Cannot import module while using custom handler

🐛 Bug report

I’m trying to use a custom crowd counting model and while serving it gives me an error while importing custom model libraries.

Python Import code is like below:

module_path = os.path.abspath(os.path.join('../../../../'))  # Parent Dir on local machine
if module_path not in sys.path:
    sys.path.append(module_path+"/SSDCNet")

...
from Network.SSDCNet import SSDCNet_classify

Error:

   File "/tmp/models/58eb8da6a1c8563374304a9fc66b2788a9c9c440/crdcnt_handler.py", line 9, in <module>
   from crowdmodel import CrowdCounter
   File "/tmp/models/58eb8da6a1c8563374304a9fc66b2788a9c9c440/crowdmodel.py", line 8, in <module>
    from Network.SSDCNet import SSDCNet_classify
ModuleNotFoundError: No module named 'Network'

Here is the log file:
server_log3.txt

Environment Details:

Operating System and version: Ubuntu 16.04

**Possible Fix: ** Instead of adding to sys.path. Adding the parent folder to PYTHONPATH before you run the torchserve command temporarily fixes the issue.

> export PYTHONPATH=/home/ubuntu/mayub/Github/SSDCNet:$PYTHONPATH 
> echo $PYTHONPATH
> torchserve --ts-config ./config.properties --start --ncs --model-store ./torchserve_model_store --models all

Since this is temporary and session based maybe a better solution is required.

Thanks !

[ghost]

@mohammedayub44 All the dependency files need to be packaged inside the mar file using the torch-model-archiver utility. TorchServe extracts the mar file in a temporary location and that is already added in the path for importing the modules.

From your code snippet it looks like you are trying to access modules which are outside the mar file.

Could you share your mar file.

how to add a folder as extra files ?

[harshbafna]

@mohammedayub44 : torch-model-archiver is just a utility tool for packaging your model and does not validates you model. It is expected that the user will supply all the other dependency files using the --extra-files flag. You can always create an egg/zip file to upload your custom python package if the number of files are more.