深度学习入门之二阶段小demo练习：使用深度学习框架（tensorflow2.0及以上、pytorch1.0及以上版本）搭建卷积神经网络，在骨架行为数据集上准确率高于60%。

根据个人经验总结的深度学习入门路线（简单快速）
https://blog.csdn.net/weixin_44414948/article/details/109704871

深度学习入门二阶段demo练习：
https://blog.csdn.net/weixin_44414948/article/details/110673660

任务要求

使用深度学习框架（tensorflow2.0及以上、pytorch1.0及以上版本）搭建卷积神经网络。

数据集：骨架行为数据集，准确率高于60%。
追加要求：同一图表绘制训练过程的loss、acc曲线，截取测试集准确率的图。

骨架行为数据集简介

该数据集是由本课题组行为检测方向同学自行采集，共有奔跑、跳跃、倚靠产品、吸烟、玩手机、背景行为共6种行为种类标签，数据为人体骨架节点（25点）序列的txt文件，共60万帧，示例如下图所示，每行共有75个点（25个关节点的xyz坐标），一行为一帧。

完整实现代码

import tensorflow as tf
from tensorflow.keras import layers, optimizers, datasets, Sequential, metrics
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt


conv_layers = [
                # 卷积层 + 池化层
    layers.Conv2D(64, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(64, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    layers.Conv2D(128, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(128, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    layers.Conv2D(256, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(256, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    layers.Conv2D(512, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(512, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),

    layers.Conv2D(512, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.Conv2D(512, kernel_size=[3, 3], padding='same', activation=tf.nn.relu),
    layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),


]

def preprocess(x, y):
    x = tf.cast(x, dtype=tf.float32) / 255.
    y = tf.cast(y, dtype=tf.int32)
    return x, y

#导入数据集
X = []
Y = []
data = pd.read_table('E:\\PycharmProjects\\untitled\\1-action_seq.txt', sep=',', header=None)
data = data.iloc[:, :75]
label = pd.read_table('E:\\PycharmProjects\\untitled\\1-label_class.txt', header= None)

for i in range(data.shape[0]):
    X.append(list(data.iloc[i, :].values))
for j in range(label.shape[0]):
    Y.append(list(label.iloc[j, :].values)) #对应的二维numpy值数组
Z = []
for i in range(len(X)):
    Z.append(X[i] + Y[i])
A = []
B = []
np.random.shuffle(Z)
for i in range(len(Z)):
    A.append(Z[i][:75])
    B.append(Z[i][-1])
x = A[:int(len(A) * 0.8)]
x = tf.reshape(x, [int(len(A) * 0.8), 25, 3, 1])
x_test = A[int(len(A) * 0.8):]
x_test = tf.reshape(x_test, [int(len(A) * 0.2)+1, 25, 3, 1])
y = B[:int(len(A) * 0.8)]
y_test = B[int(len(A) * 0.8):]
# y = tf.squeeze(y, axis=1)
# y_test = tf.squeeze(y_test, axis=1)
# print(x.shape, y.shape, x_test.shape, y_test.shape)
train_db = tf.data.Dataset.from_tensor_slices((x, y))
train_db = train_db.shuffle(1000).map(preprocess).batch(100)
test_db = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_db = test_db.map(preprocess).batch(100)
sample = next(iter(train_db))
print('sample:', sample[0].shape, sample[1].shape, tf.reduce_min(sample[0]), tf.reduce_max(sample[0]))



def main():
    #[b, 25, 3, 1] => [b, 1, 1, 512]
    conv_net = Sequential(conv_layers)
    #全连接层
    fc_net = Sequential([
        layers.Dense(256, activation=tf.nn.relu),
        layers.Dense(128, activation=tf.nn.relu),
        layers.Dense(64, activation=tf.nn.relu),
        layers.Dense(6, activation=tf.nn.relu),

    ])
    conv_net.build(input_shape=[None, 25, 3, 1])
    fc_net.build(input_shape=[None, 512])
    #优化器
    optimizer = optimizers.Adam(lr = 0.001)
    variables = conv_net.trainable_variables + fc_net.trainable_variables

    for epoch in range(2):
        for step, (x, y) in enumerate(train_db):
            # 求梯度
            with tf.GradientTape() as tape:
                # [b, 25, 3, 1] => [b, 1, 1, 512]
                out = conv_net(x)
                # squeeze 也可以叫做 reshape
                # flatten 输出成平坦的变量 => [b, 512]
                out = tf.reshape(out, [-1, 512])

                #全连接层的输出，也叫做 logits, [b, 512] => [b, 6]
                logits = fc_net(out)

                #loss的计算
                #先把y转化为onthot格式,[b] => [b, 6]
                y_onehot = tf.one_hot(y, depth=6) #因为是6分类
                loss = tf.losses.categorical_crossentropy(y_onehot, logits, from_logits=True)
                #计算均值
                loss = tf.reduce_mean(loss)

                #计算梯度
            grads = tape.gradient(loss, variables) #对所有参数进行求导， y轴为loss， x轴为 variables.

                #梯度的更新(需要设计优化器)
            optimizer.apply_gradients(zip(grads, variables))

                #打印一下
            if step % 100 == 0:
                print(epoch, step, 'loss:', float(loss))
        #做测试
        total_num = 0
        total_correct = 0
        for x, y in test_db:
            out = conv_net(x)
            out = tf.reshape(out, [-1, 512])
            logits = fc_net(out)

            prob = tf.nn.softmax(logits, axis=1)#第一根轴
            #预测最大值的位置
            pred = tf.argmax(prob, axis=1)
            pred = tf.cast(pred, dtype=tf.int32)
            #真实的值
            correct = tf.cast(tf.equal(pred, y), dtype=tf.int32)
            correct = tf.reduce_sum(correct)
            total_num += x.shape[0]
            total_correct += int(correct)
        acc = total_correct / total_num
        print(epoch, 'acc:', acc)

    # 输入两个曲线的信息
    plt.figure(figsize=(12, 8), dpi=80)
    plt.plot(epoch, acc, color='g', linestyle='-.', label='acc')
    plt.plot(epoch, loss, color='r', linestyle='--', label='loss')
    # 显示图例
    plt.legend()  # 默认loc=Best
    # 添加网格信息
    plt.grid(True, linestyle='--', alpha=0.5)  # 默认是True，风格设置为虚线，alpha为透明度
    # 添加标题
    plt.xlabel('epochs')
    plt.ylabel('loss/acc')
    plt.title('inception_Curve of loss/acc Change with epochs in GJData')
    plt.savefig('E:\\PycharmProjects\\untitled\\inception.jpg')

    plt.show()
if __name__ == '__main__':
    main()

训练过程曲线图

注：师弟做得还是挺不错的，点个赞