def train_inception_v4_stacked_lstm(width,
height,
n_lstm=7,
dropout_keep_prob=0.8,
learning_rate=0.00001,
n_epochs=200):
print('...... loading the dataset ......')
train_set_x,train_set_y,test_set_x,test_set_y = pd.load_data_set(width,height)
x = tf.placeholder(tf.float32,[None,width*height]) # input
y = tf.placeholder(tf.float32,[None,n_classes]) # label
keep_prob = tf.placeholder(tf.float32) # dropout_keep_prob
y_inception = inception_v4(x,width,height,keep_prob)
y_pred = stacked_lstm_layer(y_inception,n_lstm)
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_pred,y))
optimizer = tf.train.RMSPropOptimizer(learning_rate).minimize(cost)
correct_prediction = tf.equal(tf.argmax(y_pred,1),tf.argmax(y,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
best_acc = 0.
best_acc_epoch = 0
init = tf.initialize_all_variables()
with tf.Session() as sess:
print('...... initializating varibale ...... ')
sess.run(init)
print('...... start to training ......')
for epoch_i in range(n_epochs):
# Training
train_accuracy = 0.
for batch_i in range(n_train_example//pic_batch_size):
batch_xs = train_set_x[batch_i*pic_batch_size:(batch_i+1)*pic_batch_size]
batch_ys = train_set_y[batch_i*video_batch_size:(batch_i+1)*video_batch_size]
_,loss,acc = sess.run([optimizer, cost, accuracy],
feed_dict={
x:batch_xs,
y:batch_ys,
keep_prob:dropout_keep_prob}
)
print('epoch:{0},minibatch:{1},cost:{2},train_accuracy:{3}'.format(epoch_i,batch_i,loss,acc))
train_accuracy += acc
train_accuracy /= (n_train_example//pic_batch_size)
print('----epoch:{0},training acc = {1}'.format(epoch_i,train_accuracy))
# Validation
valid_accuracy = 0.
for batch_i in range(n_test_example//pic_batch_size):
batch_xs = test_set_x[batch_i*pic_batch_size:(batch_i+1)*pic_batch_size]
batch_ys = test_set_y[batch_i*video_batch_size:(batch_i+1)*video_batch_size]
valid_accuracy += sess.run(accuracy,
feed_dict={
x:batch_xs,
y:batch_ys,
keep_prob:1.0})
valid_accuracy /= (n_test_example//pic_batch_size)
print('epoch:{0},train_accuracy:{1},valid_accuracy:{2}'.format(epoch_i,train_accuracy,valid_accuracy))
if(train_accuracy > best_acc):
best_acc_epoch = epoch_i
best_acc = train_accuracy
print('========epoch:{0},current best accuracy = {1} in epoch_{2}'.format(epoch_i,best_acc,best_acc_epoch))
print('...... training finished ......')
print('...... best accuracy{0} @ epoch_{1}......'.format(best_acc,best_acc_epoch))
def train_inception_v3(width=256, height=256):
d_start = datetime.datetime.now()
print('...... loading the dataset ......')
train_set_x, train_set_y, test_set_x, test_set_y = pd.load_data_set(
width, height)
x = tf.placeholder(tf.float32, [None, width * height]) # input
y = tf.placeholder(tf.float32, [None, n_class]) # label
keep_prob = tf.placeholder(tf.float32) # dropout_keep_prob
y_inception = inception_v3(x, width, height)
y_pred = prediction(y_inception, keep_prob)
cost = tf.reduce_mean(
-tf.reduce_sum(y * tf.log(y_pred), reduction_indices=1))
optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)
correct_prediction = tf.equal(tf.argmax(y_pred, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
best_acc = 0.
init = tf.initialize_all_variables()
with tf.Session() as sess:
print('...... initializating varibale ...... ')
sess.run(init)
n_epochs = 100
print('...... start to training ......')
for epoch_i in range(n_epochs):
# Training
train_accuracy = 0.
for batch_i in range(n_train_example // batch_size):
batch_xs = train_set_x[batch_i * batch_size:(batch_i + 1) *
batch_size]
batch_ys = train_set_y[batch_i * batch_size:(batch_i + 1) *
batch_size]
_, loss, acc = sess.run([optimizer, cost, accuracy],
feed_dict={
x: batch_xs,
y: batch_ys,
keep_prob: dropout_keep_prob
})
print('epoch:{0},minibatch:{1},cost:{2},train_accuracy:{3}'.
format(epoch_i, batch_i, loss, acc))
train_accuracy += acc
train_accuracy /= (n_train_example // batch_size)
print('----epoch:{0},training acc = {1}'.format(
epoch_i, train_accuracy))
# Validation
valid_accuracy = 0.
for batch_i in range(n_test_example // batch_size):
batch_xs = test_set_x[batch_i * batch_size:(batch_i + 1) *
batch_size]
batch_ys = test_set_y[batch_i * batch_size:(batch_i + 1) *
batch_size]
valid_accuracy += sess.run(accuracy,
feed_dict={
x: batch_xs,
y: batch_ys,
keep_prob: 1.0
})
valid_accuracy /= (n_test_example // batch_size)
print('epoch:{0},train_accuracy:{1},valid_accuracy:{2}'.format(
epoch_i, train_accuracy, valid_accuracy))
if (train_accuracy > best_acc):
best_acc = train_accuracy
d_end = datetime.datetime.now()
print('...... training finished ......')
print('...... best accuracy:{0} ......'.format(best_acc))
print('...... running time:{0}'.format((d_end - d_start).seconds))
def train_yawn(width=256, height=256):
d_start = datetime.datetime.now()
print('...... loading the dataset ......')
train_set_x, train_set_y, test_set_x, test_set_y = pd.load_data_set(
width, height)
train_mean = np.mean(train_set_x, axis=0)
test_mean = np.mean(test_set_x, axis=0)
x = tf.placeholder(tf.float32, [None, width * height]) # input
y = tf.placeholder(tf.float32, [None, n_class]) # label
corrupt_prob = tf.placeholder(tf.float32, [1])
ae = autoencoder(x,
corrupt_prob,
width,
height,
dimensions=[5000, 300, 200])
y_pred = classify(ae['z'])
print('y_pred,shape = {}'.format(y_pred.get_shape()))
cost = ae['cost']
optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)
correct_prediction = tf.equal(tf.argmax(y_pred, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
best_acc = 0.
init = tf.initialize_all_variables()
with tf.Session() as sess:
print('...... initializating varibale ...... ')
sess.run(init)
n_epochs = 5
print('...... start to training ......')
for epoch_i in range(n_epochs):
# Training
train_accuracy = 0.
for batch_i in range(n_train_example // batch_size):
batch_xs = train_set_x[batch_i * batch_size:(batch_i + 1) *
batch_size]
batch_xs_norm = np.array(
[img - train_mean for img in batch_xs])
batch_ys = train_set_y[batch_i * batch_size:(batch_i + 1) *
batch_size]
_, loss, acc = sess.run([optimizer, cost, accuracy],
feed_dict={
x: batch_xs_norm,
y: batch_ys,
corrupt_prob: [1.0]
})
#print('epoch:{0},minibatch:{1},y_res:{2}'.format(epoch_i,batch_i,yy_res))
#print('epoch:{0},minibatch:{1},y_pred:{2}'.format(epoch_i,batch_i,yy_pred))
print('epoch:{0},minibatch:{1},cost:{2},train_accuracy:{3}'.
format(epoch_i, batch_i, loss, acc))
train_accuracy += acc
train_accuracy /= (n_train_example // batch_size)
print('----epoch:{0},training acc = {1}'.format(
epoch_i, train_accuracy))
# Validation
valid_accuracy = 0.
for batch_i in range(n_test_example // batch_size):
batch_xs = test_set_x[batch_i * batch_size:(batch_i + 1) *
batch_size]
batch_xs_norm = np.array([img - test_mean for img in batch_xs])
batch_ys = test_set_y[batch_i * batch_size:(batch_i + 1) *
batch_size]
valid_accuracy += sess.run(accuracy,
feed_dict={
x: batch_xs,
y: batch_ys,
corrupt_prob: [0.0]
})
valid_accuracy /= (n_test_example // batch_size)
print('epoch:{0},train_accuracy:{1},valid_accuracy:{2}'.format(
epoch_i, train_accuracy, valid_accuracy))
if (train_accuracy > best_acc):
best_acc = train_accuracy
# draw
n_examples = 10
test_xs = test_set_x[n_examples * 130:n_examples * 131]
test_xs_norm = np.array([img - train_mean for img in test_xs])
recon = sess.run(ae['y'],
feed_dict={
x: test_xs_norm,
corrupt_prob: [0.0]
})
fig, axs = plt.subplots(2, n_examples, figsize=(10, 2))
for example_i in range(n_examples):
axs[0][example_i].imshow(
np.reshape(test_xs[example_i, :], (width, height)))
axs[1][example_i].imshow(
np.reshape([recon[example_i, :] + train_mean],
(width, height)))
fig.show()
plt.draw()
plt.waitforbuttonpress()
d_end = datetime.datetime.now()
print('...... training finished ......')
print('...... best accuracy:{0} ......'.format(best_acc))
print('...... running time:{0} minutes ......'.format(
(d_end - d_start).seconds / 60))
def train_res_blstm(width=256,height=256):
print('...... loading the dataset ......')
train_set_x,train_set_y,test_set_x,test_set_y = pd.load_data_set(width,height)
print('...... building the model ......')
x = tf.placeholder(tf.float32,[None,width*height])
y = tf.placeholder(tf.float32,[None,n_classes])
# TensorFlow LSTM cell reuqires 2*n_hidden length(state&cell)
istate_fw = tf.placeholder(tf.float32,[None,2*n_hidden_units])
istate_bw = tf.placeholder(tf.float32,[None,2*n_hidden_units])
y_res = residual_network(x)
y_pred = blstm(y_res,istate_fw,istate_bw)
# Define loss and training functions
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_pred,y))
optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)
# Monitor Accuracy
correct_prediction = tf.equal(tf.argmax(y_pred,1),tf.argmax(y,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
best_acc = 0.
init = tf.initialize_all_variables()
# Session
with tf.Session() as sess:
print('...... initializating varibale ...... ')
sess.run(init)
n_epochs = 100
print('...... start to training ......')
for epoch_i in range(n_epochs):
# Training
train_accuracy = 0.
for batch_i in range(n_train_example//pic_batch_size):
batch_xs = train_set_x[batch_i*pic_batch_size:(batch_i+1)*pic_batch_size]
batch_ys = train_set_y[batch_i*video_batch_size:(batch_i+1)*video_batch_size]
_,loss,acc,yy = sess.run([optimizer,cost,accuracy,y_pred],
feed_dict={
x:batch_xs,
y:batch_ys,
istate_fw:np.zeros((video_batch_size,2*n_hidden_units)),
istate_bw:np.zeros((video_batch_size,2*n_hidden_units))
})
print('epoch:{0},minibatch:{1},cost:{2},train_accuracy:{3}'.format(epoch_i,batch_i,loss,acc))
train_accuracy += acc
train_accuracy /= (n_train_example//pic_batch_size)
print('epoch:{0},training acc = {1}'.format(epoch_i,train_accuracy))
# Validation
valid_accuracy = 0.
for batch_i in range(n_test_example//pic_batch_size):
batch_xs = test_set_x[batch_i*pic_batch_size:(batch_i+1)*pic_batch_size]
batch_ys = test_set_y[batch_i*video_batch_size:(batch_i+1)*video_batch_size]
valid_accuracy += sess.run(accuracy,
feed_dict={
x:batch_xs,
y:batch_ys,
istate_fw:np.zeros((video_batch_size,2*n_hidden_units)),
istate_bw:np.zeros((video_batch_size,2*n_hidden_units))
})
valid_accuracy /= (n_test_example//pic_batch_size)
print('====epoch:{0},train_accuracy:{1},valid_accuracy:{2}'.format(epoch_i,train_accuracy,valid_accuracy))
if(train_accuracy > best_acc):
best_acc = train_accuracy
print('...... training finished ......')
print('...... best accuracy{0} ......'.format(best_acc))