Question about porting from caffe to Tensorflow

[jakeoung]

Thank you very much for your excellent work! I just wonder how we can port a caffe model to Tensorflow (because I want to port another caffe model to Tensorflow)

Did you train FlowNet2 from the scratch, or use the original caffe model and port it?

[sampepose]

Hi! I rewrote the code in TF. You have to be careful since not all ops are the same between caffe and tf. For instance, deconv is slightly different, which is why I manually crop the resulting feature maps. I am reusing the caffe weights. I did not retrain (but the code is there if you want to!). Take a look at scripts/caffe to see how I did that. My advice is to write your tf model incrementally and load your transferred weights incrementally. That way you can ensure each layer is providing expected results. I found it helpful to do a sample forward pass in caffe, noting the min/max/std/mean of each layer output. Compare that with tf to ensure your implementation matches. It likely won’t be the exact same due to round off errors but it’ll be close.

…

Sent from my iPhone

On Oct 20, 2017, at 4:11 PM, koo ***@***.***> wrote: Thank you very much for your excellent work! I just wonder how we can port a caffe model to Tensorflow (because I want to port another caffe model to Tensorflow) Did you train FlowNet2 from the scratch, or use the original caffe model and port it? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

[kindloaf]

Hi @sampepose
You mentioned that for training with Tensorflow "the code is there", and I saw train.py.
Have you used train.py to train a model with similar performance as the original caffe model?