[
https://issues.apache.org/jira/browse/BEAM-14337?focusedWorklogId=772840&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-772840
]
ASF GitHub Bot logged work on BEAM-14337:
-----------------------------------------
Author: ASF GitHub Bot
Created on: 20/May/22 13:41
Start Date: 20/May/22 13:41
Worklog Time Spent: 10m
Work Description: ryanthompson591 commented on code in PR #17470:
URL: https://github.com/apache/beam/pull/17470#discussion_r878143536
##########
sdks/python/apache_beam/ml/inference/base_test.py:
##########
@@ -39,7 +41,12 @@ class FakeInferenceRunner(base.InferenceRunner):
def __init__(self, clock=None):
self._mock_clock = clock
- def run_inference(self, batch: Any, model: Any) -> Iterable[Any]:
+ def run_inference(
+ self,
+ batch: Any,
+ model: Any,
+ prediction_params: Optional[Dict[str, Any]] = None,
Review Comment:
Heres some code to demonstrate what I mean.
def pytorch_impl(prediction_params = None):
print ('got prediction_params ' + str(prediction_params))
def tf_impl():
# Does not have the prediction_params arg, should not crash when called
# but should crash if called with prediction_params argument
print('tf_impl')
mode = 'tf'
def base_do_fn(**kwargs):
if mode == 'tf':
tf_impl(**kwargs)
elif mode == 'pytorch':
pytorch_impl(**kwargs)
def base_run_inference(**kwargs):
base_do_fn(**kwargs)
base_run_inference()
print('expect pass')
mode='pytorch'
base_run_inference(prediction_params={'prediction_param': 'value'})
print('expect failure')
mode='tf'
base_run_inference(prediction_params={'prediction_param': 'value'})
----------------
What I'm saying is this method allows forwarding of any named param
whatsoever. With what you're doing we'll always be in a spot of modifying an
interface and thus having to modify every implementation of the interface going
forward for any single parameter that any single implementation wants.
I think here, and going forward, unless named parameters will apply to all
models we should not put it into the interface for all models.
Issue Time Tracking
-------------------
Worklog Id: (was: 772840)
Time Spent: 4h 40m (was: 4.5h)
> 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: 4h 40m
> 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)
