'TypeError: add_constant(): incompatible function arguments' while executing torch.Tensor.reshape()

[przybyszewskiw]

For the following code:

import torch
from torch import nn
from torch2trt import torch2trt


class Example(nn.Module):

    def __init__(self, stride=8):
        super(Example, self).__init__()
        self.stride = stride

    def grid_anchors(self, grid_size):
        grid_height, grid_width = grid_size
        shifts_x = torch.arange(
            0, grid_width * self.stride, step=self.stride, dtype=torch.float32
        )
        shifts_y = torch.arange(
            0, grid_height * self.stride, step=self.stride, dtype=torch.float32
        )
        shift_y, shift_x = torch.meshgrid(shifts_y, shifts_x)
        shift_x = shift_x.reshape(-1)
        shift_y = shift_y.reshape(-1)
        shifts = torch.stack((shift_x, shift_y, shift_x, shift_y), dim=1)
        return shifts.view(-1, 4)

    def forward(self, feature_map):
        grid_size = feature_map.shape[-2:]
        return self.grid_anchors(grid_size)


if __name__ == '__main__':
    model = Example()
    x = torch.ones(3, 400, 500)
    print(model(x))
    model_trt = torch2trt(model, [x])

I get an error:

 File "/opt/conda/lib/python3.6/site-packages/torch2trt-0.1.0-py3.6.egg/torch2trt/torch2trt.py", line 151, in trt_
TypeError: add_constant(): incompatible function arguments. The following argument types are supported:
    1. (self: tensorrt.tensorrt.INetworkDefinition, shape: tensorrt.tensorrt.Dims, weights: tensorrt.tensorrt.Weights) -> tensorrt.tensorrt.IConstantLayer

Invoked with: <tensorrt.tensorrt.INetworkDefinition object at 0x7f845d574110>, (400, 500), array([[   0.,    8.,   16., ..., 3976., 3984., 3992.],
       [   0.,    8.,   16., ..., 3976., 3984., 3992.],
       [   0.,    8.,   16., ..., 3976., 3984., 3992.],
       ...,
       [   0.,    8.,   16., ..., 3976., 3984., 3992.],
       [   0.,    8.,   16., ..., 3976., 3984., 3992.],
       [   0.,    8.,   16., ..., 3976., 3984., 3992.]], dtype=float32)

I'm using PyTorch version 1.6.0a0+9907a3e and TensorRT 7.1.2-1+cuda11.0. I was trying torch2trt both with and without plugins.

[jaybdub]

Hi @przybyszewskiw,

Thanks for sharing this issue.

There is an experimental PR to support adding custom converters for user defined methods.

#175

Because the grid map seems to be constant for a given image shape, I'm curious if this would help here.

For example,

def convert_example(ctx):
    module = ctx.method_args[0]
    feature_map = ctx.method_args[1]
    grid_size = feature_map.shape[-2:]

    # ... (insert TensorRT code to add grid as constant)

    output = ctx.method_return

    output._trt = # ...  set output _trt attribute to constant 
    
class Example(nn.Module):
 
    @tensorrt_method(convert_example)
    def forward(self, feature_map):
        # ...

Please let me know if this helps or you run into any issues.

Best,
John

[czs1886]

Got the similar error here:

TypeError: add_constant(): incompatible function arguments. The following argument types are supported:
    1. (self: tensorrt.tensorrt.INetworkDefinition, shape: tensorrt.tensorrt.Dims, weights: tensorrt.tensorrt.Weights) -> tensorrt.tensorrt.IConstantLayer

Invoked with: <tensorrt.tensorrt.INetworkDefinition object at 0x7fa66bc4aca8>, (336,), array([ 0 ..... 335])

I think mine is caused by the unsupported function torch.arange since it shows
Warning: Encountered known unsupported method torch.arange

[DuyguSerbes]

Hello, I also faced the same problem. Did you able to solve it?

  File "/media/duygu/3a331d5c-a3b3-468b-b8f7-06df52fefee6/ocakirog/tensorrt_local/torch2trt/torch2trt/torch2trt.py", line 364, in wrapper
    ret = attr(*args, **kwargs)
TypeError: add_constant(): incompatible function arguments. The following argument types are supported:
    1. (self: tensorrt.tensorrt.INetworkDefinition, shape: tensorrt.tensorrt.Dims, weights: tensorrt.tensorrt.Weights) -> tensorrt.tensorrt.IConstantLayer

[ultmaster]

For me, it was this line which induced the error.

  File "/home/xxx/python3.8/site-packages/torch/nn/modules/batchnorm.py", line 147, in forward
    self.num_batches_tracked = self.num_batches_tracked + 1  # type: ignore[has-type]

It doesn't make sense that tensorrt doesn't support batchnorm.

Then I realized that I forgot to call model.eval().