def reshape_data(X, Y):
# X=X.reshape([-1,3, 32,32]).transpose((0, 2, 3, 1))
# print("Xshape", X.shape)
ig=ImageGenerator(X, Y)
ig.translate(np.random.randint(low=-5, high=5), np.random.randint(low=-5, high=5))
# print("translate", ig.x.shape)
# ig.rotate(np.random.randint(low=-30, high=30))
# print("rotate", ig.x.shape)
ig.flip("h")
# print("flip", ig.x.shape)
ig.add_noise(np.random.uniform(0, 0.2), np.random.uniform(1, 3))
# print("noise", ig.x.shape)
return ig.x
def my_training4(X_train,
y_train,
X_val,
y_val,
conv_featmap=[6],
fc_units=[84],
conv_kernel_size=[5],
pooling_size=[2],
l2_norm=0.01,
seed=235,
learning_rate=1e-2,
epoch=20,
batch_size=245,
verbose=False,
pre_trained_model=None):
print("Building my LeNet. Parameters: ")
print("conv_featmap={}".format(conv_featmap))
print("fc_units={}".format(fc_units))
print("conv_kernel_size={}".format(conv_kernel_size))
print("pooling_size={}".format(pooling_size))
print("l2_norm={}".format(l2_norm))
print("seed={}".format(seed))
print("learning_rate={}".format(learning_rate))
# define the variables and parameter needed during training
with tf.name_scope('inputs'):
xs = tf.placeholder(shape=[None, 32, 32, 3], dtype=tf.float32)
ys = tf.placeholder(shape=[
None,
], dtype=tf.int64)
output, loss = LeNet(xs,
ys,
img_len=32,
channel_num=3,
output_size=10,
conv_featmap=conv_featmap,
fc_units=fc_units,
conv_kernel_size=conv_kernel_size,
pooling_size=pooling_size,
l2_norm=l2_norm,
seed=seed)
iters = int(X_train.shape[0] / batch_size)
print('number of batches for training: {}'.format(iters))
step = train_step(loss)
eve = evaluate(output, ys)
iter_total = 0
best_acc = 0
cur_model_name = 'lenet_{}'.format(int(time.time()))
with tf.Session() as sess:
merge = tf.summary.merge_all()
writer = tf.summary.FileWriter("log/{}".format(cur_model_name),
sess.graph)
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
# try to restore the pre_trained
if pre_trained_model is not None:
try:
print("Load the model from: {}".format(pre_trained_model))
saver.restore(sess, 'model/{}'.format(pre_trained_model))
except Exception:
print("Load model Failed!")
pass
for epc in range(epoch):
print("epoch {} ".format(epc + 1))
for itr in range(iters):
iter_total += 1
training_batch_x = X_train[itr * batch_size:(1 + itr) *
batch_size]
training_batch_y = y_train[itr * batch_size:(1 + itr) *
batch_size]
mode = epc % 5
data_augment = ImageGenerator(training_batch_x,
training_batch_y)
if mode == 0:
training_batch_x = training_batch_x
elif mode == 1:
data_augment.translate(8, 4)
training_batch_x = data_augment.x
elif mode == 2:
data_augment.flip('v')
training_batch_x = data_augment.x
elif mode == 3:
data_augment.flip('h')
training_batch_x = data_augment.x
elif mode == 4:
data_augment.add_noise(0.5, 1e-5)
_, cur_loss = sess.run([step, loss],
feed_dict={
xs: training_batch_x,
ys: training_batch_y
})
if iter_total % 100 == 0:
# do validation
valid_eve, merge_result = sess.run([eve, merge],
feed_dict={
xs: X_val,
ys: y_val
})
valid_acc = 100 - valid_eve * 100 / y_val.shape[0]
if verbose:
print(
'{}/{} loss: {} validation accuracy : {}%'.format(
batch_size * (itr + 1), X_train.shape[0],
cur_loss, valid_acc))
# save the merge result summary
writer.add_summary(merge_result, iter_total)
# when achieve the best validation accuracy, we store the model paramters
if valid_acc > best_acc:
print(
'Best validation accuracy! iteration:{} accuracy: {}%'
.format(iter_total, valid_acc))
best_acc = valid_acc
saver.save(sess, 'model/{}'.format(cur_model_name))
print(
"Traning ends. The best valid accuracy is {}. Model named {}.".format(
best_acc, cur_model_name))
def my_training_task4(X_train,
y_train,
X_val,
y_val,
conv_featmap=[6],
fc_units=[84],
conv_kernel_size=[5],
pooling_size=[2],
l2_norm=0.01,
seed=235,
learning_rate=1e-2,
epoch=20,
batch_size=245,
verbose=False,
pre_trained_model=None):
# TODO: Copy my_training function, make modifications so that it uses your
# data generator from task 4 to train.
print("Building my LeNet. Parameters: ")
print("conv_featmap={}".format(conv_featmap))
print("fc_units={}".format(fc_units))
print("conv_kernel_size={}".format(conv_kernel_size))
print("pooling_size={}".format(pooling_size))
print("l2_norm={}".format(l2_norm))
print("seed={}".format(seed))
print("learning_rate={}".format(learning_rate))
num_of_samples, height, width, channels = X_train.shape
output_size = np.unique(y_train).shape[0]
# define the variables and parameter needed during training
with tf.name_scope('inputs'):
xs = tf.placeholder(shape=[None, height, width, channels],
dtype=tf.float32)
ys = tf.placeholder(shape=[
None,
], dtype=tf.int64)
output, loss = my_LeNet(xs,
ys,
img_len=32,
channel_num=channels,
output_size=output_size,
conv_featmap=conv_featmap,
fc_units=fc_units,
conv_kernel_size=conv_kernel_size,
pooling_size=pooling_size,
l2_norm=l2_norm,
seed=seed)
iters = int(X_train.shape[0] / batch_size)
print('number of batches for training: {}'.format(iters))
step = train_step(loss)
eve = evaluate(output, ys)
iter_total = 0
best_acc = 0
cur_model_name = 'lenet_{}'.format(int(time.time()))
with tf.Session() as sess:
merge = tf.summary.merge_all()
writer = tf.summary.FileWriter("log/{}".format(cur_model_name),
sess.graph)
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
# try to restore the pre_trained
if pre_trained_model is not None:
try:
print("Load the model from: {}".format(pre_trained_model))
saver.restore(sess, 'model/{}'.format(pre_trained_model))
except Exception:
print("Load model Failed!")
pass
#bat = batch_size/4
print("Generating data")
origin = ImageGenerator(X_train, y_train)
#flip = ImageGenerator(X_train,y_train)
#flip.flip(mode='h')
#noise = ImageGenerator(X_train,y_train)
#noise.add_noise(portion=1, amplitude=1)
#rotate = ImageGenerator(X_train,y_train)
#rotate.rotate(angle=45)
translate = ImageGenerator(X_train, y_train)
generator = []
generator.append(origin.next_batch_gen(batch_size))
#generator.append(flip.next_batch_gen(batch_size))
#generator.append(noise.next_batch_gen(batch_size))
#generator.append(rotate.next_batch_gen(batch_size))
generator.append(translate.next_batch_gen(batch_size))
print("Data generation finished")
for epc in range(epoch):
print("epoch {} ".format(epc + 1))
shift = np.random.choice(np.arange(1, 11), size=2, replace=True)
translate = ImageGenerator(X_train, y_train)
translate.translate(shift_height=shift[0], shift_width=shift[1])
generator[-1] = translate.next_batch_gen(batch_size)
for itr in range(iters):
iter_total += 1
#index = np.random.choice(X_train.shape[0],batch_size)
#training_batch_x = X_train[index]
#training_batch_y = y_train[index]
training_batch_x, training_batch_y = next(
generator[itr % len(generator)])
_, cur_loss = sess.run([step, loss],
feed_dict={
xs: training_batch_x,
ys: training_batch_y
})
if iter_total % 100 == 0:
# do validation
valid_eve, merge_result = sess.run([eve, merge],
feed_dict={
xs: X_val,
ys: y_val
})
valid_acc = 100 - valid_eve * 100 / y_val.shape[0]
if verbose:
print(
'{}/{} loss: {} validation accuracy : {}%'.format(
batch_size * (itr + 1), X_train.shape[0],
cur_loss, valid_acc))
# save the merge result summary
writer.add_summary(merge_result, iter_total)
# when achieve the best validation accuracy, we store the model paramters
if valid_acc > best_acc:
print(
'Best validation accuracy! iteration:{} accuracy: {}%'
.format(iter_total, valid_acc))
best_acc = valid_acc
saver.save(sess, 'model/{}'.format(cur_model_name))
print(
"Traning ends. The best valid accuracy is {}. Model named {}.".format(
best_acc, cur_model_name))