Refactor CLIP&ErnieViL model and Add ChineseCLIP Model

[JunnYu]

PR types

Others

PR changes

Models

Description

• 重构CLIP模型，支持使用Config加载，兼容旧版model_config.json文件。（权重参数命名没有变化，只修改了Config的参数名）；
• 添加CLIP和ErnieViL的单测（其中 ErnieViL的tokenizer与Ernie一致，因此并未添加单测。）。
• 新增ChineseCLIP模型：
• ChineseCLIP 当前支持的权重

CHINESE_CLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "OFA-Sys/chinese-clip-vit-base-patch16",
    "OFA-Sys/chinese-clip-vit-huge-patch14",
    "OFA-Sys/chinese-clip-vit-large-patch14",
    "OFA-Sys/chinese-clip-vit-large-patch14-336px",
]

• CLIP当前支持的权重：

CLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "openai/clip-vit-base-patch32",  # ViT-B/32
    "openai/clip-vit-base-patch16",  # ViT-B/16
    "openai/clip-vit-large-patch14",  # ViT-L/14
    "laion/CLIP-ViT-H-14-laion2B-s32B-b79K",
    "laion/CLIP-ViT-B-32-laion2B-s34B-b79K",
    "openai/clip-rn50",  # RN50
    "openai/clip-rn101",  # RN101
    "openai/clip-rn50x4",  # RN50x4
]

• 重构ErnieViL模型，支持使用Config加载，由于该模型使用较少，因此未兼容旧版model_config.json文件，直接在community底下新增了权重文件。（方便下载，未来的新模型几乎都会从 community 目录下下载。）

ERNIE_VIL_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "PaddlePaddle/ernie_vil-2.0-base-zh",
    "PaddlePaddle/disco_diffusion_ernie_vil-2.0-base-zh",
]

• 同步修改TaskFlow中 ERNIE_VIL 相关的链接。
• 修改DPT及BIT的Config默认值，return_dict=True
• 重命名 PreTrainedModel 成 PretrainedModel，大写的T改成小写。

[paddle-bot]

Thanks for your contribution!

[JunnYu]

TODO：ERNIEVIl还需要更新一下

DONE

[codecov]

Codecov Report

Merging #4270 (f0d1394) into develop (943a0ff) will increase coverage by 0.99%.
The diff coverage is 61.32%.

@@             Coverage Diff             @@
##           develop    #4270      +/-   ##
===========================================
+ Coverage    38.67%   39.67%   +0.99%     
===========================================
  Files          421      432      +11     
  Lines        59600    60739    +1139     
===========================================
+ Hits         23052    24097    +1045     
- Misses       36548    36642      +94     

Impacted Files	Coverage Δ
paddlenlp/taskflow/taskflow.py	75.29% <ø> (ø)
paddlenlp/transformers/auto/modeling.py	71.88% <ø> (ø)
paddlenlp/transformers/auto/tokenizer.py	81.74% <ø> (ø)
paddlenlp/transformers/chineseclip/converter.py	0.00% <0.00%> (ø)
paddlenlp/transformers/clip/converter.py	0.00% <0.00%> (ø)
paddlenlp/transformers/clip/modeling.py	60.61% <ø> (+37.96%)	⬆️
...lenlp/transformers/chineseclip/image_processing.py	20.48% <20.48%> (ø)
...ddlenlp/transformers/ernie_vil/image_processing.py	20.48% <20.48%> (ø)
...ransformers/guided_diffusion_utils/resize_right.py	21.39% <33.33%> (-0.66%)	⬇️
paddlenlp/transformers/chineseclip/procesing.py	38.77% <38.77%> (ø)
... and 25 more

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[wj-Mcat]

你的这个 PR 太大了，看起来很花费时间，所以我只留了几个 comment。

至于 converter 这块我还是会继续兼容你这边的工作，保证你这边能够正常工作。