Hook for pytorch

[AceCoooool]

pyinstaller test.py is completed successfully. However, when run this exe file, there is the error like that:

File "/home/ace/.pyenv/versions/anaconda3-4.2.0/lib/python3.5/site-packages/PyInstaller/loader/pyimod03_importers.py", line 389, in load_module
    exec(bytecode, module.__dict__)
  File "site-packages/torch/__init__.py", line 302, in <module>
  File "site-packages/torch/__init__.py", line 297, in manager_path
RuntimeError: Unable to find torch_shm_manager at /home/ace/Developer/temp/dist/test/torch/lib/torch_shm_manager
Failed to execute script test

This is maily caused by path，torch_shm_manager is in the file /home/ace/.pyenv/versions/anaconda3-4.2.0/lib/python3.5/site-packages/torch/lib/torch_shm_manager. (this is because pyinstaller change the file to the current file: /home/ace/Developer/temp/dist/test)
How can I just let my torch package not change the file ? (Even I can not use this exe file in other computers)

Could anyone help me? thank you.

[AceCoooool]

I using a very silly way: add the torch package to the /home/ace/Developer/dist/test . That's work.

[iamtodor]

It takes me a while to figure out the solution from @AceCoooool
Here is the code

cnn2@cnn2:~$ cd ~/.local/lib/python3.6/site-packages

I assume your torch is placed right there

cnn2@cnn2:~/.local/lib/python3.6/site-packages$ cp -a torch/. your/dist/location

It copies all content from torch to your desired location.

Basically, I think it's a bug that should be fixed @htgoebel

[htgoebel]

@iamtodor This way work, but it is not what I'd call a solution. A solution would be to provide a hook, so others will not step into these issue.

[iamtodor]

@htgoebel yes, this way is very dirty. So that's why I ask to call it the bug and fix it

[htgoebel]

@iamtodor You are very welcome to provide a pull-request: https://pyinstaller.readthedocs.io/en/stable/hooks.html

[Microactive]

Have you solve the problem?
Now I meet the similar problem. My test.py only several lines code:
import torch
a = torch.zeros(2, 3)
print(a)

pyinstaller test.py is completed successfully. However, when run this exe file, there is the error like that
Failed to execute script test. I did as @AceCoooool did. But it can't work. there is the error like that
Failed to execute script test.

[Vishwaatpurdue]

Hi,
I have a similar issue. I used the hook mentioned here to load the torch dependencies, but the exe is not working in other windows machines (I built the exe for windows).

[KnightOfTheMoonlight]

You may try reinstall pytorch. That works for me.

[MarcinMisiurewicz]

I've run into similar problem, but none of the workarounds work. What is strange, my .exe file works when run from anaconda environment (any env: with or without torch installed), but when I try to run it from cmd outside of anaconda it gives me error:

Traceback (most recent call last):
  File "verify.py", line 1, in <module>
  File "C:\ProgramData\Anaconda3\envs\py367\lib\site-packages\PyInstaller\loader\pyimod03_importers.py", line 627, in exec_module
    exec(bytecode, module.__dict__)
  File "site-packages\torch\__init__.py", line 79, in <module>
ImportError: DLL load failed: The specified module could not be found.
[11108] Failed to execute script verify

line 1 in verify.py is import torch
Do you have any idea what causes it? Any hints on where to look for it?

Win 10
Anaconda with python 3.6.7 (on 3.7 also fails)
PyInstaller 3.4
PyTorch 1.1.0

[raviv]

I'm having a similar issue:
Same platform as @MarcinMisiurewicz, but the exe can't find pytorch unless I'm in the conda env used to build it.
Oddly, pyinstaller did copy torch to the dist folder, but the exe still can't find it.

Update
It seems to depend on how pytorch was installed in the conda env from which you run pyinstaller.
If pytorch was installed using conda (e.g. conda install pytorch torchvision cudatoolkit=10.0 -c pytorch) the resulting exe won't work unless the cudatoolkit conda package is installed.
However, if pytorch was installed using pip, the resulting exe works just fine outside of conda env.

[MarcinMisiurewicz]

@raviv Wow, it works! Thank you - it's third day I'm trying to get it work.

I'm wondering what is the source of this problem, is it pytorch for windows, or pyinstaller?

Additional note: when I'm using plain Python (without conda), with pytorch installed from pip, it also crushes, exactly like above. So it seems that the workaround is to use conda, but install pytorch with pip (???). I don't know, it is some dark wizardry here...

[raviv]

@MarcinMisiurewicz glad it worked for you.
It seems that pyinstaller doesn't pick up the dependency of pytorch from conda, on cudatoolkit10.

[MarcinMisiurewicz]

when I've checked the dependency walker it think it was both: cudatoolkit and Python37.dll. I'll check it

[raviv]

@MarcinMisiurewicz After removing pytorch from conda and installing it using pip (in coda env) I can now build an exe that runs on a clean windows machine, with CUDA toolkit and driver installed of course.

[parth52]

@raviv glad it worked for you.
It seems that pyinstaller doesn't pick up the dependency of pytorch from conda, on cudatoolkit10.

i made a simple application to calculate speed between cpu and gpu and its works fine in my anaconda environment but whenever i am trying to make .exe of that particular file which raise an error that cudatoolkit is not found do conda install cudatoolkit

my code as below temp.py file

from numba import jit,cuda
import numpy as np

to measure exec time

from timeit import default_timer as timer

normal function to run on cpu

def func(a):
for i in range(10000000):
a[i]+= 1

function optimized to run on gpu

@jit(target ="cuda")
def func2(a):
for i in range(10000000):
a[i]+= 1
if name=="main":
n = 10000000
a = np.ones(n, dtype = np.float64)
b = np.ones(n, dtype = np.float32)

start = timer() 
func2(a) 
print("with GPU:", timer()-start) 
   
  
start = timer() 
func(a) 
print("without GPU:", timer()-start)  
ll=input()
print(ll)

and error which i got as below while i am running exe file:-

cudatoolkit is not found

whereas i installed cudatoolkit=9.0 in my conda environment.

and if i try to make .exe using either conda or cmd its give me an same error

[jurneo]

You may try reinstall pytorch. That works for me.

this is also a solution for me.

[bwoodsend]

This was fixed in #3657.

[bmahlbrand]

This was fixed in #3657.

definitely still a thing in 4.4

[bwoodsend]

Open a new issue if you want us to be able do anything about it. I imagine that pytorch has changed a bit since 3 and a half years ago.