Attempting from_onnx, dimension 1 conflicts 0/1 on conv2d_transpose

#1

Hi, I’m attempting to compile CenterTrack with TVM. I’ve used the CT repo’s onnx conversion script to export a model.onnx file (I have torch==1.3 installed).

I’m now attempting to follow the Compile ONNX Models tutorial.

I’ve opened my model.onnx file in Netron to view the inputs, which are:

input_shapes = {'input.1': (1,3,736,1280), 'input.4: (1,3,736,1280), 'input.7': (1,3,736,1280)} (This is a tracking model which uses the current image, previous image frame, and a single-channel heatmap output from previous frame)

I’ve built TVM from source including both CUDA and LLVM. I’ve tried both targets on the step titled, " Compile the model with relay", however, at the line: mod, params = relay.frontend.from_onnx(onnx_model, shape_dict)

I get the following error on a conv2d_transpose layer:

  %187 = nn.relu(%186);
  %188 = nn.conv2d_transpose(%187, %dla_up.ida_0.up_1.weight, channels=1, kernel_size=[4, 4], strides=[2, 2], padding=[1, 1, 1, 1], groups=256) in particular dimension 1 conflicts 0 does not match 1; unable to unify: `Tensor[(256, 0, 4, 4), float32]` and `Tensor[(256, 1, 4, 4), float32]`; ;
  add(%188, %155)
}

Does anyone know what this means and how I can solve it?

Keras import: dimensions don't match in 3 layer example
#2

It looks like something does go wrong during the parsing. Relay is telling you that the input tensor has shape [256, 0, 4, 4], with the 0 representing the number of channels, while it expects it to have shape [256, 1, 4, 4]. So for some reason the shape propagation during parsing is ending up at 0 instead of 1. I’ll take a look and see if i can figure out what specifically is going wrong.

Can you post the arguments you used in CenterTrack’s onnx conversion script?

#3

Hmm do you know what common causes for that might be? Sure here’s how I ran the onnx conversion script, it should download weights if none are passed:

python3 -u convert_onnx.py tracking \
--exp_id onnx_001 \
--dataset mydata --dataset_version v1 \
--pre_hm --ltrb_amodal \
--same_aug --hm_disturb 0.05 \
--lost_disturb 0.4 --fp_disturb 0.1 \
--gpus 0,1 \
--input_h 736 --input_w 1280 \
--arch 'dla_34' --dla_node 'conv' \
> onnx_001.txt 2>&1 &