EDG/extrapolate.py at main · Chuge0335/EDG · GitHub

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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228

import torch
import argparse
from tqdm import tqdm
from merge.model import load_model, load_camera_model
import inspect
import os
import re
from merge.utils import rgetattr
from lvdm.modules.attention import SpatialTransformer, TemporalTransformer
from lvdm.modules.networks.openaimodel3d import ResBlock, Downsample, Upsample

@torch.inference_mode()
def dyn(
    *,
    sft_path: str,
    dpo_path: str,
    save_dir: str,
    sft_config: str,
    dpo_config: str = 'configs/inference_512_v1.0.yaml',
    alpha: float,
):

    # global args
    # keys, _, _, values = inspect.getargvalues(inspect.currentframe())
    # args = argparse.Namespace(**{k:values[k] for k in keys})
    sft_model = load_model(sft_path, sft_config).to('cuda')
    dpo_model = load_model(dpo_path, dpo_config).to('cuda')
    # assert len(sft_model.state_dict()) == len(dpo_model.state_dict())
    total = len(dpo_model.state_dict())
    for name, dpo_model_param in tqdm(dpo_model.named_parameters(), total=total):
        sft_model_param = sft_model.state_dict()[name]
        sft_model_param.data = dpo_model_param.data + alpha * (dpo_model_param.data - sft_model_param.data)

    os.makedirs(save_dir, exist_ok=True)
    torch.save({'state_dict': sft_model.state_dict()}, save_dir + '/ckpt.pt')

@torch.inference_mode()
def dyn2(
    *,
    sft_path: str,
    dpo_path: str,
    save_dir: str,
    sft_config: str,
    dpo_config: str = 'configs/inference_512_v1.0.yaml',
    alpha: float,
):

    # global args
    # keys, _, _, values = inspect.getargvalues(inspect.currentframe())
    # args = argparse.Namespace(**{k:values[k] for k in keys})
    sft_model = load_model(sft_path, sft_config).to('cuda')
    dpo_model = load_model(dpo_path, dpo_config).to('cuda')
    # assert len(sft_model.state_dict()) == len(dpo_model.state_dict())
    total = len(dpo_model.state_dict())
    for name, dpo_model_param in tqdm(dpo_model.named_parameters(), total=total):
        sft_model_param = sft_model.state_dict()[name]
        dpo_model_param.data = dpo_model_param.data + alpha * (dpo_model_param.data - sft_model_param.data)

    os.makedirs(save_dir, exist_ok=True)
    torch.save({'state_dict': dpo_model.state_dict()}, save_dir + '/ckpt.pt')

@torch.inference_mode()
def main1(
    *,
    dynamic_path: str,
    consistency_path: str,
    sft_path: str,
    pre_path: str,
    base_path: str,
    save_dir: str,
    alpha: float,
):

    # global args
    # keys, _, _, values = inspect.getargvalues(inspect.currentframe())
    # args = argparse.Namespace(**{k:values[k] for k in keys})

    sft_models = [load_model(sft_path) for sft_path in [dynamic_path, consistency_path]]
    base_model = load_model(base_path)

    total = len(base_model.state_dict())

    for name, base_model_param in tqdm(base_model.named_parameters(), total=total):
        base_model_param.data = base_model_param.data + alpha * (
            sum([(sft_model.state_dict()[name].data - base_model_param.data) for sft_model in sft_models])
        )

    os.makedirs(save_dir, exist_ok=True)
    torch.save({'state_dict': base_model.state_dict()}, save_dir + '/ckpt.pt')

# TODO1: extrapolate different ability
# TODO2: migrate the interfernece

@torch.inference_mode()
def main2(
    *,
    dynamic_path: str,
    consistency_path: str,
    sft_path: str,
    pre_path: str,
    base_path: str,
    save_dir: str,
    alpha: float,
    binary: str, # [spa, temporal, res, init_attn, out, image_proj] => [0,1,2,3,4,5]
):

    # global args
    # keys, _, _, values = inspect.getargvalues(inspect.currentframe())
    # args = argparse.Namespace(**{k:values[k] for k in keys})

    # sft_models = [load_model(sft_path) for sft_path in [dynamic_path, consistency_path]]
    dynamic_model = load_model(dynamic_path).to('cuda')
    consistency_model = load_model(pre_path).to('cuda')
    base_model = load_model(base_path).to('cuda')

    total = len(base_model.state_dict())

    for name, base_model_param in tqdm(base_model.named_parameters(), total=total):
        match = re.match(r'model\.diffusion_model\.(input|output)_blocks\.(\d+)\.(\d+)\.\w+', name)
        module = None
        if match:
            upper_name = '.'.join(name.split('.')[:5])
            module = rgetattr(base_model, upper_name)
        else:
            match2 = re.match(r'model\.diffusion_model\.middle_block\.(\d+)\.\w+', name)
            if match2:
                upper_name = '.'.join(name.split('.')[:4])
                module = rgetattr(base_model, upper_name)
        if binary[0] == '1' and isinstance(module, TemporalTransformer):
            base_model_param.data = base_model_param.data + alpha * (dynamic_model.state_dict()[name].data - base_model_param.data)
        elif binary[1] == '1' and isinstance(module, SpatialTransformer):
            base_model_param.data = base_model_param.data + alpha * (dynamic_model.state_dict()[name].data - base_model_param.data)
        elif binary[2] == '1' and isinstance(module, ResBlock):
            base_model_param.data = base_model_param.data + alpha * (dynamic_model.state_dict()[name].data - base_model_param.data)

        if binary[3] == '1' and 'diffusion_model.init_attn' in name:
            base_model_param.data = base_model_param.data + alpha * (dynamic_model.state_dict()[name].data - base_model_param.data)
        if binary[4] == '1' and 'diffusion_model.out' in name:
            base_model_param.data = base_model_param.data + alpha * (dynamic_model.state_dict()[name].data - base_model_param.data)
        if binary[5] == '1' and 'image_proj_model' in name:
            base_model_param.data = base_model_param.data + alpha * (dynamic_model.state_dict()[name].data - base_model_param.data)
        # if 'diffusion_model.fps_embedding' in name:
        #     base_model_param.data = base_model_param.data + alpha * (consistency_model.state_dict()[name].data - base_model_param.data)

    os.makedirs(save_dir, exist_ok=True)
    torch.save({'state_dict': base_model.state_dict()}, save_dir + '/ckpt.pt')


@torch.inference_mode()
def main3(
    *,
    dynamic_path: str,
    camera_path: str,
    camera_config: str,
    sft_path: str,
    pre_path: str,
    base_path: str,
    save_dir: str,
    alpha: float,
    binary: str, # [spa, temporal, res, init_attn, out, image_proj] => [0,1,2,3,4,5]
):

    # global args
    # keys, _, _, values = inspect.getargvalues(inspect.currentframe())
    # args = argparse.Namespace(**{k:values[k] for k in keys})

    # sft_models = [load_model(sft_path) for sft_path in [dynamic_path, consistency_path]]
    # dynamic_model = load_model(dynamic_path).to('cuda')
    import lvdm
    old = lvdm.modules.attention.TemporalTransformer
    lvdm.modules.attention.TemporalTransformer = TemporalTransformer2
    lvdm.modules.networks.openaimodel3d.TemporalTransformer = TemporalTransformer2

    camera_model = load_camera_model('/opt/tiger/DiT/ckpt/camera_motion_odata/epoch=6-step=14000.ckpt', temporal_selfatt_only=False, image_cross_attention=True).to('cuda')

    lvdm.modules.attention.TemporalTransformer=old
    lvdm.modules.networks.openaimodel3d.TemporalTransformer=old
    base_model = load_model(base_path).to('cuda')

    total = len(base_model.state_dict())

    for name, _ in tqdm(camera_model.named_parameters(), total=total):
        if name not in base_model.state_dict():
            continue
        camera_model.state_dict()[name].data = camera_model.state_dict()[name].data + alpha * (camera_model.state_dict()[name].data - base_model.state_dict()[name].data)
        # match = re.match(r'model\.diffusion_model\.(input|output)_blocks\.(\d+)\.(\d+)\.\w+', name)
        # module = None
        # if match:
        #     upper_name = '.'.join(name.split('.')[:5])
        #     module = rgetattr(base_model, upper_name)
        # else:
        #     match2 = re.match(r'model\.diffusion_model\.middle_block\.(\d+)\.\w+', name)
        #     if match2:
        #         upper_name = '.'.join(name.split('.')[:4])
        #         module = rgetattr(base_model, upper_name)
        # if binary[0] == '1' and isinstance(module, TemporalTransformer):
        #     base_model_param.data = base_model_param.data + alpha * (camera_path.state_dict()[name].data - base_model_param.data)
        # elif binary[1] == '1' and isinstance(module, SpatialTransformer):
        #     base_model_param.data = base_model_param.data + alpha * (camera_path.state_dict()[name].data - base_model_param.data)
        # elif binary[2] == '1' and isinstance(module, ResBlock):
        #     base_model_param.data = base_model_param.data + alpha * (camera_path.state_dict()[name].data - base_model_param.data)

        # if binary[3] == '1' and 'diffusion_model.init_attn' in name:
        #     base_model_param.data = base_model_param.data + alpha * (camera_path.state_dict()[name].data - base_model_param.data)
        # if binary[4] == '1' and 'diffusion_model.out' in name:
        #     base_model_param.data = base_model_param.data + alpha * (camera_path.state_dict()[name].data - base_model_param.data)
        # if binary[5] == '1' and 'image_proj_model' in name:
        #     base_model_param.data = base_model_param.data + alpha * (camera_path.state_dict()[name].data - base_model_param.data)
        # if 'diffusion_model.fps_embedding' in name:
        #     base_model_param.data = base_model_param.data + alpha * (consistency_model.state_dict()[name].data - base_model_param.data)

    os.makedirs(save_dir, exist_ok=True)
    torch.save({'state_dict': camera_model.state_dict()}, save_dir + '/ckpt.pt')


if __name__ == '__main__':
    import defopt
    try:
        defopt.run(eval(os.getenv('func', 'main')))
    except:
        import sys,pdb,bdb
        type, value, tb = sys.exc_info()
        if type == bdb.BdbQuit:
            exit()
        print(type,value)
        pdb.post_mortem(tb)