def rank_transfer(train_generator,
val_generator,
source_model_path,
target_model_path,
batch_size=48):
model = rank_transfer_model(source_model_path)
model.compile(
optimizer=SGD(lr=0.001, momentum=0.9), # 'adam',
# optimizer='adadelta',
loss={
# 'sub_score': 'mse',
'score1': 'binary_crossentropy',
'score2': 'binary_crossentropy',
# 'sub_score': 'mse'
},
loss_weights={
# 'sub_score': 0.5,
'score1': 0.5,
'score2': 0.5
},
metrics=['accuracy'])
early_stopping = EarlyStopping(monitor='val_loss', patience=5)
auto_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.1,
patience=5,
verbose=0,
mode='auto',
epsilon=0.0001,
cooldown=0,
min_lr=0)
monitor_bn_nan = MonitorNanOnBN()
if 'market-' in target_model_path:
train_data_cnt = 3000
val_data_cnt = 300
else:
train_data_cnt = 160
val_data_cnt = 18
safe_remove(target_model_path)
model.fit_generator(
train_generator,
steps_per_epoch=train_data_cnt / batch_size + 1,
epochs=50,
validation_data=val_generator,
validation_steps=val_data_cnt / batch_size + 1,
callbacks=[
auto_lr, # early_stopping, # monitor_bn_nan,
ModelCheckpoint(target_model_path,
monitor='val_loss',
verbose=0,
save_best_only=True,
save_weights_only=False,
mode='auto',
period=1)
])
def test_predict(net, probe_path, gallery_path, pid_path, score_path):
net = Model(inputs=[net.input], outputs=[net.get_layer('avg_pool').output])
test_f, test_info = extract_feature(gallery_path, net)
query_f, query_info = extract_feature(probe_path, net)
result, result_argsort = sort_similarity(query_f, test_f)
for i in range(len(result)):
result[i] = result[i][result_argsort[i]]
result = np.array(result)
safe_remove(pid_path)
safe_remove(score_path)
np.savetxt(pid_path, result_argsort, fmt='%d')
np.savetxt(score_path, result, fmt='%.4f')
def test_predict(net, probe_path, gallery_path, pid_path, score_path):
#net = Model(inputs=[net.input], outputs=[net.output])
#print("output shape:", net.output.shape)
test_f, test_info = extract_feature(gallery_path, net) ###19732
query_f, query_info = extract_feature(probe_path, net) ##3368
result, result_argsort = sort_similarity(query_f, test_f)
for i in range(len(result)):
result[i] = result[i][result_argsort[i]]
result = np.array(result)
safe_remove(pid_path)
safe_remove(score_path)
np.savetxt(pid_path, result_argsort, fmt='%d')
np.savetxt(score_path, result, fmt='%.4f')
def test_sepbn_predict(net_path, probe_path, gallery_path, pid_path, score_path):
model = load_model(net_path, custom_objects={'cross_entropy_loss': cross_entropy_loss})
model = Model(inputs=[model.get_layer('resnet50').get_input_at(0)[1]],
outputs=[model.get_layer('resnet50').get_output_at(0)[1]])
test_f, test_info = extract_feature(gallery_path, model)
query_f, query_info = extract_feature(probe_path, model)
result, result_argsort = sort_similarity(query_f, test_f)
for i in range(len(result)):
result[i] = result[i][result_argsort[i]]
result = np.array(result)
safe_remove(pid_path)
safe_remove(score_path)
np.savetxt(pid_path, result_argsort, fmt='%d')
np.savetxt(score_path, result, fmt='%.4f')
def rank_transfer(train_generator,
val_generator,
source_model_path,
target_model_path,
batch_size=48):
model = rank_transfer_model(source_model_path)
plot_model(model, 'rank_model.png')
model.compile(
optimizer=SGD(lr=0.001, momentum=0.9), # 'adam',
# optimizer='adam',
loss={
'sub_score': cross_entropy_loss,
#'score1': 'binary_crossentropy',
#'score2': 'binary_crossentropy'
# 'sub_score': 'mse'
},
loss_weights={
'sub_score': 1,
# 'score1': 0.5,
# 'score2': 0.5
},
# metrics=['accuracy']
)
early_stopping = EarlyStopping(monitor='val_loss', patience=3)
auto_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.1,
patience=2,
verbose=0,
mode='auto',
epsilon=0.0001,
cooldown=0,
min_lr=0)
if 'market-' in target_model_path:
train_data_cnt = 16500
val_data_cnt = 1800
else:
train_data_cnt = 1600
val_data_cnt = 180
model.fit_generator(train_generator,
steps_per_epoch=train_data_cnt / batch_size + 1,
epochs=5,
validation_data=val_generator,
validation_steps=val_data_cnt / batch_size + 1,
callbacks=[early_stopping, auto_lr])
safe_remove(target_model_path)
# model.save('simple_rank_transfer.h5')
model.save(target_model_path)
def lmp_test_predict(net, probe_path, gallery_path, pid_path, score_path):
x = net.layers[173].output
x = MaxPooling2D(pool_size=(7, 2),
data_format='channels_last',
padding='same',
name='lmp')(x)
x = AveragePooling2D(pool_size=(1, 4))(x)
net = Model(inputs=[net.input], outputs=[x])
test_f, test_info = extract_feature(gallery_path, net)
query_f, query_info = extract_feature(probe_path, net)
result, result_argsort = sort_similarity(query_f, test_f)
for i in range(len(result)):
result[i] = result[i][result_argsort[i]]
result = np.array(result)
safe_remove(pid_path)
safe_remove(score_path)
np.savetxt(pid_path, result_argsort, fmt='%d')
np.savetxt(score_path, result, fmt='%.4f')
pair_generator_by_rank_list(train_images, batch_size, similar_persons, similar_matrix, train=False),
target,
batch_size=batch_size, num_classes=class_count
)
if __name__ == '__main__':
# sources = ['cuhk_grid_viper_mix']
sources = ['cuhk']
target = 'market'
pair_model('../pretrain/cuhk_pair_pretrain.h5', 751)
for source in sources:
pair_pretrain_on_dataset(source, target)
transform_dir = '/home/cwh/coding/Market-1501'
safe_remove('pair_transfer_pid.log')
test_pair_predict('market_pair_pretrain.h5',
transform_dir + '/probe', transform_dir + '/test',
'pair_transfer_pid.log', 'pair_transfer_score.log')
market_result_eval('pair_transfer_pid.log', TEST='/home/cwh/coding/Market-1501/test',
QUERY='/home/cwh/coding/Market-1501/probe')
# sources = ['grid-cv-%d' % i for i in range(10)]
# for source in sources:
# softmax_pretrain_on_dataset(source,
# project_path='/home/cwh/coding/rank-reid',
# dataset_parent='/home/cwh/coding')
# pair_pretrain_on_dataset(source,
# project_path='/home/cwh/coding/rank-reid',
# dataset_parent='/home/cwh/coding')
