Convert model to torch script via scripting, got ---TypeError: cannot create weak reference to 'builtin_function_or_method' object

1

Hi, I am trying to convert my LSTM model to torch script, then used in a production environment, I did it according to the guide lines in < DEPLOYING A SEQ2SEQ MODEL WITH THE HYBRID FRONTEND> (https://pytorch.org/tutorials/beginner/deploy_seq2seq_hybrid_frontend_tutorial.html) , but I encountered the problem as follows:

File “/home/ner/src/model/bilstm.py”, line 76, in init
self.forward_lstm = LatticeLSTM(lstm_input, lstm_hidden, gaz_dropout, gaz_alphabet_size, gaz_emb_dim, gaz_embedding, True, HP_fix_gaz_emb, False)
File “/home/.local/lib/python2.7/site-packages/torch/jit/init.py”, line 891, in init_then_register
_create_methods_from_stubs(self, methods)
File “/home/.local/lib/python2.7/site-packages/torch/jit/init.py”, line 852, in _create_methods_from_stubs
self._create_methods(defs, rcbs, defaults)
File “/home/.local/lib/python2.7/site-packages/torch/jit/init.py”, line 603, in _try_compile_weak_script
entry = _compiled_weak_fns.get(fn)
File “/usr/lib64/python2.7/weakref.py”, line 284, in get
return self.data.get(ref(key),default)
TypeError: cannot create weak reference to ‘builtin_function_or_method’ object

here is part of my code:

class MyLSTM(torch.jit.ScriptModule):
constants = [‘left2right’,‘use_gpu’,‘hidden_dim’]
def init(self, input_dim, hidden_dim, word_drop, word_alphabet_size, word_emb_dim, pretrain_word_emb=None, left2right=True, fix_word_emb=True, use_gpu = False, use_bias = True):
super(LatticeLSTM, self).init()
skip_direction = “forward” if left2right else “backward”
print "build LatticeLSTM… “, skip_direction, “, Fix emb:”, fix_word_emb, " gaz drop:”, word_drop
self.use_gpu = use_gpu
…

@torch.jit.script_method
def forward(self, input, skip_input_list):
# type: (Tensor,Tuple[List[List[List[int]]], bool])
skip_input = skip_input_list[0]
…

Any thoughts?

2

Solved. It is because I mistakenly delete the forward function. If you encountered the same problem,check your forward function and all the functions the forward function calls, make sure to add torch.jit.script decorator to the functions and torch.jit.script_method decorator to your module’s methods.