torch2trt

torch2trt copied to clipboard

### Description: I'm experiencing a discrepancy between the inference results of my PyTorch model and the TensorRT model obtained by converting it using the torch2trt tool. ### Reproduce This can...

Thrsu

### Description: I'm experiencing a discrepancy between the inference results of my PyTorch model and the TensorRT model obtained by converting it using the torch2trt tool. ### Reproduce This can...

Thrsu

Encountered known unsupported method torch.nn.functional.pixel_shuffle

6

i convert alphapose fast_res50_256x192.pth model, but i encounter the unsupported layer ?

zlheos

Segmentation fault (core dumped)

4

My Code : # create example data, x = torch.ones((1, 3, 112, 112)).cuda() model_trt_new = TRTModule() model_trt_new.load_state_dict(torch.load('face_trt.pth')) y_trt_new = model_trt_new(x) This code can run normally, but when it is finished,...

fly520fly520

accuracy drops a lot in fp16 mode

4

my model's accuracy drops a lot when I convert it into fp16 mode, even a pretrained resnet34 experienced accuracy drop in fp16 mode. ``` import os os.environ['CUDA_MODULE_LOADING'] = 'LAZY' from...

itachi1232gg

Is there a plan to support python 3.10?

1

Model export is failing on: ```bash File "/venv/lib/python3.10/site-packages/torch2trt-0.4.0-py3.10.egg/torch2trt/torch2trt.py", line 310, in wrapper converter["converter"](ctx) File "/venv/lib/python3.10/site-packages/torch2trt-0.4.0-py3.10.egg/torch2trt/converters/interpolate.py", line 81, in convert_interpolate_trt7 if not isinstance(scales, collections.Sequence): AttributeError: module 'collections' has no attribute 'Sequence'...

jakubhejhal

Python deprecated collections.Sequence in favor of collections.abc.Sequence. The update checks for the Python version and uses collections.abc instead of collections for Python versions over 3.9

teodorToshkov

Can't convert Vision Transformer from timm - TypeError: add_constant(): incompatible function arguments.

4

The following code: ```python import torch from torch2trt import torch2trt import timm vit_model = timm.create_model( model_name="vit_base_patch16_224_dino", pretrained=True, ) vit_model.to("cuda") vit_model.eval() input_shape = (4, 3, 224, 224) input_data = torch.rand(input_shape, dtype=torch.float32).cuda()...

jakubhejhal

Problem with AdaptiveMaxPool2d, likely with stride/kernel_size

2

I want to convert a model to trt but I have a problem with `nn.AdaptiveMaxPool2d`. It seems like it only works for specific `kernel_size`. In my case its `(384, 384)`...

gabe-scorebreak

Hi guys, I had a jetson AGX Xavier on which I developed an action recognition app using your torch2trt and trt_pose, and it worked well. This was in January 2021....

lweingart