def rank_transfer_model(pair_model_path):
pair_model = load_model(pair_model_path)
base_model = pair_model.layers[2]
base_model = Model(inputs=base_model.get_input_at(0),
outputs=[base_model.get_output_at(0)],
name='resnet50')
img0 = Input(shape=(224, 224, 3), name='img_0')
img1 = Input(shape=(224, 224, 3), name='img_1')
img2 = Input(shape=(224, 224, 3), name='img_2')
feature0 = Flatten()(base_model(img0))
feature1 = Flatten()(base_model(img1))
feature2 = Flatten()(base_model(img2))
dis1 = Lambda(eucl_dist, name='square1')([feature0, feature1])
dis2 = Lambda(eucl_dist, name='square2')([feature0, feature2])
score1 = Dense(1, activation='sigmoid', name='score1')(dis1)
score2 = Dense(1, activation='sigmoid', name='score2')(dis2)
sub_score = Lambda(sub, name='sub_score')([score1, score2])
model = Model(inputs=[img0, img1, img2], outputs=[sub_score])
model.get_layer('score1').set_weights(
pair_model.get_layer('bin_out').get_weights())
model.get_layer('score2').set_weights(
pair_model.get_layer('bin_out').get_weights())
plot_model(model, to_file='rank_model.png')
for layer in base_model.layers:
layer.trainable = False
return model
def rank_transfer_model(pair_model_path):
pair_model = load_model(pair_model_path)
base_model = pair_model.layers[2]
base_model = Model(inputs=base_model.get_input_at(0), outputs=[base_model.get_output_at(0)], name='resnet50')
# base_model = ResNet50(weights='imagenet', include_top=False, input_tensor=Input(shape=(224, 224, 3)))
# base_model = Model(inputs=[base_model.input], outputs=[base_model.output], name='resnet50')
# for layer in base_model.layers[:len(base_model.layers)/2]:
# layer.trainable = False
for layer in base_model.layers:
if isinstance(layer, BatchNormalization):
layer.trainable = False
print 'to layer: %d' % (len(base_model.layers)/3*2)
img0 = Input(shape=(224, 224, 3), name='img_0')
img1 = Input(shape=(224, 224, 3), name='img_1')
img2 = Input(shape=(224, 224, 3), name='img_2')
feature0 = Flatten()(base_model(img0))
feature1 = Flatten()(base_model(img1))
feature2 = Flatten()(base_model(img2))
dis1 = Lambda(eucl_dist, name='square1')([feature0, feature1])
dis2 = Lambda(eucl_dist, name='square2')([feature0, feature2])
score1 = Dense(1, activation='sigmoid', name='score1')(dis1)
score2 = Dense(1, activation='sigmoid', name='score2')(dis2)
sub_score = Lambda(sub, name='sub_score')([score1, score2])
model = Model(inputs=[img0, img1, img2], outputs=[sub_score])
# model = Model(inputs=[img0, img1, img2], outputs=[sub_score])
model.get_layer('score1').set_weights(pair_model.get_layer('bin_out').get_weights())
model.get_layer('score2').set_weights(pair_model.get_layer('bin_out').get_weights())
plot_model(model, to_file='rank_model.png')
print(model.summary())
return model
def pair_model(source_model_path, num_classes):
pair_model = load_model(source_model_path)
# base_model = ResNet50(weights='imagenet', include_top=False, input_tensor=Input(shape=(224, 224, 3)))
base_model = pair_model.layers[2]
base_model = Model(inputs=base_model.get_input_at(0), outputs=[base_model.get_output_at(0)], name='resnet50')
img1 = Input(shape=(224, 224, 3), name='img_1')
img2 = Input(shape=(224, 224, 3), name='img_2')
feature1 = Lambda(lambda x: K.l2_normalize(x,axis=1))(Flatten()(base_model(img1)))
feature2 = Lambda(lambda x: K.l2_normalize(x,axis=1))(Flatten()(base_model(img2)))
dis = Lambda(eucl_dist, name='square')([feature1, feature2])
# judge = Lambda(dis_sigmoid, name='bin_out')(dis)
judge = Dense(1, activation='sigmoid', name='bin_out')(Dropout(0.8)(dis))
model = Model(inputs=[img1, img2], outputs=[judge])
model.get_layer('bin_out').set_weights(pair_model.get_layer('bin_out').get_weights())
plot_model(model, to_file='model_combined.png')
# for layer in base_model.layers[:-10]:
# layer.trainable = False
# for layer in base_model.layers:
# layer.trainable = True
return model
def two_stage_rank_transfer_2dataset(source_pair_model_path, target_train_list, target_model_path, target_train_path,
rank_pid_path, rank_score_path):
train_images = reid_img_prepare(target_train_list, target_train_path)
batch_size = 16
similar_persons = np.genfromtxt(rank_pid_path, delimiter=' ')
# if 'cross' in rank_pid_path:
# similar_persons = similar_persons - 1
similar_matrix = np.genfromtxt(rank_score_path, delimiter=' ')
rank_transfer(
triplet_generator_by_rank_list(train_images, batch_size, similar_persons, similar_matrix, train=True),
triplet_generator_by_rank_list(train_images, batch_size, similar_persons, similar_matrix, train=False),
source_pair_model_path,
target_model_path,
batch_size=batch_size
)
if __name__ == '__main__':
pair_model = load_model('../pretrain/cuhk_pair_pretrain.h5')
base_model = pair_model.layers[2]
base_model = Model(inputs=base_model.get_input_at(0), outputs=[base_model.get_output_at(0)], name='resnet50')
print isinstance(base_model.layers[-20], Conv2D)
rank_transfer_2dataset('../pretrain/cuhk_pair_pretrain.h5', '../dataset/market_train.list',
'rank_transfer_test.h5',
'/content/Market-1501/train',
'/content/TFusion/TrackViz/data/cuhk_market-train/cross_filter_pid.log',
'/content/TFusion/TrackViz/data/cuhk_market-train/cross_filter_score.log')
similar_matrix = np.genfromtxt(rank_score_path, delimiter=' ')
rank_transfer(triplet_generator_by_rank_list(train_images,
batch_size,
similar_persons,
similar_matrix,
train=True),
triplet_generator_by_rank_list(train_images,
batch_size,
similar_persons,
similar_matrix,
train=False),
source_pair_model_path,
target_model_path,
batch_size=batch_size)
if __name__ == '__main__':
pair_model = load_model('../pretrain/cuhk_pair_pretrain.h5')
base_model = pair_model.layers[2]
base_model = Model(inputs=base_model.get_input_at(0),
outputs=[base_model.get_output_at(0)],
name='resnet50')
print isinstance(base_model.layers[-20], Conv2D)
rank_transfer_2dataset(
'../pretrain/cuhk_pair_pretrain.h5', '../dataset/market_train.list',
'rank_transfer_test.h5', '/home/cwh/coding/Market-1501/train',
'/home/cwh/coding/TrackViz/data/cuhk_market-train/cross_filter_pid.log',
'/home/cwh/coding/TrackViz/data/cuhk_market-train/cross_filter_score.log'
)