Some experiments with Tensorflow

• Based on https://github.com/tensorflow/tfjs-examples/tree/master/simple-object-detection

Annotation tool

• Use https://github.com/microsoft/VoTT

Notes on TF.JS on windows

• Start with the simple-object-detection from the TensorflowJS aka tfjs examples
• The @tensorflow/tfjs-node and @tensorflow/tfjs-node-gpu dependencies in package json use the outdated 1.3.x version, upgrade that to 1.7.4, which is the latest version for node 10.
• According to this issue you should run node v10.16.3
• Running on node v12 doesn't seem to work: building the canvas dependency on windows breaks mysteriously

Using the GPU

• Make sure your GPU is supported and make sure to have the latest drivers

• Install Visual Studio 2017 community edition (it has to be 2017 because otherwise CUDA 10.0 won't install - and it has to be CUDA 10.0 because tfjs seems to have a hardcoded dependency on the exact dll names of CUDA 10.0)

• Configure node-gyp to use VS2017

• Install CUDA Toolkit 10.0 (exact version) (see here and here)

  • Note: the PATHs to add are slightly different between CUDA 10.0 and CUDA 10.1, adapt accordingly

• Install cuDNN v7.6.5.32 (exact version)

• Set the environment variable to allow sharing the GPU memory: TF_FORCE_GPU_ALLOW_GROWTH=true as documented here

  • You might get the following error otherwise

  2020-08-31 00:23:11.623446: E tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_ALLOC_FAILED 2020-08-31 00:23:11.629529: E tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_ALLOC_FAILED (node:6636) UnhandledPromiseRejectionWarning: Error: Invalid TF_Status: 2 Message: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.