Different behavior when trace model. #19349
Description
🐛 Bug
When I use torch.jit.trace() to trace a model, it gives me different results with same code.
To Reproduce
I write a simple test example
import torch.nn as nn
import torch
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
print()
def set_test_name(self, name):
self.test_name = name
def forward(self, input):
print(self.test_name, ':', input.shape[2], '\n')
output = input
return output
if __name__ == '__main__':
model = Net()
model.eval()
inputs = torch.ones(1, 1, 32, 160)
model.set_test_name('predict')
output = model(inputs)
model.set_test_name('trace')
traced = torch.jit.trace(model, inputs)
traced.save('stm.pt')Expected behavior
it should print this
predict : 32
trace : 32
trace : 32
trace : 32
yet it gives me :
C:\ProgramData\Anaconda3\lib\site-packages\torch\tensor.py:427: RuntimeWarning: Iterating over a tensor might cause the trace to be incorrect. Passing a tensor of different shape won't change the number of iterations executed (and might lead to errors or silently give incorrect results).
predict : 32
trace : tensor(32)
'incorrect results).', category=RuntimeWarning)
trace : tensor(32)
trace : 32
As you can see, the type of input.shape[2] is different.
Environment
- PyTorch Version (e.g., 1.0): 1.0.1
- OS (e.g., Linux): Win10
- How you installed PyTorch (
conda,pip, source): conda - Python version: 3.7.1
cc @suo