[NEW Features] feature_extraction and processor support from_pretrained

[JunnYu]

PR types

New features

PR changes

APIs

Description

• CLIP and ErnieViL add get_input_embeddings and set_input_embeddings.
• CLIP and ErnieViL Processor support from_pretrained and save_pretrained.
• CLIP and ErnieViL FeatureExtraction support from_pretrained and save_pretrained.

here are some test cases.

from paddlenlp.transformers import CLIPTextModel
text_encoder = CLIPTextModel.from_pretrained("openai/clip-vit-base-patch32")
print(text_encoder.get_input_embeddings().weight.shape)
text_encoder.resize_token_embeddings(32)
print(text_encoder.get_input_embeddings().weight.shape)
# [49408, 512]
# [32, 512]

from paddlenlp.transformers import ErnieViLFeatureExtractor, CLIPFeatureExtractor
feature_extractor = CLIPFeatureExtractor.from_pretrained(
    "openai/clip-vit-base-patch32"
)  # Download feature_extraction_config from bos and cache.
feature_extractor.save_pretrained("./test/saved_model/")
feature_extractor = CLIPFeatureExtractor.from_pretrained(
    "./test/saved_model/"
)  # E.g. feature_extractor (or model) was saved using *save_pretrained('./test/saved_model/')*
feature_extractor = CLIPFeatureExtractor.from_pretrained("./test/saved_model/preprocessor_config.json")
feature_extractor, unused_kwargs = CLIPFeatureExtractor.from_pretrained(
    "openai/clip-vit-base-patch32", foo=False, return_unused_kwargs=True
)
assert unused_kwargs == {"foo": False}


feature_extractor = ErnieViLFeatureExtractor.from_pretrained(
    "ernie_vil-2.0-base-zh"
)  # Download feature_extraction_config from bos and cache.
feature_extractor.save_pretrained("./test/saved_model/")
feature_extractor = ErnieViLFeatureExtractor.from_pretrained(
    "./test/saved_model/"
)  # E.g. feature_extractor (or model) was saved using *save_pretrained('./test/saved_model/')*
feature_extractor = ErnieViLFeatureExtractor.from_pretrained("./test/saved_model/preprocessor_config.json")
feature_extractor, unused_kwargs = ErnieViLFeatureExtractor.from_pretrained(
    "ernie_vil-2.0-base-zh", foo=False, return_unused_kwargs=True
)
assert unused_kwargs == {"foo": False}

from paddlenlp.transformers import ErnieViLProcessor, CLIPProcessor
processor = CLIPProcessor.from_pretrained(
    "openai/clip-vit-base-patch32"
)  # Download processor from bos and cache.
processor.save_pretrained("./test/saved_model/")
processor = CLIPProcessor.from_pretrained(
    "./test/saved_model/"
)  

processor = ErnieViLProcessor.from_pretrained(
    "ernie_vil-2.0-base-zh"
)  # Download processor from bos and cache.
processor.save_pretrained("./test/saved_model/")
processor = ErnieViLProcessor.from_pretrained(
    "./test/saved_model/"
)  

[guoshengCS]

LGTM