TypeError: upsample_bilinear2d() got an unexpected keyword argument 'align_corners' (occurred when translating upsample_bilinear2d)

[lizhen2017]

My pytorch version is 0.4.0; I know this is caused by the vesion of onnx and pytorch, how about your upsample.

[YukiNagato]

Bilinear upsample operation in pytorch can not be perfectly reimplemented in caffe, so right now only nearest upsample is supported. @lizhen2017

[lizhen2017]

@YukiNagato I solved this problem by edit pytorch source code torch/onnx/symbolic.py
I add a parameter for this function :align_corners
def upsample_bilinear2d(g, input, output_size, align_corners):
if align_corners:
return _unimplemented("upsample_bilinear2d", "align_corners == True")
w_scale = float(output_size[-1]) / input.type().sizes()[-1]
h_scale = float(output_size[-2]) / input.type().sizes()[-2]
return g.op("Upsample", input, width_scale_f=w_scale,
height_scale_f=h_scale, mode_s="bilinear")
and modify _operators.py
if mode=="bilinear":
layer = myf("Deconvolution", node_name, [input_name], [output_name],
convolution_param=dict(
num_output=channels,
kernel_size=2 * factor - factor % 2,
stride=factor,
pad=pad,
group=channels,
bias_term=False,
weight_filler=dict(type="bilinear")
)
The program can run normally.
Above all, I referred to the source code for pytorch 0.4.1

[YukiNagato]

Yes, but it should be noted that Deconvolution with bilinear filter is not equal to pytorch's upsample_bilinear2d. If you really need bilinear upsample, you can use Interp operation instead (https://github.com/hszhao/PSPNet/blob/master/src/caffe/util/interp.cpp). It is equal to upsample_bilinear2d with align_corners=True. Howerver, i haven't find any caffe code equal to upsample_bilinear2d with align_corners=False.