dynamic shape inference with onnx model

[jonas-doevenspeck]

After exporting the model to onnx, onnxruntime fails to do inference with inputs that have a different shape than the one used to export the model.

To reproduce:

import onnxruntime
import torch
import numpy as np

model = SwinIR(
    upscale=4,
    in_chans=3,
    img_size=64,
    window_size=8,
    img_range=1.0,
    depths=[6,6,6,6,6,6],
    embed_dim=180,
    num_heads=[6,6,6,6,6,6],
    mlp_ratio=2,
    upsampler="pixelshuffle",
    resi_connection="1conv",
)

model.load_state_dict('001_classicalSR_DF2K_s64w8_SwinIR-M_x4.pth')

torch.onnx.export(model, torch.randn(1, 3, 64, 64), 'swin_ir.onnx', dynamic_axes = {"input": [0,2,3],"output": [0,2,3]}, verbose=False, opset_version=17, input_names=["input"], output_names=["output"])

ort_session = onnxruntime.InferenceSession('swin_ir.onnx')

x = np.random.randn(1, 3, 72, 72).astype(np.float32)
ort_inputs = {ort_session.get_inputs()[0].name: x}
ort_outs = ort_session.run(None, ort_inputs)

error message:

[E:onnxruntime:, sequential_executor.cc:514 ExecuteKernel] Non-zero status code returned while running Reshape node. Name:'/layers.0/residual_group/blocks.1/attn/Reshape_1' Status Message: [/onnxruntime_src/onnxruntime/core/providers/cpu/tensor/reshape_helper.h:40](https://vscode-remote+ssh-002dremote-002b10-002e30-002e2-002e76.vscode-resource.vscode-cdn.net/onnxruntime_src/onnxruntime/core/providers/cpu/tensor/reshape_helper.h:40) onnxruntime::ReshapeHelper::ReshapeHelper(const onnxruntime::TensorShape&, onnxruntime::TensorShapeVector&, bool) gsl::narrow_cast<int64_t>(input_shape.Size()) == size was false. The input tensor cannot be reshaped to the requested shape. Input shape:{88,6,64,64}, requested shape:{1,64,6,64,64}

environment: pytorch==1.13.1, onnxruntime==1.15.1, onnx==1.12.0

[pg-alfredlee]

Hi @jonas-doevenspeck , do you have any luck now on the onnx model inference?

[SherryXieYuchen]

Hi, I fix it by editing network_swinir.py:
1、replace
B = int(windows.shape[0] / (H * W / window_size / window_size))
x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
return x

by
B = windows.shape[0] / (H * W / window_size / window_size)
x = windows.view(B.numel(), H // window_size, W // window_size, window_size, window_size, -1)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B.numel(), H, W, -1)
return x

2、replace
# W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
if self.input_resolution == x_size:
attn_windows = self.attn(x_windows, mask=self.attn_mask) # nWB, window_sizewindow_size, C
else:
attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device))

by
# W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device))

then do the pytorch2onnx, dynamic input should work when inference with onnx model. Hope it helps.