SageMaker Studio PyTorch 1.8 kernel has no PyTorch, Numpy, or Matplotlib module

Question

I'm working with SageMaker studio with the following options:
 - kernel: PyTorch 1.8 Python 3.6 GPU optimized.
 - instance: ml.g4dn.xlarge

When running `import torch` `numpy`, `matplotlib` or `PIL`, I'm getting the `No module named 'X'` error. No matter when using `pip install` in a cell above, it will not be imported. Is this a problem only I am encountering with the new PyTorch 1.8 kernel? It also happens with the CPU-optimized version. However, PyTorch 1.6 kernel does not throw an error.

When running `conda list`, I get the output without any of the previously mentioned modules.

Answer

I have followed your explanation to recreate the `No module named 'X'`. After instance lunch, I could import `numpy`, `matplotlib` and `PIL` but could not `import torch` because `torch` is not initially installed.

I followed these steps to install and import torch:
   * Shut down your instance completely.
* Attach the kernel to your notebook again
* Lunch your `ml.g4dn.xlarge` instance again
* `pip3 install torch`
* `import torch`

Answer

No module named 'X' should not be the expected behavior. Thanks for reporting this issue and sorry for the inconvenience it caused.

Is your use case flexible to use other PyTorch versions prior to 1.8? If yes, please try other versions. 
If PT1.8 is the only choice, please try following workaround for unblocking(while service team is working on the fix).

Two options:  
 1 - executed following in a notebook cell  
```
# switch the python execution to /usr/local/bin/python in the kernel.json file. 
!sed 's|^ *"python",|  "/usr/local/bin/python",|g' /usr/local/share/jupyter/kernels/python3/kernel.json>/tmp/kernel.json; cp -f /tmp/kernel.json /usr/local/share/jupyter/kernels/python3/kernel.json;
```

2 - directly execute following shell command in kernel image specific terminal(not the global terminal).  
```
sed 's|^ *"python",|  "/usr/local/bin/python",|g' /usr/local/share/jupyter/kernels/python3/kernel.json>/tmp/kernel.json; cp -f /tmp/kernel.json /usr/local/share/jupyter/kernels/python3/kernel.json;`
```

The above command is only needed once per kernel gateway app. 
After above, please restart the kernel. You can verify using following command in a notebook cell. The  '/usr/local/bin/python' should be shown as python executable. 
```
import sys
print(sys.executable)
```

SageMaker Studio PyTorch 1.8 kernel has no PyTorch, Numpy, or Matplotlib module

Conteúdo relevante