crawlingcub opened a new issue, #12632:
URL: https://github.com/apache/tvm/issues/12632
The SoftPlus layer produces different outputs for given input. Maybe some
numerical instability?
Let me know if you need more info.
Code:
```python
import torch
import tvm
from tvm import relay
from tvm.contrib.download import download_testdata
from tvm.contrib import graph_executor
import os
import pickle as pkl
import numpy as np
import sys
DEVICE='cuda'
model=torch.nn.Softplus()
pt_inp=torch.Tensor(pkl.load(open(os.path.join(sys.argv[1], "data.pkl"),
"rb")))
pt_inp=pt_inp.to(DEVICE)
model.to(DEVICE)
with torch.no_grad():
pt_out = model(pt_inp).cpu().numpy()
model.to('cpu')
pt_inp=pt_inp.to('cpu')
input_name = "input0"
target = tvm.target.cuda()
dev = tvm.gpu(0)
scripted_model = torch.jit.trace(model, pt_inp).eval()
input_data = pt_inp.numpy()
shape_list = [(input_name, input_data.shape)]
mod, params = relay.frontend.from_pytorch(scripted_model, shape_list)
with tvm.transform.PassContext(opt_level=3):
lib = relay.build(mod, target=target, params=params)
m = graph_executor.GraphModule(lib["default"](dev))
m.set_input(input_name, tvm.nd.array(pt_inp.numpy()))
m.run()
output = torch.tensor(m.get_output(0).asnumpy())
tvm_out = m.get_output(0).asnumpy()
print(np.max(np.abs(tvm_out - pt_out)), np.mean(np.abs(tvm_out - pt_out)))
print(pt_out[0][20:30])
print(tvm_out[0][20:30])
#print(pt_out)
```
Find data
[here](https://drive.google.com/drive/folders/19PZYIsBbXBcWDcMDR81sGXs4Oww7IqtG?usp=sharing)
### Expected behavior
Outputs should be same
### Actual behavior
Output:
```
inf inf
[ 54.105717 15.969946 36.069256 13.273791 19.722383 31.054518
44.89835 15.706258 114.94983 67.98223 ]
[54.105713 15.969945 36.069256 13.27379 19.722383 31.054516 44.89835
15.706258 inf 67.98223 ]
```
### Environment
Any environment details, such as: Operating System, TVM version, etc
```
torch==1.12.1
torchvision==0.13.1
python 3.8.13
Ubuntu 18.04
TVM 1afd05939
```
--
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: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
