input_img = tf.placeholder(dtype='float',
shape=[None, 128, 128, 3],
name="Inputs")
with tf.name_scope("Target") as scope:
target_labels = tf.placeholder(dtype='float',
shape=[None, 2],
name="Targets")
nb = NetworkBuilder()
#with tf.name_scope("ModelWithCustomLayer") as scope:
with tf.name_scope("ModCosh") as scope:
model = input_img
model = nb.attach_cosh_layer(model)
model = nb.attach_conv_layer(model, 32, summary='True')
model = nb.attach_relu_layer(model)
model = nb.attach_conv_layer(model, 32, summary='True')
model = nb.attach_relu_layer(model)
model = nb.attach_pooling_layer(model)
model = nb.attach_cosh_layer(model)
model = nb.attach_conv_layer(model, 64, summary=True)
model = nb.attach_relu_layer(model)
model = nb.attach_conv_layer(model, 64, summary=True)
model = nb.attach_relu_layer(model)
model = nb.attach_pooling_layer(model)
model = nb.attach_cosh_layer(model)
model = nb.attach_conv_layer(model, 128, summary=True)
model = nb.attach_relu_layer(model)
def _setup(self, input, is_training):
num_classes = setting.num_classes
nb = NetworkBuilder(is_training)
keep_prob = tf.cond(is_training, lambda: 0.3, lambda: 1.0)
model = input
model = nb.attach_conv_layer(model,
output_size=32,
feature_size=[3, 3],
strides=[1, 1, 1, 1],
padding=1,
use_bias=True)
model = nb.attach_batch_norm_layer(model)
model = nb.attach_relu_layer(model)
self.conv_3x3_out = model
model = nb.attach_conv_layer(model,
output_size=64,
feature_size=[7, 7],
strides=[1, 2, 2, 1],
padding=3,
use_bias=True)
model = nb.attach_batch_norm_layer(model)
model = nb.attach_relu_layer(model)
self.conv_7x7_out = model
model = nb.attach_pooling_layer(model,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding=0)
self.max_pool_out = model
model = nb.attach_resnet_block_2(model,
d1=64,
d2=256,
strides=[1, 1, 1, 1])
model = nb.attach_resnet_block_1(model, d1=64, d2=256)
model = nb.attach_resnet_block_1(model, d1=64, d2=256)
self.m_b1_out = model
model = nb.attach_resnet_block_2(model,
d1=128,
d2=512,
strides=[1, 2, 2, 1])
model = nb.attach_resnet_block_1(model, d1=128, d2=512)
model = nb.attach_resnet_block_1(model, d1=128, d2=512)
model = nb.attach_multiscale_block_1(model,
d1=128,
d2=512,
d3=128,
p=1,
d=2)
self.m_b2_out = model
#skip branch
skip_connection = model
# self.skip = model
model = nb.attach_resnet_block_2(model,
d1=256,
d2=1024,
strides=[1, 2, 2, 1])
model = nb.attach_resnet_block_1(model, d1=256, d2=1024)
model = nb.attach_multiscale_block_1(model,
d1=256,
d2=1024,
d3=256,
p=1,
d=2)
model = nb.attach_multiscale_block_1(model,
d1=256,
d2=1024,
d3=256,
p=1,
d=4)
model = nb.attach_multiscale_block_1(model,
d1=256,
d2=1024,
d3=256,
p=1,
d=8)
model = nb.attach_multiscale_block_1(model,
d1=256,
d2=1024,
d3=256,
p=1,
d=16)
self.m_b3_out = model
model = nb.attach_multiscale_block_2(model,
d1=512,
d2=2048,
d3=512,
p=2,
d=4)
model = nb.attach_multiscale_block_1(model,
d1=512,
d2=2048,
d3=512,
p=2,
d=8)
model = nb.attach_multiscale_block_1(model,
d1=512,
d2=2048,
d3=512,
p=2,
d=16)
model = nb.add_dropout(model, keep_prob)
self.m_b4_out = model
model = nb.attach_conv_layer(model,
output_size=num_classes,
feature_size=[1, 1],
strides=[1, 1, 1, 1],
padding=0,
use_bias=True)
model = nb.attach_batch_norm_layer(model)
self.out = model
model = nb.attach_conv_transpose_layer(model,
output_size=num_classes * 2,
feature_size=[4, 4],
strides=[2, 2],
padding=1,
use_bias=True,
scale=2)
# model = nb.attach_batch_norm_layer(model)
#adding missing batch norm layer
self.deconv_up1 = model
#is saale ko check kr, kahi naatak na kr de baad mein
skip_connection = nb.attach_conv_layer(skip_connection, output_size=num_classes*2, feature_size=[1,1], strides=[1,1,1,1],\
padding=0, use_bias=True)
skip_connection = nb.attach_batch_norm_layer(skip_connection)
self.skip = skip_connection
model = nb.attach_element_wise_sum(model, skip_connection)
self.up1 = model
model = nb.attach_conv_transpose_layer(model,
output_size=num_classes,
feature_size=[16, 16],
strides=[8, 8],
padding=4,
use_bias=True,
scale=8)
model = nb.attach_batch_norm_layer(model)
self.deconv_up2 = model
#could be removed
# model = nb.attach_cropping_layer(input, model)
# self.cropped_deconv_up2 = model
#could be removed
model = nb.attach_conv_layer(model, output_size=num_classes, feature_size=[1,1], strides=[1,1,1,1], padding="SAME",\
use_bias=False)
self.up2 = model
return model
def train():
dg = DataSetGenerator("training_face_rec")
MAX_LABELS = len(dg.data_labels)
dg.save_labels()
with tf.name_scope("Input") as scope:
input_img = tf.placeholder(dtype='float', shape=[None, 128, 128, 3], name="input")
with tf.name_scope("Target") as scope:
target_labels = tf.placeholder(dtype='float', shape=[None, MAX_LABELS], name="Targets")
with tf.name_scope("Keep_prob_input") as scope:
keep_prob = tf.placeholder(dtype='float',name='keep_prob')
nb = NetworkBuilder()
with tf.name_scope("ModelV2") as scope:
model = input_img
model = nb.attach_conv_layer(model, 32, summary=True)
model = nb.attach_relu_layer(model)
model = nb.attach_conv_layer(model, 32, summary=True)
model = nb.attach_relu_layer(model)
model = nb.attach_pooling_layer(model)
model = nb.attach_conv_layer(model, 64, summary=True)
model = nb.attach_relu_layer(model)
model = nb.attach_conv_layer(model, 64, summary=True)
model = nb.attach_relu_layer(model)
model = nb.attach_pooling_layer(model)
model = nb.attach_conv_layer(model, 128, summary=True)
model = nb.attach_relu_layer(model)
model = nb.attach_conv_layer(model, 128, summary=True)
model = nb.attach_relu_layer(model)
model = nb.attach_pooling_layer(model)
model = nb.flatten(model)
model = nb.attach_dense_layer(model, 200, summary=True)
model = nb.attach_sigmoid_layer(model)
model = nb.attach_dense_layer(model, 32, summary=True)
model = nb.attach_sigmoid_layer(model)
model = nb.attach_dense_layer(model, MAX_LABELS)
prediction = nb.attach_softmax_layer(model)
with tf.name_scope("Optimization") as scope:
global_step = tf.Variable(0, name='global_step', trainable=False)
cost = tf.nn.softmax_cross_entropy_with_logits(logits=model, labels=target_labels)
cost = tf.reduce_mean(cost)
tf.summary.scalar("cost", cost)
optimizer = tf.train.AdamOptimizer().minimize(cost, global_step=global_step)
with tf.name_scope('accuracy') as scope:
correct_pred = tf.equal(tf.argmax(prediction, 1), tf.argmax(target_labels, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
tf.summary.scalar("accuracy", accuracy)
# dg = DataSetGenerator("../data/twins")
test_x , test_y = dg.load_test_images(image_size=(128,128))
epochs = 10
batchSize = 10
saver = tf.train.Saver()
model_save_path="./saved model v2/"
model_name='model'
is_Train = True
with tf.Session() as sess:
summaryMerged = tf.summary.merge_all()
filename = "./summary_log/run" + datetime.datetime.now().strftime("%Y-%m-%d--%H-%M-%s")
# setting global steps
tf.global_variables_initializer().run()
if os.path.exists(model_save_path+'checkpoint'):
# saver = tf.train.import_meta_graph('./saved '+modelName+'/model.ckpt.meta')
saver.restore(sess, tf.train.latest_checkpoint(model_save_path))
if is_Train:
writer = tf.summary.FileWriter(filename+"_train", sess.graph)
test_writer = tf.summary.FileWriter(filename+"_test", sess.graph)
start_time = datetime.datetime.now()
for epoch in range(epochs):
batches = dg.get_mini_batches(batchSize,(128,128), allchannel=True)
for imgs ,labels in batches:
# X_train, X_test, y_train, y_test = train_test_split(imgs, labels, test_size=0.33, random_state=42)
imgs=np.divide(imgs, 255)
error, sumOut, acu, steps,_ = sess.run([cost, summaryMerged, accuracy,global_step,optimizer],
feed_dict={input_img: imgs, target_labels: labels})
writer.add_summary(sumOut, steps)
print("epoch=", epoch,"steps=",steps, "Total Samples Trained=", steps*batchSize, "err=", error, "accuracy=", acu)
if steps % 10 == 0: # Record summaries and test-set accuracy
error, sumOut, acu, steps = sess.run([cost,summaryMerged, accuracy,global_step], feed_dict={input_img: test_x, target_labels: test_y})
test_writer.add_summary(sumOut, steps)
print("Test : error= ", error, "epoch=", epoch,"steps=",steps, "accuracy=", acu)
if steps % 100 == 0:
print("Saving the model")
saver.save(sess, model_save_path+model_name, global_step=steps)
saver.save(sess, model_save_path+model_name)
sess.close()
end_time = datetime.datetime.now()
print(diff_time(start_time,end_time))