def conTestImageBoundaryGenerator(filepath, batch_size, depth, num_classes, modality):
X_test = data.load_con_video_list(filepath)
X_teidx = np.asarray(np.arange(0, len(X_test)), dtype=np.int32)
while 1:
for X_indices,_ in minibatches(X_teidx, X_teidx,
batch_size, shuffle=False):
# Read data for each batch
video_label = []
idx = X_indices[0]
video_path = X_test[idx].split(' ')[0]
segcnt = len(X_test[idx].split(' '))
starti = endi = 0
for i in range(1, segcnt):
seginfo = X_test[idx].split(' ')[i]
starti = int(seginfo.split(',')[0])
if starti <= endi:
starti = endi + 1
endi = int(seginfo.split(',')[1].split(':')[0])
label = int(seginfo.split(',')[1].split(':')[1])-1
for j in range(starti, endi+1):
video_label.append(label)
if endi != len(video_label):
print 'invalid: endi - %d, len(video_label) - %d'%(endi, len(video_label))
video_fcnt = len(video_label)
if len(video_label)<=depth:
video_olen = len(video_label)
else:
video_olen = depth
is_training = False # Testing
if modality==0: #RGB
X_data_t,y_label = data.prepare_con_rgb_data(video_path, video_fcnt, video_olen, video_label, is_training)
if modality==1: #Depth
X_data_t,y_label = data.prepare_con_depth_data(video_path, video_fcnt, video_olen, video_label, is_training)
if modality==2: #Flow
X_data_t,y_label = data.prepare_con_flow_data(video_path, video_fcnt, video_olen, video_label, is_training)
y_bound = np.zeros((len(y_label),), dtype=np.int32)
for idx in range(2,len(y_label)-2):
if y_label[idx-1]==y_label[idx] and y_label[idx+1]==y_label[idx+2] and y_label[idx]!=y_label[idx+1]:
y_bound[idx-1]=1
y_bound[idx]=1
y_bound[idx+1]=1
y_bound[idx+2]=1
y_bound[0]=y_bound[1]=1
y_bound[len(y_label)-1]=y_bound[len(y_label)-2]=1
yield (np.reshape(X_data_t,(1,video_olen,112,112,3)), y_bound)
def conTestImageGenerator(filepath, batch_size, depth, num_classes, modality):
X_test = data.load_con_video_list(filepath)
X_teidx = np.asarray(np.arange(0, len(X_test)), dtype=np.int32)
while 1:
for X_indices,_ in minibatches(X_teidx, X_teidx,
batch_size, shuffle=False):
# Read data for each batch
image_path = []
image_fcnt = []
image_olen = []
image_start = []
is_training = []
y_label_t = []
for data_a in range(batch_size):
X_index_a = X_indices[data_a]
# Read data for each batch
idx = X_indices[data_a]
video_path = X_test[idx].split(' ')[0]
starti = int(X_test[idx].split(' ')[1].split(',')[0])
endi = int(X_test[idx].split(' ')[1].split(',')[1].split(':')[0])
label = int(X_test[idx].split(' ')[1].split(',')[1].split(':')[1])-1
image_path.append(video_path)
image_fcnt.append(endi-starti+1)
image_olen.append(depth)
image_start.append(starti)
is_training.append(False) # Testing
y_label_t.append(label)
image_info = zip(image_path,image_fcnt,image_olen,image_start,is_training)
if modality==0: #RGB
X_data_t = threading_data([_ for _ in image_info],
data.prepare_iso_rgb_data)
elif modality==1: #Depth
X_data_t = threading_data([_ for _ in image_info],
data.prepare_iso_depth_data)
elif modality==2: #Flow
X_data_t = threading_data([_ for _ in image_info],
data.prepare_iso_flow_data)
y_hot_label_t = keras.utils.to_categorical(y_label_t, num_classes=num_classes)
yield (X_data_t, y_hot_label_t)
from datetime import datetime
RGB = 0
nb_epoch = 10
init_epoch = 0
depth = 32
batch_size = 8
num_classes = 249
weight_decay = 0.00005
dataset_name = 'congr_rcm_rgb'
training_datalist = './dataset_splits/ConGD/train_rgb_isolist.txt'
testing_datalist = './dataset_splits/ConGD/valid_rgb_isolist.txt'
model_prefix = '.'
weights_file = '%s/trained_models/rcm/%s_weights.{epoch:02d}-{val_loss:.2f}.h5'%(model_prefix,dataset_name)
train_data = data.load_con_video_list(training_datalist)
train_steps = len(train_data)/batch_size
test_data = data.load_con_video_list(testing_datalist)
test_steps = len(test_data)/batch_size
print 'nb_epoch: %d - depth: %d - batch_size: %d - weight_decay: %.6f' %(nb_epoch, depth, batch_size, weight_decay)
def lr_polynomial_decay(global_step):
learning_rate = 0.001
end_learning_rate=0.000001
decay_steps=train_steps*nb_epoch
power = 0.9
p = float(global_step)/float(decay_steps)
lr = (learning_rate - end_learning_rate)*np.power(1-p, power)+end_learning_rate
if global_step>0:
curtime = '%s' % datetime.now()
info = ' - lr: %.6f @ %s %d' %(lr, curtime.split('.')[0], global_step)
