def test_custom_load_tf_weights(self):
model, output_loading_info = TFBertForTokenClassification.from_pretrained(
"jplu/tiny-tf-bert-random", output_loading_info=True
)
self.assertEqual(sorted(output_loading_info["unexpected_keys"]), [])
for layer in output_loading_info["missing_keys"]:
self.assertTrue(layer.split("_")[0] in ["dropout", "classifier"])
def create_and_check_bert_for_token_classification(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.num_labels = self.num_labels
model = TFBertForTokenClassification(config=config)
inputs = {
"input_ids": input_ids,
"attention_mask": input_mask,
"token_type_ids": token_type_ids,
}
result = model(inputs)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))