深度学习CNN端到端字符识别

# -*- coding: utf-8 -*- """ Created on Wed Aug 30 09:43:06 2017 @author: WeiZhe Train data existing at files """ import numpy as np import tensorflow as tf import random # Load Data p2pimgs=np.load(r'E:\Program\PythonProgram\p2p\data\imgsave.npy') p2plabels=np.load(r'E:\Program\PythonProgram\p2p\data\labelsave.npy') #%% # weight init def weight_variable(shape, strName): initial = tf.truncated_normal(shape, stddev=0.1) return tf.Variable(initial, name=strName) # bias init def bias_variable(shape, strName): initial = tf.constant(0.1, shape=shape) return tf.Variable(initial, name=strName) # convolution def conv2d(x, W): return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') # max pooling def max_pool_2x2(x): return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') #%% Model setup # placeholder x = tf.placeholder("float", [None, 90*140]) y_ = tf.placeholder("float", [None,6]) keep_prob = tf.placeholder("float") def CNN(): # first layer, conv&pool W_conv1 = weight_variable([3, 4, 1, 32], 'W_conv1') b_conv1 = bias_variable([32], 'b_conv1') x_image = tf.reshape(x,[-1,90,140,1]) h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) h_pool1 = max_pool_2x2(h_conv1) out1 = tf.nn.dropout(h_pool1,keep_prob) # batches*20*15 # second layer, con&pool W_conv2 = weight_variable([3, 4, 32, 64], 'W_conv2') b_conv2 = bias_variable([64], 'b_conv2') h_conv2 = tf.nn.relu(conv2d(out1, W_conv2) + b_conv2) h_pool2 = max_pool_2x2(h_conv2) out2 = tf.nn.dropout(h_pool2,keep_prob) # batches*10*8 # third layer # W_conv3 = weight_variable([3, 4, 64, 64], 'W_conv3') # b_conv3 = bias_variable([64], 'b_conv3') # # h_conv3 = tf.nn.relu(conv2d(out2, W_conv3) + b_conv3) # h_pool3 = max_pool_2x2(h_conv3) # out3 = tf.nn.dropout(h_pool3,keep_prob) # fourth layer # W_conv4 = weight_variable([3, 3 , 64, 64], 'W_conv4') # b_conv4 = bias_variable([64], 'b_conv4') # # h_conv4 = tf.nn.relu(conv2d(out3, W_conv4) + b_conv4) # h_pool4 = max_pool_2x2(h_conv4) # out4 = tf.nn.dropout(h_pool4,keep_prob) # Fully connected layer W_fc1 = weight_variable([23*35*64, 1024], 'W_fc1') b_fc1 = bias_variable([1024], 'b_fc1') out2_flat = tf.reshape(out2, [-1, 23*35*64]) h_fc1 = tf.nn.relu(tf.matmul(out2_flat, W_fc1) + b_fc1) h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) # batches*1024 # output W_fc2 = weight_variable([1024, 6], 'W_fc2') b_fc2 = bias_variable([6], 'b_fc2') # out=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2) out = tf.matmul(h_fc1_drop, W_fc2) + b_fc2 return out def Train_CNN(): # train and test y_conv=CNN() #cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv)) loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=y_conv, labels=y_)) train_step = tf.train.AdamOptimizer(1.0e-4).minimize(loss) correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1)) accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float")) saver = tf.train.Saver() #%% Run sess=tf.Session() sess.run(tf.global_variables_initializer()) for i in range(10000): NUM=random.sample(range(p2pimgs.shape[0]),64) batch_xs=np.array([p2pimgs[Ind] for Ind in NUM]) batch_ys=np.array([p2plabels[Ind] for Ind in NUM]) sess.run(train_step,feed_dict={x: batch_xs, y_: batch_ys, keep_prob: 0.5}) if i%100 == 0: NUM=random.sample(range(p2pimgs.shape[0]),100) batch_xs_test=np.array([p2pimgs[Ind] for Ind in NUM]) batch_ys_test=np.array([p2plabels[Ind] for Ind in NUM]) print("step %d, training accuracy %f"%(i,sess.run(accuracy,feed_dict={x:batch_xs_test, y_: batch_ys_test, keep_prob: 1.0}))) savePath=saver.save(sess,'./model/p2p.model') print('Save Path: ',savePath) sess.close() if __name__=='__main__': Train_CNN()