def read_test_data(path):
num_clips = int(sys.argv[1])
labels = pd.read_csv(os.path.join(path,'Label_Map','label.txt'))
all_clips_name = rd.read_dataset(path,labels,'Test',seed=66,balance=False)
mean_image = np.load(os.path.join(path,'Data','Train','mean_image.npy'))
clip_Y,clip_X = rd.read_minibatch(0,num_clips,all_clips_name,mean_image,'Test')
return clip_Y,clip_X
def training_net():
all_clips_name, mean_image = load_clip_name(parameters.path, 'Train', True)
clip_X, clip_Y, isTraining = net_placeholder(None)
logits, Softmax_output = VTN.vtn(clip_X, isTraining)
loss = net_loss(clip_Y, logits)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_step = tf.train.AdamOptimizer(
parameters.LEARNING_RATE).minimize(loss)
Saver = tf.train.Saver(var_list=tf.global_variables())
with tf.Session() as sess:
writer = tf.summary.FileWriter(os.path.join(parameters.path, 'Model'),
sess.graph)
init_var_op = tf.global_variables_initializer()
sess.run(init_var_op)
for i in range(parameters.TRAIN_STEPS):
Y, X = rd.read_minibatch(i, parameters.BATCH_SIZE, all_clips_name,
mean_image, 'Train')
feed_dict = {clip_X: X, clip_Y: Y, isTraining: True}
_, loss_ = sess.run([train_step, loss], feed_dict=feed_dict)
if i % 100 == 0:
# accuracy on val clips
val_acc = val(sess, clip_X, clip_Y, isTraining, Softmax_output,
1, parameters.path, 'Val', False)
print('\nVal_accuracy = %g\n' % (val_acc))
# Way 1 : saving checkpoint model
if sys.argv[1] == 'CHECKPOINT':
if os.path.exists(
parameters.CHECKPOINT_MODEL_SAVE_PATH) and (i
== 0):
shutil.rmtree(parameters.CHECKPOINT_MODEL_SAVE_PATH)
Saver.save(
sess,
os.path.join(parameters.CHECKPOINT_MODEL_SAVE_PATH,
parameters.MODEL_NAME + str(i)))
# Way 2 : saving pb model
elif sys.argv[1] == 'PB':
if os.path.exists(parameters.PB_MODEL_SAVE_PATH):
shutil.rmtree(parameters.PB_MODEL_SAVE_PATH)
save_inference_model.save_model(
sess, parameters.PB_MODEL_SAVE_PATH, clip_X,
Softmax_output, isTraining)
else:
print(
'The argument is incorrect for the way saving model!')
sys.exit(0)
print('===>Step %d: loss = %g ' % (i, loss_))
def val(sess,clip_X,clip_Y,Softmax_output,test_batch_size,path,status,balance):
all_clips_name,mean_image = load_clip_name(path,status,balance)
acc_count = 0
for j in range(len(all_clips_name)):
if (j*test_batch_size)>len(all_clips_name):
break
Y,X = rd.read_minibatch(j,test_batch_size,all_clips_name,mean_image,status)
feed_dict = {clip_X:X,clip_Y:Y}
softmax = sess.run(Softmax_output,feed_dict=feed_dict)
# Compute clip-level accuracy
for one_output,one_clip_Y in zip(softmax,Y):
if np.argmax(one_output) == np.argmax(one_clip_Y):
acc_count += 1
accuracy = (acc_count/(len(all_clips_name)*1.0))
return accuracy