Example of an absolute path:
  # dataset_file: /projects/pytorch/

  input_dtype: uint8
  # The data type of input images. Two values are supported:
  # - uint8. In that case, all input images should be NumPy arrays with the np.uint8 data type and values in the range
  #   [0, 255].
  # - float32. In that case, all input images should be NumPy arrays with the np.float32 data type and values in the
  #   range [0.0, 1.0].

  - LongestMaxSize:
      max_size: 256
  - PadIfNeeded:
      min_height: 256
      min_width: 256
      border_mode: 0
      value: [0, 0, 0]
  # A list of preprocessing augmentations that will be applied to each image before applying augmentations from
  # a policy. A preprocessing augmentation should be defined as `key`: `value`, where `key` is the name of augmentation
  # from Albumentations, and `value` is a dictionary with augmentation parameters. The found policy will also apply
  # those preprocessing augmentations before applying the main augmentations.
  # Here is an example of an augmentation pipeline that first pads an image to the size 512x512 pixels, then resizes
  # the resulting image to the size 256x256 pixels and finally crops a random patch with the size 224x224 pixels.
  #  preprocessing:
  #    - PadIfNeeded:
  #        min_height: 512
  #        min_width: 512
  #    - Resize:
  #        height: 256
  #        width: 256
  #    - RandomCrop:
  #        height: 224
  #        width: 224

    mean: [0.485, 0.456, 0.406]
    std: [0.229, 0.224, 0.225]
  # Normalization values for images. For each image, the search pipeline will subtract `mean` and divide by `std`.
  # Normalization is applied after transforms defined in `preprocessing`. Note that regardless of `input_dtype`,
  # the normalization function will always receive a `float32` input with values in the range [0.0, 1.0], so you should
  # define `mean` and `std` values accordingly. ImageNet normalization is used by default.

    batch_size: 32
    shuffle: true
    num_workers: 8
    pin_memory: true
    drop_last: true
  # Parameters for the PyTorch DataLoader. Please refer to the PyTorch documentation for the description of parameters -

    _target_: torch.optim.Adam
    lr: 1e-3
    betas: [0, 0.999]
  # Optimizer configuration for the main (either Classification or Semantic Segmentation) Model

    _target_: torch.optim.Adam
    lr: 1e-3
    betas: [0, 0.999]
  # Optimizer configuration for Policy Model

seed: 42 # Random seed. If the value is not null, it will be passed to `seed_everything` -

    dir: ${config_dir:}/outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
    # Path to the directory that will contain all outputs produced by the search algorithm. `${config_dir:}` contains
    # path to the directory with the `search.yaml` config file. Please refer to the Hydra documentation for more
    # information -

  # Configuration for PyTorch Lightning Trainer. You can read more about Trainer and its arguments at
  max_epochs: 50
  # Number of epochs to search for augmentation parameters.
  # More detailed description -

  benchmark: true
  # If true enables cudnn.benchmark.
  # More detailed description -

  gpus: 1
  # Number of GPUs to train on. Set to `0` or None` to use CPU for training.
  # More detailed description -