Converting model to Pytorch

[gullalc]

Hi @arashno ,

I need to convert your res152 recognition model to pytorch. I am using mmdnn for conversion, but I cannot find the output node which needs to be given to the mmdnn conversion command. I have visualized the graph using the meta file, and there is no name that looks like output, softmax or predictions.

Can you or anyone else help with the conversion?

[arashno]

Hi, For the recognition only model, the output node of the ResNet-152 model is the fully connected layer at resnet.py line 53. I think the name of the output node should be: 'output:0'. For the model with identifying, counting, and additional attributes outputs, the name of the species recognition output node should be 'output0:0' Best, Arash

[gullalc]

I have tried output:0 with the following command: mmconvert -sf tensorflow -in snapshot-55.meta -iw snapshot-55 --dstNode output:0 -df pytorch -om serengeti_resnet_152.pth

but it gives me the error that node is not in the graph. AssertionError: output:0 is not in graph

[arashno]

Actually, I am going to release a new version of this repo with new features in a few days. I will check convertibility to PyTorch for the new version. Please wait until I release the new version. Thanks, Arash

[arashno]

Unfortunately, I was unable to convert the Tensorflow models to PyTorch. But, the good news is that I will release a PyTorch version in 1-2 weeks. I finished the coding and now I am doing testing and preparing the trained models now. Arash

[r-barnes]

Just wondering what became of this?

[r-barnes]

@arashno : Did you ever release a PyTorch version?

[r-barnes]

@gullalc : It appears that valid node names for mmconvert include

 'output0/weights/Initializer/truncated_normal/shape'
 'output0/weights/Initializer/truncated_normal/mean'
 'output0/weights/Initializer/truncated_normal/stddev'
 'output0/weights/Initializer/truncated_normal/TruncatedNormal'
 'output0/weights/Initializer/truncated_normal/mul'
 'output0/weights/Initializer/truncated_normal'
 'output0/weights'
 'output0/weights/Assign'
 'output0/weights/read'
 'output0/weights/Regularizer/l2_regularizer/L2Loss'
 'output0/weights/Regularizer/l2_regularizer'
 'output0/biases/Initializer/Const'
 'output0/biases'
 'output0/biases/Assign'
 'output0/biases/read'
 'output0/weights/Momentum/Initializer/Const'
 'output0/weights/Momentum'
 'output0/weights/Momentum/Assign'
 'output0/weights/Momentum/read'
 'output0/biases/Momentum/Initializer/Const'
 'output0/biases/Momentum'
 'output0/biases/Momentum/Assign'
 'output0/biases/Momentum/read'

[arashno]

Hi, Yes, the PyTorch version is available at https://github.com/arashno/pytorch_multigpu_csv But, unfortunately, I didn't have enough computational resources to train a full model. If you have access to enough computational power, I am up to improving the PyTorch repository.