Tensorflow Huge Model Support (HMS)

This library is designed to speed up huge model training on unified memory. It takes a computation graph built by the user, conducts analysis, implements group execution and prefetch by editing the graph. A callback hook is provided to easily apply HMS on a tf.keras model.

Publications

Chen, CL., Chen, CC., Yu, WH. et al. An annotation-free whole-slide training approach to pathological classification of lung cancer types using deep learning. Nat Commun 12, 1193 (2021). https://doi.org/10.1038/s41467-021-21467-y

Chuang, WY., Chen, CC., Yu, WH. et al. Identification of nodal micrometastasis in colorectal cancer using deep learning on annotation-free whole-slide images. Mod Pathol (2021). https://doi.org/10.1038/s41379-021-00838-2

License

Copyright (C) 2021 aetherAI Co., Ltd. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode).

Requirements

• Tensorflow v1 (tensorflow-gpu==1.15.3)
• GCC >= 7

Installation

To install HMS, simply run the following commands:

[CUDA_PATH=YOU_CUDA_PATH] pip install .

, where CUDA_PATH is /usr/local/cuda by default.

Usage

HMS can be simply applied on tf.keras model by a callback function, as described below.

1. Import HMS tf_keras module.

from tensorflow_huge_model_support.tf_keras import init, HMSTFKerasCallback

2. Call init before model building(, and after horovod initializes).

Without horovod:

init()

With horovod:

import horovod.tensorflow.keras as hvd
hvd.init()
init(hvd=hvd)

3. Define a HMSKerasCallback.

hms_callback = HMSTFKerasCallback(
    hvd=hvd,
    default_batch_size=DEFAULT_BATCH_SIZE
)

, where hvd can be skipped if not using Horovod.

4. Pass the callback to the Keras fit or fit_generator function.

model.fit_generator(..., callbacks=[hms_callback] + OTHER_CALLBACKS, ...)

Note: Don't forget to add hvd.callbacks.BroadcastGlobalVariablesCallback(0) in the callback list if using Horovod.