def infer(data_type):
proposal_data = {'results': {}, 'version': "VERSION 1.0"}
json_data = load_json("data/meta.json")
database = json_data['database']
dict_dim = 2048
class_num = 2
window_lens = [30]
window_stride = 30
model_path = "model/video_lstm30.tar.gz"
prob_layer = lstm_net(dict_dim, class_num, is_infer=True)
# initialize PaddlePaddle
paddle.init(use_gpu=False, trainer_count=1)
# load the trained models
if os.path.exists(model_path):
with gzip.open(model_path, 'r') as f:
parameters = paddle.parameters.Parameters.from_tar(f)
index = 0
for video in database.keys():
dataSet = database[video]["subset"]
if dataSet != data_type:
continue
with open("data/" + dataSet + "/" + str(video) + ".pkl", 'rb') as f:
video_fea = cPickle.load(f)
print index, video
index += 1
video_len = np.shape(video_fea)[0]
this_vid_proposals = []
for pos in range(0, video_len, window_stride):
inputs = []
for window_len in window_lens:
if pos + window_len < video_len:
inputs.append([video_fea[pos:pos + window_len]])
probs = paddle.infer(
output_layer=prob_layer, parameters=parameters, input=inputs, field="value")
if len(probs) <= 0:
continue
max_probs = np.max(probs[0])
window_index = 0
label = np.argmax(probs[0])
if label == 0:
continue
proposal = {
'label': label,
'score': int(max_probs * 100) / 100.0,
'segment': [pos, pos + window_lens[window_index]],
}
this_vid_proposals += [proposal]
proposal_data['results'][video] = this_vid_proposals
with open("res/" + data_type + "_len.json", 'w') as fobj:
json.dump(proposal_data, fobj)
def stat():
file_path = 'data/meta.json'
json_data = load_json(file_path)
database = json_data['database']
data = []
for video in database.keys():
vd = []
if database[video]['subset'] != 'validation':
continue
for i in database[video]['annotations']:
st, ed = i['segment']
vd.append(int((ed - st) / 10.0) * 10)
mlen = int(np.mean(vd) / 10) * 10
data.append(mlen)
print Counter(data).keys()
def plot(dataset):
file_path = 'data/meta.json'
json_data = load_json("data/meta.json")
database = json_data['database']
data_x = []
data_y = []
for video in database.keys():
if database[video]['subset'] != 'validation':
continue
fig, ax = plt.subplots()
gt_x, gt_y = get_gt_data(video)
pr_x, pr_y = get_pr_data(video, dataset)
bd_x, bd_y = get_bd_data(video, dataset)
line1, = ax.plot(gt_x, gt_y)
line2, = ax.plot(pr_x, pr_y, '--', linewidth=1)
plt.savefig('image/' + dataset + '/' + video + '.jpg')
def get_gt_data(video):
file_path = 'data/meta.json'
json_data = load_json(file_path)
database = json_data['database']
data_x = []
data_y = []
for i in database[video]['annotations']:
st, ed = i['segment']
data_x.append(st)
data_y.append(0)
data_x.append(st)
data_y.append(1)
data_x.append(ed)
data_y.append(1)
data_x.append(ed)
data_y.append(0)
return data_x, data_y
def get_bd_data(video, dataset):
file_path = 'res/' + dataset + '.json'
json_data = load_json(file_path)
database = json_data['results']
data_x = []
data_y = []
for i in database[video]:
st, ed = i['segment']
data_x.append(st)
data_y.append(0)
data_x.append(st)
data_y.append(1)
data_x.append(ed)
data_y.append(1)
data_x.append(ed)
data_y.append(0)
return data_x, data_y
def infer(data_type):
proposal_data = {'results': {}, 'version': "VERSION 1.0"}
json_data = load_json("res/" + data_type + ".json")
database = json_data['results']
for video in database.keys():
new_seg = []
new_score = []
this_vid_proposals = []
last_ed = 0
last_st = 0
width = 0
for index, i in enumerate(database[video]):
st, ed = i['segment']
score = i['score']
if score < 0.5:
continue
if st - last_ed < 30 and width <= 6 and last_ed != 0:
width += 1
last_ed = ed
continue
if last_ed != 0:
new_seg.append([last_st, last_ed])
new_score.append(i['score'])
last_st = st
last_ed = ed
width = 0
for i, sore in zip(new_seg, new_score):
last_st, last_ed = i
if last_ed - last_st < 150:
if last_ed - last_st <= 30:
continue
last_st = (last_st + last_ed) / 2 - 80
last_ed = (last_st + last_ed) / 2 + 80
proposal = {
'score': sore,
'segment': [last_st, last_ed],
}
this_vid_proposals += [proposal]
proposal_data['results'][video] = this_vid_proposals
with open("res/" + data_type + "_refine.json", 'w') as fobj:
json.dump(proposal_data, fobj)
def infer(video, st, ed):
proposal_data = {'results': {}, 'version': "VERSION 1.0"}
json_data = load_json("data/meta.json")
database = json_data['database']
dict_dim = 2048
class_num = 2
dataSet = 'validation'
model_path = "model/video_lstm30.tar.gz"
prob_layer = lstm_net(dict_dim, class_num, is_infer=True)
# initialize PaddlePaddle
paddle.init(use_gpu=False, trainer_count=1)
# load the trained models
if os.path.exists(model_path):
with gzip.open(model_path, 'r') as f:
parameters = paddle.parameters.Parameters.from_tar(f)
with open("data/" + dataSet + "/" + str(video) + ".pkl", 'rb') as f:
video_fea = cPickle.load(f)
probs = paddle.infer(output_layer=prob_layer,
parameters=parameters,
input=[[video_fea[int(st):int(ed)]]],
field="value")
print probs