def run_training():
initial_steps=0
max_steps=30000
train_clip_path="./val_training_samples.txt" # turn-tap是专门对提议进行回归的，它的提议采用的clip pyramid的形式，也就是人为首先预设出来很多proposal
background_path="./background_samples.txt"
train_featmap_dir="./path_to_features_val/" # turn-tap使用的是denseflow特征
test_featmap_dir="./path_to_features_test/"
test_clip_path="./test_swin.txt" # 间隔为16帧，时序窗口金字塔分别由{1,2,4,8,16,32}个units组成,共408013个clips
test_video_length_info={} # 测试视频时长信息{'视频名':时长}
with open("./thumos14_video_length_test.txt") as f:
for l in f:
test_video_length_info[l.rstrip().split(" ")[0]]=int(l.rstrip().split(" ")[2]) # 获取视频时长（以帧为单位）
train_video_length_info={}
with open("./thumos14_video_length_val.txt") as f:
for l in f:
train_video_length_info[l.rstrip().split(" ")[0]]=int(l.rstrip().split(" ")[2])
# 初始化模型获取数据集加载器
model=turn_model.TURN_Model(batch_size,
train_video_length_info,
ctx_num,
unit_feature_size,
unit_size,
lambda_reg,
lr,
train_clip_path,
background_path,
test_clip_path,
train_featmap_dir,
test_featmap_dir)
with tf.Graph().as_default()：
def training():
initial_steps = 0
max_steps = 20000
model = turn_model.TURN_Model()
with tf.Graph().as_default():
loss_cls_reg, vs_train_op, vs_eval_op, loss_reg = model.construct_model()
tf.summary.scalar('loss', loss_cls_reg)
# Create a session for running Ops on the Graph.
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.2)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
# Run the Op to initialize the variables.
init = tf.global_variables_initializer()
sess.run(init)
saver = tf.train.Saver()
# merged = tf.summary.merge_all()
# writer = tf.summary.FileWriter("logs/", sess.graph)
start_time = time.time()
for step in xrange(max_steps):
feed_dict = model.fill_feed_dict_train_reg()
_, loss_v, loss_reg_v = sess.run([vs_train_op, loss_cls_reg, loss_reg], feed_dict=feed_dict)
if step % 50 == 0:
duration = time.time() - start_time
start_time = time.time()
pt,epoch,train_set_length=model.train_set.get_training_info()
# rs = sess.run(merged)
# writer.add_summary(rs, step)
# Print status to stdout.
print('Step %d (%d/%d) Epoch %d total loss = %.2f, regression loss = %.2f(%.3f sec)' % (
step, pt,train_set_length,epoch,loss_v, loss_reg_v, duration))
if step % 1000 == 0:
probs = []
labels = []
pred = []
for i in range(1000):
# for i in range(model.val_set):
movie_name, gt_start, gt_end, clip_start, clip_end, feat,unit_duration = model.val_set.get_sample(i)
feed_dict = {model.visual_featmap_ph_test: feat}
outputs = sess.run(vs_eval_op, feed_dict=feed_dict)
pred_start, pred_end, action_score = postprocess(clip_start, clip_end, outputs,unit_duration)
# print(movie_name, gt_start, gt_end, pred_start, pred_end, action_score)
if gt_start == gt_end == 0:
label = 0
else:
label = 1
if action_score > 0.5:
pred.append(1)
else:
pred.append(0)
probs.append(action_score)
labels.append(label)
# print(pred,probs,labels)
print("Accuracy", accuracy_score(y_pred=pred, y_true=labels))
print("Loss", log_loss(y_pred=probs, y_true=labels))
saver.save(sess, os.path.join(ActivityNet.MODEL_DIR, 'turn_activity_net.model'))
def run_training():
initial_steps = 0
max_steps = 30000
train_clip_path = "./val_training_samples.txt"
background_path = "./background_samples.txt"
train_featmap_dir = "./path_to_features_val/"
test_featmap_dir = "./path_to_features_test/"
test_clip_path = "./test_swin.txt"
test_video_length_info = {}
with open("./thumos14_video_length_test.txt") as f:
for l in f:
test_video_length_info[l.rstrip().split(" ")[0]] = int(
l.rstrip().split(" ")[2])
train_video_length_info = {}
with open("./thumos14_video_length_val.txt") as f:
for l in f:
train_video_length_info[l.rstrip().split(" ")[0]] = int(
l.rstrip().split(" ")[2])
model = turn_model.TURN_Model(batch_size, train_video_length_info, ctx_num,
unit_feature_size, unit_size, lambda_reg, lr,
train_clip_path, background_path,
test_clip_path, train_featmap_dir,
test_featmap_dir)
with tf.Graph().as_default():
loss_cls_reg, vs_train_op, vs_eval_op, loss_reg = model.construct_model(
)
# Create a session for running Ops on the Graph.
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.2)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
# Run the Op to initialize the variables.
init = tf.initialize_all_variables()
sess.run(init)
for step in xrange(max_steps):
start_time = time.time()
feed_dict = model.fill_feed_dict_train_reg()
_, loss_v, loss_reg_v = sess.run(
[vs_train_op, loss_cls_reg, loss_reg], feed_dict=feed_dict)
duration = time.time() - start_time
if step % 5 == 0:
# Print status to stdout.
print(
'Step %d: total loss = %.2f, regression loss = %.2f(%.3f sec)'
% (step, loss_v, loss_reg_v, duration))
if (step + 1) % test_steps == 0:
print "Start to test:-----------------\n"
do_eval_slidingclips(sess, vs_eval_op, model,
test_video_length_info, step + 1)
def testing():
model = turn_model.TURN_Model()
with tf.Graph().as_default():
sess = tf.Session()
# First let's load meta graph and restore weights
# saver = tf.train.import_meta_graph(os.path.join(THUMOS14.MODEL_DIR,'turn_tap_model.meta'))
loss_cls_reg, vs_train_op, vs_eval_op, loss_reg = model.construct_model(test=True)
init = tf.global_variables_initializer()
sess.run(init)
saver = tf.train.Saver()
saver.restore(sess, os.path.join(ActivityNet.MODEL_DIR, 'turn_tap_model'))
# utils.show_tensor(os.path.join(ActivityNet.MODEL_DIR, 'turn_tap_model'))
result = {}
probs = []
labels = []
pred = []
for i in range(len(model.test_set.samples)):
# for i in range(50):
movie_name, gt_start, gt_end, clip_start, clip_end, feat,unit_duration = model.test_set.get_sample(i)
feed_dict = {
model.visual_featmap_ph_test: feat
}
outputs = sess.run(vs_eval_op, feed_dict=feed_dict)
pred_start, pred_end, action_score = postprocess(clip_start, clip_end, outputs,unit_duration)
# print(movie_name,gt_start,gt_end,pred_start,pred_end,action_score)
_list = result.get(movie_name, [])
_list.append((pred_start, pred_end, action_score))
result[movie_name] = _list
if gt_start == gt_end == 0:
label = 0
else:
label = 1
if action_score > 0.5:
pred.append(1)
else:
pred.append(0)
if action_score == 0:
action_score = 1e-8
elif action_score == 1:
action_score = 0.9999
probs.append(action_score)
labels.append(label)
# print(pred,probs,labels)
print("Accuracy", accuracy_score(y_pred=pred, y_true=labels))
print("Loss", log_loss(y_pred=probs, y_true=labels))
def training():
initial_steps=0
max_steps=10000
model=turn_model.TURN_Model()
print(model.init_placeholder())
with tf.Graph().as_default():
loss_cls_reg,vs_train_op,vs_eval_op, loss_reg=model.construct_model()
# Create a session for running Ops on the Graph.
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction = 0.2)
sess = tf.Session(config = tf.ConfigProto(gpu_options = gpu_options))
# Run the Op to initialize the variables.
init = tf.global_variables_initializer()
sess.run(init)
saver = tf.train.Saver()
start_time = time.time()
for step in xrange(max_steps):
feed_dict = model.fill_feed_dict_train_reg()
_, loss_v, loss_reg_v = sess.run([vs_train_op,loss_cls_reg, loss_reg], feed_dict=feed_dict)
if step % 50 == 0:
duration = time.time() - start_time
# Print status to stdout.
print('Step %d: total loss = %.2f, regression loss = %.2f(%.3f sec)' % (step, loss_v, loss_reg_v, duration))
start_time = time.time()
if step %1000==0:
for i in range(15,25):
movie_name, gt_start, gt_end, clip_start, clip_end, feat = model.test_set.get_sample(i)
feed_dict = {
model.visual_featmap_ph_test: feat
}
outputs = sess.run(vs_eval_op, feed_dict=feed_dict)
pred_start, pred_end, action_score = output_process(clip_start, clip_end, outputs)
print(movie_name, gt_start, gt_end, pred_start, pred_end, action_score)