another 'Tensor' object has no attribute '_trt'

Hi Trying out EfficientNet_b1 for jetson

seems like I get the following warning and error

Warning: Encountered known unsupported method torch.nn.functional.has_torch_function_variadic
Warning: Encountered known unsupported method torch.nn.functional.linear

  File "model_convert.py", line 24, in <module>
    model_trt = torch2trt(model, [x])
  File "/usr/local/lib/python3.6/dist-packages/torch2trt-0.2.0-py3.6.egg/torch2trt/torch2trt.py", line 546, in torch2trt
  File "/usr/local/lib/python3.6/dist-packages/torch2trt-0.2.0-py3.6.egg/torch2trt/torch2trt.py", line 407, in mark_outputs
AttributeError: 'Tensor' object has no attribute '_trt'

I am aware SiLU is not implemented for onnx so I have changed them to ReLU but it seems like there is another error

[jaybdub]

Hi @ooodragon94,

I was able to convert the model by registering converters for torch.nn.functional.silu and torch.nn.functional.linear

Below is the program I used to convert the model.

import timm
import torch
from torch2trt import torch2trt, tensorrt_converter, get_arg, trt, add_missing_trt_tensors, torch_dtype_to_trt

# REGISTER NEW CONVERTERS
@tensorrt_converter('torch.nn.functional.silu')
def convert_silu(ctx):
    input = get_arg(ctx, 'input', pos=0, default=None)
    output = ctx.method_return
    input_trt = add_missing_trt_tensors(ctx.network, [input])[0]
    
    layer = ctx.network.add_activation(input_trt, trt.ActivationType.SIGMOID)
    layer = ctx.network.add_elementwise(input_trt, layer.get_output(0), trt.ElementWiseOperation.PROD)
    
    output._trt = layer.get_output(0)
    
@tensorrt_converter('torch.nn.functional.linear')
def convert_Linear(ctx):
    input = ctx.method_args[0]
    weight = get_arg(ctx, 'weight', 1, None)
    bias = get_arg(ctx, 'bias', 2, None)
    input_trt = add_missing_trt_tensors(ctx.network, [input])[0]
    output = ctx.method_return

    # reshape to ...xNx1x1
    layer = ctx.network.add_shuffle(input_trt)
    layer.reshape_dims = tuple(input_trt.shape) + (1, 1) 

    bias_trt = trt.Weights(torch_dtype_to_trt(weight.dtype))
    if bias is not None:
        bias_trt = bias.detach().cpu().numpy()
        
    # add fully connected
    layer = ctx.network.add_fully_connected(
        input=layer.get_output(0),
        num_outputs=int(weight.shape[0]),
        kernel=weight.detach().cpu().numpy(),
        bias=bias_trt)

    # reshape back to N
    layer = ctx.network.add_shuffle(layer.get_output(0))
    layer.reshape_dims = tuple(output.shape[1:])

    output._trt = layer.get_output(0)


# CONVERT MODEL
model = timm.models.efficientnet_b1(pretrained=True).cuda().eval()
data = torch.zeros(1, 3, 224, 224).cuda()
model_trt = torch2trt(model, [data])

# TEST
data = torch.randn(1, 3, 224, 224).cuda()
print(torch.max(torch.abs(model_trt(data) - model(data)))

Please let me know if this works for you and I’ll work on a pull-request to integrate these converters into the project.

Best, John

[ooodragon94]

Your explanation rocks. Thank you so much for your kind and thorough response!