Awesome

Accelerated DL & RL

PyTorch framework for Deep Learning research and development. It was developed with a focus on reproducibility, fast experimentation and code/ideas reusing. Being able to research/develop something new, rather than write another regular train loop. <br/> Break the cycle - use the Catalyst!

Project manifest. Part of PyTorch Ecosystem. Part of Catalyst Ecosystem:

• Alchemy - Experiments logging & visualization
• Catalyst - Accelerated Deep Learning Research and Development
• Reaction - Convenient Deep Learning models serving

Catalyst at AI Landscape.

Catalyst.Segmentation

Note: this repo uses advanced Catalyst Config API and could be a bit out-of-day right now. Use Catalyst's minimal examples section for a starting point and up-to-day use cases, please.

You will learn how to build image segmentation pipeline with transfer learning using the Catalyst framework.

Goals

1. Install requirements
2. Prepare data
3. Run: raw data → production-ready model
4. Get results
5. Customize own pipeline

1. Install requirements

Using local environment:

pip install -r requirements/requirements.txt

Using docker:

This creates a build catalyst-segmentation with the necessary libraries:

make docker-build

2. Get Dataset

Try on open datasets

export DATASET="isbi"

rm -rf data/
mkdir -p data

if [[ "$DATASET" == "isbi" ]]; then
    # binary segmentation
    # http://brainiac2.mit.edu/isbi_challenge/
    download-gdrive 1uyPb9WI0t2qMKIqOjFKMv1EtfQ5FAVEI isbi_cleared_191107.tar.gz
    tar -xf isbi_cleared_191107.tar.gz &>/dev/null
    mv isbi_cleared_191107 ./data/origin
elif [[ "$DATASET" == "voc2012" ]]; then
    # semantic segmentation
    # http://host.robots.ox.ac.uk/pascal/VOC/voc2012/
    wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
    tar -xf VOCtrainval_11-May-2012.tar &>/dev/null
    mkdir -p ./data/origin/images/; mv VOCdevkit/VOC2012/JPEGImages/* $_
    mkdir -p ./data/origin/raw_masks; mv VOCdevkit/VOC2012/SegmentationClass/* $_
fi

Use your own dataset

Data structure

Make sure, that final folder with data has the required structure:

/path/to/your_dataset/
        images/
            image_1
            image_2
            ...
            image_N
        raw_masks/
            mask_1
            mask_2
            ...
            mask_N

Data location

• The easiest way is to move your data:

  mv /path/to/your_dataset/* /catalyst.segmentation/data/origin
  

  In that way you can run pipeline with default settings.

• If you prefer leave data in /path/to/your_dataset/

  • In local environment:

    • Link directory

      ln -s /path/to/your_dataset $(pwd)/data/origin
      

    • Or just set path to your dataset DATADIR=/path/to/your_dataset when you start the pipeline.

  • Using docker

    You need to set:

       -v /path/to/your_dataset:/data \ #instead default  $(pwd)/data/origin:/data
    

    in the script below to start the pipeline.

3. Segmentation pipeline

Fast&Furious: raw data → production-ready model

The pipeline will automatically guide you from raw data to the production-ready model.

We will initialize Unet model with a pre-trained ResNet-18 encoder. During current pipeline model will be trained sequentially in two stages.

Run in local environment:

CUDA_VISIBLE_DEVICES=0 \
CUDNN_BENCHMARK="True" \
CUDNN_DETERMINISTIC="True" \
WORKDIR=./logs \
DATADIR=./data/origin \
IMAGE_SIZE=256 \
CONFIG_TEMPLATE=./configs/templates/binary.yml \
NUM_WORKERS=4 \
BATCH_SIZE=256 \
bash ./bin/catalyst-binary-segmentation-pipeline.sh

Run in docker:

export LOGDIR=$(pwd)/logs
docker run -it --rm --shm-size 8G --runtime=nvidia \
   -v $(pwd):/workspace/ \
   -v $LOGDIR:/logdir/ \
   -v $(pwd)/data/origin:/data \
   -e "CUDA_VISIBLE_DEVICES=0" \
   -e "USE_WANDB=1" \
   -e "LOGDIR=/logdir" \
   -e "CUDNN_BENCHMARK='True'" \
   -e "CUDNN_DETERMINISTIC='True'" \
   -e "WORKDIR=/logdir" \
   -e "DATADIR=/data" \
   -e "IMAGE_SIZE=256" \
   -e "CONFIG_TEMPLATE=./configs/templates/binary.yml" \
   -e "NUM_WORKERS=4" \
   -e "BATCH_SIZE=256" \
   catalyst-segmentation ./bin/catalyst-binary-segmentation-pipeline.sh

Run in local environment:

CUDA_VISIBLE_DEVICES=0 \
CUDNN_BENCHMARK="True" \
CUDNN_DETERMINISTIC="True" \
WORKDIR=./logs \
DATADIR=./data/origin \
IMAGE_SIZE=256 \
CONFIG_TEMPLATE=./configs/templates/semantic.yml \
NUM_WORKERS=4 \
BATCH_SIZE=256 \
bash ./bin/catalyst-semantic-segmentation-pipeline.sh

Run in docker:

export LOGDIR=$(pwd)/logs
docker run -it --rm --shm-size 8G --runtime=nvidia \
   -v $(pwd):/workspace/ \
   -v $LOGDIR:/logdir/ \
   -v $(pwd)/data/origin:/data \
   -e "CUDA_VISIBLE_DEVICES=0" \
   -e "USE_WANDB=1" \
   -e "LOGDIR=/logdir" \
   -e "CUDNN_BENCHMARK='True'" \
   -e "CUDNN_DETERMINISTIC='True'" \
   -e "WORKDIR=/logdir" \
   -e "DATADIR=/data" \
   -e "IMAGE_SIZE=256" \
   -e "CONFIG_TEMPLATE=./configs/templates/semantic.yml" \
   -e "NUM_WORKERS=4" \
   -e "BATCH_SIZE=256" \
   catalyst-segmentation ./bin/catalyst-semantic-segmentation-pipeline.sh

The pipeline is running and you don’t have to do anything else, it remains to wait for the best model!

Visualizations

You can use W&B account for visualisation right after pip install wandb:

wandb: (1) Create a W&B account
wandb: (2) Use an existing W&B account
wandb: (3) Don't visualize my results

Tensorboard also can be used for visualisation:

tensorboard --logdir=/catalyst.segmentation/logs

4. Results

All results of all experiments can be found locally in WORKDIR, by default catalyst.segmentation/logs. Results of experiment, for instance catalyst.segmentation/logs/logdir-191107-094627-2f31d790, contain:

checkpoints

• The directory contains all checkpoints: best, last, also of all stages.
• best.pth and last.pht can be also found in the corresponding experiment in your W&B account.

configs

• The directory contains experiment`s configs for reproducibility.

logs

• The directory contains all logs of experiment.
• Metrics also logs can be displayed in the corresponding experiment in your W&B account.

code

• The directory contains code on which calculations were performed. This is necessary for complete reproducibility.

5. Customize own pipeline

For your future experiments framework provides powerful configs allow to optimize configuration of the whole pipeline of segmentation in a controlled and reproducible way.

• Common settings of stages of training and model parameters can be found in catalyst.segmentation/configs/_common.yml.

  • model_params: detailed configuration of models, including:
    • model, for instance ResnetUnet
    • detailed architecture description
    • using pretrained model
  • stages: you can configure training or inference in several stages with different hyperparameters. In our example:
    • optimizer params
    • first learn the head(s), then train the whole network

• The CONFIG_TEMPLATE with other experiment`s hyperparameters, such as data_params and is here: catalyst.segmentation/configs/templates/binary.yml. The config allows you to define:

  • data_params: path, batch size, num of workers and so on
  • callbacks_params: Callbacks are used to execute code during training, for example, to get metrics or save checkpoints. Catalyst provide wide variety of helpful callbacks also you can use custom.

You can find much more options for configuring experiments in catalyst documentation.