EN1_h_conv4 = conv2d(EN1_a_pool3, EN1_W_conv4) + EN1_b_conv4 #outsize = batch*1*2*80 EN1_a_conv4 = tf.nn.tanh(EN1_h_conv4) #outsize = batch*1*2*80 ## flatten layer EN1_x_flat = tf.reshape(EN1_a_conv4, [-1,160]) #outsize = batch*160 ## fully connected layer 1 EN1_W_fc1 = weight_variable([160,100]) EN1_b_fc1 = bias_variable([100]) EN1_h_fc1 = tf.matmul(EN1_x_flat, EN1_W_fc1) + EN1_b_fc1 #outsize = batch*100 EN1_a_fc1 = tf.nn.relu(EN1_h_fc1) #outsize = batch*100 EN1_a_fc1_dropout = tf.nn.dropout(EN1_a_fc1, keep_prob) #dropout layer 1 ## fully connected layer 2 EN1_W_fc2 = nl.weight_variable([100,6]) EN1_b_fc2 = nl.bias_variable([6]) EN1_h_fc2 = tf.matmul(EN1_a_fc1_dropout, EN1_W_fc2) + EN1_b_fc2 #outsize = batch*6 EN1_a_fc2 = tf.nn.relu(EN1_h_fc2) #outsize = batch*6 #regularization and loss function original_cost = tf.reduce_mean(tf.pow(y - EN1_a_fc2, 2)) tv = tf.trainable_variables() #L2 regularization regularization_cost = 2 * tf.reduce_sum([ tf.nn.l2_loss(v) for v in tv ]) #2 is a hyper parameter cost = original_cost + regularization_cost Optimizer = tf.train.AdamOptimizer(0.0001).minimize(cost) init = tf.global_variables_initializer() #average accuracy every batch accuracy = tf.reduce_mean((y - EN1_a_fc2), 0) #average accuracy every batch testaccuracy = np.zeros([27,6], dtype = np.float32) #save the model saver = tf.train.Saver()
newlandmarks[0] = (1-rx) / 2 * 39 #ratio x
newlandmarks[1] = (1-ry) / 2 * 39 #ratio y
#one dimension which represents one grey picture, set the first dimension as index
x_test[14390+i*10+j,:,:] = imagematrix
y_test[14390+i*10+j,:] = newlandmarks
## RE31
x = tf.placeholder(tf.float32, shape=[None,15,15], name='x') #input imagematrix_data to be fed
y = tf.placeholder(tf.float32, shape=[None,2], name='y') #correct output to be fed
keep_prob = tf.placeholder(tf.float32, name='keep_prob') #keep_prob parameter to be fed
x_image = tf.reshape(x, [-1,15,15,1])
## convolutional layer 1, kernel 4*4, insize 1, outsize 20
W_conv1 = nl.weight_variable([4,4,1,20])
b_conv1 = nl.bias_variable([20])
h_conv1 = nl.conv_layer(x_image, W_conv1) + b_conv1 #outsize = batch*12*12*20
a_conv1 = tf.nn.tanh(h_conv1) #outsize = batch*12*12*20
## max pooling layer 1
h_pool1 = nl.max_pool_22_layer(a_conv1) #outsize = batch*6*6*20
a_pool1 = tf.nn.tanh(h_pool1) #outsize = batch*6*6*20
## convolutional layer 2, kernel 3*3, insize 20, outsize 40
W_conv2 = nl.weight_variable([3,3,20,40])
b_conv2 = nl.bias_variable([40])
h_conv2 = nl.conv_layer(a_pool1, W_conv2) + b_conv2 #outsize = batch*4*4*40
a_conv2 = tf.nn.tanh(h_conv2) #outsize = batch*4*4*40
## max pooling layer 2
h_pool2 = nl.max_pool_22_layer(a_conv2) #outsize = batch*2*2*40
x_test[i + 11439, :, :] = imagematrix
y_test[i + 11439, :] = newlandmarks
## F1
x = tf.placeholder(tf.float32, shape=[None, 39, 39],
name='x') #input imagematrix_data to be fed
y = tf.placeholder(tf.float32, shape=[None, 10],
name='y') #correct output to be fed
keep_prob = tf.placeholder(tf.float32,
name='keep_prob') #keep_prob parameter to be fed
x_image = tf.reshape(x, [-1, 39, 39, 1])
## convolutional layer 1, kernel 4*4, insize 1, outsize 20
W_conv1 = nl.weight_variable([4, 4, 1, 20])
b_conv1 = nl.bias_variable([20])
h_conv1 = nl.conv_layer(x_image, W_conv1) + b_conv1 #outsize = batch*36*36*20
a_conv1 = tf.nn.relu(h_conv1) #outsize = batch*36*36*20
## max pooling layer 1
h_pool1 = nl.max_pool_22_layer(a_conv1) #outsize = batch*18*18*20
a_pool1 = tf.nn.relu(h_pool1) #outsize = batch*18*18*20
## convolutional layer 2, kernel 3*3, insize 20, outsize 40
W_conv2 = nl.weight_variable([3, 3, 20, 40])
b_conv2 = nl.bias_variable([40])
h_conv2 = nl.conv_layer(a_pool1, W_conv2) + b_conv2 #outsize = batch*16*16*40
a_conv2 = tf.nn.relu(h_conv2) #outsize = batch*16*16*40
## max pooling layer 2
h_pool2 = nl.max_pool_22_layer(a_conv2) #outsize = batch*8*8*40
