Semantic Segmentation on Describable Textures Dataset

This repository was created as a homework assignment during my Msc. studies in Data Science.

The data set consists of images with different structures which have to be classified. However, instead of a simple label a mask must be predicted (semantic segmentation). For the first network, only conv layers may be used. For the second network the choice of the network architecture is free. I decided to use a U-Net, which was developed by the University of Freiburg, Germany.

Setup

Create the conda environment and activate it:

conda env create -f environment.yml
conda activate describable-textures-dataset

Download the images:

python dtd_loader_color_patches.py

Arguments:

• --tiled: Set to True if you want to use the larger colored data set

Also both dataset (tiled and not tiled) can be downloaded. However, in the file config.yaml must be specified which dataset shall be used for training and testing.

Mask Distribution

The following plot shows the distribution of the mask:

This distribution is only considered if the BCE with logits loss is used.

General Settings

Some general configurations are specified in the file config.yaml:

• device: either cuda or cpu
• num_classes: Number of classes, must be 47 for the dtd dataset
• max_num_epoch: Max. number of epochs (or number of epochs without early stopping)
• loss: The loss function, must be one of cross-entropy, dice or bce-with-logits
• tiled: True if the tiled dataset shall be used

Train a model

python train.py [OPTIONS]

Arguments:

• --learning_rate: The learning rate
• --batch_size: The batch size
• --model_name: Name of the model, one of 'simple_fcn', 'simple_u-net' or 'pretrained_u-net'
• --wandb: Set to True if you want to use wandb.ai, default is using Tensorboard
• --early_stopping: Set to True if you want to use early stopping

Run Sweep

With wandb, it is also possible to run sweeps. First, define the model_name in the File sweep.yaml and then execute:

wandb sweep sweep.yaml
wandb agent your-sweep-id

Run a trained model

python evaluate.py [OPTIONS]

Arguments:

• --model_name: Name of the model, one of 'simple_fcn', 'simple_u-net' or 'pretrained_u-net'
• --accuracy Set to True if the Top-1 accuracy on the test set shall be calculated
• --plot Set to True if some predictions shall be plotted