重新整理文件夹结构，并添加CNN训练代码。

tools/TrainCNN/backward.py

@@ -0,0 +1,123 @@
+import tensorflow as tf
+from progressive.bar import Bar
+import generate
+import forward
+
+
+def save_kernal(fp, val):
+    print(val.shape[2], file=fp)
+    print(val.shape[3], file=fp)
+    print(val.shape[1], file=fp)
+    print(val.shape[0], file=fp)
+    for in_channel in range(val.shape[2]):
+        for out_channel in range(val.shape[3]):
+            for row in range(val.shape[0]):
+                for col in range(val.shape[1]):
+                    print(val[row][col][in_channel][out_channel], file=fp)
+
+
+def save_weight_mat(fp, val):
+    print(val.shape[0], file=fp)
+    print(val.shape[1], file=fp)
+    for row in range(val.shape[0]):
+        for col in range(val.shape[1]):
+            print(val[row][col], file=fp)
+
+
+def save_bias(fp, val):
+    print(val.shape[0], file=fp)
+    for i in range(val.shape[0]):
+        print(val[i], file=fp)
+
+
+def save_para(folder, paras):
+    with open(folder + "/conv1_w", "w") as fp:
+        save_kernal(fp, paras[0])
+    with open(folder + "/conv1_b", "w") as fp:
+        save_bias(fp, paras[1])
+    with open(folder + "/conv2_w", "w") as fp:
+        save_kernal(fp, paras[2])
+    with open(folder + "/conv2_b", "w") as fp:
+        save_bias(fp, paras[3])
+    with open(folder + "/fc1_w", "w") as fp:
+        save_weight_mat(fp, paras[4])
+    with open(folder + "/fc1_b", "w") as fp:
+        save_bias(fp, paras[5])
+    with open(folder + "/fc2_w", "w") as fp:
+        save_weight_mat(fp, paras[6])
+    with open(folder + "/fc2_b", "w") as fp:
+        save_bias(fp, paras[7])
+
+
+STEPS = 30000
+BATCH = 10
+LEARNING_RATE_BASE  = 0.01
+LEARNING_RATE_DECAY = 0.99
+MOVING_AVERAGE_DECAY = 0.99
+
+
+def train(dataset, show_bar=False):
+    test_images, test_labels = dataset.all_test_sets()
+
+    x = tf.placeholder(tf.float32, [None, forward.SRC_ROWS, forward.SRC_COLS, forward.SRC_CHANNELS])
+    y_= tf.placeholder(tf.float32, [None, forward.OUTPUT_NODES])
+    nodes, vars = forward.forward(0.001)
+    y = nodes[-1]
+
+    ce  = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=tf.argmax(y_, 1))
+    cem = tf.reduce_mean(ce)
+    loss= cem + tf.add_n(tf.get_collection("losses"))
+
+    global_step = tf.Variable(0, trainable=False)
+    learning_rate = tf.train.exponential_decay(
+        LEARNING_RATE_BASE,
+        global_step,
+        len(dataset.train_sets) / BATCH,
+        LEARNING_RATE_DECAY,
+        staircase=False)
+    train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step)
+
+    ema = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)
+    ema_op = ema.apply(tf.trainable_variables())
+    with tf.control_dependencies([train_step, ema_op]):
+        train_op = tf.no_op(name='train')
+
+    correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
+    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
+    acc = 0
+
+    with tf.Session() as sess:
+        init_op = tf.global_variables_initializer()
+        sess.run(init_op)
+
+        if show_bar:
+            bar = Bar(max_value=STEPS, width=u'50%')
+            bar.cursor.clear_lines(1)
+            bar.cursor.save()
+
+        for i in range(STEPS):
+            images_samples, labels_samples = dataset.sample_train_sets(BATCH)
+
+            _, loss_value, step = sess.run(
+                [train_op, loss, global_step],
+                feed_dict={x: images_samples, y_: labels_samples}
+            )
+
+            if i % 100 == 0:
+                if i % 1000 == 0:
+                    acc = sess.run(accuracy, feed_dict={x: test_images, y_: test_labels})
+
+            if show_bar:
+                bar.title = "step: %d, loss: %f, acc: %f" % (step, loss_value, acc)
+                bar.cursor.restore()
+                bar.draw(value=i+1)
+
+        vars_val = sess.run(vars)
+        save_para("paras", vars_val)
+        # nodes_val = sess.run(nodes, feed_dict={x:test})
+        # return vars_val, nodes_val
+
+
+if __name__ == "__main__":
+    dataset = generate.DataSet("images")
+    train(dataset, show_bar=True)

tools/TrainCNN/forward.py

@@ -0,0 +1,98 @@
+import tensorflow as tf
+
+
+def get_weight(shape, regularizer=None):
+    w = tf.Variable(tf.truncated_normal(shape, stddev=0.1))
+    if regularizer is None:
+        tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(regularizer)(w))
+    return w
+
+
+def get_bias(shape):
+    b = tf.Variable(tf.zeros(shape))
+    return b
+
+
+def conv2d(x, w):
+    return tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding="VALID")
+
+
+def avg_pool_2x2(x):
+    return tf.nn.avg_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="VALID")
+
+
+def max_pool_2x2(x):
+    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="VALID")
+
+
+# 原图像行数
+SRC_ROWS = 36
+
+# 原图像列数
+SRC_COLS = 48
+
+# 原图像通道数
+SRC_CHANNELS = 1
+
+# 第一层卷积核大小
+CONV1_KERNAL_SIZE = 5
+
+# 第一层卷积输出通道数
+CONV1_OUTPUT_CHANNELS = 8
+
+# 第二层卷积核大小
+CONV2_KERNAL_SIZE = 3
+
+# 第二层卷积输出通道数
+CONV2_OUTPUT_CHANNELS = 16
+
+# 第一层全连接宽度
+FC1_OUTPUT_NODES = 32
+
+# 第二层全连接宽度（输出标签类型数）
+FC2_OUTPUT_NODES = 8
+
+# 输出标签类型数
+OUTPUT_NODES = FC2_OUTPUT_NODES
+
+
+def forward(x, regularizer=None):
+    vars = []
+    nodes = []
+
+    conv1_w = get_weight(
+        [CONV1_KERNAL_SIZE, CONV1_KERNAL_SIZE, 1, CONV1_OUTPUT_CHANNELS]
+    )
+    conv1_b = get_bias([CONV1_OUTPUT_CHANNELS])
+    conv1   = tf.nn.relu(tf.nn.bias_add(conv2d(x, conv1_w), conv1_b))
+    pool1   = avg_pool_2x2(conv1)
+    vars.extend([conv1_w, conv1_b])
+    nodes.extend([conv1, pool1])
+
+    conv2_w = get_weight(
+        [CONV2_KERNAL_SIZE, CONV2_KERNAL_SIZE, CONV1_OUTPUT_CHANNELS, CONV2_OUTPUT_CHANNELS]
+    )
+    conv2_b = get_bias([CONV2_OUTPUT_CHANNELS])
+    conv2 = tf.nn.relu(tf.nn.bias_add(conv2d(pool1, conv2_w), conv2_b))
+    pool2 = avg_pool_2x2(conv2)
+    vars.extend([conv2_w, conv2_b])
+    nodes.extend([conv2, pool2])
+
+    pool_shape = pool2.get_shape().as_list()
+    node = pool_shape[1] * pool_shape[2] * pool_shape[3]
+    reshaped = tf.reshape(pool2, [-1, node])
+
+    fc1_w = get_weight([node, FC1_OUTPUT_NODES], regularizer)
+    fc1_b = get_bias([FC1_OUTPUT_NODES])
+    fc1   = tf.nn.relu(tf.matmul(reshaped, fc1_w) + fc1_b)
+    vars.extend([fc1_w, fc1_b])
+    nodes.extend([fc1])
+
+    fc2_w = get_weight([FC1_OUTPUT_NODES, FC2_OUTPUT_NODES], regularizer)
+    fc2_b = get_bias([FC2_OUTPUT_NODES])
+    fc2   = tf.nn.softmax(tf.matmul(fc1, fc2_w) + fc2_b)
+    vars.extend([fc2_w, fc2_b])
+    nodes.extend([fc2])
+
+    return nodes, vars
+

tools/TrainCNN/generate.py

@@ -0,0 +1,49 @@
+import numpy as np
+import os
+import cv2
+import random
+from forward import OUTPUT_NODES
+
+
+class DataSet:
+    def __init__(self, folder):
+        self.train_sets = []
+        self.test_sets = []
+        self.generate_data_sets(folder)
+
+    def generate_data_sets(self, folder):
+        def file2nparray(name):
+            image = cv2.imread(name)
+            return image[:, :, 0]
+
+        def id2label(id):
+            a = np.zeros([OUTPUT_NODES, 1])
+            a[id] = 1
+            return a[:]
+
+        sets = []
+        for i in range(OUTPUT_NODES):
+            dir = "%s/%d" % (folder, i)
+            files = os.listdir(dir)
+            for file in files:
+                sets.append([file2nparray("%s/%s" % (dir, file)), id2label(i)])
+        sets = np.array(sets)
+        np.random.shuffle(sets)
+        length = len(sets)
+        self.train_sets = sets[:length*3//4]
+        self.test_sets = sets[length*3//4:]
+
+    def sample_train_sets(self, length):
+        samples = []
+        labels = []
+        for i in range(length):
+            id = random.randint(0, length-1)
+            samples.append(self.train_sets[id][0])
+            labels.append(self.train_sets[id][1])
+        return np.array(samples), np.array(labels)
+
+    def all_train_sets(self):
+        return self.train_sets[:, 0, :, :], self.train_sets[:, 1, :, :]
+
+    def all_test_sets(self):
+        return self.test_sets[:, 0, :, :], self.test_sets[:, 1, :, :]

tools/monitor.sh

@@ -0,0 +1,11 @@
+#!/bin/sh
+
+exe=$1
+while true; do
+	state=`ps aux | grep "$1" | grep -v grep | grep -v $0`
+	if [ ! "$state" ]; then
+		exec $exe &
+		echo "restart $exe"
+	fi
+	sleep 2
+done