def iou(params0, params1):
row0, col0, rad0 = params0
row1, col1, rad1 = params1
shape0 = Point(row0, col0).buffer(rad0)
shape1 = Point(row1, col1).buffer(rad1)
return (shape0.intersection(shape1).area / shape0.union(shape1).area)
def Trainer(sess, training_size = 5000, batch_size = 32, validation_data_size = 1000, epoches = 100, scope_name = 'vars'):
"""
This function does the training using default or user-defined paremeters
sess: The tensorflow session
train_size: the # of batches used in the training
batch_size: the # of samples in each batch
validation_data_size: the # of samples used in validation
epoches: epoches of training
scope_name: the variable scope for trainable variables
"""
inputs = tf.placeholder(tf.float32, shape = (None, 200, 200, 1))
outputs = tf.placeholder(tf.float32, shape = (None, 3))
train_loss, logits, train_op = ConvNet(inputs, outputs)
# save the checkpoints
saved_variables = tf.global_variables()
saver = tf.train.Saver(saved_variables)
#variables initialization
sess.run(tf.global_variables_initializer())
# model training/validation, save the best-validated checkpoint
maximal_accu = -1
for epoch in range(epoches):
# training steps
losses = []
for batch in range(training_size):
# generate a batch of samples
for idx in range(batch_size):
params, img = noisy_circle(200, 50, 2)
img = np.reshape(img, (1, 200, 200, 1))
params = np.reshape(np.array(params), (1, 3))
if idx == 0:
X = img
Y = params
else:
X = np.concatenate((X, img), axis = 0)
Y = np.concatenate((Y, params), axis = 0)
with tf.variable_scope(scope_name, reuse = True):
loss, predicts, _ = sess.run([train_loss, logits, train_op], feed_dict = {inputs: X, outputs: Y})
losses.append(loss)
print('The training loss at epoch: ' + str(epoch) + ' is: ' + str(np.mean(np.array(losses))))
# validation steps
predictions = ConvNet(inputs, outputs, mode = 'predict')
ious = []
for batch in range(validation_data_size):
params, img = noisy_circle(200, 50, 2)
img = np.reshape(img, (1, 200, 200, 1))
X = img
Y = np.reshape(np.array(params), (1,3))
with tf.variable_scope(scope_name, reuse = True):
predict = sess.run(predictions, feed_dict = {inputs: X, outputs: Y})
row0, col0, rad0 = np.squeeze(Y).tolist()
row1, col1, rad1 = np.squeeze(predict).tolist()
shape0 = Point(row0, col0).buffer(rad0)
shape1 = Point(row1, col1).buffer(rad1)
iou = shape0.intersection(shape1).area / shape0.union(shape1).area
ious.append(iou)
validation_accu = (np.array(ious) > 0.7).mean()
print('The validation accuracy at epoch: ' + str(epoch) + ' is: ' + str(validation_accu))
if validation_accu >= maximal_accu:
maximal_accu = validation_accu
print('----------------------------------------Better checkpoint found at epoch: ' + str(epoch))
os.system('rm checkpoints/*')
saver.save(sess, 'checkpoints/dump', global_step=epoch)
