Electron-Diffraction-CNN

A small ML Training and Evaluation Framework for reliable experimentation.

Setup

This project targets a GPU-enabled Linux Workstation. Additional work may be required for testing on other operating systems or on CPUs.

We target Python 3.6+.

Create/Activate a virtual environment (via anaconda, virtualenv, or pyenv) Recommended: Anaconda
pip install -e .

Dev Install

Development installations are recommended if you'd like to contribute to this library.

pip install -e .[dev]

Testing

python setup.py test

Thereafter, coverage reporting can be found at docs/cov/

How to Run Experiments

Collecting data into a manifest file

This project includes a make_manifest.py script that assumes the following:

All of your EBSD data exists under a single directory (likely called data) on your filesystem.
The folders that contain image data (.tiff files) also contain a Grain List.txt file with metadata about the images in the folder.
The folder structure for the images, while arbitrary, will not change at the time experiments are run.

Before training starts, first create a manifest file to represent your dataset.

Splitting data manifest into groups

We also include a split_manifest.py script whose purpose is to divide the EBSD image dataset into train, test, and validation groups.

After creating a manifest file and before starting experimenting, please use this script to mutate the manifest file for repeatable experimentation.

This script should add a column (likely called _split) that will be used in the training process.

Train a Model

Define a model through subclassing either a Regression or Classification type in the src/vecchio/models.py file. Thereafter, train your model in a repeatble way using the train.py or regression_train.py scripts below. Consider saving the exact command that is run, because the specific arguments will be integral in the evaluation phase. See documentation below for arguments and example uages.

Evaluate the model performance.

For classification or regression, please make use of the eval.py and regression_eval.py scripts (docs below).

Docs

`make_manifest.py --help`

usage: make_manifest.py [-h] [-o OUTPUT] source_directory

Produce an EBSD data manifest file from a root directory.

positional arguments:
  source_directory      Root directory to begin parsing data

optional arguments:
  -h, --help            show this help message and exit
  -o OUTPUT, --output OUTPUT
                        /path/to/output.csv Default: `manifest.csv`

Example usage:

python src/make_manifest.py data/

python src/make_manifest.py data/ -o my_manifest.csv

`split_manifest.py --help`

usage: split_manifest.py [-h] [-ts TEST_SIZE] [-vs VAL_SIZE] [-s SEED]
                         [-sm {shuffle,stratified-shuffle}] [-o OUTPUT]
                         manifest label_column

Split the data in the manifest file into test, train, and validation.

positional arguments:
  manifest              path/to/manifest.csv
  label_column          Name of column considered the label or `y`.

optional arguments:
  -h, --help            show this help message and exit
  -ts TEST_SIZE, --test-size TEST_SIZE
                        Ratio of dataset to include in test set
  -vs VAL_SIZE, --validation-size VAL_SIZE
                        Ratio of dataset to include in validation set
  -s SEED, --seed SEED  Random seed
  -sm {shuffle,stratified-shuffle}, --split-method {shuffle,stratified-shuffle}
  -o OUTPUT, --output OUTPUT
                        (optional) /path/to/mutated/copy/of/manifest.csv

Example usage:

python src/split_manifest.py manifest.csv 'Phase'

python src/split_manifest.py manifest.csv 'Phase' -s 42 -sm 'stratified-shuffle'

python src/split_manifest.py manifest.csv 'Phase' -o split_manifest.csv -vs 0.1 -sm 'stratified-shuffle'

`train.py --help`

usage: train.py [-h] [--weight-classes] [-bs BATCH_SIZE] [-e EPOCHS]
                [-md MIN_DELTA] [-p PATIENCE] [-o OUTPUT]
                {XceptionClassifier,ResNet50Model} manifest label_column

Train a Classification ML Model

positional arguments:
  {XceptionClassifier,ResNet50Model}
                        The model to use for classification.
  manifest              path/to/manifest.csv
  label_column          Column name for label (y) from manifest file.

optional arguments:
  -h, --help            show this help message and exit
  --weight-classes      Flag that turns on class balancing (should not be used
                        with regression).
  -bs BATCH_SIZE, --batch-size BATCH_SIZE
                        Training batch size (default: 32)
  -e EPOCHS, --epochs EPOCHS
                        Number of epochs (default: 10000)
  -md MIN_DELTA, --min-delta MIN_DELTA
                        Minimum change int the monitored quantity to qualify
                        as an improvement. Default: 0.001
  -p PATIENCE, --patience PATIENCE
                        Number of epochs with no improvement after which
                        training will be stopped Default: 25
  -o OUTPUT, --output OUTPUT
                        (optional) path/to/output/directory/

Example Usage:

python src/train.py XceptionClassifier manifest.csv Phase -e 1

`eval.py --help`

usage: eval.py [-h] [-o OUTPUT] [-bs BATCH_SIZE]
               trained_model manifest label_column

Evaluate a Classification ML Model

positional arguments:
  trained_model         /path/to/model_checkpoint.h5
  manifest              path/to/manifest.csv
  label_column          Column name for label (y) from manifest file.

optional arguments:
  -h, --help            show this help message and exit
  -o OUTPUT, --output OUTPUT
                        (optional) path/to/output/directory/
  -bs BATCH_SIZE, --batch-size BATCH_SIZE
                        Training batch size (default: 32)

Example usage:

python src/eval.py models/XceptionClassifier-0.0.0/2019-04-07-06-37/model_checkpoint.h5 manifest.csv Phase -o class_eval.csv

FAQ

Q: Error: Failed to load the native TensorFlow runtime. (I don't have a GPU / Haven't set up GPU Drivers) A: Try installing a non-gpu version of tensorflow: pip install tensorflow

krkaufma / electron-diffraction-cnn Goto Github PK