File size: 5,289 Bytes
f8f62f3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import argparse

import torch
import open_clip
import pandas as pd
from fvcore.nn import FlopCountAnalysis, flop_count_str, ActivationCountAnalysis


parser = argparse.ArgumentParser(description='OpenCLIP Profiler')

# benchmark specific args
parser.add_argument('--model', metavar='NAME', default='',
                    help='model(s) to profile')
parser.add_argument('--results-file', default='', type=str, metavar='FILENAME',
                    help='Output csv file for results')


def profile_fvcore(
        model,
        image_input_size=(3, 224, 224),
        text_input_size=(77,),
        batch_size=1,
        detailed=False,
        force_cpu=False
):
    if force_cpu:
        model = model.to('cpu')
    device, dtype = next(model.parameters()).device, next(model.parameters()).dtype
    example_image_input = torch.ones((batch_size,) + image_input_size, device=device, dtype=dtype)
    example_text_input = torch.ones((batch_size,) + text_input_size, device=device, dtype=torch.int64)
    fca = FlopCountAnalysis(model, (example_image_input, example_text_input))
    aca = ActivationCountAnalysis(model, (example_image_input, example_text_input))
    if detailed:
        fcs = flop_count_str(fca)
        print(fcs)
    return fca.total(), aca.total()


def profile_fvcore_text(
        model,
        text_input_size=(77,),
        batch_size=1,
        detailed=False,
        force_cpu=False
):
    if force_cpu:
        model = model.to('cpu')
    device = next(model.parameters()).device
    example_input = torch.ones((batch_size,) + text_input_size, device=device, dtype=torch.int64)
    fca = FlopCountAnalysis(model, example_input)
    aca = ActivationCountAnalysis(model, example_input)
    if detailed:
        fcs = flop_count_str(fca)
        print(fcs)
    return fca.total(), aca.total()


def profile_fvcore_image(
        model,
        image_input_size=(3, 224, 224),
        batch_size=1,
        detailed=False,
        force_cpu=False
):
    if force_cpu:
        model = model.to('cpu')
    device, dtype = next(model.parameters()).device, next(model.parameters()).dtype
    example_input = torch.ones((batch_size,) + image_input_size, device=device, dtype=dtype)
    fca = FlopCountAnalysis(model, example_input)
    aca = ActivationCountAnalysis(model, example_input)
    if detailed:
        fcs = flop_count_str(fca)
        print(fcs)
    return fca.total(), aca.total()


def count_params(model):
    return sum([m.numel() for m in model.parameters()])


def profile_model(model_name):
    model = open_clip.create_model(model_name, force_custom_text=True, pretrained_hf=False)
    model.eval()
    if torch.cuda.is_available():
        model = model.cuda()

    if isinstance(model.visual.image_size, (tuple, list)):
        image_input_size = (3,) + tuple(model.visual.image_size[-2:])
    else:
        image_input_size = (3, model.visual.image_size, model.visual.image_size)
    text_input_size = (77,)

    results = {}
    results['model'] = model_name
    results['image_size'] = image_input_size[1]

    model_cfg = open_clip.get_model_config(model_name)
    if model_cfg:
        vision_cfg = open_clip.CLIPVisionCfg(**model_cfg['vision_cfg'])
        text_cfg = open_clip.CLIPTextCfg(**model_cfg['text_cfg'])
        results['image_width'] = int(vision_cfg.width)
        results['text_width'] = int(text_cfg.width)
        results['embed_dim'] = int(model_cfg['embed_dim'])
    else:
        results['image_width'] = 0
        results['text_width'] = 0
        results['embed_dim'] = 0

    retries = 2
    while retries:
        retries -= 1
        try:
            macs, acts = profile_fvcore(
                model, image_input_size=image_input_size, text_input_size=text_input_size, force_cpu=not retries)

            image_macs, image_acts = profile_fvcore_image(
                model.visual, image_input_size=image_input_size, force_cpu=not retries)

            text_macs, text_acts = profile_fvcore_text(
                model.text, text_input_size=text_input_size, force_cpu=not retries)

            results['gmacs'] = round(macs / 1e9, 2)
            results['macts'] = round(acts / 1e6, 2)
            results['mparams'] = round(count_params(model) / 1e6, 2)
            results['image_gmacs'] = round(image_macs / 1e9, 2)
            results['image_macts'] = round(image_acts / 1e6, 2)
            results['image_mparams'] = round(count_params(model.visual) / 1e6, 2)
            results['text_gmacs'] = round(text_macs / 1e9, 2)
            results['text_macts'] = round(text_acts / 1e6, 2)
            results['text_mparams'] = round(count_params(model.text) / 1e6, 2)
        except RuntimeError as e:
            pass
    return results


def main():
    args = parser.parse_args()

    # FIXME accept a text file name to allow lists of models in txt/csv
    if args.model == 'all':
        parsed_model = open_clip.list_models()
    else:
        parsed_model = args.model.split(',')

    results = []
    for m in parsed_model:
        row = profile_model(m)
        results.append(row)

    df = pd.DataFrame(results, columns=results[0].keys())
    df = df.sort_values('gmacs')
    print(df)
    if args.results_file:
        df.to_csv(args.results_file, index=False)


if __name__ == '__main__':
    main()