Deep Learning Model Generation

Introduction

To do

How to run

Clone this repository, create a python env for the project (optional) and activate it. Then install all the dependencies with pip

git clone git@github.com:riccardolops/deep_learning_model-mesh_generation.git
cd deep_learning_model-mesh_generation
python -m venv env
source env/bin/activate
pip install -r requirements.txt

Run

Run the project as follows:

python main.py [-h] -c CONFIG [--verbose]

arguments:
  -h, --help            show this help message and exit
  -c CONFIG, --config CONFIG
                        the config file used to run the experiment
  --verbose             To log also to stdout

E.g. to run the generation experiment, execute:

python main.py --config configs/gen-training.yaml

YAML config files

You can find the config files used to obtain the best result in the config folder. Two files are needed: experiment.yaml, augmentations.yaml. For both the two tasks, the best config file is provided:

• gen-training.yaml for the network which, from the sparse annotation, generate the dense labels
• seg-pretrain.yaml which train the segmentation network only over the generated labels
• seg-finetuning.yaml which train the segmentation network over the real dense labels Execute main.py with these 3 configs in this order to reproduce our results

Checkpoints

To do: Download the pre-trained checkpoints here

experiment.yaml

experiment.yaml describe each part of the project, like the network/loss/optimizer, how to load data and so on:

# title of the experiment
title: canal_generator_train
# Where to output everything, in this path a folder with
# the same name as the title is created containing checkpoints,
# logs and a copy of the config used
project_dir: '/path/to/results'
seed: 47

# which experiment to execute: Segmentation or Generation
experiment:
  name: Generation

data_loader:
  dataset: /path/to/dataset
  # null to use training_set, generated to used the generated dataset
  training_set: null
  # which augmentations to use, see: augmentations.yaml
  augmentations: configs/augmentations.yaml
  background_suppression: 0
  batch_size: 2
  labels:
    BACKGROUND: 0
    INSIDE: 1
  mean: 0.08435
  num_workers: 8
  # shape of a single patch
  patch_shape:
  - 120
  - 120
  - 120
  # reshape of the whole volume before extracting the patches
  resize_shape:
  - 168
  - 280
  - 360
  sampler_type: grid
  grid_overlap: 0
  std: 0.17885
  volumes_max: 2100
  volumes_min: 0
  weights:
  - 0.000703
  - 0.999

# which network to use
model:
  name: PosPadUNet3D

loss:
  name: Jaccard

lr_scheduler:
  name: Plateau

optimizer:
  learning_rate: 0.1
  name: SGD

trainer:
  # Reload the last checkpoints?
  reload: True
  checkpoint: /path/to/checkpoints/last.pth
  # train the network
  do_train: True
  # do a single test of the network with the loaded checkpoints
  do_test: False
  # generate the synthetic dense dataset
  do_inference: False
  epochs: 100

augmentations.yaml

augmentations.yaml defines which type of augmentations use during training. Two different augmentations files have been used, one for the segmentation task, one for the generation task. The file should follow this structure:

RandomAffine:
  scales: !!python/tuple [0.5, 1.5]
  degrees: !!python/tuple [10, 10]
  isotropic: false
  image_interpolation: linear
  p: 0.5
RandomFlip:
  axes: 2
  flip_probability: 0.7