objLoaderVID.loaderVID() train_vid_videos = objLoaderVID.get_videos() vid_images = 0 for vid_idx in xrange(len(train_vid_videos)): video = train_vid_videos[vid_idx] annos = video.annotations vid_images += len(annos) total_image_size = vid_images logger.info('total training VID images size is: ' + str(vid_images)) # debug # cur_batch = data_reader(train_vid_videos) # network initialization tracknet = goturn_net_coord.TRACKNET(BATCH_SIZE) tracknet.build() # learning policy global_step = tf.Variable(0, trainable=False, name="global_step") learning_rate = tf.train.piecewise_constant( global_step, [tf.cast(v, tf.int32) for v in POLICY['step_values']], POLICY['learning_rates']) train_step = tf.train.AdamOptimizer(learning_rate, POLICY['momentum'], POLICY['momentum2']).minimize( tracknet.loss_wdecay, global_step=global_step) # summary merged_summary = tf.summary.merge_all() sess = tf.Session()
'/home/jaehyuk/code/github/vot-toolkit/tracker/examples/python/checkpoints/checkpoint.ckpt-29449', '/home/jaehyuk/code/github/vot-toolkit/tracker/examples/python/checkpoints/checkpoint.ckpt-19633', '/home/jaehyuk/code/github/vot-toolkit/tracker/examples/python/checkpoints/checkpoint.ckpt-9817' ] for checkpoint in cklist: if os.path.exists(checkpoint + '.meta'): ckpt = checkpoint break imagefile = handle.frame() if not imagefile: sys.exit(0) bbox_estim = bbox_estimator(False, logger) tracknet = goturn_net_coord.TRACKNET(BATCH_SIZE, train=False) tracknet.build() sess = tf.Session() init = tf.global_variables_initializer() init_local = tf.local_variables_initializer() sess.run(init) sess.run(init_local) coord = tf.train.Coordinator() # start the threads tf.train.start_queue_runners(sess=sess, coord=coord) ### ckpt if not os.path.exists(ckpt_dir): os.makedirs(ckpt_dir)
step = 1 for st in stlist: if os.path.exists(st): step = int(st.split('.')[1]) break # debug # step = 50 imagefile = handle.frame() if not imagefile: sys.exit(0) bbox_estim = bbox_estimator(False, logger) tracknet = goturn_net_coord.TRACKNET(BATCH_SIZE, train=True, online=True) tracknet.build() # TODO check trainiable_variables tvars = tf.trainable_variables() g_vars = [ var for var in tvars if 'fc1_image' or 'fc2_image' or 'fc3_image' or 'fc4_image' in var.name ] # g_vars = [var for var in tvars if 'fc1_image' in var.name] train_step = tf.train.AdamOptimizer(1e-8).minimize(tracknet.loss) sess = tf.Session() init = tf.global_variables_initializer() init_local = tf.local_variables_initializer() sess.run(init)