A sample Azure Machine Learning services project for Transfer Learning-based custom image recognition by utilizing Keras.
- An Azure account (free trials are available).
- An installed copy of Azure Machine Learning Workbench with a workspace created. (Create Azure Machine Learning Preview accounts and install Azure Machine Learning Workbench)
- This example could be run on any compute context.
This project uses Training dataset, Test dataset and Validation dataset.
For preparing this directory structure,
run train_test_split.py
like below on your dataset and one whole dataset is split into 3 directories (training, testing, validation).
python -m train_test_split --image_dir data/my_photos --output_dir data/my_split --pct_test 10 --pct_validation 20 --seed 1337
Then compress them into one zip file.
Store it under container in Azure Blob Storage and add the following references to your .runconfig file to load dataset into your Azure Machine Learning compute target automatically for your training:
EnvironmentVariables:
"STORAGE_ACCOUNT_NAME": "<YOUR_AZURE_STORAGE_ACCOUNT_NAME>"
"STORAGE_ACCOUNT_KEY": "<YOUR_AZURE_STORAGE_ACCOUNT_KEY>"
"CONTAINER_NAME" : "<YOUR_STORAGE_CONTAINER_NAME>"
"ZIPFILE_NAME" : "<YOUR_ZIPPED_DATASET_NAME>"
This documentation tells you how to set up GPU VMs for Azure Machine Learning Services.
You need to modify the auto-generated runconfig file after the registration of your VM.
Run train_keras.py
like below from Azure Machine Learning.
az ml experiment submit -c YOUR_VM_TARGET ./train_keras.py --gpu 2 --use_weights True --score True --learning_rates 0.001 0.0005 0.00002 --epochs 20 10 10 --model_type Xception
All of run histories, logs and trained models are managed by Azure Machine Learning Services.
There are three scripts in this module - one for splitting your image data, one for training your model, and the final for scoring your model and getting a picture of the confusion matrix.
train_test_split.py
- Splits an existing directory of image files (with subdirectories named for the labels)
- Uses hashing on the filenames to ensure that even if you add more files, existing files will hash into the same buckets they did previously
train_keras.py
- Trains a Keras model for image recognition based on Transfer Learning. See Keras Applications for a sense of the pre-built models that are available.
- As new models are launched or you develop/find them, it's easy to add them to this script and use them as base models for your own transfer-learned versions.
- Models are automatically named based on the hyperparameters given, and can also have a timestamp appended. However, this can be overridden if needed.
- Data augmentation flags are supported, and include the methods supported by Keras (flipping horizontally, vertically, rotation, zooming, and shearing).
- You can bundle training and scoring if desired.
- As part of the training, a Markdown file describing the model is produced. If you are also doing scoring, details on the performance will also be included.
score_keras.py
- Takes an existing trained model and runs multiple passes over the test-set of images, generating a confusion matrix and precision, recall and F-score values.
- All scripts support the
--help
flag to give you details on usage, and should guide you in required vs. optional parameters. If you have any issues, please feel free to reach out.
python -m train_test_split --image_dir data/my_photos --output_dir data/my_split --pct_test 10 --pct_validation 20 --seed 1337
- Splits images in
data/my_photos
into output directorydata/my_split
- 10% of the images in each class go into
testing
, 20% intovalidation
, and the other 70% intotraining
- Splits images in
az ml experiment submit -c YOUR_VM_TARGET ./train_keras.py --model_type InceptionV3 --batch_size 8 --learning_rates 0.001 0.001 0.0005 --epochs 1 1 1 --use_weights True --score True --gpu 1
- Trains a new model based on a pre-trained InceptionV3 instance.
- This script needs to be run through Azure Machine Learning Services to manage trained models, logs and output files automatically by it.
- Since
output_model
is not specified, the name is imputed from the hyperparameters. - Uses class-weights based on inverse class distribution.
- Scores the results and writes out confusion matrix and description markdown file.
python -m score_keras --model_dir ./models --model_name InceptionV3_avg-1024_relu_lr001-001-0005_wts_e1-1-1 --num_batches 20 --image_dir ./data/my_split
- Scores an existing model (
./models/InceptionV3_avg-1024_relu_lr001-001-0005_wts_e1-1-1.h5
) - Uses images from
./data/my_split/testing
for evaluation. - Stores confusion matrix and scores into
./models
, into a.png
and.csv
respectively.
- Scores an existing model (
See LICENSE
. Overall code is licensed under MIT license. train_test_split.py
is Apache v2.0 licensed because it is adapted from TensorFlow code (see comments in that file).