TRT Error

[elohimarthur]

Hi! dear professor!
I want to try to convert my model into TensorRT with torch2trt, I have fixed some Warning like unsupport op torch.zeros and torch.flip(I am not sure I'm right). However, I meet error below and how can I fix it?

[11/24/2022-14:44:24] [TRT] [E] 4: [graphShapeAnalyzer.cpp::analyzeShapes::1294] Error Code 4: Miscellaneous (IShuffleLayer encoder.2:1:SHUFFLE:GPU: reshape wildcard -1 has infinite number of solutions or no solution. Reshaping [4,10000] to [0,-1,0,1].)
[11/24/2022-14:44:24] [TRT] [E] 4: [graphShapeAnalyzer.cpp::analyzeShapes::1294] Error Code 4: Miscellaneous (IShuffleLayer encoder.2:1:SHUFFLE:GPU: reshape wildcard -1 has infinite number of solutions or no solution. Reshaping [4,10000] to [0,-1,0,1].)
[11/24/2022-14:44:24] [TRT] [E] 4: [graphShapeAnalyzer.cpp::analyzeShapes::1294] Error Code 4: Miscellaneous (IShuffleLayer encoder.2:1:SHUFFLE:GPU: reshape wildcard -1 has infinite number of solutions or no solution. Reshaping [4,10000] to [0,-1,0,1].)
[11/24/2022-14:44:24] [TRT] [E] 4: [graphShapeAnalyzer.cpp::analyzeShapes::1294] Error Code 4: Miscellaneous (IShuffleLayer encoder.2:1:SHUFFLE:GPU: reshape wildcard -1 has infinite number of solutions or no solution. Reshaping [4,10000] to [0,-1,0,1].)
[11/24/2022-14:44:24] [TRT] [E] 4: [graphShapeAnalyzer.cpp::analyzeShapes::1294] Error Code 4: Miscellaneous (IShuffleLayer encoder.2:1:SHUFFLE:GPU: reshape wildcard -1 has infinite number of solutions or no solution. Reshaping [4,10000] to [0,-1,0,1].)
[11/24/2022-14:44:24] [TRT] [E] 4: [graphShapeAnalyzer.cpp::analyzeShapes::1294] Error Code 4: Miscellaneous (IShuffleLayer encoder.2:1:SHUFFLE:GPU: reshape wildcard -1 has infinite number of solutions or no solution. Reshaping [4,10000] to [0,-1,0,1].)
[11/24/2022-14:44:24] [TRT] [E] 4: [graphShapeAnalyzer.cpp::analyzeShapes::1294] Error Code 4: Miscellaneous (IShuffleLayer encoder.2:1:SHUFFLE:GPU: reshape wildcard -1 has infinite number of solutions or no solution. Reshaping [4,10000] to [0,-1,0,1].)
[11/24/2022-14:44:24] [TRT] [E] 4: [graphShapeAnalyzer.cpp::analyzeShapes::1294] Error Code 4: Miscellaneous (IShuffleLayer encoder.2:1:SHUFFLE:GPU: reshape wildcard -1 has infinite number of solutions or no solution. Reshaping [4,10000] to [0,-1,0,1].)
[11/24/2022-14:44:24] [TRT] [E] 4: [graphShapeAnalyzer.cpp::analyzeShapes::1294] Error Code 4: Miscellaneous (IShuffleLayer encoder.2:1:SHUFFLE:GPU: reshape wildcard -1 has infinite number of solutions or no solution. Reshaping [4,10000] to [0,-1,0,1].)
[11/24/2022-14:44:24] [TRT] [E] 3: encoder.2:1:SHUFFLE:GPU: has zero placeholder that refers to invalid axis 2.
[11/24/2022-14:44:24] [TRT] [E] 4: [network.cpp::validate::2917] Error Code 4: Internal Error (Layer encoder.2:1:SHUFFLE:GPU failed validation)

[elohimarthur]

My op of Torch.flip like below:
Do I make it right?

from torch2trt.torch2trt import *
from torch2trt.module_test import add_module_test


@tensorrt_converter('torch.flip')
@tensorrt_converter('torch.Tensor.flip')

def convert_flip(ctx):
    input = ctx.method_args[0]
    input_trt = add_missing_trt_tensors(ctx.network, [input])[0]
    output = ctx.method_return

    permutation = ctx.method_args[1]

    layer = ctx.network.add_shuffle(input_trt)
    layer.second_transpose = tuple(permutation)

    output._trt = layer.get_output(0)

class TorchFlip(torch.nn.Module):

    def __init__(self, *args):
        super(TorchFlip, self).__init__()
        self.args = args
        
    def forward(self, x):
        return torch.flip(x, *self.args)

class TorchFlip(torch.nn.Module):
    
    def __init__(self, *args):
        super(TorchFlip, self).__init__()
        self.args = args
        
    def forward(self, x):
        return x.flip(*self.args)

@add_module_test(torch.float32, torch.device('cuda'), [(1, 3, 3)])
@add_module_test(torch.float32, torch.device('cuda'), [(1, 3, 3, 3)])
@add_module_test(torch.float32, torch.device('cuda'), [(2, 3, 3, 3)], max_batch_size=2)

def test_torch_flip_0_0():
    return TorchFlip(1)

test_torch_flip_0_0()

[jaybdub]

Hi @elohimarthur ,

Thanks for reaching out!

I believe to use torch.flip, the slice method needs to be used instead shuffle.

I've added a PR to enable torch.flip and torch.Tensor.flip here

#822

Please let me know if this works for your use case or you have any issues.

Best,
John