Merge branch 'datawhalechina:main' into main

Author: ironartisan

docs/ch03/ch03.md

@@ -193,7 +193,7 @@ $$
   $$
   其中 $h_{ii}$ 是对应的Hessian矩阵的对角元素，定义为：
   $$
-  h_{ii} = \frac{\partial^2 L}{\partial w_i \partial w_j}
+   h_{i i}=\frac{\partial^2 L}{\partial w_i^2}
   $$
 
 - **剪枝原则**：

docs/ch04/ch04.md

@@ -183,7 +183,7 @@ $$
 Z = round(q_{max}-\frac{r_{max}}{S})
 $$ 
 
-  其中，$r_{max}$ 和 $r_{max}$分别表示浮点数中的最小值和最大值,$q_{max}$ 和 $q_{min}$分别表示定点数中的最小值和最大值。
+  其中，$r_{min}$ 和 $r_{max}$分别表示浮点数中的最小值和最大值,$q_{min}$ 和 $q_{max}$分别表示定点数中的最小值和最大值。
 
 ![图4-10 线性量化](images/linear.png)
 
@@ -547,4 +547,4 @@ $$
 
 - [Model Quantization 1: Basic Concepts](https://medium.com/@florian_algo/model-quantization-1-basic-concepts-860547ec6aa9)
 - [Model Quantization 3: Timing and Granularity](https://blog.gopenai.com/model-quantization-3-timing-and-granularity-a0978c6e58d4)
-- [A Visual Guide to Quantization](https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-quantization#%C2%A7symmetric-quantization)
+- [A Visual Guide to Quantization](https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-quantization#%C2%A7symmetric-quantization)

docs/ch07/ch07.md

@@ -1,7 +1,7 @@
 # 第7章 项目实践
 
   在前面的章节里，我们分别学习了剪枝、量化、神经网络架构搜索及知识蒸馏等模型压缩技术，那么你能融合两种以上的技术对模型进行压缩吗？
-> 任务：根据前面已学内容，选择一种实际应用场景，使用两种及以上技术对模型进行压缩并对比效果~
+> 任务：基于前面已学知识，选择一种实际应用场景，不限框架和方法，使用两种及以上技术对模型进行压缩并对比前后效果~
 
 ## 7.1 总结
 

docs/notebook/INSTALL.md

@@ -5,6 +5,6 @@
 ```
 conda create -n compression python=3.10
 conda activate compression
-pip install - r requirements.txt
+pip install -r requirements.txt
 ```
 

docs/notebook/ch02/1.mnist_classify.ipynb

@@ -511,7 +511,7 @@
     "        transforms.Grayscale(num_output_channels=1),  # Convert to grayscale if needed\n",
     "        transforms.Resize((28, 28)),  # Resize to match MNIST dimensions\n",
     "        transforms.ToTensor(),  # Convert image to tensor\n",
-    "        transforms.Normalize((0.5,), (0.5,))  # Normalize as per model's training\n",
+    "        transforms.Normalize((0.1307,), (0.3081,))  # Normalize as per model's training\n",
     "    ])\n",
     "    image = Image.open(image_path)\n",
     "    image = transform(image).unsqueeze(0)  # Add batch dimension\n",

docs/notebook/ch03/2.pruning_criteria.ipynb

@@ -389,11 +389,11 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 49,
+   "execution_count": null,
    "metadata": {},
    "outputs": [],
    "source": [
-    "# 定义一个LeNet网络\n",
+    "# 为避免前面的操作影响后续结果，重新定义一个LeNet网络，和前面一致\n",
     "class LeNet(nn.Module):\n",
     "    def __init__(self, num_classes=10):\n",
     "        super(LeNet, self).__init__()\n",
@@ -451,13 +451,20 @@
     "model.load_state_dict(checkpoint)"
    ]
   },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "基于梯度幅度的修剪标准"
+   ]
+  },
   {
    "cell_type": "code",
-   "execution_count": 52,
+   "execution_count": null,
    "metadata": {},
    "outputs": [],
    "source": [
-    "# 基于梯度幅度的修剪标准\n",
+    "# 修剪整个模型的权重，传入整个模型\n",
     "def gradient_magnitude_pruning(model, percentile):\n",
     "    for name, param in model.named_parameters():\n",
     "        if 'weight' in name:\n",
@@ -467,10 +474,11 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 53,
+   "execution_count": null,
    "metadata": {},
    "outputs": [],
    "source": [
+    "# 修剪局部模型权重，传入某一层的权重\n",
     "@torch.no_grad()\n",
     "def gradient_magnitude_pruning(weight, gradient, percentile=0.5):\n",
     "    num_elements = weight.numel()\n",
@@ -491,7 +499,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 54,
+   "execution_count": null,
    "metadata": {},
    "outputs": [
     {
@@ -514,7 +522,7 @@
     }
    ],
    "source": [
-    "# 使用示例\n",
+    "# 使用示例，这里以fc2层的权重为例\n",
     "percentile = 0.5\n",
     "gradient_magnitude_pruning(model.fc2.weight, gradients['fc2.weight'], percentile)"
    ]

docs/notebook/requirements.txt

@@ -1,4 +1,4 @@
-numpy==2.1.1
+numpy==1.24.3
 matplotlib==3.9.2
 tqdm==4.66.5
 jupyter==1.1.1
@@ -8,4 +8,4 @@ torchprofile==0.0.4
 torchsummary==1.5.1
 fast-pytorch-kmeans
 scipy
-datasets
+datasets