def test(itr, dataset, args, model, logger, device):
done = False
instance_logits_stack = []
element_logits_stack = []
labels_stack = []
while not done:
if dataset.currenttestidx % 100 == 0:
print('Testing test data point %d of %d' %
(dataset.currenttestidx, len(dataset.testidx)))
features, labels, done = dataset.load_data(is_training=False)
features = torch.from_numpy(features).float().to(device)
with torch.no_grad():
_, element_logits = model(Variable(features),
is_training=False) # (1,T,20)
tmp = F.softmax(torch.mean(torch.topk(element_logits,
k=int(np.ceil(len(features) /
8)),
dim=0)[0],
dim=0),
dim=0).cpu().data.numpy()
element_logits = element_logits.cpu().data.numpy() # (1,T,20)
instance_logits_stack.append(tmp)
element_logits_stack.append(element_logits)
labels_stack.append(labels)
instance_logits_stack = np.array(instance_logits_stack)
labels_stack = np.array(labels_stack)
dmap, iou = dmAP(element_logits_stack, dataset.path_to_annotations, args)
if args.dataset_name == 'Thumos14':
test_set = sio.loadmat('test_set_meta.mat')['test_videos'][0]
for i in range(np.shape(labels_stack)[0]):
if test_set[i]['background_video'] == 'YES':
labels_stack[i, :] = np.zeros_like(labels_stack[i, :])
cmap = cmAP(instance_logits_stack, labels_stack) # (212,20), (212,20)
print('Classification map %f' % cmap)
print('Detection map @ %f = %f' % (iou[0], dmap[0]))
print('Detection map @ %f = %f' % (iou[1], dmap[1]))
print('Detection map @ %f = %f' % (iou[2], dmap[2]))
print('Detection map @ %f = %f' % (iou[3], dmap[3]))
print('Detection map @ %f = %f' % (iou[4], dmap[4]))
logger.log_value('Test Classification mAP', cmap, itr)
for item in list(zip(dmap, iou)):
logger.log_value('Test Detection mAP @ IoU = ' + str(item[1]), item[0],
itr)
utils.write_to_file(args.dataset_name, dmap, cmap, itr)
def test(dataset, args, itr):
# Placeholders
feature_seq = tf.placeholder(tf.float32,
[1, args.max_seqlen, args.feature_size])
seq_len = tf.cast(
tf.reduce_sum(tf.sign(tf.reduce_max(tf.abs(feature_seq), axis=2)),
axis=1), tf.int32)
fseq = feature_seq[:, :tf.reduce_max(seq_len), :]
sgn = tf.sign(tf.reduce_sum(tf.abs(fseq), keep_dims=True, axis=2))
seq_len = tf.cast(
tf.reduce_sum(tf.sign(tf.reduce_max(tf.abs(fseq), axis=2)), axis=1),
tf.int32)
k = tf.cast(tf.ceil(tf.cast(seq_len, tf.float32) / 8), tf.int32)
with tf.device('/gpu:0'):
with tf.variable_scope('Fully_Connected', reuse=True):
fc_W = _variable_with_weight_decay(
'fc_w', [args.feature_size, args.feature_size], 0.0005)
fc_b = _variable_with_weight_decay('fc_b', [args.feature_size],
0.0000)
feature = tf.matmul(
fseq, tf.tile(tf.expand_dims(fc_W, 0), [1, 1, 1])) + fc_b
feature = tf.nn.relu(feature)
feature = tf.nn.dropout(feature, 1.0)
with tf.variable_scope('Attention', reuse=True) as an:
atn_W = _variable_with_weight_decay(
'atn_w', [args.feature_size, args.num_class], 0.0005)
atn_b = _variable_with_weight_decay('atn_b', [args.num_class],
0.0000)
temporal_logits = tf.matmul(
feature, tf.tile(tf.expand_dims(atn_W, 0), [1, 1, 1])) + atn_b
# Initialize everything
init = tf.global_variables_initializer()
config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(init)
saver = tf.train.Saver()
saver.restore(
sess, tf.train.latest_checkpoint('./ckpt/' + args.model_name + '/'))
# Test
element_logit_stack = []
instance_logit_stack = []
label_stack = []
done = False
while not done:
features, label, done = dataset.load_data(is_training=False)
element_logit = []
features = np.concatenate([
features,
np.zeros((args.max_seqlen - len(features) % args.max_seqlen,
args.feature_size))
],
axis=0)
for i in range(0, len(features), args.max_seqlen):
tmp = sess.run(temporal_logits,
feed_dict={
feature_seq:
np.expand_dims(features[i:i + args.max_seqlen],
axis=0)
})
if len(element_logit) == 0:
element_logit = np.squeeze(tmp)
else:
element_logit = np.concatenate(
[element_logit, np.squeeze(tmp)], axis=0)
element_logit = np.array(element_logit)
instance_logit = np.mean(np.sort(
element_logit,
axis=0)[::-1][:max(1, int(np.shape(element_logit)[0] / 8)), :],
axis=0)
instance_logit = np.exp(instance_logit)
instance_logit /= np.sum(instance_logit)
instance_logit_stack.append(instance_logit)
element_logit_stack.append(element_logit)
label_stack.append(label)
instance_logit_stack = np.array(instance_logit_stack)
label_stack = np.array(label_stack)
dmap, iou = dmAP(element_logit_stack, dataset.path_to_annotations)
if args.dataset_name == 'Thumos14':
test_set = sio.loadmat('test_set_meta.mat')['test_videos'][0]
for i in range(np.shape(label_stack)[0]):
if test_set[i]['background_video'] == 'YES':
label_stack[i, :] = np.zeros_like(label_stack[i, :])
cmap = cmAP(instance_logit_stack, label_stack)
np.save('package-1.npy', (instance_logit_stack, label_stack))
print('Classification map %f' % cmap)
print('Detection map @ %f = %f' % (iou[0], dmap[0]))
print('Detection map @ %f = %f' % (iou[1], dmap[1]))
print('Detection map @ %f = %f' % (iou[2], dmap[2]))
print('Detection map @ %f = %f' % (iou[3], dmap[3]))
print('Detection map @ %f = %f' % (iou[4], dmap[4]))
utils.write_to_file(args.dataset_name, dmap, cmap, itr)
def test(itr, dataset, args, model, logger, device):
done = False
# flow,rgb,tcam的综合分类结果
instance_logits_stack = [] #(2380,100)
# tcam的分类结果
tcam_stack = [] #(2380,T,100)
# label标签
labels_stack = [] #(2380,100)
while not done:
if dataset.currenttestidx % 100 ==0:
print('Testing test data point %d of %d' %(dataset.currenttestidx, len(dataset.testidx)))
features, labels, done = dataset.load_data(is_training=False)
seq_len = np.sum(np.max(np.abs(features), axis=2) > 0, axis=1)
features = torch.from_numpy(features).float().to(device)
with torch.no_grad():
_, logits_f, _, logits_r, tcam, _ = model(Variable(features), device, is_training=False, seq_len=torch.from_numpy(seq_len).to(device))
logits_f, logits_r, tcam = logits_f[0], logits_r[0], tcam[0]
topk = int(np.ceil(len(features[0])/8))
tmp = F.softmax(torch.mean(torch.topk(logits_f, k=topk, dim=0)[0], dim=0), dim=0).cpu().data.numpy()
tmp += F.softmax(torch.mean(torch.topk(logits_r, k=topk, dim=0)[0], dim=0), dim=0).cpu().data.numpy()
tmp += F.softmax(torch.mean(torch.topk(tcam, k=topk, dim=0)[0], dim=0), dim=0).cpu().data.numpy()
tcam = tcam.cpu().data.numpy()
instance_logits_stack.append(tmp)
tcam_stack.append(tcam)
labels_stack.append(labels)
instance_logits_stack = np.array(instance_logits_stack) #(video_num,20)
labels_stack = np.array(labels_stack) # (video_num,20)
if args.dataset_name.find('Thumos14')!= -1 and args.num_class == 101:
test_set = sio.loadmat('test_set_meta.mat')['test_videos'][0]
bg_vid = 0
for i in range(np.shape(labels_stack)[0]):
if test_set[i]['background_video'] == 'YES':
bg_vid += 1
labels_stack[i,:] = np.zeros_like(labels_stack[i,:])
cmap = cmAP(instance_logits_stack, labels_stack)
print(cmap)
# tcam_stack : (video_num,T,20), dataset.path_to_annotations: GT 标注 , None, None
dmap, iou = dmAP(tcam_stack, dataset.path_to_annotations, args.activity_net, valid_id=dataset.lst_valid)
print('Classification map %f' %cmap)
for k in range(len(dmap)):
print('Detection map @ %f = %f' %(iou[k], dmap[k]))
print('Mean Detection map = %f' %(np.mean(dmap)))
dmap += [np.mean(dmap)]
logger.log_value('Test Classification mAP', cmap, itr)
for item in list(zip(dmap,iou)):
logger.log_value('Test Detection mAP @ IoU = ' + str(item[1]), item[0], itr)
utils.write_to_file(args.dataset_name + args.model_name, dmap, cmap, itr)
if args.activity_net:
return dmap.pop()
else:
return dmap[4]