Ask a Question

Describe the question

Use the timm repository generate the onnx file, and use onnx function shape inference ( infer_shapes ), the flatten operator and expand operater appear the unknown value in tensor dim , Input is static [1 ,3, 224, 224].

System information

• OS Platform and Distribution (e.g. Linux Ubuntu 20.04):Linux Ubuntu 20.04
• ONNX version (e.g. 1.13): 1.17
• Python version: 3.10
• Protobuf version: 5.27

Reproduction instructions

• Describe the code to reproduce the behavior.

import onnx
import timm

model = timm.create_model('vit_base_patch16_224', pretrained=True)
torch.onnx.export(model, dummy_input, 'model.onnx', 
                  opset_version=17,
                  input_names=['input'], output_names=['output'])
inferred_model= onnx.load('model.onnx')
inferred_model = onnx.shape_inference.infer_shapes(inferred_model)
save_model(inferred_model, "infer_model.onnx")
...

Expected Behaviors

The expand and flatten operator should appear output shape , not unknow dim value.

Another question

The flatten opearor surported in opset_version[21, 13, 11, 9, 1], why when i use 17 opset version, the flatten operator decompose (shape/ slice/ concat/ reshape ) operator, i think it should be a fusion operator. Expand operator also appear this situation, it decompose of ( Mul Equal Where ) base operators.