Introduce Class Usage Logging (#1243)

Summary:
Pull Request resolved: #1243

This builds on https://pytorch.org/docs/stable/notes/large_scale_deployments.html#api-usage-logging to log when certain classes are initialized.

In this diff, I only covered bert/roberta models (and not even necessarily all of them) to prove that the system works

This will not affect OSS users, as for them, the logger is a no-op.

Reviewed By: snisarg

Differential Revision: D19744167

fbshipit-source-id: f07845d1c1bd1f493c0e084661ab138b48b56ad2

Author: m3rlin45

pytext/models/bert_classification_models.py

@@ -32,6 +32,7 @@
     TransformerSentenceEncoderBase,
 )
 from pytext.utils.label import get_label_weights
+from pytext.utils.usage import log_class_usage
 
 
 class NewBertModel(BaseModel):
@@ -120,6 +121,7 @@ def __init__(self, encoder, decoder, output_layer, stage=Stage.TRAIN) -> None:
         self.output_layer = output_layer
         self.stage = stage
         self.module_list = [encoder, decoder]
+        log_class_usage(__class__)
 
 
 class BertPairwiseModel(BasePairwiseModel):
@@ -162,6 +164,7 @@ def __init__(
         self.encoder1 = encoder1
         self.encoder2 = encoder2
         self.encoders = [encoder1, encoder2]
+        log_class_usage(__class__)
 
     @classmethod
     def _create_encoder(

pytext/models/bert_regression_model.py

@@ -9,6 +9,7 @@
 from pytext.models.bert_classification_models import NewBertModel
 from pytext.models.module import create_module
 from pytext.models.output_layers import RegressionOutputLayer
+from pytext.utils.usage import log_class_usage
 
 
 class NewBertRegressionModel(NewBertModel):
@@ -37,3 +38,7 @@ def from_config(cls, config: Config, tensorizers: Dict[str, Tensorizer]):
         )
         output_layer = RegressionOutputLayer.from_config(config.output_layer)
         return cls(encoder, decoder, output_layer)
+
+    def __init__(self, encoder, decoder, output_layer) -> None:
+        super().__init__(encoder, decoder, output_layer)
+        log_class_usage(__class__)

pytext/models/roberta.py

@@ -27,6 +27,7 @@
 )
 from pytext.torchscript.module import get_script_module_cls
 from pytext.utils.file_io import PathManager
+from pytext.utils.usage import log_class_usage
 from torch.serialization import default_restore_location
 
 
@@ -72,6 +73,7 @@ def __init__(self, config: Config, output_encoded_layers: bool, **kwarg) -> None
         assert config.pretrained_encoder.load_path, "Load path cannot be empty."
         self.encoder = create_module(config.pretrained_encoder)
         self.representation_dim = self.encoder.encoder.token_embedding.weight.size(-1)
+        log_class_usage(__class__)
 
     def _embedding(self):
         # used to tie weights in MaskedLM model
@@ -128,6 +130,7 @@ def __init__(self, config: Config, output_encoded_layers: bool, **kwarg) -> None
                 self.load_state_dict(roberta_state)
 
         self.representation_dim = self._embedding().weight.size(-1)
+        log_class_usage(__class__)
 
     def _embedding(self):
         # used to tie weights in MaskedLM model
@@ -200,6 +203,7 @@ def __init__(self, encoder, decoder, output_layer, stage=Stage.TRAIN) -> None:
         self.module_list = [encoder, decoder]
         self.output_layer = output_layer
         self.stage = stage
+        log_class_usage(__class__)
 
     def arrange_model_inputs(self, tensor_dict):
         tokens, pad_mask, segment_labels, positions, _ = tensor_dict["tokens"]

pytext/utils/usage.py

@@ -0,0 +1,11 @@
+#!/usr/bin/env python3
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import torch
+
+
+def log_class_usage(klass):
+    identifier = "PyText"
+    if klass and hasattr(klass, "__name__"):
+        identifier += f".{klass.__name__}"
+    torch._C._log_api_usage_once(identifier)