refactor: merge multiple project into one and create new project

dl-exp/exp2/datasets/.gitignore

@@ -0,0 +1,2 @@
+# Ignore datasets file
+mnist.npz

dl-exp/exp2/models/.gitignore

@@ -0,0 +1,3 @@
+# Ignore every saved model files
+*.pth
+*.ckpt

dl-exp/exp2/modified/dataset.py

@@ -0,0 +1,80 @@
+from pathlib import Path
+import numpy
+import torch
+from torch.utils.data import DataLoader, Dataset
+from torchvision.transforms import v2 as tvtrans
+import settings
+
+class MnistDataset(Dataset):
+    """适配PyTorch的自定义Dataset，用于加载MNIST数据。"""
+
+    shape: int
+    transform: tvtrans.Transform
+    images_data: numpy.ndarray
+    labels_data: torch.Tensor
+
+    def __init__(self, images: numpy.ndarray, labels: numpy.ndarray, transform: tvtrans.Transform):
+        images_len: int = images.shape[0]
+        labels_len: int = labels.shape[0]
+        assert (images_len == labels_len)
+        self.shape = images_len
+
+        self.images_data = images
+        self.labels_data = torch.from_numpy(labels)
+        self.transform = transform
+
+    def __getitem__(self, index):
+        return self.transform(self.images_data[index]), self.labels_data[index]
+
+    def __len__(self):
+        return self.shape
+
+
+class MnistDataLoaders:
+    """包含适配PyTorch的训练数据Loader和测试数据Loader的类。"""
+
+    train_loader: DataLoader
+    test_loader: DataLoader
+
+    def __init__(self, batch_size: int):
+        dataset = numpy.load(settings.MNIST_DATASET_PATH)
+
+        # 所有图片均为黑底白字
+        # 6万张训练图片：60000x28x28。标签只有第一维。
+        train_images: numpy.ndarray = dataset['x_train']
+        train_labels: numpy.ndarray = dataset['y_train']
+        # 1万张测试图片：10000x28x28。标签只有第一维。
+        test_images: numpy.ndarray = dataset['x_test']
+        test_labels: numpy.ndarray = dataset['y_test']
+
+        # 定义数据转换器
+        trans = tvtrans.Compose([
+            # 从uint8转换为float32并自动归一化到0-1区间
+            # YYC MARK: 下面这个被标outdated了，换下面两个替代。
+            # tvtrans.ToTensor(),
+            tvtrans.ToImage(),
+            tvtrans.ToDtype(torch.float32, scale=True),
+
+            # 为了符合后面图像的输入颜色通道条件，要在最后挤出一个新的维度
+            # YYC MARK: 上面这两步已经帮我们自动挤出那个灰度通道了。
+            # tvtrans.Lambda(lambda x: x.unsqueeze(-1))
+
+            # 这个特定的标准化参数 (0.1307, 0.3081) 是 MNIST 数据集的标准化参数，这些数值是MNIST训练集的全局均值和标准差。
+            # 这种标准化有助于模型训练时的数值稳定性和收敛速度。
+            # YYC MARK: 但我不想用，反正最后训练的也收敛。
+            # tvtrans.Normalize((0.1307,), (0.3081,)),
+        ])
+
+        # 创建数据集
+        train_dataset = MnistDataset(train_images, train_labels, 
+                                     transform=trans)
+        test_dataset = MnistDataset(test_images, test_labels, 
+                                    transform=trans)
+
+        # 赋值到自身
+        self.train_loader = DataLoader(dataset=train_dataset, 
+                                       batch_size=batch_size, 
+                                       shuffle=False)
+        self.test_loader = DataLoader(dataset=test_dataset, 
+                                      batch_size=batch_size, 
+                                      shuffle=False)

dl-exp/exp2/modified/model.py

@@ -0,0 +1,40 @@
+import torch
+import torch.nn.functional as F
+
+class Cnn(torch.nn.Module):
+    """卷积神经网络模型"""
+
+    def __init__(self):
+        super(Cnn, self).__init__()
+
+        self.conv1 = torch.nn.Conv2d(1, 32, kernel_size=(3, 3))
+        self.pool1 = torch.nn.MaxPool2d(kernel_size=(2, 2))
+        self.conv2 = torch.nn.Conv2d(32, 64, kernel_size=(3, 3))
+        self.pool2 = torch.nn.MaxPool2d(kernel_size=(2, 2))
+        self.conv3 = torch.nn.Conv2d(64, 64, kernel_size=(3, 3))
+        self.flatten = torch.nn.Flatten()
+        # 28x28过第一轮卷积后变为26x26，过第一轮池化后变为13x13
+        # 过第二轮卷积后变为11x11，过第二轮池化后变为5x5
+        # 过第三轮卷积后变为3x3。
+        # 最后一轮卷积核个数为64。
+        self.fc1 = torch.nn.Linear(64 * 3 * 3, 64)
+        self.fc2 = torch.nn.Linear(64, 10)
+
+    def forward(self, x):
+        x = F.relu(self.conv1(x))
+        x = self.pool1(x)
+        x = F.relu(self.conv2(x))
+        x = self.pool2(x)
+        x = F.relu(self.conv3(x))
+        x = self.flatten(x)
+        x = F.relu(self.fc1(x))
+        x = self.fc2(x)
+        # YYC MARK:
+        # 绝对不要在这里用F.softmax(x, dim=1)输出！
+        # 由于这些代码是从tensorflow里转换过来的，
+        # tensorflow的loss function是接受possibility作为交叉熵计算的，
+        # 而pytorch要求接受logits，即模型softmax之前的参数作为交叉熵计算。
+        # 所以这里直接输出模型结果。
+        return x
+
+

dl-exp/exp2/modified/predict.py

@@ -0,0 +1,135 @@
+from pathlib import Path
+import sys
+import numpy
+import torch
+import torch.nn.functional as F
+from PIL import Image, ImageFile
+import matplotlib.pyplot as plt
+from model import Cnn
+import settings
+
+sys.path.append(str(Path(__file__).resolve().parent.parent.parent))
+import gpu_utils
+
+
+class PredictResult:
+    """预测的结果"""
+
+    possibilities: torch.Tensor
+    """每个数字不同的概率"""
+
+    def __init__(self, possibilities: torch.Tensor):
+        """
+        创建预测结果。
+
+        :param possibilities: 传入的tensor表示每个数字不同的概率，是经过softmax后的数值。
+        其shape为二维。dim 0为batch，应当只有一维；dim 1为每个数字对应的概率。
+        """
+        self.possibilities = possibilities
+
+    def chosen_number(self) -> int:
+        """
+        获取最终选定的数字
+        
+        :return: 以当前概率分布，推测的最终数字。
+        """
+        # 输出出来是10个数字各自的可能性，所以要选取最高可能性的那个对比
+        # 在dim=1上找最大的那个，就选那个。dim=0是批次所以不管他。
+        return self.possibilities.argmax(1).item()
+
+    def number_possibilities(self) -> list[float]:
+        """
+        获取每个数字出现的概率
+        
+        :return: 返回一个具有10个元素的列表，列表的每一项表示当前index所代表数字的概率。
+        """
+        return list(self.possibilities[0][i].item() for i in range(10))
+
+class Predictor:
+    device: torch.device
+    model: Cnn
+
+    def __init__(self):
+        self.device = gpu_utils.get_gpu_device()
+        self.model = Cnn().to(self.device)
+
+        # 加载保存好的模型参数
+        self.model.load_state_dict(torch.load(settings.SAVED_MODEL_PATH))
+
+    def __predict_tensor(self, in_data: torch.Tensor) -> PredictResult:
+        """
+        其它预测函数都要使用的预测后端。其它预测函数将数据处理成Tensor，然后传递给此函数进行实际预测。
+        
+        :param in_data: 传入的tensor，该tensor的shape必须是28x28，dtype为float32。
+        :return: 预测结果。
+        """
+        # 上传tensor到GPU
+        in_data = in_data.to(self.device)
+        # 为了满足要求，要在第一维度挤出2下
+        # 一次是灰度通道，一次是批次。
+        # 相当于batch size = 1的计算
+        in_data = in_data.unsqueeze(0).unsqueeze(0)
+        # 开始预测，由于模型输出的是没有softmax的数值，因此最后还需要softmax一下，
+        with torch.no_grad():
+            out_data = self.model(in_data)
+            out_data = F.softmax(out_data, dim=-1)
+            return PredictResult(out_data)
+       
+
+    def predict_sketchpad(self, image: list[list[bool]]) -> PredictResult:
+        """
+        以sketchpad的数据进行预测。
+
+        :param image: 该列表的shape必须为28x28。
+        :return: 预测结果。
+        """
+        input = torch.tensor(image, dtype=torch.float32)
+        assert(input.dim() == 2)
+        assert(input.size(0) == 28)
+        assert(input.size(1) == 28)
+
+        return self.__predict_tensor(input)
+
+    def predict_image(self, image: ImageFile.ImageFile) -> PredictResult:
+        """
+        以Pillow图像的数据进行预测。
+        
+        :param image: Pillow图像数据。该图像必须为28x28大小。
+        :return: 预测结果。
+        """
+        # 确保图像为灰度图像，以及宽高合适
+        grayscale_image = image.convert('L')
+        assert(grayscale_image.width == 28)
+        assert(grayscale_image.height == 28)
+        # 转换为numpy数组。注意这里的numpy数组是只读的，所以要先拷贝一份
+        numpy_data = numpy.reshape(grayscale_image, (28, 28), copy=True)
+        # 转换到Tensor，设置dtype
+        data = torch.from_numpy(numpy_data).float()
+        # 归一化到255，又因为图像输入是白底黑字，需要做转换。
+        data.div_(255.0).sub_(1).mul_(-1)
+
+        return self.__predict_tensor(data)
+
+def main():
+    predictor = Predictor()
+
+    # 遍历测试目录中的所有图片，并处理。
+    test_dir = Path(__file__).resolve().parent.parent / 'test_images'
+    for image_path in test_dir.glob('*.png'):
+        if image_path.is_file():
+            print(f'Predicting {image_path} ...')
+            image = Image.open(image_path)
+            rv = predictor.predict_image(image)
+
+            print(f'Predict digit: {rv.chosen_number()}')
+            plt.figure(f'Image - {image_path}')
+            plt.imshow(image)
+            plt.axis('on')
+            plt.title(f'Predict digit: {rv.chosen_number()}')
+            plt.show()
+
+
+if __name__ == "__main__":
+    gpu_utils.print_gpu_availability()
+    main()
+    

dl-exp/exp2/modified/settings.py

@@ -0,0 +1,12 @@
+from pathlib import Path
+
+MNIST_DATASET_PATH: Path = Path(__file__).resolve().parent.parent / 'datasets' / 'mnist.npz'
+"""MNIST数据集文件的路径"""
+
+SAVED_MODEL_PATH: Path = Path(__file__).resolve().parent.parent / 'models' / 'cnn.pth'
+"""训练好的模型保存的位置"""
+
+N_EPOCH: int = 5
+"""训练时的epoch次数"""
+N_BATCH_SIZE: int = 1000
+"""训练时的batch size"""

dl-exp/exp2/modified/sketchpad.py

@@ -0,0 +1,190 @@
+from pathlib import Path
+import sys
+import typing
+import tkinter as tk
+from predict import PredictResult, Predictor
+
+sys.path.append(str(Path(__file__).resolve().parent.parent.parent))
+import gpu_utils
+
+
+class SketchpadApp:
+    IMAGE_HW: typing.ClassVar[int] = 28
+    PIXEL_HW: typing.ClassVar[int] = 15
+
+    def __init__(self, root: tk.Tk, predictor: Predictor):
+        self.root = root
+        self.root.title("看图说数")
+        
+        # 创建画板框架
+        canvas_frame = tk.Frame(root)
+        canvas_frame.pack(pady=10)
+        # 创建图像大小的画板
+        self.canvas_pixel_count = SketchpadApp.IMAGE_HW
+        self.canvas_pixel_size = SketchpadApp.PIXEL_HW  # 每个像素的大小
+        canvas_hw = self.canvas_pixel_count * self.canvas_pixel_size
+        self.canvas_width = canvas_hw
+        self.canvas_height = canvas_hw
+        self.canvas = tk.Canvas(
+            canvas_frame, 
+            width=self.canvas_width, 
+            height=self.canvas_height, 
+            bg='black'
+        )
+        self.canvas.pack()
+        # 存储画板状态。False表示没有画（黑色），True表示画了（白色）。
+        self.canvas_data = [[False for _ in range(self.canvas_pixel_count)] for _ in range(self.canvas_pixel_count)]
+        # 绑定鼠标事件
+        self.canvas.bind("<B1-Motion>", self.paint)
+        self.canvas.bind("<Button-1>", self.paint)
+        # 绘制初始网格
+        self.draw_grid()
+        
+        # 创建表格框架
+        table_frame = tk.Frame(root)
+        table_frame.pack(pady=10)
+        # 表头数据
+        header_words = ("猜测的数字", ) + tuple(f'为{i}的概率' for i in range(10))
+        # 创建表头
+        for col, header in enumerate(header_words):
+            header_label = tk.Label(
+                table_frame,
+                text=header,
+                relief="solid",
+                borderwidth=1,
+                width=12,
+                height=2,
+                bg="lightblue"
+            )
+            header_label.grid(row=0, column=col, sticky="nsew")
+        # 创建第二行（显示数值的行）
+        self.value_labels = []
+        for col in range(len(header_words)):
+            value_label = tk.Label(
+                table_frame,
+                text="0.00",  # 默认显示0.00
+                relief="solid",
+                borderwidth=1,
+                width=12,
+                height=2,
+                bg="white"
+            )
+            value_label.grid(row=1, column=col, sticky="nsew")
+            self.value_labels.append(value_label)
+        # 设置第一列的特殊样式（猜测的数字）
+        self.value_labels[0].config(text="N/A", bg="lightyellow")
+        # 清空样式
+        self.clear_table()
+
+        # 创建按钮框架
+        button_frame = tk.Frame(root)
+        button_frame.pack(pady=10)
+        # 执行按钮
+        execute_button = tk.Button(
+            button_frame, 
+            text="执行", 
+            command=self.execute,
+            bg='lightgreen',
+            width=10
+        )
+        execute_button.pack(side=tk.LEFT, padx=5)
+        # 重置按钮
+        reset_button = tk.Button(
+            button_frame, 
+            text="重置", 
+            command=self.reset,
+            bg='lightcoral',
+            width=10
+        )
+        reset_button.pack(side=tk.LEFT, padx=5)
+        # 设置用于执行的predictor
+        self.predictor = predictor
+        
+    # region: 画板部分
+
+    canvas: tk.Canvas
+    canvas_data: list[list[bool]]
+    canvas_width: int
+    canvas_height: int
+
+    def draw_grid(self):
+        """绘制网格线"""
+        for i in range(self.canvas_pixel_count + 1):
+            # 垂直线
+            self.canvas.create_line(
+                i * self.canvas_pixel_size, 0,
+                i * self.canvas_pixel_size, self.canvas_height,
+                fill='lightgray'
+            )
+            # 水平线
+            self.canvas.create_line(
+                0, i * self.canvas_pixel_size,
+                self.canvas_width, i * self.canvas_pixel_size,
+                fill='lightgray'
+            )
+    
+    def paint(self, event):
+        """处理鼠标绘制事件"""
+        # 计算点击的网格坐标
+        col = event.x // self.canvas_pixel_size
+        row = event.y // self.canvas_pixel_size
+        
+        # 确保坐标在有效范围内
+        if 0 <= col < self.canvas_pixel_count and 0 <= row < self.canvas_pixel_count:
+            # 更新网格状态
+            if self.canvas_data[row][col] != True:
+                self.canvas_data[row][col] = True
+                
+                # 绘制黑色矩形
+                x1 = col * self.canvas_pixel_size
+                y1 = row * self.canvas_pixel_size
+                x2 = x1 + self.canvas_pixel_size
+                y2 = y1 + self.canvas_pixel_size
+                
+                self.canvas.create_rectangle(x1, y1, x2, y2, fill='white', outline='')
+
+    # endregion
+
+    # region: 表格部分
+
+    value_labels: list[tk.Label]
+
+    def show_in_table(self, result: PredictResult):
+        self.value_labels[0].config(text=str(result.chosen_number()))
+
+        number_possibilities = result.number_possibilities()
+        for index, label in enumerate(self.value_labels[1:]):
+            label.config(text=f'{number_possibilities[index]:.4f}')
+
+    def clear_table(self):
+        for label in self.value_labels:
+            label.config(text='N/A')
+
+    # endregion
+
+    # region: 按钮部分
+
+    predictor: Predictor
+
+    def execute(self):
+        """执行按钮功能 - 将画板数据传递给后端"""
+        prediction = self.predictor.predict_sketchpad(self.canvas_data)
+        self.show_in_table(prediction)
+    
+    def reset(self):
+        """重置按钮功能 - 清空画板"""
+        self.canvas.delete("all")
+        self.canvas_data = [[0 for _ in range(self.canvas_pixel_count)] for _ in range(self.canvas_pixel_count)]
+        self.draw_grid()
+        self.clear_table()
+
+    # endregion
+
+if __name__ == "__main__":
+    gpu_utils.print_gpu_availability()
+    
+    predictor = Predictor()
+
+    root = tk.Tk()
+    app = SketchpadApp(root, predictor)
+    root.mainloop()

dl-exp/exp2/modified/train.py

@@ -0,0 +1,85 @@
+from pathlib import Path
+import sys
+import typing
+import torch
+import torchinfo
+import ignite.engine
+import ignite.metrics
+from ignite.engine import Engine, Events
+from ignite.handlers.tqdm_logger import ProgressBar
+from dataset import MnistDataLoaders
+from model import Cnn
+import settings
+
+sys.path.append(str(Path(__file__).resolve().parent.parent.parent))
+import gpu_utils
+
+
+class Trainer:
+    """核心训练器"""
+
+    device: torch.device
+    data_source: MnistDataLoaders
+    model: Cnn
+
+    trainer: Engine
+    evaluator: Engine
+    pbar: ProgressBar
+
+    def __init__(self):
+        # 创建训练设备，模型和数据加载器。
+        self.device = gpu_utils.get_gpu_device()
+        self.model = Cnn().to(self.device)
+        self.data_source = MnistDataLoaders(batch_size=settings.N_BATCH_SIZE)
+        # 展示模型结构。批次为指定批次数量，通道只有一个灰度通道，大小28x28。
+        torchinfo.summary(self.model, (settings.N_BATCH_SIZE, 1, 28, 28))
+
+        # 优化器和损失函数
+        optimizer = torch.optim.Adam(self.model.parameters(), eps=1e-7)
+        criterion = torch.nn.CrossEntropyLoss()
+        # 创建训练器
+        self.trainer = ignite.engine.create_supervised_trainer(
+            self.model, optimizer, criterion, self.device)
+        # 将训练器关联到进度条
+        self.pbar = ProgressBar(persist=True)
+        self.pbar.attach(self.trainer, output_transform=lambda loss: {"loss": loss})
+
+        # 创建测试的评估器的评估量
+        evaluator_metrics = {
+            # 这个Accuracy要的是logits，而不是possibilities，
+            # 所以依然是不需要softmax处理后的结果。
+            "accuracy": ignite.metrics.Accuracy(device=self.device),
+            "loss": ignite.metrics.Loss(criterion, device=self.device)
+        }
+        # 创建测试评估器
+        self.evaluator = ignite.engine.create_supervised_evaluator(
+            self.model, metrics=evaluator_metrics, device=self.device)
+        
+    def train_model(self):
+        # 训练模型
+        self.trainer.run(self.data_source.train_loader, max_epochs=settings.N_EPOCH)
+
+    def save_model(self):
+        # 确保保存模型的文件夹存在。
+        settings.SAVED_MODEL_PATH.parent.mkdir(parents=True, exist_ok=True)
+        # 仅保存模型参数
+        torch.save(self.model.state_dict(), settings.SAVED_MODEL_PATH)
+        print(f'Model was saved into: {settings.SAVED_MODEL_PATH}')
+
+    def test_model(self):
+        # 测试模型并输出结果
+        self.evaluator.run(self.data_source.test_loader)
+        metrics = self.evaluator.state.metrics
+        print(f"Accuracy: {metrics['accuracy']:.4f} Loss: {metrics['loss']:.4f}")
+
+
+def main():
+    trainer = Trainer()
+    trainer.train_model()
+    trainer.save_model()
+    trainer.test_model()
+
+
+if __name__ == "__main__":
+    gpu_utils.print_gpu_availability()
+    main()

dl-exp/exp2/source/predict.py

@@ -0,0 +1,40 @@
+from pathlib import Path
+import tensorflow as tf
+from PIL import Image
+import numpy as np
+import matplotlib.pyplot as plt
+
+from train import CNN
+
+class Predict(object):
+    def __init__(self):
+        latest = tf.train.latest_checkpoint('./ckpt')
+        self.cnn = CNN()
+        # 恢复网络权重
+        self.cnn.model.load_weights(latest)
+
+    def predict(self, image_path):
+        # 以黑白方式读取图片
+        img = Image.open(image_path).convert('L')
+        img = np.reshape(img, (28, 28, 1)) / 255.
+        x = np.array([1 - img])
+        y = self.cnn.model.predict(x)
+
+        # 因为x只传入了一张图片，取y[0]即可
+        # np.argmax()取得最大值的下标，即代表的数字
+        print(image_path)
+#        print(y[0])
+        print('        -> Predict digit', np.argmax(y[0]))
+        plt.figure("Image")  # 图像窗口名称
+        plt.imshow(img)
+        plt.axis('on')  # 关掉坐标轴为 off
+        plt.title(np.argmax(y[0]))  # 图像题目 # 必须有这个，要不然无法显示
+        plt.show()
+
+
+if __name__ == "__main__":
+    app = Predict()
+    images_dir = Path(__file__).resolve().parent.parent / 'test_images'
+    app.predict(images_dir / '0.png') 
+    app.predict(images_dir / '1.png')
+    app.predict(images_dir / '4.png')

dl-exp/exp2/source/train.py

@@ -0,0 +1,56 @@
+from pathlib import Path
+import tensorflow as tf
+from tensor.keras import datasets, layers, models
+
+class CNN(object):
+    def __init__(self):
+        model = models.Sequential()
+        # 第1层卷积，卷积核大小为3*3，32个，28*28为待训练图片的大小
+        model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)))
+        model.add(layers.MaxPooling2D((2, 2)))
+        # 第2层卷积，卷积核大小为3*3，64个
+        model.add(layers.Conv2D(64, (3, 3), activation='relu'))
+        model.add(layers.MaxPooling2D((2, 2)))
+        # 第三层卷积，卷积核大小为3*3，64个
+        model.add(layers.Conv2D(64, (3, 3), activation='relu'))
+        model.add(layers.Flatten())
+        model.add(layers.Dense(64, activation='relu'))
+        model.add(layers.Dense(10, activation='softmax'))
+        model.summary()
+        self.model = model
+
+class DataSource(object):
+    def __init__(self):
+        # mnist数据集存储的位置，如何不存在将自动下载
+        data_path = Path(__file__).resolve().parent.parent / 'datasets' / 'mnist.npz'
+        (train_images, train_labels), (test_images,
+                                       test_labels) = datasets.mnist.load_data(path=data_path)
+        # 6万张训练图片，1万张测试图片
+        train_images = train_images.reshape((60000, 28, 28, 1))
+        test_images = test_images.reshape((10000, 28, 28, 1))
+        # 像素值映射到 0 - 1 之间
+        train_images, test_images = train_images / 255.0, test_images / 255.0
+        self.train_images, self.train_labels = train_images, train_labels
+        self.test_images, self.test_labels = test_images, test_labels
+
+class Train:
+    def __init__(self):
+        self.cnn = CNN()
+        self.data = DataSource()
+    def train(self):
+        check_path = Path(__file__).resolve().parent.parent / 'models' / 'cnn.ckpt'
+        # period 每隔5epoch保存一次
+        save_model_cb = tf.keras.callbacks.ModelCheckpoint(
+            str(check_path), save_weights_only=True, verbose=1, period=5)
+        self.cnn.model.compile(optimizer='adam',
+                               loss='sparse_categorical_crossentropy',
+                               metrics=['accuracy'])
+        self.cnn.model.fit(self.data.train_images, self.data.train_labels,
+                           epochs=5, batch_size=1000, callbacks=[save_model_cb])
+        test_loss, test_acc = self.cnn.model.evaluate(
+            self.data.test_images, self.data.test_labels)
+        print("准确率: %.4f, 共测试了%d张图片 " % (test_acc, len(self.data.test_labels)))
+
+if __name__ == "__main__":
+    app = Train()
+    app.train()

dl-exp/exp2/test_images/.gitignore

@@ -0,0 +1,2 @@
+# Ignore all test images
+*.png