[Relay][ONNX] KeyError: ‘W’ when compiling from onnx

[mnboos]

Dear tvm team

I'm aware that I'm not supposed to use this issue tracker as a first instance. However, in the community was no reaction and in the meantime the model I uploaded was deleted. So I added it here as an attachment, so it won't get lost anymore.
yolo_failing.zip

The attached model leads to an exception during the compilation:

 File "C:\src\xyz\src\tvm\compile_model.py", line 78, in compile_model
    relay_func, params = relay.frontend.from_onnx(onnx_model, shape=input_shape_dict)
  File "C:\Users\bmar\AppData\Roaming\Python\Python36\site-packages\tvm-0.6.dev0-py3.6-win-amd64.egg\tvm\relay\frontend\onnx.py", line 1139, in from_onnx
    sym, params = g.from_onnx(graph, opset)
  File "C:\Users\bmar\AppData\Roaming\Python\Python36\site-packages\tvm-0.6.dev0-py3.6-win-amd64.egg\tvm\relay\frontend\onnx.py", line 960, in from_onnx
    inputs = [self._nodes[self._renames.get(i, i)] for i in node.input]
  File "C:\Users\bmar\AppData\Roaming\Python\Python36\site-packages\tvm-0.6.dev0-py3.6-win-amd64.egg\tvm\relay\frontend\onnx.py", line 960, in <listcomp>
    inputs = [self._nodes[self._renames.get(i, i)] for i in node.input]
KeyError: 'W'

[tqchen]

cross ref https://discuss.tvm.ai/t/relay-onnx-keyerror-w-when-compiling-from-onnx/2102/3

Given that you have a good idea about what is happening, can you attempt a fix to the problem?

[mnboos]

I will clarify with my boss and come back to you.

[aitest11]

When I load RNN onnx model to relay, I have encountered similar issue:

Traceback (most recent call last):
File "from_onnx.py", line 96, in
sym, params = relay.frontend.from_onnx(onnx_model, shape_dict, dtype='float32')
File "/tvm/python/tvm/relay/frontend/onnx.py", line 1144, in from_onnx
sym, params = g.from_onnx(graph, opset)
File "/tvm/python/tvm/relay/frontend/onnx.py", line 965, in from_onnx
inputs = [self._nodes[self._renames.get(i, i)] for i in node.input]
File "//tvm/python/tvm/relay/frontend/onnx.py", line 965, in
inputs = [self._nodes[self._renames.get(i, i)] for i in node.input]
KeyError: ' '

[jroesch]

Could you send the parameters to the from_onnx call?

For example with:

import onnx
from tvm import relay

yolo = onnx.load("/Users/jroesch/Downloads/yolo_failing.onnx")
import pdb; pdb.set_trace()
# What shapes do we pass?
relay.frontend.from_onnx(yolo)

[jroesch]

cc @zhreshold could you look into this?

[zhreshold]

I actually digged into and found some bugs.
Here's the fix: #3143

@mnboos
However, the provided yolo_failing.zip is not a valid onnx model, the nodes are not in topological order which will cause failure when iterating throught the nodes in onnx.graph.
Please let me know if you have a proper model and I can add it to the unittest.

[mnboos]

@zhreshold Thanks for your efforts. The model was created with the official keras2onnx converter, I'll file an issue there. On the other hand it would be nice if TVM could either first sort topologically if necessary or provide a useful error message.

[jroesch]

Is it possible to just directly use the Keras importer in that case?

[mnboos]

Theoretically yes, but we're looking into TVM and TVM handles onnx as input, which lead to the error above. Using another library would just be avoiding a symptom, whereas I'm more a friend of solving problem at its root. Don't get me wrong, I don't want to make accusations, I'm just saying there is room for improvement in terms of robustness in TVM (when it comes to handling (invalid) onnx). Since the ONNX specification states, that the graph is topologically sorted, it would be sufficient for TVM to just reject the ONNX with an error like "This graph is not topologically sorted and thus invalid."

[tqchen]

Close by #3143, we welcome contribution for a topo correctness checker in onnx importer