kompressor 0.1.0

~60x faster, byte-exact rANS entropy coder for CompressAI mbt2018_mean
Documentation 
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

import time
import math
import sys
import torch
import torch.nn.functional as F
from compressai.zoo import mbt2018_mean

from thin_model import ThinMeanScaleHyperprior

DEVICE = "cuda"
torch.backends.cudnn.benchmark = True


def pad(x, factor=64):
    h, w = x.size(2), x.size(3)
    new_h = math.ceil(h / factor) * factor
    new_w = math.ceil(w / factor) * factor
    pad_h = new_h - h
    pad_w = new_w - w
    return F.pad(x, (0, pad_w, 0, pad_h))


def count_params(model):
    return sum(p.numel() for p in model.parameters())


def bench_forward(model, x, n_warmup=10, n_iters=50):
    model.eval()
    with torch.no_grad():
        for _ in range(n_warmup):
            _ = model.g_s(model.g_a(x))
        torch.cuda.synchronize()
        t0 = time.perf_counter()
        for _ in range(n_iters):
            y = model.g_a(x)
            x_hat = model.g_s(y)
        torch.cuda.synchronize()
        t1 = time.perf_counter()
    return (t1 - t0) / n_iters * 1000.0  # ms


def bench_ga(model, x, n_warmup=10, n_iters=50):
    model.eval()
    with torch.no_grad():
        for _ in range(n_warmup):
            _ = model.g_a(x)
        torch.cuda.synchronize()
        t0 = time.perf_counter()
        for _ in range(n_iters):
            y = model.g_a(x)
        torch.cuda.synchronize()
        t1 = time.perf_counter()
    return (t1 - t0) / n_iters * 1000.0


def bench_gs(model, y, n_warmup=10, n_iters=50):
    model.eval()
    with torch.no_grad():
        for _ in range(n_warmup):
            _ = model.g_s(y)
        torch.cuda.synchronize()
        t0 = time.perf_counter()
        for _ in range(n_iters):
            x_hat = model.g_s(y)
        torch.cuda.synchronize()
        t1 = time.perf_counter()
    return (t1 - t0) / n_iters * 1000.0


def bench_full_codec(model, x, n_warmup=3, n_iters=10):
    """Real entropy coding round trip (CPU-bound arithmetic coding included)."""
    model.eval()
    with torch.no_grad():
        for _ in range(n_warmup):
            out = model.compress(x)
            _ = model.decompress(out["strings"], out["shape"])
        torch.cuda.synchronize()

        t0 = time.perf_counter()
        for _ in range(n_iters):
            out = model.compress(x)
        torch.cuda.synchronize()
        t_compress = (time.perf_counter() - t0) / n_iters * 1000.0

        t0 = time.perf_counter()
        for _ in range(n_iters):
            _ = model.decompress(out["strings"], out["shape"])
        torch.cuda.synchronize()
        t_decompress = (time.perf_counter() - t0) / n_iters * 1000.0

    bpp = sum(len(s[0]) for s in out["strings"][:1]) # y strings, list-of-list
    total_bytes = sum(len(s) for slist in out["strings"] for s in slist)
    return t_compress, t_decompress, total_bytes


def run_for_resolution(h, w, quality):
    print(f"\n=== Resolution {w}x{h}, quality={quality} ===")
    x = torch.rand(1, 3, h, w, device=DEVICE)
    xp = pad(x)
    print(f"padded to {xp.shape[3]}x{xp.shape[2]}")

    baseline = mbt2018_mean(quality=quality, pretrained=True).to(DEVICE)
    baseline.update(force=True)
    n_params_base = count_params(baseline)

    thin = ThinMeanScaleHyperprior(N=64, M=96).to(DEVICE)
    thin.update(force=True)
    n_params_thin = count_params(thin)

    print(f"baseline params: {n_params_base/1e6:.2f}M  |  thin params: {n_params_thin/1e6:.2f}M "
          f"({n_params_base/n_params_thin:.1f}x fewer)")

    # forward-only (no entropy coding) timings
    base_fwd = bench_forward(baseline, xp)
    thin_fwd = bench_forward(thin, xp)
    print(f"[forward g_a+g_s, no entropy coding]  baseline: {base_fwd:.2f} ms   thin: {thin_fwd:.2f} ms   "
          f"speedup: {base_fwd/thin_fwd:.2f}x")

    base_ga = bench_ga(baseline, xp)
    thin_ga = bench_ga(thin, xp)
    print(f"[g_a only]  baseline: {base_ga:.2f} ms   thin: {thin_ga:.2f} ms")

    y_base = baseline.g_a(xp)
    y_thin = thin.g_a(xp)
    base_gs = bench_gs(baseline, y_base)
    thin_gs = bench_gs(thin, y_thin)
    print(f"[g_s only]  baseline: {base_gs:.2f} ms   thin: {thin_gs:.2f} ms")

    # full real codec round trip incl. actual arithmetic coding (CPU-bound part included)
    base_c, base_d, base_bytes = bench_full_codec(baseline, xp)
    thin_c, thin_d, thin_bytes = bench_full_codec(thin, xp)
    print(f"[full compress()] baseline: {base_c:.2f} ms   thin: {thin_c:.2f} ms   speedup: {base_c/thin_c:.2f}x")
    print(f"[full decompress()] baseline: {base_d:.2f} ms   thin: {thin_d:.2f} ms   speedup: {base_d/thin_d:.2f}x")
    print(f"[full round trip] baseline: {base_c+base_d:.2f} ms   thin(untrained,rate meaningless): {thin_c+thin_d:.2f} ms")
    print(f"bytes (informational, thin is untrained so its rate is not meaningful): "
          f"baseline={base_bytes}  thin={thin_bytes}")


if __name__ == "__main__":
    torch.set_num_threads(24)
    for (h, w) in [(512, 512), (1080, 1920)]:
        run_for_resolution(h, w, quality=3)