Created by: GT-ZhangAcer

PaddlePaddle环境：2.0Beta-CPU-pip清华源安装 系统以及Python环境：MAC+Python3.7

代码部分

case1 Fluid版本代码 AI Studio

import paddle.fluid as fluid
import numpy as np
import PIL.Image as Image

data_path = "/home/aistudio/work"
save_model_path = "./model"


# 构建Reader

def switch_reader(is_val: bool = False):
    def reader():
        # 读取标签数据
        with open(data_path + "/OCR_100P.txt", 'r') as f:
            labels = f.read()
        # 判断是否是验证集
        if is_val:
            index_range = range(500, 800)
        else:
            index_range = range(1, 500)
        # 抽取数据使用迭代器返回
        for index in index_range:
            im = Image.open(data_path + "/" + str(index) + ".jpg").convert('L')  # 使用Pillow读取图片
            im = np.array(im).reshape(1, 1, 30, 15).astype(np.float32)  # NCHW格式
            im /= 255  # 归一化以提升训练效果
            lab = labels[index - 1]  # 因为循环中i是从1开始迭代的，所有这里需要减去1
            yield im, int(lab)

    return reader  # 注意！此处不需要带括号


# 划分mini_batch
batch_size = 128
train_reader = fluid.io.batch(reader=switch_reader(), batch_size=batch_size)
val_reader = fluid.io.batch(reader=switch_reader(is_val=True), batch_size=batch_size)

# 定义网络输入格式
img = fluid.data(name="img", shape=[-1, 1, 30, 15], dtype="float32")
label = fluid.data(name='label', shape=[-1, 1], dtype='int64')

# 定义网络
hidden = fluid.layers.fc(input=img, size=200, act='relu')
# 第二个全连接层，激活函数为ReLU
hidden = fluid.layers.fc(input=hidden, size=200, act='relu')
# 以softmax为激活函数的全连接输出层，输出层的大小必须为数字的个数10
net_out = fluid.layers.fc(input=hidden, size=10, act='softmax')
# 使用API来计算正确率
acc = fluid.layers.accuracy(input=net_out, label=label)
# 将上方的结构克隆出来给测试程序使用
eval_prog = fluid.default_main_program().clone(for_test=True)
# 定义损失函数
loss = fluid.layers.cross_entropy(input=net_out, label=label)
avg_loss = fluid.layers.mean(loss)

# 定义优化方法
sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.01)
sgd_optimizer.minimize(avg_loss)

# 定义执行器
place = fluid.CPUPlace()
exe = fluid.Executor(place)

# 数据传入设置
feeder = fluid.DataFeeder(place=place, feed_list=[img, label])
prog = fluid.default_startup_program()
exe.run(prog)

# 开始训练
Epoch = 10
for i in range(Epoch):
    batch_loss = None
    batch_acc = None
    # 训练集 只看loss来判断模型收敛情况
    for batch_id, data in enumerate(train_reader()):
        outs = exe.run(
            feed=feeder.feed(data),
            fetch_list=[loss])
        batch_loss = np.average(outs[0])
    # 验证集 只看准确率来判断收敛情况
    for batch_id, data in enumerate(val_reader()):
        outs = exe.run(program=eval_prog,
                       feed=feeder.feed(data),
                       fetch_list=[acc])
        batch_acc = np.average(outs[0])
    print("Epoch:", i, "\tLoss:{:3f}".format(batch_loss), "\tAcc:{:2f} %".format(batch_acc * 100))

# 保存模型
fluid.io.save_inference_model(save_model_path, ['img'], [net_out], exe)

该版本在第10个EPOCH后就可以达到较高的准确率

case2 使用2.0Beta的Hapi进行搭建

import os

import paddle as pp
import PIL.Image as Image
import numpy as np

from paddle.io import Dataset

USE_GPU = False
DATA_PATH = "./data"
CLASSIFY_NUM = 10
IMAGE_SHAPE_C = 1
IMAGE_SHAPE_H = 30
IMAGE_SHAPE_W = 15


class Reader(Dataset):
    def __init__(self, data_path, add_label: bool = True, is_val: bool = False):
        """
        数据读取器
        :param data_path: 数据集所在路径
        :param add_label: 是否添加label
        :param is_val: 是否为评估模式
        """
        super().__init__()
        self.add_label = add_label

        self.img_list = [os.path.join(data_path, str(i) + ".jpg") for i in range(1, 800)]
        with open(os.path.join(data_path, "OCR_100P.txt"), 'r') as f:
            label_list = f.read()

        self.img_list = self.img_list[:500] if not is_val \
            else self.img_list[500:800]
        self.label_list = label_list[:500] if not is_val \
            else label_list[500:800]

        print("正在打印", "训练" if not is_val else "验证", "的示例数据，请检查是否对应")
        for sample_index in range(3):
            self.print_sample(sample_index)

    def __getitem__(self, index):
        """
        获取一组数据
        :param index: 文件索引号
        :return:
        """
        # 第一步打开图像文件并获取label值
        img = Image.open(self.img_list[index]).convert('L')
        label = self.label_list[index]
        # 第二步做图像增强 - 因本次图片较小，不适合做增强所以跳过
        # 第三步转换为Numpy的array格式，并且使用浮点型数据格式，shape转化为CHW（数量、通道数、高、宽）
        img = np.array(img, dtype="float32").reshape(1 * 30 * 15)
        label = np.array([label], dtype="int64")
        # 第四步进行归一化操作方便收敛
        img /= 255
        return img, label if self.add_label else img

    def print_sample(self, index: int = 0):
        print("文件名", self.img_list[index], "\t标签值", self.label_list[index])

    def __len__(self):
        return len(self.img_list)


class Net(pp.nn.Layer):
    def __init__(self):
        super().__init__()
        self.layer1 = pp.nn.Linear(in_features=IMAGE_SHAPE_C * IMAGE_SHAPE_H * IMAGE_SHAPE_W, out_features=200)
        self.layer2 = pp.nn.Linear(in_features=200, out_features=200)
        self.layer3 = pp.nn.Linear(in_features=200, out_features=CLASSIFY_NUM)

    # 定义网络结构的前向计算过程
    def forward(self, inputs):
        layer1 = pp.nn.functional.relu(self.layer1(inputs))
        layer2 = pp.nn.functional.relu(self.layer2(layer1))
        layer3 = self.layer3(layer2)
        out = pp.nn.functional.softmax(layer3)
        return out


input_define = pp.static.InputSpec(shape=[-1, IMAGE_SHAPE_C * IMAGE_SHAPE_H * IMAGE_SHAPE_W],
                                   dtype="float32",
                                   name="img")
label_define = pp.static.InputSpec(shape=[-1, 1],
                                   dtype="int64",
                                   name="label")

model = pp.Model(Net(), inputs=input_define, labels=label_define)
optimizer = pp.optimizer.SGD(learning_rate=0.01, parameters=model.parameters())
model.prepare(optimizer=optimizer,
              loss=pp.nn.loss.CrossEntropyLoss(),
              metrics=pp.metric.Accuracy(topk=(1, 2)))

model.fit(train_data=Reader(data_path=DATA_PATH),
          batch_size=16,
          epochs=100,
          save_dir="output/",
          save_freq=10)

该版本在Batch_size为128时效果为：

该版本在Batch_size为1时效果为：

没能收敛好，感觉可能是自己代码某个地方漏掉了什么，希望可以得到大佬帮助~