shape=(config.batch_size, config.input_steps))
Y_end = tf.placeholder(tf.float32,
shape=(config.batch_size, config.input_steps))
LR = tf.placeholder(tf.float32)
optimizer, loss, TEM_trainable_variables = TEM_Train(
X_feature, Y_action, Y_start, Y_end, LR, config)
""" Init tf"""
model_saver = tf.train.Saver(var_list=TEM_trainable_variables,
max_to_keep=80)
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
tf_config.log_device_placement = True
sess = tf.InteractiveSession(config=tf_config)
tf.global_variables_initializer().run()
train_dict, val_dict, test_dict = TEM_load_data.getDatasetDict()
train_data_dict = TEM_load_data.getFullData("train")
val_data_dict = TEM_load_data.getFullData("val")
train_info = {
"cost": [],
"loss_action": [],
"loss_start": [],
"loss_end": [],
"l2": []
}
val_info = {
"cost": [],
"loss_action": [],
"loss_start": [],
"loss_end": [],
self.input_steps=100
if __name__ == "__main__":
config = Config()
X_feature = tf.placeholder(tf.float32, shape=(config.batch_size,config.input_steps,config.n_inputs))
tem_scores=TEM_inference(X_feature,config)
model_saver=tf.train.Saver(var_list=tf.trainable_variables(),max_to_keep=80)
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
tf_config.log_device_placement =True
sess=tf.InteractiveSession(config=tf_config)
tf.global_variables_initializer().run()
model_saver.restore(sess,"models/TEM/tem_model_best")
video_dict= TEM_load_data.load_json("./data/activitynet_annotations/anet_anno_action.json")
batch_result_action=[]
batch_result_start=[]
batch_result_end=[]
batch_result_xmin=[]
batch_result_xmax=[]
batch_video_list=TEM_load_data.getBatchListTest(video_dict,config.batch_size,shuffle=False)
for idx in range(len(batch_video_list)):
batch_anchor_xmin,batch_anchor_xmax,batch_anchor_feature=TEM_load_data.getProposalDataTest(batch_video_list[idx],video_dict)
out_scores=sess.run(tem_scores,feed_dict={X_feature:batch_anchor_feature})
batch_result_action.append(out_scores["anchors_action"])
batch_result_start.append(out_scores["anchors_start"])
batch_result_end.append(out_scores["anchors_end"])
if __name__ == '__main__':
opt = parse_arguments()
batch_size = opt.batchsize
if opt.experiment == None:
opt.experiment = './pytorch_models'
output_root = '../../output'
else:
output_root = os.path.join('../../output/', opt.experiment)
opt.experiment = os.path.join('./pytorch_models', opt.experiment)
# video_dict = TEM_load_data.load_json("./data/activitynet_annotations/anet_anno_action.json")
gt_path = '../../datasets/virat/bsn_dataset/stride_100_interval_300/gt_annotations.pkl'
with open(gt_path, 'rb') as input_file:
video_dict = pickle.load(input_file)
feat_dict = TEM_load_data.load_whole_features()
batch_result_action = []
batch_result_start = []
batch_result_end = []
batch_result_xmin = []
batch_result_xmax = []
batch_video_list = TEM_load_data.getBatchListTest(video_dict,
batch_size,
shuffle=False)
tem = TEM(embedsize=opt.embedsize, hiddensize=opt.hiddensize)
model_path = os.path.join(opt.experiment, 'TEM/tem_model_best.pth')
tem.load_state_dict(torch.load(model_path))
tem.cuda()