KeyError in from_pytorch()

Questions
1

I tried to implement from_pytorch.py on x86_64 linux(Ubuntu) following the tutorial of Compile PyTorch Models. However, it raises a KeyError: ‘279’ at this line inputs = _get_op_inputs(op_node, outputs, output_index_map) where

def _get_op_inputs(op_node, outputs, output_index_map):
  input_names = [output_index_map[name]
               for name in _get_input_names(op_node)]
  return [outputs[name] for name in input_names]

It raises the error when dealing with op_node %281 : int[] = prim::ListConstruct(%279, %280), scope: __module.layer2/__module.layer2.0/__module.layer2.0.downsample/__module.layer2.0.downsample.0 It seems like 279 or 280 is not included in output_index_map, which raises an KeyError. torch version=1.4.0, torchvision version=0.5.0 I wonder why this happens and how to solve it. Thank you.

2

Thank you for trying it out, can you post your script?

3

Thank you for your reply. The script is the same as the tutorial from_pytorch.py It works on one computer but has bugs above on another computer. I wonder if it is the configuration.

4

I tried the tutorial now on torch 1.4 and torchvision 0.5, it works. Maybe add print(graph) below to see if the graph has changed between your two envs.

5

I compared the graphs of two envs and found they are the same. And the output_index_map is the same, the only difference is the orders of keys. The env that does not work also has KeyError: 'input.5' , and I suppose that all input.x(except input.1) will have KeyError under that env. But I don’t quite understand why the env that works for from_pytorch.py could find key ‘input.5’ in output_index_map. Since I print(output_index_map) and found no key named ‘input.5’. Is there any conversion of the keys? Thanks

6

output_index_map is gradually updated to have more keys during conversion, so even if when you do print(output_index_map) the key is not there, it might be added later.

7

Thanks. I did not realize that output_index_map is gradually updated. I think probably it is the order of operators dictionary in parse_operators() function matters. Each node_name has dependencies of the previous ones. So if output_index_map is updated out of order, there will be a KeyError. Though I did not figure out why the operators dictionary has random orders. Thanks again for your help.

8

hmm that’s a good point. The output of graph.nodes() is topologically sorted, but we are making a dict out of it, that may break the topological order guarantee.

Can you change ops dict at

to be a list of (node_name, node)? That will guarantee the topological order.

You also need to remove .items() here

9

ok I found the answer to this issue. In Python 3.6 and above, the order of key is insertion order, see https://stackoverflow.com/questions/39980323/are-dictionaries-ordered-in-python-3-6

So our impl is broken for py 3.5 and below, can you confirm if one of your python version is 3.5 or lower?

cc @alexwong @zhiics @jwfromm

10

My python3 version is 3.5.2. Thanks for your help!

11

You can send a PR to fix this, I can merge asap