-
Notifications
You must be signed in to change notification settings - Fork 143
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Showing
9 changed files
with
427 additions
and
15 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,45 @@ | ||
| { | ||
| "stream_0": { | ||
| "emb_size": 768, | ||
| "feedforward_size": 3072, | ||
| "hidden_size": 768, | ||
| "hidden_act": "gelu", | ||
| "heads_num": 12, | ||
| "layers_num": 12, | ||
| "max_seq_length": 77, | ||
| "embedding": ["word", "pos"], | ||
| "encoder": "transformer", | ||
| "mask": "causal", | ||
| "remove_embedding_layernorm": true, | ||
| "layernorm_positioning": "pre", | ||
| "pooling": "last" | ||
| }, | ||
|
|
||
| "stream_1": { | ||
| "emb_size": 1024, | ||
| "feedforward_size": 4096, | ||
| "hidden_size": 1024, | ||
| "hidden_act": "gelu_fast", | ||
| "heads_num": 16, | ||
| "layers_num": 24, | ||
| "max_seq_length": 257, | ||
| "embedding": ["patch", "pos"], | ||
| "encoder": "transformer", | ||
| "mask": "fully_visible", | ||
| "remove_embedding_layernorm": false, | ||
| "layernorm_positioning": "pre", | ||
| "pooling": "first" | ||
| }, | ||
|
|
||
| "data_processor": "clip", | ||
| "embedding": ["dual"], | ||
| "encoder": "dual", | ||
| "target": ["clr"], | ||
| "image_height": 224, | ||
| "image_width": 224, | ||
| "patch_size": 14, | ||
| "feature_size": 768, | ||
| "projection": true, | ||
| "tie_weights": false, | ||
| "dropout": 0.0 | ||
| } |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,7 @@ | ||
| { | ||
| "pad_token": "<|endoftext|>", | ||
| "unk_token": "<|endoftext|>", | ||
| "cls_token": "<|startoftext|>", | ||
| "sep_token": "<|endoftext|>", | ||
| "mask_token": "<|endoftext|>" | ||
| } |
125 changes: 125 additions & 0 deletions
125
scripts/convert_clip_from_huggingface_to_tencentpretrain.py
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,125 @@ | ||
| import argparse | ||
| import collections | ||
| import torch | ||
|
|
||
|
|
||
| def convert_clip_transformer(input_model, output_model, layers_num): | ||
|
|
||
| for i in range(layers_num): | ||
| output_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.linear_layers.0.weight"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".self_attn.q_proj.weight"] | ||
| output_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.linear_layers.0.bias"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".self_attn.q_proj.bias"] | ||
|
|
||
| output_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.linear_layers.1.weight"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".self_attn.k_proj.weight"] | ||
| output_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.linear_layers.1.bias"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".self_attn.k_proj.bias"] | ||
|
|
||
| output_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.linear_layers.2.weight"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".self_attn.v_proj.weight"] | ||
| output_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.linear_layers.2.bias"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".self_attn.v_proj.bias"] | ||
|
|
||
| output_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.final_linear.weight"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".self_attn.out_proj.weight"] | ||
| output_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.final_linear.bias"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".self_attn.out_proj.bias"] | ||
|
|
||
| output_model["encoder.encoder_0.transformer." + str(i) + ".layer_norm_1.gamma"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".layer_norm1.weight"] | ||
| output_model["encoder.encoder_0.transformer." + str(i) + ".layer_norm_1.beta"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".layer_norm1.bias"] | ||
|
|
||
| output_model["encoder.encoder_0.transformer." + str(i) + ".feed_forward.linear_1.weight"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".mlp.fc1.weight"] | ||
| output_model["encoder.encoder_0.transformer." + str(i) + ".feed_forward.linear_1.bias"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".mlp.fc1.bias"] | ||
| output_model["encoder.encoder_0.transformer." + str(i) + ".feed_forward.linear_2.weight"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".mlp.fc2.weight"] | ||
| output_model["encoder.encoder_0.transformer." + str(i) + ".feed_forward.linear_2.bias"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".mlp.fc2.bias"] | ||
|
|
||
| output_model["encoder.encoder_0.transformer." + str(i) + ".layer_norm_2.gamma"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".layer_norm2.weight"] | ||
| output_model["encoder.encoder_0.transformer." + str(i) + ".layer_norm_2.beta"] = \ | ||
| input_model["text_model.encoder.layers." + str(i) + ".layer_norm2.bias"] | ||
|
|
||
|
|
||
| output_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.linear_layers.0.weight"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".self_attn.q_proj.weight"] | ||
| output_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.linear_layers.0.bias"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".self_attn.q_proj.bias"] | ||
|
|
||
| output_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.linear_layers.1.weight"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".self_attn.k_proj.weight"] | ||
| output_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.linear_layers.1.bias"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".self_attn.k_proj.bias"] | ||
|
|
||
| output_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.linear_layers.2.weight"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".self_attn.v_proj.weight"] | ||
| output_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.linear_layers.2.bias"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".self_attn.v_proj.bias"] | ||
|
|
||
| output_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.final_linear.weight"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".self_attn.out_proj.weight"] | ||
| output_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.final_linear.bias"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".self_attn.out_proj.bias"] | ||
|
|
||
| output_model["encoder.encoder_1.transformer." + str(i) + ".layer_norm_1.gamma"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".layer_norm1.weight"] | ||
| output_model["encoder.encoder_1.transformer." + str(i) + ".layer_norm_1.beta"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".layer_norm1.bias"] | ||
|
|
||
| output_model["encoder.encoder_1.transformer." + str(i) + ".feed_forward.linear_1.weight"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".mlp.fc1.weight"] | ||
| output_model["encoder.encoder_1.transformer." + str(i) + ".feed_forward.linear_1.bias"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".mlp.fc1.bias"] | ||
| output_model["encoder.encoder_1.transformer." + str(i) + ".feed_forward.linear_2.weight"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".mlp.fc2.weight"] | ||
| output_model["encoder.encoder_1.transformer." + str(i) + ".feed_forward.linear_2.bias"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".mlp.fc2.bias"] | ||
|
|
||
| output_model["encoder.encoder_1.transformer." + str(i) + ".layer_norm_2.gamma"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".layer_norm2.weight"] | ||
| output_model["encoder.encoder_1.transformer." + str(i) + ".layer_norm_2.beta"] = \ | ||
| input_model["vision_model.encoder.layers." + str(i) + ".layer_norm2.bias"] | ||
|
|
||
| def main(): | ||
| parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter) | ||
| parser.add_argument("--input_model_path", type=str, default="models/input_model.bin", | ||
| help=".") | ||
| parser.add_argument("--output_model_path", type=str, default="models/output_model.bin", | ||
| help=".") | ||
| parser.add_argument("--layers_num", type=int, default=12, help=".") | ||
|
|
||
| args = parser.parse_args() | ||
|
|
||
| input_model = torch.load(args.input_model_path, map_location="cpu") | ||
|
|
||
| output_model = collections.OrderedDict() | ||
|
|
||
| output_model["embedding.dual.embedding_0.word.embedding.weight"] = input_model["text_model.embeddings.token_embedding.weight"] | ||
| output_model["embedding.dual.embedding_0.pos.embedding.weight"] = input_model["text_model.embeddings.position_embedding.weight"] | ||
| output_model["embedding.dual.embedding_1.patch.cls_emb"] = input_model["vision_model.embeddings.class_embedding"].unsqueeze(0).unsqueeze(0) | ||
| output_model["embedding.dual.embedding_1.patch.projection.weight"] = input_model["vision_model.embeddings.patch_embedding.weight"] | ||
| output_model["embedding.dual.embedding_1.pos.embedding.weight"] = input_model["vision_model.embeddings.position_embedding.weight"] | ||
|
|
||
| output_model["embedding.dual.stream_1_layer_norm.gamma"] = input_model["vision_model.pre_layrnorm.weight"] | ||
| output_model["embedding.dual.stream_1_layer_norm.beta"] = input_model["vision_model.pre_layrnorm.bias"] | ||
|
|
||
| convert_clip_transformer(input_model, output_model, args.layers_num) | ||
|
|
||
| output_model["encoder.encoder_0.layer_norm.gamma"] = input_model["text_model.final_layer_norm.weight"] | ||
| output_model["encoder.encoder_0.layer_norm.beta"] = input_model["text_model.final_layer_norm.bias"] | ||
| output_model["encoder.encoder_1.layer_norm.gamma"] = input_model["vision_model.post_layernorm.weight"] | ||
| output_model["encoder.encoder_1.layer_norm.beta"] = input_model["vision_model.post_layernorm.bias"] | ||
| output_model["target.clr.logit_scale"] = input_model["logit_scale"] | ||
| output_model["target.clr.encoder_0_projection"] = input_model["text_projection.weight"].T | ||
| output_model["target.clr.encoder_1_projection"] = input_model["visual_projection.weight"].T | ||
|
|
||
| torch.save(output_model, args.output_model_path) | ||
|
|
||
|
|
||
| if __name__ == "__main__": | ||
| main() |
125 changes: 125 additions & 0 deletions
125
scripts/convert_clip_from_tencentpretrain_to_huggingface.py
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,125 @@ | ||
| import argparse | ||
| import collections | ||
| import torch | ||
|
|
||
|
|
||
| def convert_clip_transformer(input_model, output_model, layers_num): | ||
|
|
||
| for i in range(12): | ||
| output_model["text_model.encoder.layers." + str(i) + ".self_attn.q_proj.weight"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.linear_layers.0.weight"] | ||
| output_model["text_model.encoder.layers." + str(i) + ".self_attn.q_proj.bias"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.linear_layers.0.bias"] | ||
|
|
||
| output_model["text_model.encoder.layers." + str(i) + ".self_attn.k_proj.weight"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.linear_layers.1.weight"] | ||
| output_model["text_model.encoder.layers." + str(i) + ".self_attn.k_proj.bias"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.linear_layers.1.bias"] | ||
|
|
||
| output_model["text_model.encoder.layers." + str(i) + ".self_attn.v_proj.weight"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.linear_layers.2.weight"] | ||
| output_model["text_model.encoder.layers." + str(i) + ".self_attn.v_proj.bias"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.linear_layers.2.bias"] | ||
|
|
||
| output_model["text_model.encoder.layers." + str(i) + ".self_attn.out_proj.weight"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.final_linear.weight"] | ||
| output_model["text_model.encoder.layers." + str(i) + ".self_attn.out_proj.bias"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".self_attn.final_linear.bias"] | ||
|
|
||
| output_model["text_model.encoder.layers." + str(i) + ".layer_norm1.weight"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".layer_norm_1.gamma"] | ||
| output_model["text_model.encoder.layers." + str(i) + ".layer_norm1.bias"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".layer_norm_1.beta"] | ||
|
|
||
| output_model["text_model.encoder.layers." + str(i) + ".mlp.fc1.weight"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".feed_forward.linear_1.weight"] | ||
| output_model["text_model.encoder.layers." + str(i) + ".mlp.fc1.bias"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".feed_forward.linear_1.bias"] | ||
| output_model["text_model.encoder.layers." + str(i) + ".mlp.fc2.weight"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".feed_forward.linear_2.weight"] | ||
| output_model["text_model.encoder.layers." + str(i) + ".mlp.fc2.bias"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".feed_forward.linear_2.bias"] | ||
|
|
||
| output_model["text_model.encoder.layers." + str(i) + ".layer_norm2.weight"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".layer_norm_2.gamma"] | ||
| output_model["text_model.encoder.layers." + str(i) + ".layer_norm2.bias"] = \ | ||
| input_model["encoder.encoder_0.transformer." + str(i) + ".layer_norm_2.beta"] | ||
|
|
||
| for i in range(12): | ||
| output_model["vision_model.encoder.layers." + str(i) + ".self_attn.q_proj.weight"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.linear_layers.0.weight"] | ||
| output_model["vision_model.encoder.layers." + str(i) + ".self_attn.q_proj.bias"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.linear_layers.0.bias"] | ||
|
|
||
| output_model["vision_model.encoder.layers." + str(i) + ".self_attn.k_proj.weight"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.linear_layers.1.weight"] | ||
| output_model["vision_model.encoder.layers." + str(i) + ".self_attn.k_proj.bias"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.linear_layers.1.bias"] | ||
|
|
||
| output_model["vision_model.encoder.layers." + str(i) + ".self_attn.v_proj.weight"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.linear_layers.2.weight"] | ||
| output_model["vision_model.encoder.layers." + str(i) + ".self_attn.v_proj.bias"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.linear_layers.2.bias"] | ||
|
|
||
| output_model["vision_model.encoder.layers." + str(i) + ".self_attn.out_proj.weight"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.final_linear.weight"] | ||
| output_model["vision_model.encoder.layers." + str(i) + ".self_attn.out_proj.bias"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".self_attn.final_linear.bias"] | ||
|
|
||
| output_model["vision_model.encoder.layers." + str(i) + ".layer_norm1.weight"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".layer_norm_1.gamma"] | ||
| output_model["vision_model.encoder.layers." + str(i) + ".layer_norm1.bias"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".layer_norm_1.beta"] | ||
|
|
||
| output_model["vision_model.encoder.layers." + str(i) + ".mlp.fc1.weight"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".feed_forward.linear_1.weight"] | ||
| output_model["vision_model.encoder.layers." + str(i) + ".mlp.fc1.bias"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".feed_forward.linear_1.bias"] | ||
| output_model["vision_model.encoder.layers." + str(i) + ".mlp.fc2.weight"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".feed_forward.linear_2.weight"] | ||
| output_model["vision_model.encoder.layers." + str(i) + ".mlp.fc2.bias"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".feed_forward.linear_2.bias"] | ||
|
|
||
| output_model["vision_model.encoder.layers." + str(i) + ".layer_norm2.weight"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".layer_norm_2.gamma"] | ||
| output_model["vision_model.encoder.layers." + str(i) + ".layer_norm2.bias"] = \ | ||
| input_model["encoder.encoder_1.transformer." + str(i) + ".layer_norm_2.beta"] | ||
|
|
||
| def main(): | ||
| parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter) | ||
| parser.add_argument("--input_model_path", type=str, default="models/input_model.bin", | ||
| help=".") | ||
| parser.add_argument("--output_model_path", type=str, default="models/output_model.bin", | ||
| help=".") | ||
| parser.add_argument("--layers_num", type=int, default=12, help=".") | ||
|
|
||
| args = parser.parse_args() | ||
|
|
||
| input_model = torch.load(args.input_model_path, map_location="cpu") | ||
|
|
||
| output_model = collections.OrderedDict() | ||
|
|
||
| output_model["text_model.embeddings.token_embedding.weight"] = input_model["embedding.dual.embedding_0.word.embedding.weight"] | ||
| output_model["text_model.embeddings.position_embedding.weight"] = input_model["embedding.dual.embedding_0.pos.embedding.weight"] | ||
| output_model["vision_model.embeddings.class_embedding"] = input_model["embedding.dual.embedding_1.patch.cls_emb"].squeeze().squeeze() | ||
| output_model["vision_model.embeddings.patch_embedding.weight"] = input_model["embedding.dual.embedding_1.patch.projection.weight"] | ||
| output_model["vision_model.embeddings.position_embedding.weight"] = input_model["embedding.dual.embedding_1.pos.embedding.weight"] | ||
|
|
||
| output_model["vision_model.pre_layrnorm.weight"] = input_model["embedding.dual.stream_1_layer_norm.gamma"] | ||
| output_model["vision_model.pre_layrnorm.bias"] = input_model["embedding.dual.stream_1_layer_norm.beta"] | ||
|
|
||
| convert_clip_transformer(input_model, output_model, args.layers_num) | ||
|
|
||
| output_model["text_model.final_layer_norm.weight"] = input_model["encoder.encoder_0.layer_norm.gamma"] | ||
| output_model["text_model.final_layer_norm.bias"] = input_model["encoder.encoder_0.layer_norm.beta"] | ||
| output_model["vision_model.post_layernorm.weight"] = input_model["encoder.encoder_1.layer_norm.gamma"] | ||
| output_model["vision_model.post_layernorm.bias"] = input_model["encoder.encoder_1.layer_norm.beta"] | ||
| output_model["logit_scale"] = input_model["target.clr.logit_scale"] | ||
| output_model["text_projection.weight"] = input_model["target.clr.encoder_0_projection"].T | ||
| output_model["visual_projection.weight"] = input_model["target.clr.encoder_1_projection"].T | ||
|
|
||
| torch.save(output_model, args.output_model_path) | ||
|
|
||
|
|
||
| if __name__ == "__main__": | ||
| main() |
Oops, something went wrong.