leovinus2001
leovinus2001
Same issue 4.0b3 and PyTorch 1.6.0
Still an issue with coremltools 4.0b3 and PyTorch 1.6.0
Setting bidirectional to False i.e. uni directional makes conversion work. I believe this is your clue. >Mismatch: 50% Seems to me this is not a converter issue but a PyTorch...
Mhh. How about you use torch.manual_seed(0) # deterministic output and reduce the hidden_size to 2 and just print out the values of the h0/c0 and ex_hn/ex_cn for clues?
Out of curiosity, I was digging a bit deeper. Issue still unsolved but more intriguing and with a little extra insight I hope that helps. Hope you do not mind....
TL;DR I suspect the converter ObjC/C++ code makes a mistake with reformats of the h_n/c_n before you see the predict result in Python, but I cannot confirm it atm. That...
The story continues based on the previous message: >In even more detail, backtrace for > coreml_results = mlmodel.predict({"x": input.numpy()}, useCPUOnly=True) >is roughly > >We call > >a) coremltools//models/model.py L302 >...
Let's summarize what this bug seems to mean for Bidirectional LSTMs in CoreML and your Apps. Is this about correct? 1) y = blstm(x) without input h_0/c_0 nor use of...
PS: This issue sounds an awful lot like issue #263 which is two years old, ugh. In #263, the suggested workaround is ok but this mangling of the hn/cn outputs...