Conversation
| for checkpoint in TOKENIZER_CHECKPOINTS | ||
| ] | ||
| self.tokenizers = [BertTokenizer.from_pretrained(checkpoint) for checkpoint in TOKENIZER_CHECKPOINTS] | ||
| self.tf_tokenizers = [TFBertTokenizer.from_pretrained(checkpoint) for checkpoint in TOKENIZER_CHECKPOINTS] |
There was a problem hiding this comment.
Note to probably-Amy here: I'm not testing the old slow Bert tokenizer here because the fast one seems generally better maintained and is fully supported in TFLite now.
| model.save(save_path) | ||
| loaded_model = tf.keras.models.load_model(save_path) | ||
| loaded_output = loaded_model(test_inputs) | ||
| model.export(save_path) |
There was a problem hiding this comment.
This is the recommended TF/Keras way to export models for inference now, and the old approach was always a bit shaky.
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
This should be ready to go now, and finally fixes the remaining CI issues after the |
amyeroberts
left a comment
amyeroberts
left a comment
There was a problem hiding this comment.
Thanks for the continued TF tidy-up!
| """ | ||
| if name_scope is not None: | ||
| if not tf_name.startswith(name_scope): | ||
| if not tf_name.startswith(name_scope) and "final_logits_bias" not in tf_name: |
There was a problem hiding this comment.
Why the check on "final_logits_bias" here?
There was a problem hiding this comment.
No good reason, unfortunately! BART and derived models put the final_logits_bias weights outside the main model name scope, and I'm not entirely sure why (but changing it now might invalidate old user checkpoints)
3 participants
This PR hopefully fixes the last remaining issues from the
build()PR and gets the CI back to normal!