Gluon-ts: Why does the PyTorchPredictor still need to import MXNet

[chendhui]

https://github.com/awslabs/gluon-ts/tree/master/src/gluonts/torch
This sub-package reimplements the Predictor method using Pytorch. I think the goal is to become irrelevant to the MXNet during the inference stage, and we can also use the Predictor API in a Pytorch-only environment.

However, when I trained a model and saved it. I use the PyTorchPredictor API as Predictor. But the runtime environment also needs to import the MXNet. I am wondering what is the significance of PyTorchPredictor.

[lostella]

But the runtime environment also needs to import the MXNet.

@chendhui what do you mean by that? There's an example notebook using this class, which runs fine without having MXNet installed, if I remember correctly

[PascalIversen]

Hey @chendhui, thanks for opening the issue.
Just checked, there is some dependency left:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-1-602532f91f38> in <module>
      1 from gluonts.dataset.field_names import FieldName
----> 2 from gluonts.dataset.loader import TrainDataLoader
      3 print(sys.modules.keys())
      4 print()
      5 print()

~/Projects/gts/gluon-ts/src/gluonts/dataset/loader.py in <module>
     26 from gluonts.dataset.util import MPWorkerInfo
     27 from gluonts.itertools import batcher, cyclic, pseudo_shuffled
---> 28 from gluonts.transform import Transformation, TransformedDataset
     29 
     30 logger = logging.getLogger(__name__)

~/Projects/gts/gluon-ts/src/gluonts/transform/__init__.py in <module>
     64 ]
     65 
---> 66 from ._base import (
     67     AdhocTransform,
     68     Chain,

~/Projects/gts/gluon-ts/src/gluonts/transform/_base.py in <module>
     17 
     18 # First-party imports
---> 19 from gluonts.core.component import validated
     20 from gluonts.dataset.common import DataEntry
     21 from gluonts.runtime_params import GLUONTS_MAX_IDLE_TRANSFORMS

~/Projects/gts/gluon-ts/src/gluonts/core/component.py in <module>
    178 
    179 
--> 180 @equals.register(mx.gluon.HybridBlock)
    181 def equals_representable_block(
    182     this: mx.gluon.HybridBlock, that: mx.gluon.HybridBlock

AttributeError: module 'mxnet' has no attribute 'gluon'

[lostella]

Right! I think I have a local branch where I had started reorganizing code and imports to get rid of these things. I'll try to rebase it, hopefully not much changed since then

[lostella]

@chendhui the desired goal, as you say, is to have the package independent of MXNet unless MXNet-based models are used -- however this is still in progress, so thanks for pointing that out 👍

[lostella]

@chendhui this should be fixed in the master branch now, please let me know if you keep seeing the issue

[chendhui]

Thank you very much for your reply. Looking forward to the fixed codes.