def __init__(self):
self.set_gpu(1)
self.img_path = ''
self.model_path = './model/MCNN_model/v1-2050'
# crop_size = 256
# place holder位置保持器(定义变量)
self.input_img_placeholder = tf.placeholder(
tf.float32, shape=([None, None, None, 3]))
self.density_map_placeholder = tf.placeholder(tf.float32,
shape=(None, None, None,
1))
self.inference_density_map = multi_column_cnn(
self.input_img_placeholder)
def infer():
set_gpu(1)
img_path = 'D:\\YourZhouProject\\mcnn_project\\pytorch_mcnn\\part_A_final\\test_data\\images\\IMG_78.jpg'
model_path = 'D:\\YourZhouProject\\mcnn_project\\tf_mcnn\\work\\ckpts2\\mcnn\\v1-1425'
# crop_size = 256
ori_crowd_img = cv.imread(img_path)
# ori_crowd_img = cv.resize(ori_crowd_img, (256, 256))
# h, w = ori_crowd_img.shape[0], ori_crowd_img.shape[1]
img = ori_crowd_img.reshape(
(ori_crowd_img.shape[0], ori_crowd_img.shape[1],
ori_crowd_img.shape[2]))
# place holder位置保持器
input_img_placeholder = tf.placeholder(tf.float32,
shape=([None, None, None, 3]))
density_map_placeholder = tf.placeholder(tf.float32,
shape=(None, None, None, 1))
inference_density_map = multi_column_cnn(input_img_placeholder)
saver = tf.train.Saver()
time_star = time.time()
with tf.Session() as sess:
saver.restore(sess, model_path)
result = sess.run(
inference_density_map,
feed_dict={input_img_placeholder: [(img - 127.5) / 128]})
time_over = time.time() - time_star
print(time_over)
num = result.sum()
print(num)
dmap_img = result[0, :, :, 0]
final_img = image_processing(dmap_img)
final_img = image_add_heatmap(ori_crowd_img, final_img, 0.5)
cv.putText(final_img, "P : " + str(int(num)), (50, 50),
cv.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv.imshow("really", final_img)
cv.waitKey(0)
cv.destroyAllWindows()
def infer():
set_gpu(0)
cap = cv.VideoCapture("D:/YourZhouDownloads/FFPUT/school_2/school_01.mp4")
# img_path = 'D:\\YourZhouProject\\mcnn_project\\pytorch_mcnn\\part_A_final\\test_data\\images\\IMG_45.jpg'
model_path = 'D:\\YourZhouProject\\mcnn_project\\tf_mcnn\\work\\ckpts\\mcnn\\v1-2200'
# crop_size = 256
# place holder位置保持器
input_img_placeholder = tf.placeholder(tf.float32,
shape=([None, None, None, 3]))
density_map_placeholder = tf.placeholder(tf.float32,
shape=(None, None, None, 1))
inference_density_map = multi_column_cnn(input_img_placeholder)
while True:
ret, image = cap.read()
ori_crowd_img = cv.flip(image, 1)
# ori_crowd_img = cv.imread(img_path)
# h, w = ori_crowd_img.shape[0], ori_crowd_img.shape[1]
img = ori_crowd_img.reshape(
(ori_crowd_img.shape[0], ori_crowd_img.shape[1],
ori_crowd_img.shape[2]))
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_path)
result = sess.run(
inference_density_map,
feed_dict={input_img_placeholder: [(img - 127.5) / 128]})
print(result.sum())
dmap_img = result[0, :, :, 0]
# utils.show_density_map(dmap_img)
final_img = image_processing(dmap_img)
final_img = image_add_heatmap(ori_crowd_img, final_img, 0.3)
cv.imshow("really", final_img)
if (cv.waitKey(1) == 27):
cv.destroyAllWindows()
break
def test():
cfig = ConfigFactory()
#set_gpu(0)
dataset = 'A'
# training dataset
img_root_dir = cfig.data_root_dir + r'part_' + dataset + r'_final/train_data/images/'
gt_root_dir = cfig.data_root_dir + r'part_' + dataset + r'_final/train_data/ground_truth/'
# testing dataset
val_img_root_dir = cfig.data_root_dir + r'part_' + dataset + r'_final/test_data/images/'
val_gt_root_dir = cfig.data_root_dir + r'part_' + dataset + r'_final/test_data/ground_truth/'
# place holder
input_img_placeholder = tf.placeholder(tf.float32,
shape=(None, None, None, 3))
density_map_placeholder = tf.placeholder(tf.float32,
shape=(None, None, None, 1))
# network generation
inference_density_map = multi_column_cnn(input_img_placeholder)
# density map loss
density_map_loss = 0.5 * tf.reduce_sum(
tf.square(tf.subtract(density_map_placeholder, inference_density_map)))
# jointly training
joint_loss = density_map_loss
# optimizer = tf.train.MomentumOptimizer(configs.learing_rate, momentum=configs.momentum).minimize(joint_loss)
# adam optimizer
# optimizer = tf.train.AdamOptimizer(cfig.lr).minimize(joint_loss)
init = tf.global_variables_initializer()
file_path = cfig.log_router
# training log route
if not os.path.exists(file_path):
os.makedirs(file_path)
# model saver route
if not os.path.exists(cfig.ckpt_router):
os.makedirs(cfig.ckpt_router)
log = open(cfig.log_router + cfig.name + r'_training.logs',
mode='a+',
encoding='utf-8')
saver = tf.train.Saver(max_to_keep=cfig.max_ckpt_keep)
ckpt = tf.train.get_checkpoint_state(cfig.ckpt_router)
# start session
sess = tf.Session()
if ckpt and ckpt.model_checkpoint_path:
print('load model', ckpt.model_checkpoint_path)
saver.restore(sess, ckpt.model_checkpoint_path)
else:
sess.run(init)
data_loader = ImageDataLoader(img_root_dir,
gt_root_dir,
shuffle=False,
downsample=False,
pre_load=False)
data_loader_val = ImageDataLoader(val_img_root_dir,
val_gt_root_dir,
shuffle=False,
downsample=False,
pre_load=False)
absolute_error = 0.0
square_error = 0.0
file_index = 1
for blob in data_loader_val:
img, gt_dmp, gt_count = blob['data'], blob['gt_density'], blob[
'crowd_count']
feed_dict = {
input_img_placeholder: (img - 127.5) / 128,
density_map_placeholder: gt_dmp
}
inf_dmp, loss = sess.run([inference_density_map, joint_loss],
feed_dict=feed_dict)
print(blob['fname'], gt_count.sum(), inf_dmp.sum(), loss)
#print(absolute_error,square_error)
absolute_error = absolute_error + np.abs(
np.subtract(gt_count.sum(), inf_dmp.sum())).mean()
square_error = square_error + np.power(
np.subtract(gt_count.sum(), inf_dmp.sum()), 2).mean()
file_index = file_index + 1
show_map(img[0, :, :, 0])
show_density_map(inf_dmp[0, :, :, 0])
show_density_map(gt_dmp[0, :, :, 0])
mae = absolute_error / data_loader_val.num_samples
rmse = np.sqrt(square_error / data_loader_val.num_samples)
print(str('MAE_' + str(mae) + '_MSE_' + str(rmse)))
def train():
set_gpu(0)
dataset = 'A'
# training dataset训练数据集
img_root_dir = r'D:/people_all/ShanghaiTech/part_' + dataset + r'_final/train_data/images/'
gt_root_dir = r'D:/people_all/ShanghaiTech/part_' + dataset + r'_final/train_data/ground_truth/'
# testing dataset测试数据集
val_img_root_dir = r'D:/people_all/ShanghaiTech/part_' + dataset + r'_final/test_data/images/'
val_gt_root_dir = r'D:/people_all/ShanghaiTech/part_' + dataset + r'_final/test_data/ground_truth/'
# training dataset file list训练数据集文件列表
img_file_list = os.listdir(img_root_dir)
gt_img_file_list = os.listdir(gt_root_dir)
# testing dataset file listtesting dataset file list
val_img_file_list = os.listdir(val_img_root_dir)
val_gt_file_list = os.listdir(val_gt_root_dir)
cfig = ConfigFactory()
# place holder位置保持器
input_img_placeholder = tf.placeholder(tf.float32,
shape=(None, None, None, 3))
density_map_placeholder = tf.placeholder(tf.float32,
shape=(None, None, None, 1))
# network generation网络生成
inference_density_map = multi_column_cnn(input_img_placeholder)
# density map loss密度图损失
density_map_loss = 0.5 * tf.reduce_sum(
tf.square(tf.subtract(density_map_placeholder, inference_density_map)))
# jointly training联合训练
joint_loss = density_map_loss
# optimizer = tf.train.MomentumOptimizer(configs.learing_rate, momentum=configs.momentum).minimize(joint_loss)
# adam optimizer ADAM优化器
optimizer = tf.train.AdamOptimizer(cfig.lr).minimize(joint_loss)
init = tf.global_variables_initializer()
file_path = cfig.log_router
# training log route训练日志路径
if not os.path.exists(file_path):
os.makedirs(file_path)
# model saver route模型保护程序路由
if not os.path.exists(cfig.ckpt_router):
os.makedirs(cfig.ckpt_router)
log = open(cfig.log_router + cfig.name + r'_training.logs',
mode='a+',
encoding='utf-8')
saver = tf.train.Saver(max_to_keep=cfig.max_ckpt_keep)
ckpt = tf.train.get_checkpoint_state(cfig.ckpt_router)
# start session开始会话
sess = tf.Session()
if ckpt and ckpt.model_checkpoint_path:
print('load model')
saver.restore(sess, ckpt.model_checkpoint_path)
sess.run(init)
# 开始训练
for i in range(cfig.total_iters):
# 训练
for file_index in range(len(img_file_list)):
img_path = img_root_dir + img_file_list[file_index]
gt_path = gt_root_dir + 'GT_' + img_file_list[file_index].split(
r'.')[0] + '.mat.'
img, gt_dmp, gt_count = read_crop_train_data(img_path,
gt_path,
scale=4)
feed_dict = {
input_img_placeholder: (img - 127.5) / 128,
density_map_placeholder: gt_dmp
}
_, inf_dmp, loss = sess.run(
[optimizer, inference_density_map, joint_loss],
feed_dict=feed_dict)
format_time = str(
time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
format_str = 'step %d, joint loss=%.5f, inference= %.5f, gt=%d'
log_line = format_time, img_file_list[file_index], format_str % (
i * len(img_file_list) + file_index, loss, inf_dmp.sum(),
gt_count)
log.writelines(str(log_line) + '\n')
print(log_line)
# covert graph to pb file
# tf.train.write_graph(sess.graph_def, "./", 'graph.pb' + str(file_index), as_text=True)
# test whether add new operation dynamically
# sess.graph.finalize()
# saver.save(sess, cfig.ckpt_router + '/v1', global_step=i)
if i % 25 == 0:
saver.save(sess, cfig.ckpt_router + '/v1', global_step=i)
val_log = open(cfig.log_router + cfig.name + r'_validating_' +
str(i) + '_.logs',
mode='w',
encoding='utf-8')
absolute_error = 0.0
square_error = 0.0
# validating验证
for file_index in range(len(val_img_file_list)):
img_path = val_img_root_dir + val_img_file_list[file_index]
gt_path = val_gt_root_dir + 'GT_' + val_img_file_list[
file_index].split(r'.')[0] + '.mat'
# gt_path = val_gt_root_dir + 'GT_' + val_img_file_list[file_index].split(r'.')[0]
img, gt_dmp, gt_count = read_test_data(img_path,
gt_path,
scale=4)
feed_dict = {
input_img_placeholder: (img - 127.5) / 128,
density_map_placeholder: gt_dmp
}
_, inf_dmp, loss = sess.run(
[optimizer, inference_density_map, joint_loss],
feed_dict=feed_dict)
format_time = str(
time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
format_str = 'step %d, joint loss=%.5f, inference= %.5f, gt=%d'
absolute_error = absolute_error + np.abs(
np.subtract(gt_count, inf_dmp.sum())).mean()
square_error = square_error + np.power(
np.subtract(gt_count, inf_dmp.sum()), 2).mean()
log_line = format_time, val_img_file_list[
file_index], format_str % (file_index, loss, inf_dmp.sum(),
gt_count)
val_log.writelines(str(log_line) + '\n')
print(log_line)
mae = absolute_error / len(val_img_file_list)
rmse = np.sqrt(absolute_error / len(val_img_file_list))
val_log.writelines(
str('MAE_' + str(mae) + '_MSE_' + str(rmse)) + '\n')
val_log.close()
print(str('MAE_' + str(mae) + '_MSE_' + str(rmse)))
def train():
cfig = ConfigFactory()
#set_gpu(0)
dataset = 'A'
# training dataset
img_root_dir = cfig.data_root_dir + r'part_' + dataset + r'_final/train_data/images/'
gt_root_dir = cfig.data_root_dir + r'part_' + dataset + r'_final/train_data/ground_truth/'
# testing dataset
val_img_root_dir = cfig.data_root_dir + r'part_' + dataset + r'_final/test_data/images/'
val_gt_root_dir = cfig.data_root_dir + r'part_' + dataset + r'_final/test_data/ground_truth/'
cfig = ConfigFactory()
# place holder
input_img_placeholder = tf.placeholder(tf.float32,
shape=(None, None, None, 3))
density_map_placeholder = tf.placeholder(tf.float32,
shape=(None, None, None, 1))
# network generation
inference_density_map = multi_column_cnn(input_img_placeholder)
# density map loss
density_map_loss = 0.5 * tf.reduce_sum(
tf.square(tf.subtract(density_map_placeholder, inference_density_map)))
# jointly training
joint_loss = density_map_loss
# optimizer = tf.train.MomentumOptimizer(configs.learing_rate, momentum=configs.momentum).minimize(joint_loss)
# adam optimizer
optimizer = tf.train.AdamOptimizer(cfig.lr).minimize(joint_loss)
init = tf.global_variables_initializer()
file_path = cfig.log_router
# training log route
if not os.path.exists(file_path):
os.makedirs(file_path)
# model saver route
if not os.path.exists(cfig.ckpt_router):
os.makedirs(cfig.ckpt_router)
log = open(cfig.log_router + cfig.name + r'_training.logs',
mode='a+',
encoding='utf-8')
saver = tf.train.Saver(max_to_keep=cfig.max_ckpt_keep)
ckpt = tf.train.get_checkpoint_state(cfig.ckpt_router)
# start session
sess = tf.Session()
if ckpt and ckpt.model_checkpoint_path:
print('load model, ckpt.model_checkpoint_path')
saver.restore(sess, ckpt.model_checkpoint_path)
else:
sess.run(init)
data_loader = ImageDataLoader(img_root_dir,
gt_root_dir,
shuffle=True,
downsample=True,
pre_load=True)
data_loader_val = ImageDataLoader(val_img_root_dir,
val_gt_root_dir,
shuffle=False,
downsample=False,
pre_load=True)
# start training
for i in range(cfig.start_iters, cfig.total_iters):
# training
index = 1
for blob in data_loader:
img, gt_dmp, gt_count = blob['data'], blob['gt_density'], blob[
'crowd_count']
feed_dict = {
input_img_placeholder: (img - 127.5) / 128,
density_map_placeholder: gt_dmp
}
_, inf_dmp, loss = sess.run(
[optimizer, inference_density_map, joint_loss],
feed_dict=feed_dict)
format_time = str(
time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
format_str = 'step %d, joint loss=%.5f, inference= %.5f, gt=%d'
log_line = format_time, blob['fname'], format_str % (
i * data_loader.num_samples + index, loss, inf_dmp.sum(),
gt_count)
log.writelines(str(log_line) + '\n')
print(log_line)
index = index + 1
if i % 50 == 0:
val_log = open(cfig.log_router + cfig.name + r'_validating_' +
str(i) + '_.logs',
mode='w',
encoding='utf-8')
absolute_error = 0.0
square_error = 0.0
file_index = 1
for blob in data_loader_val:
img, gt_dmp, gt_count = blob['data'], blob['gt_density'], blob[
'crowd_count']
feed_dict = {
input_img_placeholder: (img - 127.5) / 128,
density_map_placeholder: gt_dmp
}
inf_dmp, loss = sess.run([inference_density_map, joint_loss],
feed_dict=feed_dict)
format_time = str(
time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
format_str = 'step %d, joint loss=%.5f, inference= %.5f, gt=%d'
absolute_error = absolute_error + np.abs(
np.subtract(gt_count, inf_dmp.sum())).mean()
square_error = square_error + np.power(
np.subtract(gt_count, inf_dmp.sum()), 2).mean()
log_line = format_time, blob['fname'], format_str % (
file_index, loss, inf_dmp.sum(), gt_count)
val_log.writelines(str(log_line) + '\n')
print(log_line)
file_index = file_index + 1
mae = absolute_error / data_loader_val.num_samples
rmse = np.sqrt(square_error / data_loader_val.num_samples)
val_log.writelines(
str('MAE_' + str(mae) + '_MSE_' + str(rmse)) + '\n')
val_log.close()
print(str('MAE_' + str(mae) + '_MSE_' + str(rmse)))
saver.save(sess, cfig.ckpt_router + '/v1', global_step=i + 1)