File: lstm_comparison.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 (58 lines) | stat: -rw-r--r-- 2,076 bytes parent folder | download
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




from caffe2.python import workspace, core, lstm_benchmark, utils
from copy import copy

@utils.debug
def Compare(args):
    results = []
    num_iters = 1000
    args.gpu = True
    with core.DeviceScope(core.DeviceOption(workspace.GpuDeviceType, 0)):
        for batch_size in [64, 128, 256]:
            for seq_length in [20, 100]:
                for hidden_dim in [40, 100, 400, 800]:
                    args.batch_size = batch_size
                    args.seq_length = seq_length
                    args.hidden_dim = hidden_dim
                    args.data_size = batch_size * seq_length * num_iters
                    args.iters_to_report = num_iters // 3

                    args.implementation = 'own'
                    t_own = lstm_benchmark.Benchmark(args)
                    workspace.ResetWorkspace()
                    args.implementation = 'cudnn'
                    t_cudnn = lstm_benchmark.Benchmark(args)
                    workspace.ResetWorkspace()
                    results.append((copy(args), float(t_own), float(t_cudnn)))
                    print(args)
                    print("t_cudnn / t_own: {}".format(t_cudnn / t_own))

    for args, t_own, t_cudnn in results:
        print("{}: cudnn time: {}, own time: {}, ratio: {}".format(
            str(args), t_cudnn, t_own, t_cudnn / t_own))

    ratio_sum = 0
    for args, t_own, t_cudnn in results:
        ratio = float(t_cudnn) / t_own
        ratio_sum += ratio
        print("hidden_dim: {}, seq_lengths: {}, batch_size: {}, num_layers: {}:"
              " cudnn time: {}, own time: {}, ratio: {}".format(
                  args.hidden_dim, args.seq_length, args.batch_size,
                  args.num_layers, t_cudnn, t_own, ratio))

    print("Ratio average: {}".format(ratio_sum / len(results)))


if __name__ == '__main__':
    args = lstm_benchmark.GetArgumentParser().parse_args()

    workspace.GlobalInit([
        'caffe2',
        '--caffe2_log_level=0',
        '--caffe2_print_blob_sizes_at_exit=0',
        '--caffe2_gpu_memory_tracking=1'])

    Compare(args)