"Conv: dilations must be specified for 2 axes" when bundling a 3D conv network, with or without dilations specified #3773
Description
I used 'torch.onnx.export' to convert my pytorch model to ONNX. When I run the bundlization on my 3D convnet with:
../bin/x-model-builder -input-tensor-dims=1,2,1320,14,8 -output-tensor-names=out -model-input-name=in -model ce_no_dilation.onnx -emit-bundle=. -backend=CPU -network-name=bundle_ce_no_dilation
I get the following error
../lib/Importer/ONNXModelLoader.cpp:1835 got an unexpected ErrorValue: location: ../lib/Importer/ONNXModelLoader.cpp:574 message: Conv: dilations must be specified for 2 axes.
Any thoughts on what could be causing this? I'm not sure if to look at the original model the ONNX conversion or the Glow bundling process?
my ONNX model graph:
graph torch-jit-export ( %in[FLOAT, 1x2x1320x14x8] ) optional inputs with matching initializers ( %learned_0[FLOAT, 20x2x7x5x5] %learned_1[FLOAT, 20] %learned_2[FLOAT, 20x20x5x5x5] %learned_3[FLOAT, 20] %learned_4[FLOAT, 50x20x3x3x3] %learned_5[FLOAT, 50] %learned_6[FLOAT, 100x50x3x3x3] %learned_7[FLOAT, 100] %learned_8[FLOAT, 2x100x1x1x1] %learned_9[FLOAT, 2] ) { %11 = Conv[dilations = [1, 1, 1], group = 1, kernel_shape = [7, 5, 5], pads = [3, 2, 2, 3, 2, 2], strides = [1, 1, 1]](%in, %learned_0, %learned_1) %12 = Relu(%11) %13 = Conv[dilations = [1, 1, 1], group = 1, kernel_shape = [5, 5, 5], pads = [2, 2, 2, 2, 2, 2], strides = [1, 1, 1]](%12, %learned_2, %learned_3) %14 = Relu(%13) %15 = Conv[dilations = [1, 1, 1], group = 1, kernel_shape = [3, 3, 3], pads = [1, 1, 1, 1, 1, 1], strides = [1, 1, 1]](%14, %learned_4, %learned_5) %16 = Relu(%15) %17 = Conv[dilations = [1, 1, 1], group = 1, kernel_shape = [3, 3, 3], pads = [1, 1, 1, 1, 1, 1], strides = [1, 1, 1]](%16, %learned_6, %learned_7) %18 = Relu(%17) %out = Conv[dilations = [1, 1, 1], group = 1, kernel_shape = [1, 1, 1], pads = [0, 0, 0, 0, 0, 0], strides = [1, 1, 1]](%18, %learned_8, %learned_9) return %out }