[GitHub] [tvm] Kevin-XiongC opened a new issue, #11383: [Bug] Numerical inconsistency for roi_align in Relay

[GitBox <gi...@apache.org>]

Kevin-XiongC opened a new issue, #11383:
URL: https://github.com/apache/tvm/issues/11383

   I am converting roi_align from Paddle as frontend. However, I encountered the numerical issue the same as PyTorch's like below
   
   ![image](https://user-images.githubusercontent.com/22584786/169289356-da5ce45b-25ce-4336-87a3-7acde74a644a.png)
   I wonder whether I missed something or this should be a bug in fact.
   
   ### Expected behavior
   
   The results of each entry should be numerically close between the compiled model and Pytorch version.
   
   ### Actual behavior
   
   Some results's differences are greater than 0.1
   
   ### Environment
   
   Ubuntu20.04, tvm 0.9dev, python3.8,llvm 13
   
   ### Steps to reproduce
   
   ``` python
   import warnings
   warnings.filterwarnings('ignore')
   import os
   import unittest
   import numpy as np
   import tvm
   from tvm import relay
   from tvm.contrib import graph_executor
   from collections import OrderedDict
   import sys
   import torch
   import torchvision
   N = 1
   C = 3
   H = 3
   W = 3
   num = 1
   aligned = True
   target = tvm.target.Target("llvm")
   
   np.random.seed(1)
   dev = tvm.cpu(0)
   data = np.random.rand(1, C, H, W)
   a, b = np.random.rand(num, 1), np.random.rand(num, 1)
   x_min = np.minimum(a, b)*H
   x_max = np.maximum(a, b)*H
   a, b = np.random.rand(num, 1), np.random.rand(num, 1)
   y_min = np.minimum(a, b)*W
   y_max = np.maximum(a, b)*W
   boxes = np.concatenate([x_min, y_min, x_max, y_max], axis=1)
   x = np.pad(boxes, pad_width=((0, 0), (1, 0)), mode='constant')
   m = torchvision.ops.RoIAlign(
       output_size=3, spatial_scale=1, sampling_ratio=-1, aligned=aligned)
   res_torch = m.forward(
       torch.Tensor(data), torch.Tensor(x))
   
   print("torch:", res_torch.numpy().ravel())
   scripted_model = torch.jit.trace(m, example_inputs=[torch.Tensor(data), torch.Tensor(
       x)]).eval()
   torch_mod, torch_params = relay.frontend.from_pytorch(
       scripted_model, [('input', data.shape), ('rois', x.shape)])
   with tvm.transform.PassContext(opt_level=3):
       lib = relay.build(torch_mod, target=target, params=torch_params)
   gen_module = graph_executor.GraphModule(lib['default'](dev))
   map_inputs = {"input": data, "rois": x}
   gen_module.set_input(**map_inputs)
   gen_module.run()
   res_torch_tvm = gen_module.get_output(0).numpy()
   print("torch_tvm:", res_torch_tvm.ravel())
   
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org