Hugging Face's logo

microsoft
/
LLM2CLIP-EVA02-B-16

PyTorch
clip
custom_code
Model card Files Community
4
LLM2CLIP-EVA02-B-16
File size: 8,654 Bytes 
25bf357
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
 
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# Part of the code was taken from:
# https://github.com/huggingface/transformers/blob/main/src/transformers/models/clap/convert_clap_original_pytorch_to_hf.py

import argparse

import torch
from PIL import Image
from transformers import AutoModel, AutoConfig
from transformers import  CLIPImageProcessor, pipeline, CLIPTokenizer
from configuration_evaclip import EvaCLIPConfig
from modeling_evaclip import EvaCLIPModel


KEYS_TO_MODIFY_MAPPING = {
    "cls_token":"embeddings.class_embedding",
    "pos_embed":"embeddings.position_embedding.weight",
    "patch_embed.proj":"embeddings.patch_embedding",
    ".positional_embedding":".embeddings.position_embedding.weight",
    ".token_embedding":".embeddings.token_embedding",
    # "text.text_projection":"text_projection.weight",
    "mlp.c_fc":"mlp.fc1",
    "mlp.c_proj":"mlp.fc2",
    "mlp.w1":"mlp.fc1",
    "mlp.w2":"mlp.fc2",
    "mlp.w3":"mlp.fc3",
    ".proj.":".out_proj.",
    # "q_bias":"q_proj.bias",
    # "v_bias":"v_proj.bias",
    "out.":"out_proj.",
    "norm1":"layer_norm1",
    "norm2":"layer_norm2",
    "ln_1":"layer_norm1",
    "ln_2":"layer_norm2",
    ".attn":".self_attn",
    "norm.":"post_layernorm.",
    "ln_final":"final_layer_norm",
    "visual.blocks":"vision_model.encoder.layers",
    # "text.transformer.resblocks":"text_model.encoder.layers",
    "visual.head":"visual_projection",
    "visual.":"vision_model.",
    # "text.":"text_model.",

}

def rename_state_dict(state_dict):
    model_state_dict = {}

    for key, value in state_dict.items():
        # check if any key needs to be modified
        for key_to_modify, new_key in KEYS_TO_MODIFY_MAPPING.items():
            if key_to_modify in key:
                key = key.replace(key_to_modify, new_key)
        if "text_projection" in key:
            model_state_dict[key] = value.T
        elif "attn.qkv" in key:
            # split qkv into query key and value
            mixed_qkv = value
            qkv_dim = mixed_qkv.size(0) // 3

            query_layer = mixed_qkv[:qkv_dim]
            key_layer = mixed_qkv[qkv_dim : qkv_dim * 2]
            value_layer = mixed_qkv[qkv_dim * 2 :]

            model_state_dict[key.replace("qkv", "q_proj")] = query_layer
            model_state_dict[key.replace("qkv", "k_proj")] = key_layer
            model_state_dict[key.replace("qkv", "v_proj")] = value_layer

        elif "attn.in_proj" in key:
            # split qkv into query key and value
            mixed_qkv = value
            qkv_dim = mixed_qkv.size(0) // 3

            query_layer = mixed_qkv[:qkv_dim]
            key_layer = mixed_qkv[qkv_dim : qkv_dim * 2]
            value_layer = mixed_qkv[qkv_dim * 2 :]

            model_state_dict[key.replace("in_proj_", "q_proj.")] = query_layer
            model_state_dict[key.replace("in_proj_", "k_proj.")] = key_layer
            model_state_dict[key.replace("in_proj_", "v_proj.")] = value_layer

        elif "class_embedding" in key:
            model_state_dict[key] = value[0,0,:]
        elif "vision_model.embeddings.position_embedding" in key:
            model_state_dict[key] = value[0,:,:]

        else:
            model_state_dict[key] = value

    return model_state_dict

# This requires having a clone of https://github.com/baaivision/EVA/tree/master/EVA-CLIP as well as the right conda env
# Part of the code is copied from https://github.com/baaivision/EVA/blob/master/EVA-CLIP/README.md "Usage" section
def getevaclip(checkpoint_path, input_pixels, captions):
    from eva_clip import create_model_and_transforms, get_tokenizer
    model_name = "EVA02-CLIP-bigE-14-plus"
    model, _, _ = create_model_and_transforms(model_name, checkpoint_path, force_custom_clip=True)
    tokenizer = get_tokenizer(model_name)
    text = tokenizer(captions)

    with torch.no_grad():
        text_features = model.encode_text(text)
        image_features = model.encode_image(input_pixels)
        image_features_normed = image_features / image_features.norm(dim=-1, keepdim=True)
        text_features_normed = text_features / text_features.norm(dim=-1, keepdim=True)

        label_probs = (100.0 * image_features_normed @ text_features_normed.T).softmax(dim=-1)

    return  label_probs

def save_model_and_config(pytorch_dump_folder_path, hf_model, transformers_config):
    hf_model.save_pretrained(pytorch_dump_folder_path, safe_serialization=False)
    transformers_config.save_pretrained(pytorch_dump_folder_path)

def check_loaded_model(pytorch_dump_folder_path, processor, image):
    # hf_config = AutoConfig.from_pretrained(pytorch_dump_folder_path, trust_remote_code=True)
    # hf_model = AutoModel.from_pretrained(pytorch_dump_folder_path, config=hf_config, trust_remote_code=True)
    hf_model = AutoModel.from_pretrained(pytorch_dump_folder_path, trust_remote_code=True)
    
    processor = CLIPImageProcessor.from_pretrained("openai/clip-vit-base-patch16")
    image_path = 'LLM2CLIP-EVA02-B-16/CLIP.png'
    image = Image.open(image_path)
    input_pixels = processor(images=image, return_tensors="pt").pixel_values
    with torch.no_grad():
        image_features = hf_model.get_image_features(input_pixels)
    print(image_features.shape)
    
    
    # detector = pipeline(model=hf_model, task="zero-shot-image-classification", tokenizer = tokenizer, image_processor=processor)
    # detector_probs = detector(image, candidate_labels=captions)
    # print(f"text_probs loaded hf_model using pipeline: {detector_probs}")

def convert_evaclip_checkpoint(checkpoint_path, pytorch_dump_folder_path, config_path, image_path, save=False):
    processor = CLIPImageProcessor.from_pretrained("openai/clip-vit-base-patch16")
    image = Image.open(image_path)
    input_pixels = processor( images=image, return_tensors="pt", padding=True).pixel_values

    # This requires having a clone of https://github.com/baaivision/EVA/tree/master/EVA-CLIP as well as the right conda env
    # original_evaclip_probs = getevaclip(checkpoint_path, input_pixels, captions)
    # print(f"original_evaclip label probs: {original_evaclip_probs}")

    transformers_config = EvaCLIPConfig.from_pretrained(config_path)
    hf_model = EvaCLIPModel(transformers_config)
    pt_model_state_dict = torch.load(checkpoint_path)['module']
    state_dict = rename_state_dict(pt_model_state_dict)

    hf_model.load_state_dict(state_dict, strict=False)

    with torch.no_grad():
        image_features = hf_model.get_image_features(input_pixels)
        # text_features = hf_model.get_text_features(input_ids)
        image_features /= image_features.norm(dim=-1, keepdim=True)
        # text_features /= text_features.norm(dim=-1, keepdim=True)
        
    print(image_features.shape)
    # label_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1)
    # print(f"hf_model label probs: {label_probs}")

    if save:
        save_model_and_config(pytorch_dump_folder_path, hf_model, transformers_config)
    
    check_loaded_model(pytorch_dump_folder_path, processor, image)

    # hf_model.push_to_hub("ORGANIZATION_NAME/EVA02_CLIP_E_psz14_plus_s9B")

    


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--pytorch_dump_folder_path", default="LLM2CLIP-EVA02-B-16" ,type=str, help="Path to the output PyTorch model.")
    parser.add_argument("--checkpoint_path", default="/blob/hwq/data/tune_logs/T_vitB224_512x8_lr1e-5_Rcc12mR_Rcc3m_RIM15v1_RIM15v2_Ryfcc15m_20ep-2024_10_10-23/checkpoints/epoch_20/mp_rank_00_model_states.pt", type=str, help="Path to checkpoint" )
    parser.add_argument("--config_path", default='LLM2CLIP-EVA02-B-16', type=str, help="Path to hf config.json of model to convert")
    parser.add_argument("--image_path", default='LLM2CLIP-EVA02-B-16/CLIP.png', type=str, help="Path to image")
    parser.add_argument("--save", default=False, type=str, help="Path to image")

    args = parser.parse_args()

    convert_evaclip_checkpoint(args.checkpoint_path, args.pytorch_dump_folder_path, args.config_path, args.image_path, args.save)