{"id":40529,"date":"2025-05-29T09:38:11","date_gmt":"2025-05-29T01:38:11","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/40529.html"},"modified":"2025-05-29T09:38:11","modified_gmt":"2025-05-29T01:38:11","slug":"day-35","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/40529.html","title":{"rendered":"Day 35"},"content":{"rendered":"

\u6a21\u578b\u53ef\u89c6\u5316\u4e0e\u63a8\u7406 \u77e5\u8bc6\u70b9\u56de\u987e：<\/p>\n

\u4e09\u79cd\u4e0d\u540c\u7684\u6a21\u578b\u53ef\u89c6\u5316\u65b9\u6cd5：\u63a8\u8350torchinfo\u6253\u5370summary+\u6743\u91cd\u5206\u5e03\u53ef\u89c6\u5316 \u8fdb\u5ea6\u6761\u529f\u80fd：\u624b\u52a8\u548c\u81ea\u52a8\u5199\u6cd5，\u8ba9\u6253\u5370\u7ed3\u679c\u66f4\u52a0\u7f8e\u89c2 \u63a8\u7406\u7684\u5199\u6cd5：\u8bc4\u4f30\u6a21\u5f0f \u6a21\u578b\u7ed3\u6784\u53ef\u89c6\u5316 \u7406\u89e3\u4e00\u4e2a\u6df1\u5ea6\u5b66\u4e60\u7f51\u7edc\u6700\u91cd\u8981\u76842\u70b9：<\/p>\n

1. \u4e86\u89e3\u635f\u5931\u5982\u4f55\u5b9a\u4e49\u7684，\u77e5\u9053\u635f\u5931\u4ece\u4f55\u800c\u6765—-\u628a\u62bd\u8c61\u7684\u4efb\u52a1\u901a\u8fc7\u635f\u5931\u51fd\u6570\u91cf\u5316\u51fa\u6765<\/p>\n

2. \u4e86\u89e3\u53c2\u6570\u603b\u91cf，\u5373\u77e5\u9053\u6bcf\u4e00\u5c42\u7684\u8bbe\u8ba1—\u5c42\u8bbe\u8ba1\u51b3\u5b9a\u53c2\u6570\u603b\u91cf<\/p>\n

\u4e3a\u4e86\u4e86\u89e3\u53c2\u6570\u603b\u91cf，\u6211\u4eec\u9700\u8981\u77e5\u9053\u5c42\u8bbe\u8ba1，\u4ee5\u53ca\u6bcf\u4e00\u5c42\u53c2\u6570\u7684\u6570\u91cf\u3002\u4e0b\u9762\u4ecb\u7ecd1\u51e0\u4e2a\u5c42\u53ef\u89c6\u5316\u5de5\u5177：<\/p>\n

1. nn.model\u81ea\u5e26\u7684\u65b9\u6cd5<\/p>\n

# nn.Module \u7684\u5185\u7f6e\u529f\u80fd，\u76f4\u63a5\u8f93\u51fa\u6a21\u578b\u7ed3\u6784
\nprint(model) <\/p>\n

\u8fd9\u662f\u6700\u57fa\u7840\u3001\u6700\u7b80\u5355\u7684\u65b9\u6cd5，\u4f1a\u76f4\u63a5\u6253\u5370\u6a21\u578b\u5bf9\u8c61，\u5b83\u4f1a\u8f93\u51fa\u6a21\u578b\u7684\u7ed3\u6784，\u663e\u793a\u6a21\u578b\u4e2d\u5404\u4e2a\u5c42\u7684\u540d\u79f0\u548c\u53c2\u6570\u4fe1\u606f<\/p>\n

# nn.Module \u7684\u5185\u7f6e\u529f\u80fd，\u8fd4\u56de\u6a21\u578b\u7684\u53ef\u8bad\u7ec3\u53c2\u6570\u8fed\u4ee3\u5668
\nfor name, param in model.named_parameters():
\n print(f"Parameter name: {name}, Shape: {param.shape}") <\/p>\n

\u53ef\u4ee5\u5c06\u6a21\u578b\u4e2d\u5e26\u6709weight\u7684\u53c2\u6570（\u5373\u6743\u91cd）\u63d0\u53d6\u51fa\u6765，\u5e76\u8f6c\u4e3a numpy \u6570\u7ec4\u5f62\u5f0f，\u5bf9\u5176\u8ba1\u7b97\u7edf\u8ba1\u5206\u5e03，\u5e76\u4e14\u7ed8\u5236\u53ef\u89c6\u5316\u56fe\u8868<\/p>\n

# \u63d0\u53d6\u6743\u91cd\u6570\u636e
\nimport numpy as np
\nweight_data = {}
\nfor name, param in model.named_parameters():
\n if 'weight' in name:
\n weight_data[name] = param.detach().cpu().numpy()<\/p>\n

# \u53ef\u89c6\u5316\u6743\u91cd\u5206\u5e03
\nfig, axes = plt.subplots(1, len(weight_data), figsize=(15, 5))
\nfig.suptitle('Weight Distribution of Layers')<\/p>\n

for i, (name, weights) in enumerate(weight_data.items()):
\n # \u5c55\u5e73\u6743\u91cd\u5f20\u91cf\u4e3a\u4e00\u7ef4\u6570\u7ec4
\n weights_flat = weights.flatten()<\/p>\n

# \u7ed8\u5236\u76f4\u65b9\u56fe
\n axes[i].hist(weights_flat, bins=50, alpha=0.7)
\n axes[i].set_title(name)
\n axes[i].set_xlabel('Weight Value')
\n axes[i].set_ylabel('Frequency')
\n axes[i].grid(True, linestyle='–', alpha=0.7)<\/p>\n

plt.tight_layout()
\nplt.subplots_adjust(top=0.85)
\nplt.show()<\/p>\n

# \u8ba1\u7b97\u5e76\u6253\u5370\u6bcf\u5c42\u6743\u91cd\u7684\u7edf\u8ba1\u4fe1\u606f
\nprint("\\\\n=== \u6743\u91cd\u7edf\u8ba1\u4fe1\u606f ===")
\nfor name, weights in weight_data.items():
\n mean = np.mean(weights)
\n std = np.std(weights)
\n min_val = np.min(weights)
\n max_val = np.max(weights)
\n print(f"{name}:")
\n print(f" \u5747\u503c: {mean:.6f}")
\n print(f" \u6807\u51c6\u5dee: {std:.6f}")
\n print(f" \u6700\u5c0f\u503c: {min_val:.6f}")
\n print(f" \u6700\u5927\u503c: {max_val:.6f}")
\n print("-" * 30) <\/p>\n

\u5bf9\u6bd4 fc1.weight \u548c fc2.weight \u7684\u7edf\u8ba1\u4fe1\u606f ，\u53ef\u4ee5\u53d1\u73b0\u5b83\u4eec\u7684\u5747\u503c\u3001\u6807\u51c6\u5dee\u3001\u6700\u503c\u7b49\u5b58\u5728\u5dee\u5f02\u3002\u8fd9\u53cd\u6620\u4e86\u4e0d\u540c\u5c42\u5728\u6a21\u578b\u4e2d\u7684\u4f5c\u7528\u4e0d\u540c\u3002\u6743\u91cd\u7edf\u8ba1\u4fe1\u606f\u53ef\u4ee5\u4e3a\u8d85\u53c2\u6570\u8c03\u6574\u63d0\u4f9b\u53c2\u8003\u3002<\/p>\n

2.torchsummary\u5e93\u7684summary\u65b9\u6cd5<\/h6>\n

# pip install torchsummary -i https:\/\/pypi.tuna.tsinghua.edu.cn\/simple
\nfrom torchsummary import summary
\n# \u6253\u5370\u6a21\u578b\u6458\u8981，\u53ef\u4ee5\u653e\u7f6e\u5728\u6a21\u578b\u5b9a\u4e49\u540e\u9762
\nsummary(model, input_size=(4,)) <\/p>\n

\u00a0\u8be5\u65b9\u6cd5\u4e0d\u663e\u793a\u8f93\u5165\u5c42\u7684\u5c3a\u5bf8，\u56e0\u4e3a\u8f93\u5165\u7684\u795e\u7ecf\u7f51\u662f\u81ea\u5df1\u8bbe\u7f6e\u7684，\u6240\u4ee5\u4e0d\u9700\u8981\u663e\u793a\u8f93\u5165\u5c42\u7684\u5c3a\u5bf8\u3002\u4f46\u662f\u5728\u4f7f\u7528\u8be5\u65b9\u6cd5\u65f6，input_size=(4,) \u53c2\u6570\u662f\u5fc5\u9700\u7684，\u56e0\u4e3a PyTorch \u9700\u8981\u77e5\u9053\u8f93\u5165\u6570\u636e\u7684\u5f62\u72b6\u624d\u80fd\u63a8\u65ad\u6a21\u578b\u5404\u5c42\u7684\u8f93\u51fa\u5f62\u72b6\u548c\u53c2\u6570\u6570\u91cf\u3002<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u8fd9\u662f\u56e0\u4e3aPyTorch \u7684\u6a21\u578b\u5728\u5b9a\u4e49\u65f6\u662f\u52a8\u6001\u7684，\u5b83\u4e0d\u4f1a\u9884\u5148\u77e5\u9053\u8f93\u5165\u6570\u636e\u7684\u5177\u4f53\u5f62\u72b6\u3002nn.Linear(4, 10) \u53ea\u5b9a\u4e49\u4e86 \u201c\u8f93\u5165\u7ef4\u5ea6\u662f 4，\u8f93\u51fa\u7ef4\u5ea6\u662f 10\u201d，\u4f46\u4e0d\u77e5\u9053\u8f93\u5165\u7684\u6279\u91cf\u5927\u5c0f\u548c\u5176\u4ed6\u7ef4\u5ea6，\u6bd4\u5982\u5377\u79ef\u5c42\u9700\u8981\u77e5\u9053\u8f93\u5165\u7684\u901a\u9053\u6570\u3001\u9ad8\u5ea6\u3001\u5bbd\u5ea6\u7b49\u4fe1\u606f\u3002—-\u5e76\u975e\u6240\u6709\u8f93\u5165\u6570\u636e\u90fd\u662f\u7ed3\u6784\u5316\u6570\u636e<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u56e0\u6b64，\u8981\u751f\u6210\u6a21\u578b\u6458\u8981（\u5982\u6bcf\u5c42\u7684\u8f93\u51fa\u5f62\u72b6\u3001\u53c2\u6570\u6570\u91cf），\u5fc5\u987b\u63d0\u4f9b\u4e00\u4e2a\u793a\u4f8b\u8f93\u5165\u5f62\u72b6，\u8ba9 PyTorch \u201c\u8fd0\u884c\u201d \u4e00\u6b21\u6a21\u578b，\u4ece\u800c\u63a8\u65ad\u51fa\u5404\u5c42\u7684\u4fe1\u606f\u3002<\/p>\n

summary \u51fd\u6570\u7684\u6838\u5fc3\u903b\u8f91\u662f：<\/p>\n

1. \u521b\u5efa\u4e00\u4e2a\u4e0e input_size \u5f62\u72b6\u5339\u914d\u7684\u865a\u62df\u8f93\u5165\u5f20\u91cf（\u901a\u5e38\u586b\u5145\u96f6）<\/p>\n

2. \u5c06\u865a\u62df\u8f93\u5165\u4f20\u9012\u7ed9\u6a21\u578b，\u6267\u884c\u4e00\u6b21\u524d\u5411\u4f20\u64ad（\u4f46\u4e0d\u8ba1\u7b97\u68af\u5ea6）<\/p>\n

3. \u8bb0\u5f55\u6bcf\u4e00\u5c42\u7684\u8f93\u5165\u548c\u8f93\u51fa\u5f62\u72b6，\u4ee5\u53ca\u53c2\u6570\u6570\u91cf<\/p>\n

4. \u751f\u6210\u53ef\u8bfb\u7684\u6458\u8981\u62a5\u544a<\/p>\n

\u6784\u5efa\u795e\u7ecf\u7f51\u7edc\u7684\u65f6\u5019<\/p>\n

1. \u8f93\u5165\u5c42\u4e0d\u9700\u8981\u5199：x\u591a\u5c11\u4e2a\u7279\u5f81 \u8f93\u5165\u5c42\u5c31\u6709\u591a\u5c11\u795e\u7ecf\u5143<\/p>\n

2. \u9690\u85cf\u5c42\u9700\u8981\u5199，\u4ece\u7b2c\u4e00\u4e2a\u9690\u85cf\u5c42\u53ef\u4ee5\u770b\u51fa\u7279\u5f81\u7684\u4e2a\u6570<\/p>\n

3. \u8f93\u51fa\u5c42\u7684\u795e\u7ecf\u5143\u548c\u4efb\u52a1\u6709\u5173，\u6bd4\u5982\u5206\u7c7b\u4efb\u52a1，\u8f93\u51fa\u5c42\u67093\u4e2a\u795e\u7ecf\u5143，\u4e00\u4e2a\u5bf9\u5e94\u6bcf\u4e2a\u7c7b\u522b<\/p>\n

\u53ef\u5b66\u4e60\u53c2\u6570\u8ba1\u7b97<\/p>\n

1. Linear-1\u5bf9\u5e94self.fc1 = nn.Linear(4, 10)，\u8868\u660e\u524d\u4e00\u5c42\u67094\u4e2a\u795e\u7ecf\u5143，\u8fd9\u4e00\u5c42\u670910\u4e2a\u795e\u7ecf\u5143，\u6bcf2\u4e2a\u795e\u7ecf\u5143\u4e4b\u95f4\u9760\u7740\u7ebf\u76f8\u8fde，\u6240\u6709\u67094*10\u4e2a\u6743\u91cd\u53c2\u6570+10\u4e2a\u504f\u7f6e\u53c2\u6570=50\u4e2a\u53c2\u6570<\/p>\n

2. relu\u5c42\u4e0d\u6d89\u53ca\u53ef\u5b66\u4e60\u53c2\u6570，\u53ef\u4ee5\u628a\u5b83\u548c\u524d\u4e00\u4e2a\u7ebf\u6027\u5c42\u770b\u6210\u4e00\u5c42，\u56fe\u4e0a\u4e5f\u662f\u8fd9\u4e2a\u542b\u4e49<\/p>\n

3. Linear-3\u5c42\u5bf9\u5e94\u4ee3\u7801 self.fc2 = nn.Linear(10,3),10*3\u4e2a\u6743\u91cd\u53c2\u6570+3\u4e2a\u504f\u7f6e=33\u4e2a\u53c2\u6570<\/p>\n

\u603b\u53c2\u657083\u4e2a，\u5360\u7528\u5185\u5b58\u51e0\u4e4e\u4e3a0<\/p>\n

1.3 torchinfo\u5e93\u7684summary\u65b9\u6cd5 \u00a0torchinfo \u662f\u63d0\u4f9b\u6bd4 torchsummary \u66f4\u8be6\u7ec6\u7684\u6a21\u578b\u6458\u8981\u4fe1\u606f，\u5305\u62ec\u6bcf\u5c42\u7684\u8f93\u5165\u8f93\u51fa\u5f62\u72b6\u3001\u53c2\u6570\u6570\u91cf\u3001\u8ba1\u7b97\u91cf\u7b49\u3002<\/p>\n

# pip install torchinfo -i https:\/\/pypi.tuna.tsinghua.edu.cn\/simple
\nfrom torchinfo import summary
\nsummary(model, input_size=(4, )) <\/p>\n

\u8fdb\u5ea6\u6761\u529f\u80fd tqdm\u8fd9\u4e2a\u5e93\u975e\u5e38\u9002\u5408\u7528\u5728\u5faa\u73af\u4e2d\u89c2\u5bdf\u8fdb\u5ea6\u3002\u5c24\u5176\u5728\u6df1\u5ea6\u5b66\u4e60\u8fd9\u79cd\u8bad\u7ec3\u662f\u5faa\u73af\u7684\u573a\u666f\u4e2d\u3002\u4ed6\u6700\u6838\u5fc3\u7684\u903b\u8f91\u5982\u4e0b<\/p>\n

1. \u521b\u5efa\u4e00\u4e2a\u8fdb\u5ea6\u6761\u5bf9\u8c61，\u5e76\u4f20\u5165\u603b\u8fed\u4ee3\u6b21\u6570\u3002\u4e00\u822c\u7528with\u8bed\u53e5\u521b\u5efa\u5bf9\u8c61，\u8fd9\u6837\u5bf9\u8c61\u4f1a\u5728with\u8bed\u53e5\u7ed3\u675f\u540e\u81ea\u52a8\u9500\u6bc1，\u4fdd\u8bc1\u8d44\u6e90\u91ca\u653e\u3002with\u662f\u5e38\u89c1\u7684\u4e0a\u4e0b\u6587\u7ba1\u7406\u5668，\u8fd9\u6837\u7684\u4f7f\u7528\u65b9\u5f0f\u8fd8\u6709\u7528with\u6253\u5f00\u6587\u4ef6，\u7ed3\u675f\u540e\u4f1a\u81ea\u52a8\u5173\u95ed\u6587\u4ef6\u3002<\/p>\n

2. \u66f4\u65b0\u8fdb\u5ea6\u6761，\u901a\u8fc7pbar.update(n)\u6307\u5b9a\u6bcf\u6b21\u524d\u8fdb\u7684\u6b65\u6570n（\u9002\u7528\u4e8e\u975e\u56fa\u5b9a\u6b65\u957f\u7684\u5faa\u73af）\u3002<\/p>\n

1.\u624b\u52a8\u66f4\u65b0<\/p>\n

from tqdm import tqdm # \u5148\u5bfc\u5165tqdm\u5e93
\nimport time # \u7528\u4e8e\u6a21\u62df\u8017\u65f6\u64cd\u4f5c<\/p>\n

# \u521b\u5efa\u4e00\u4e2a\u603b\u6b65\u6570\u4e3a10\u7684\u8fdb\u5ea6\u6761
\nwith tqdm(total=10) as pbar: # pbar\u662f\u8fdb\u5ea6\u6761\u5bf9\u8c61\u7684\u53d8\u91cf\u540d
\n # pbar \u662f progress bar（\u8fdb\u5ea6\u6761）\u7684\u7f29\u5199，\u7ea6\u5b9a\u4fd7\u6210\u7684\u547d\u540d\u4e60\u60ef\u3002
\n for i in range(10): # \u5faa\u73af10\u6b21（\u5bf9\u5e94\u8fdb\u5ea6\u6761\u768410\u6b65）
\n time.sleep(0.5) # \u6a21\u62df\u6bcf\u6b21\u5faa\u73af\u8017\u65f60.5\u79d2
\n pbar.update(1) # \u6bcf\u6b21\u5faa\u73af\u540e，\u8fdb\u5ea6\u6761\u524d\u8fdb1\u6b65
\nfrom tqdm import tqdm
\nimport time<\/p>\n

# \u521b\u5efa\u8fdb\u5ea6\u6761\u65f6\u6dfb\u52a0\u63cf\u8ff0（desc）\u548c\u5355\u4f4d（unit）
\nwith tqdm(total=5, desc="\u4e0b\u8f7d\u6587\u4ef6", unit="\u4e2a") as pbar:
\n # \u8fdb\u5ea6\u6761\u8fd9\u4e2a\u5bf9\u8c61，\u53ef\u4ee5\u8bbe\u7f6e\u63cf\u8ff0\u548c\u5355\u4f4d
\n # desc\u662f\u63cf\u8ff0，\u5728\u5de6\u4fa7\u663e\u793a
\n # unit\u662f\u5355\u4f4d，\u5728\u8fdb\u5ea6\u6761\u53f3\u4fa7\u663e\u793a
\n for i in range(5):
\n time.sleep(1)
\n pbar.update(1) # \u6bcf\u6b21\u5faa\u73af\u8fdb\u5ea6+1<\/p>\n

unit \u53c2\u6570\u7684\u6838\u5fc3\u4f5c\u7528\u662f\u660e\u786e\u8fdb\u5ea6\u6761\u4e2d\u6bcf\u4e2a\u8fdb\u5ea6\u5355\u4f4d\u7684\u542b\u4e49，\u4f7f\u53ef\u89c6\u5316\u4fe1\u606f\u66f4\u5177\u53ef\u8bfb\u6027\u3002\u5728\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u4e2d，\u5e38\u7528\u7684\u5355\u4f4d\u5305\u62ec：<\/p>\n