File: embedding_generation_benchmark.py

package info (click to toggle)
pytorch 1.13.1%2Bdfsg-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 139,252 kB
  • sloc: cpp: 1,100,274; python: 706,454; ansic: 83,052; asm: 7,618; java: 3,273; sh: 2,841; javascript: 612; makefile: 323; xml: 269; ruby: 185; yacc: 144; objc: 68; lex: 44
file content (196 lines) | stat: -rw-r--r-- 5,256 bytes parent folder | download | duplicates (2)
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
## @package embedding_generation_benchmark
# Module caffe2.python.embedding_generation_benchmark





from caffe2.proto import caffe2_pb2
from caffe2.python import workspace, core, utils, model_helper

import argparse
import numpy as np
import time

import logging

logging.basicConfig()
log = logging.getLogger("embedding_generation_benchmark")
log.setLevel(logging.DEBUG)


def generate_data(T, batch_size, max_seq_length):
    '''
    Fill a queue with input data
    '''
    log.info("Generating T={} batches".format(T))

    generate_input_init_net = core.Net('generate_input_init')
    queue = generate_input_init_net.CreateBlobsQueue(
        [], "inputqueue", num_blobs=1, capacity=T,
    )
    workspace.RunNetOnce(generate_input_init_net)

    generate_input_net = core.Net('generate_input')
    generate_input_net.EnqueueBlobs([queue, "scratch"], ["scratch"])
    np.random.seed(2603)

    for t in range(T):
        if (t % (max(10, T // 10)) == 0):
            log.info("Generating data {}/{}".format(t, T))
        X = np.tile(np.arange(max_seq_length), [batch_size, 1]).transpose()
        workspace.FeedBlob("scratch", X)
        workspace.RunNetOnce(generate_input_net.Proto())

    log.info("Finished data generation")
    return queue


def generate_embedding_table(vocab_size, embedding_size):
    log.info("Generating embedding table with dimensions {}"
             .format([vocab_size, embedding_size]))

    generate_table_net = core.Net('generate_table')
    table = generate_table_net.GaussianFill(
        [],
        ['embedding_table'],
        shape=[vocab_size, embedding_size],
    )

    workspace.RunNetOnce(generate_table_net)
    return table


def create_model(args, queue, embedding_table, embedding_size):
    model = model_helper.ModelHelper(name='embedding_generation_bench')
    input_blob = model.net.DequeueBlobs(queue, 'input_data')

    if args.implementation == 'sinusoid':
        model.net.SinusoidPositionEncoding(
            [input_blob],
            ['output'],
            embedding_size=embedding_size
        )
    else:
        model.net.Gather(
            [embedding_table, input_blob],
            ['output'],
        )

    return model


def Caffe2EmbeddingGeneration(args):
    T = args.data_size // args.batch_size

    queue = generate_data(T, args.batch_size, args.seq_length)

    embedding_table = None
    if args.implementation == 'table':
        embedding_table = generate_embedding_table(
            args.seq_length,
            args.embedding_size,
        )

    model = create_model(args, queue, embedding_table, args.embedding_size)

    workspace.RunNetOnce(model.param_init_net)
    workspace.CreateNet(model.net)

    start_time = time.time()
    num_iters = T
    total_iters = 0

    # Run the Benchmark
    log.info("------ Warming up ------")
    workspace.RunNet(model.net.Proto().name)

    log.info("------ Starting benchmark ------")
    start_time = time.time()
    last_time = time.time()
    for iteration in range(1, num_iters, args.iters_to_report):
        iters_once = min(args.iters_to_report, num_iters - iteration)
        total_iters += iters_once
        workspace.RunNet(model.net.Proto().name, iters_once)

        new_time = time.time()
        log.info(
            "Iter: {} / {}. Embeddings Generated Per Second: {}k.".format(
                iteration,
                num_iters,
                (iters_once * args.batch_size * args.seq_length) /
                (new_time - last_time) // 100 / 10,
            )
        )
        last_time = new_time

    total_per_sec = (num_iters - 1) * args.batch_size * args.seq_length
    total_per_sec = total_per_sec / (time.time() - start_time) // 100 / 10

    log.info("Done. Total embeddings generated per second " +
             "excluding 1st iteration: {}k".format(total_per_sec))

    return time.time() - start_time


@utils.debug
def Benchmark(args):
    return Caffe2EmbeddingGeneration(args)


def GetArgumentParser():
    parser = argparse.ArgumentParser(
        description="Embedding generation benchmark."
    )

    parser.add_argument(
        "--embedding_size",
        type=int,
        default=512,
        help="Embedding size",
    )
    parser.add_argument(
        "--batch_size",
        type=int,
        default=16,
        help="The batch size."
    )
    parser.add_argument(
        "--data_size",
        type=int,
        default=10000,
        help="Number of sequences to generate"
    )
    parser.add_argument(
        "--seq_length",
        type=int,
        default=128,
        help="Max sequence length"
    )
    parser.add_argument(
        "--iters_to_report",
        type=int,
        default=20,
        help="Number of iterations to report progress"
    )
    parser.add_argument(
        "--implementation",
        type=str,
        default="sinusoid",
        help="'table' or 'sinusoid'",
    )
    return parser


if __name__ == '__main__':
    args, extra_args = GetArgumentParser().parse_known_args()

    workspace.GlobalInit([
        'caffe2',
        '--caffe2_log_level=0',
        '--caffe2_print_blob_sizes_at_exit=0'] + extra_args)

    device = core.DeviceOption(caffe2_pb2.CPU)

    with core.DeviceScope(device):
        Benchmark(args)