TensorGate

---

Version	License	Linux CPU	Linux GPU	Windows CPU	Windows GPU	Mac GPU

TensorGate is an open-source framework based on TensorFlow deep learning open source library. It offers a set of upper layer demo for a variety of deep learning, such as image classification, recognition, segmentation, GAN network, etc. Also, it offeres a set of analyizer tools to parse the log data.

Requirements

• python > 3.6
• tensorflow >= 1.7.0
• CUDA == 9.0
• cuDNN == 7.0
• python-opencv
• json

Usage

# running to train mnist dataset on GPU-0
$ python main.py 0 -dataset=mnist
# load config file to train mnist on GPU-0
$ python main.py 0 -dataset=mnist -extra=demo.json

Directory

• <asserts> used by example code
• <demo> pre-setting config file
• <gate> provide critical functions for running gate framework
  • <config> provide config file for a variety of datasets
    • <dataset> a series of dataset config to index
      • mnist.py MNIST classification
    • config_base.py config base class
    • config_params.py config params to set parameters
    • config_factory.py config factory to index the dataset
  • <data> data index and prefetech method
    • <tfqueue> using tensorflow queue and batch prefetch method
    • <custom> customed data model by using placeholder
    • data_utils.py a unified tools for all data model
    • data_factory.py data factory to index the data model
  • <layer> customed loss/net/ops assemble
  • <net> collect a variety of network models
    • <custom> customed network for specific task
    • <deepfuse> multi-layer weight shared network model
    • <nets> slim net model
    • <vae> varational auto-encoder model zoos
    • net_factory.py net factory to index the model
    • net_model.py model parameter config assemble
  • <solver> offer the training tools and snapshot
    • learning_rate.py provide different learning rate decay method
    • optimizer.py provide various optimizer
    • snapshot.py dumping the ckpt in according to a certain of iterations
    • summary.py record the running info
    • updater.py updater gradient to weights
  • <util> system utils
    • checkpoint.py load and operate tensorflow ckpt
    • devicequery.py help to search/show device info
    • filesystem.py folder/file operation
    • logger.py event helper
    • profiler.py detecter of network model
    • similarity.py a cosine-metric method
    • string.py string/tensor-string operation
    • variable.py search/print/select variables in the network
    • heatmap.py generate heatmap image for visual data
  • <processing> data processing method
    • <slim> including slim processing method and corresponding interface
    • processing_vision.py vision data processing
    • processing_audio.py audio data processing
    • processing_text.py text type data processing
  • context.py the running context to manage the app
  • env.py a global output control center
• <samples> offer some examples for current deep learning tasks
  • <vision> several common vision tasks
    • classification.py
    • regression.py
    • mask_rcnn.py
    • vanilla_gan.py
    • vanilla_vae.py
  • <avec2014> research on AVEC2014 (multi-modal) depressive recognition tasks
  • <kinship> research on Kinship recognition tasks
  • <trafficflow> research on TrafficFlow dataset tasks
• <tools> some external tools to analyze the data and log event files
  • drawer.py provide a set of drawing tools by using log file
  • dataset.py provide tool to generate train/val/test file for specific data folder
• main.py start and initialize the system
• pipline.py execute multi-task at once
• compile.py packege gate into a fold in .bin or .py without debug info

To-Do

• (05/15/18) Merge drawer into gate
• (05/12/18) Resume pipline & fix a bug of config-base re-write
• (05/12/18) Merge drawer in
• (05/08/18) Review Kinface related issues
• (05/08/18) Review AVEC2014-CNN/HEATMAP/FLOW/BICNN/Audio-NET
• Reconstruct gate framework to make more flexible
• Add mask-rcnn trainig module
• Add mask-rcnn inference module
• Add mask-rcnn visualization method
• (04/08/18) Add batchnorm params to update collections (fixed BN un-trained)
• (03/08/18) Add GradCAM, guidedCAM, guided backpropagation
• (03/08/18) Add heatmap for AVEC2014-Image
• (03/08/18) Package heatmap as a class
• Update slim model to TensorGate
• Update net factory logic: argscope in the head of net model
• Package functions with class
• Learning rate: add cosine, linear cosine, noisy linear cosine, inverse time
• Optimizer: add proximal, proximal adagrad
• Env: summary, logger, compiler
• Move classical model to ./example
• Move project model to ./issue
• Separate preprocessing method by input format
• Re-construct data layer
• Move a part of classical method to example folder

License

Apache License 2.0