FlanChanXwO commited on
Commit
ea12560
·
verified ·
1 Parent(s): 4f0c3ac

Upload scripts/train_fixed.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. scripts/train_fixed.py +166 -3
scripts/train_fixed.py CHANGED
@@ -1,3 +1,166 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:a12c6edcd23ae7951bc1e5b51e4519ccfdcca17917eb3f658b05957cd451e3cf
3
- size 6418
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """训练南+定长4位数字验证码分类器,导出 ONNX。
3
+ 模型: CNN 骨干 + 4 个位置头,每头 10 类(0-9)。
4
+ 输入 (1,1,64,160) 灰度/255,输出 (B,4,10) logits。
5
+ 用法: train_fixed.py <labels.json> <图片目录> <输出.onnx> [val_frac] [epochs]
6
+ """
7
+ import json
8
+ import os
9
+ import random
10
+ import sys
11
+
12
+ import numpy as np
13
+ import torch
14
+ import torch.nn as nn
15
+ from PIL import Image
16
+
17
+ W, H = 160, 64
18
+ NDIGITS = 4
19
+ RGB = "rgb" in sys.argv
20
+
21
+
22
+ class Dataset:
23
+ def __init__(self, items, aug=False):
24
+ self.items = items
25
+ self.aug = aug
26
+
27
+ def __len__(self):
28
+ return len(self.items)
29
+
30
+ def __getitem__(self, i):
31
+ path, lbl = self.items[i]
32
+ mode = "RGB" if RGB else "L"
33
+ im = Image.open(path).convert(mode).resize((W, H), Image.BILINEAR)
34
+ if self.aug:
35
+ im = augment(im)
36
+ a = np.asarray(im, dtype=np.float32) / 255.0
37
+ if RGB:
38
+ x = torch.from_numpy(a).permute(2, 0, 1) # (C,H,W)
39
+ else:
40
+ x = torch.from_numpy(a).unsqueeze(0)
41
+ t = torch.tensor([int(c) for c in lbl], dtype=torch.long)
42
+ return x, t
43
+
44
+
45
+ def augment(im):
46
+ import random as R
47
+ from PIL import Image
48
+ if R.random() < 0.8:
49
+ im = im.rotate(R.uniform(-6, 6), resample=Image.BILINEAR, fillcolor=255)
50
+ scale = R.uniform(0.92, 1.08)
51
+ tx, ty = R.uniform(-3, 3), R.uniform(-2, 2)
52
+ w, h = im.size
53
+ a, b, c, d, e, f = (scale, 0, -w * scale / 2 + w / 2 + tx,
54
+ 0, scale, -h * scale / 2 + h / 2 + ty)
55
+ im = im.transform((w, h), Image.AFFINE, (a, b, c, d, e, f),
56
+ resample=Image.BILINEAR, fillcolor=255)
57
+ # 可选轻微噪声(默认关,避免盖掉验证码原有干扰线)
58
+ if os.environ.get("AUG_NOISE", "0") == "1":
59
+ arr = np.asarray(im, dtype=np.float32)
60
+ arr += np.random.normal(0, 5, arr.shape)
61
+ return Image.fromarray(np.clip(arr, 0, 255).astype(np.uint8), im.mode)
62
+ return im
63
+
64
+
65
+ class FixedNet(nn.Module):
66
+ def __init__(self):
67
+ super().__init__()
68
+ in_ch = 3 if RGB else 1
69
+ self.features = nn.Sequential(
70
+ nn.Conv2d(in_ch, 32, 3, padding=1), nn.BatchNorm2d(32), nn.ReLU(), nn.MaxPool2d(2),
71
+ nn.Conv2d(32, 64, 3, padding=1), nn.BatchNorm2d(64), nn.ReLU(), nn.MaxPool2d(2),
72
+ nn.Conv2d(64, 128, 3, padding=1), nn.BatchNorm2d(128), nn.ReLU(), nn.MaxPool2d(2),
73
+ nn.Conv2d(128, 256, 3, padding=1), nn.BatchNorm2d(256), nn.ReLU(),
74
+ nn.Conv2d(256, 256, 3, padding=1), nn.BatchNorm2d(256), nn.ReLU(),
75
+ )
76
+ # 保留空间位置:20列特征按4个数字位置各池化
77
+ self.heads = nn.ModuleList([nn.Linear(256, 10) for _ in range(NDIGITS)])
78
+
79
+ def forward(self, x):
80
+ x = self.features(x) # (B,256,8,20)
81
+ x = x.mean(dim=2) # 高度池化 (B,256,20)
82
+ # 20列 -> 4个位置(每位置5列),保持左右顺序
83
+ B, C, T = x.shape
84
+ per = T // NDIGITS
85
+ x = x.view(B, C, NDIGITS, per).mean(dim=3) # (B,256,4)
86
+ x = x.permute(0, 2, 1) # (B,4,256)
87
+ return torch.stack([self.heads[p](x[:, p]) for p in range(NDIGITS)], dim=1) # (B,4,10)
88
+
89
+
90
+ def collate(batch):
91
+ xs = torch.stack([b[0] for b in batch])
92
+ ts = torch.stack([b[1] for b in batch])
93
+ return xs, ts
94
+
95
+
96
+ def main():
97
+ labels = json.load(open(sys.argv[1]))
98
+ imgdir = sys.argv[2]
99
+ out = sys.argv[3] if len(sys.argv) > 3 else "/tmp/sp_captest/captcha_fixed.onnx"
100
+ val_frac = float(sys.argv[4]) if len(sys.argv) > 4 else 0.1
101
+ epochs = int(sys.argv[5]) if len(sys.argv) > 5 else 400
102
+ use_aug = sys.argv[6] != "0" if len(sys.argv) > 6 else True
103
+
104
+ all_items = [(f"{imgdir}/{fn}", v["label"]) for fn, v in labels.items()
105
+ if v.get("label") and len(v["label"]) == NDIGITS]
106
+ random.Random(42).shuffle(all_items)
107
+ n_val = int(len(all_items) * val_frac)
108
+ train_ds = Dataset(all_items[n_val:], aug=use_aug)
109
+ val_ds = Dataset(all_items[:n_val], aug=False)
110
+ tr = torch.utils.data.DataLoader(train_ds, batch_size=16, shuffle=True, collate_fn=collate)
111
+ va = torch.utils.data.DataLoader(val_ds, batch_size=16, shuffle=False, collate_fn=collate)
112
+ print(f"train={len(train_ds)} val={len(val_ds)} total={len(all_items)}")
113
+
114
+ device = "cpu"
115
+ model = FixedNet().to(device)
116
+ # 断点续训: 从已有 checkpoint 加载权重
117
+ resume_pt = out + ".pt"
118
+ start_ep = 0
119
+ if os.path.exists(resume_pt):
120
+ model.load_state_dict(torch.load(resume_pt))
121
+ start_ep = int(sys.argv[7]) if len(sys.argv) > 7 else 0
122
+ print(f"从 checkpoint 续训, 起始 epoch={start_ep}")
123
+ opt = torch.optim.Adam(model.parameters(), lr=1e-3)
124
+ sched = torch.optim.lr_scheduler.ReduceLROnPlateau(opt, factor=0.5, patience=10)
125
+ crit = nn.CrossEntropyLoss()
126
+
127
+ best = 0.0
128
+ for ep in range(start_ep, start_ep + epochs):
129
+ model.train()
130
+ tot = 0.0
131
+ for x, t in tr:
132
+ x, t = x.to(device), t.to(device)
133
+ logits = model(x) # (B,4,10)
134
+ loss = sum(crit(logits[:, p], t[:, p]) for p in range(NDIGITS))
135
+ opt.zero_grad(); loss.backward(); opt.step()
136
+ tot += loss.item() * x.size(0)
137
+ model.eval()
138
+ acc = n = 0
139
+ with torch.no_grad():
140
+ for x, t in va:
141
+ p = model(x).argmax(-1) # (B,4)
142
+ acc += (p == t).all(dim=1).sum().item()
143
+ n += x.size(0)
144
+ acc /= max(1, n)
145
+ sched.step(acc)
146
+ if acc > best:
147
+ best = acc
148
+ torch.save(model.state_dict(), out + ".pt")
149
+ if ep % 10 == 0 or ep == epochs - 1:
150
+ print(f"epoch {ep} loss={tot/len(train_ds):.3f} val_acc={acc:.2%} best={best:.2%}", flush=True)
151
+ if not os.path.exists(out + ".pt"): # 从未有更好验证精度时,保存最后权重
152
+ torch.save(model.state_dict(), out + ".pt")
153
+ print(f"DONE best_val_acc={best:.2%}")
154
+
155
+ model.load_state_dict(torch.load(out + ".pt"))
156
+ model.eval()
157
+ dummy = torch.randn(1, 3 if RGB else 1, H, W)
158
+ torch.onnx.export(model, dummy, out,
159
+ input_names=["input"], output_names=["logits"],
160
+ dynamic_axes={"input": {0: "batch"}, "logits": {0: "batch"}},
161
+ opset_version=13, external_data=False)
162
+ print(f"ONNX exported -> {out}")
163
+
164
+
165
+ if __name__ == "__main__":
166
+ main()