[
https://issues.apache.org/jira/browse/BEAM-14337?focusedWorklogId=772900&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-772900
]
ASF GitHub Bot logged work on BEAM-14337:
-----------------------------------------
Author: ASF GitHub Bot
Created on: 20/May/22 16:20
Start Date: 20/May/22 16:20
Worklog Time Spent: 10m
Work Description: TheNeuralBit commented on code in PR #17470:
URL: https://github.com/apache/beam/pull/17470#discussion_r878324657
##########
sdks/python/apache_beam/ml/inference/base.py:
##########
@@ -96,7 +107,9 @@ def expand(self, pcoll: beam.PCollection) ->
beam.PCollection:
pcoll
# TODO(BEAM-14044): Hook into the batching DoFn APIs.
| beam.BatchElements()
- | beam.ParDo(_RunInferenceDoFn(self._model_loader, self._clock)))
+ | beam.ParDo(
+ _RunInferenceDoFn(
+ self._model_loader, self._prediction_params, self._clock)))
Review Comment:
> But after a long discussion we settled on making these a side input.
This is not a side input, this is a constant parameter. A side input comes
from a PCollection wrapped in a PCollectionView and is passed to the ParDo:
https://beam.apache.org/documentation/programming-guide/#side-inputs, and Beam
handles passing actual data to the DoFn's process method.
That's what I'm trying to suggest, making it a side input would justify a
change to base RunInference, because it means you'd need to change the `ParDo`
creation.
Issue Time Tracking
-------------------
Worklog Id: (was: 772900)
Time Spent: 5h 50m (was: 5h 40m)
> Support **kwargs for PyTorch models.
> ------------------------------------
>
> Key: BEAM-14337
> URL: https://issues.apache.org/jira/browse/BEAM-14337
> Project: Beam
> Issue Type: Sub-task
> Components: sdk-py-core
> Reporter: Anand Inguva
> Assignee: Andy Ye
> Priority: P2
> Time Spent: 5h 50m
> Remaining Estimate: 0h
>
> Some models in Pytorch instantiating from torch.nn.Module, has extra
> parameters in the forward function call. These extra parameters can be passed
> as Dict or as positional arguments.
> Example of PyTorch models supported by Hugging Face ->
> [https://huggingface.co/bert-base-uncased]
> [Some torch models on Hugging
> face|https://github.com/huggingface/transformers/blob/main/src/transformers/models/bert/modeling_bert.py]
> Eg:
> [https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertModel]
> {code:java}
> inputs = {
> input_ids: Tensor1,
> attention_mask: Tensor2,
> token_type_ids: Tensor3,
> }
> model = BertModel.from_pretrained("bert-base-uncased") # which is a
> # subclass of torch.nn.Module
> outputs = model(**inputs) # model forward method should be expecting the keys
> in the inputs as the positional arguments.{code}
>
> [Transformers|https://pytorch.org/hub/huggingface_pytorch-transformers/]
> integrated in Pytorch is supported by Hugging Face as well.
>
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
