Open FengLiXue opened 4 years ago
@FengLiXue I was running into the same issue and after observing the pretrained_r3d
graph vs a model from a local training session on the same data it appears the pre-trained version is using a different (possibly older and simpler) network. It is also possible the provided r3d.tfrecords
may not be synchronized with the provided model since the timestamps appear as if they were uploaded months apart. I tried generating my own using create_tfrecord.py
but ran into more errors when attempting to load them into a training session.
Original network:
@zlzeng Can you elaborate a bit on the state of this code? It is great research but the repo seems a bit fragmented. It seems there are several people here interested in helping to update the repo if you could fill in a few missing details.
Hey @nini-lxz, can you offer any additional details? Again, awesome work and thank you for sharing!
Hi @alfarok @FengLiXue did you manage to find an answer ? I noticed the same issue when I try to use the prediction code after training...
Hey @SebastienButler, I never received a response so I assumed this repo was no longer being maintained. I ended up writing a modified version of the network architecture in TensorFlow 2/Keras and upgraded to Python 3.7. Interestingly the authors released a 2nd paper earlier this year continuing their research. Possibly there are plans to release updated supporting code?
In the demo.py to get the inputs in the graph when infer using the 'graph.get_tensor_by_name('inputs:0')'. But I cannot find where define this in the main.py when build the graph.