fix linear_to_conv2d_map to work with other distilbert model types

ane_transformers/huggingface/distilbert.py

@@ -2,6 +2,7 @@
 # For licensing see accompanying LICENSE.md file.
 # Copyright (C) 2022 Apple Inc. All Rights Reserved.
 #
+import re
 
 from ane_transformers.reference.layer_norm import LayerNormANE
 
@@ -520,14 +521,26 @@ def forward(
         return ((loss, ) + output) if loss is not None else output
 
 
+_LINEAR_TO_CONV2D_LAYERS = [
+    "q_lin.weight",
+    "k_lin.weight",
+    "v_lin.weight",
+    "out_lin.weight",
+    "lin1.weight",
+    "lin2.weight",
+    "classifier.weight",
+    "pre_classifier.weight",
+    "vocab_transform.weight",
+    "vocab_projector.weight",
+    "qa_outputs.weight",
+]
+
+
 def linear_to_conv2d_map(state_dict, prefix, local_metadata, strict,
                          missing_keys, unexpected_keys, error_msgs):
     """ Unsqueeze twice to map nn.Linear weights to nn.Conv2d weights
     """
     for k in state_dict:
-        is_internal_proj = all(substr in k for substr in ['lin', '.weight'])
-        is_output_proj = all(substr in k
-                             for substr in ['classifier', '.weight'])
-        if is_internal_proj or is_output_proj:
+        if any(k.endswith(layer) for layer in _LINEAR_TO_CONV2D_LAYERS):
             if len(state_dict[k].shape) == 2:
                 state_dict[k] = state_dict[k][:, :, None, None]

ane_transformers/huggingface/test_distilbert.py

@@ -9,7 +9,6 @@
 import logging
 import numpy as np
 import unittest
-import time
 
 import torch
 
@@ -32,6 +31,10 @@
     ("This is not what I expected!", "NEGATIVE"),
 ])
 
+MASKED_LM_MODEL = 'distilbert-base-uncased'
+QUESTION_ANSWERING_MODEL = 'distilbert-base-uncased-distilled-squad'
+TOKEN_CLASSIFICATION_MODEL = 'elastic/distilbert-base-uncased-finetuned-conll03-english'
+MULTIPLE_CHOICE_MODEL = 'Gladiator/distilbert-base-uncased_swag_mqa'
 
 class TestDistilBertForSequenceClassification(unittest.TestCase):
     """
@@ -191,5 +194,56 @@ def test_coreml_conversion_and_speedup(self):
             )
 
 
+class TestDistilBertLoadState(unittest.TestCase):
+    """
+    Test load_state_dict compatibility.
+    """
+
+    test_params = (
+        (
+            MASKED_LM_MODEL,
+            transformers.AutoModelForMaskedLM,
+            ane_transformers.DistilBertForMaskedLM,
+        ),
+        (
+            QUESTION_ANSWERING_MODEL,
+            transformers.AutoModelForQuestionAnswering,
+            ane_transformers.DistilBertForQuestionAnswering,
+        ),
+        (
+            TOKEN_CLASSIFICATION_MODEL,
+            transformers.AutoModelForTokenClassification,
+            ane_transformers.DistilBertForTokenClassification,
+        ),
+        (
+            MULTIPLE_CHOICE_MODEL,
+            transformers.AutoModelForMultipleChoice,
+            ane_transformers.DistilBertForMultipleChoice,
+        ),
+    )
+
+    def test_load_state(self):
+        for model_name, auto_model_cls, ane_model_cls in self.test_params:
+            with self.subTest(ane_model_cls=ane_model_cls):
+                try:
+                    # Instantiate the reference model from an exemplar pre-trained
+                    # model hosted on huggingface.co/models
+                    reference_model = auto_model_cls.from_pretrained(
+                        model_name,
+                        return_dict=False,
+                        torchscript=True,
+                    ).eval()
+                except Exception as e:
+                    raise RuntimeError(
+                        "Failed to download reference model from huggingface.co/models!"
+                    ) from e
+                logger.info("Downloaded reference model from huggingface.co/models")
+
+                # Initialize an ANE equivalent model and restore the checkpoint
+                test_model = ane_model_cls(reference_model.config).eval()
+                test_model.load_state_dict(reference_model.state_dict())
+                logger.info("Initialized and restored test model")
+
+
 if __name__ == "__main__":
     unittest.main()