def __init__(self, data_dir, transformation, device, num_bin):
args.data_dir = data_dir
args.filename_list = data_dir
args.dataset = 'AFLW2000'
self.transformations = transformation
self.device = device
self.num_bin = num_bin
idx_tensor = [idx for idx in xrange(67)]
self.idx_tensor = torch.FloatTensor(idx_tensor).to(self.device)
if args.dataset == 'Pose_300W_LP':
pose_dataset = datasets.Pose_300W_LP(args.data_dir,
args.filename_list,
self.transformations)
elif args.dataset == 'Pose_300W_LP_random_ds':
pose_dataset = datasets.Pose_300W_LP_random_ds(
args.data_dir, args.filename_list, self.transformations)
elif args.dataset == 'AFLW2000':
pose_dataset = datasets.AFLW2000(args.data_dir, args.filename_list,
self.transformations)
elif args.dataset == 'AFLW2000_ds':
pose_dataset = datasets.AFLW2000_ds(args.data_dir,
args.filename_list,
self.transformations)
elif args.dataset == 'BIWI':
pose_dataset = datasets.BIWI(args.data_dir, args.filename_list,
self.transformations)
elif args.dataset == 'AFLW':
pose_dataset = datasets.AFLW(args.data_dir, args.filename_list,
self.transformations)
elif args.dataset == 'AFLW_aug':
pose_dataset = datasets.AFLW_aug(args.data_dir, args.filename_list,
self.transformations)
elif args.dataset == 'AFW':
pose_dataset = datasets.AFW(args.data_dir, args.filename_list,
self.transformations)
else:
print 'Error: not a valid dataset name'
sys.exit()
self.test_loader = torch.utils.data.DataLoader(
dataset=pose_dataset, batch_size=args.batch_size, num_workers=2)
transforms.Scale(224),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]) # rgb模式
if args.dataset == 'Pose_300W_LP':
pose_dataset = datasets.Pose_300W_LP(args.data_dir, args.filename_list,
transformations)
elif args.dataset == 'Pose_300W_LP_random_ds':
pose_dataset = datasets.Pose_300W_LP_random_ds(args.data_dir,
args.filename_list,
transformations)
elif args.dataset == 'AFLW2000':
pose_dataset = datasets.AFLW2000(args.data_dir, args.filename_list,
transformations)
elif args.dataset == 'AFLW2000_ds':
pose_dataset = datasets.AFLW2000_ds(args.data_dir, args.filename_list,
transformations)
elif args.dataset == 'BIWI':
pose_dataset = datasets.BIWI(args.data_dir, args.filename_list,
transformations)
elif args.dataset == 'AFLW':
pose_dataset = datasets.AFLW(args.data_dir, args.filename_list,
transformations)
elif args.dataset == 'AFLW_aug':
pose_dataset = datasets.AFLW_aug(args.data_dir, args.filename_list,
transformations)
elif args.dataset == 'AFW':
pose_dataset = datasets.AFW(args.data_dir, args.filename_list,
transformations)
load_filtered_state_dict(model, saved_state_dict)
print 'Loading data.'
transformations = transforms.Compose([transforms.Scale(240),
transforms.RandomCrop(224), transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
if args.dataset == 'Pose_300W_LP':
pose_dataset = datasets.Pose_300W_LP(args.data_dir, args.filename_list, transformations, bin_width_degrees)
elif args.dataset == 'Pose_300W_LP_random_ds':
pose_dataset = datasets.Pose_300W_LP_random_ds(args.data_dir, args.filename_list, transformations, bin_width_degrees)
elif args.dataset == 'Synhead':
pose_dataset = datasets.Synhead(args.data_dir, args.filename_list, transformations, bin_width_degrees)
elif args.dataset == 'AFLW2000':
pose_dataset = datasets.AFLW2000(args.data_dir, args.filename_list, transformations, bin_width_degrees)
elif args.dataset == 'BIWI':
pose_dataset = datasets.BIWI(args.data_dir, args.filename_list, transformations, bin_width_degrees)
elif args.dataset == 'AFLW':
pose_dataset = datasets.AFLW(args.data_dir, args.filename_list, transformations, bin_width_degrees)
elif args.dataset == 'AFLW_aug':
pose_dataset = datasets.AFLW_aug(args.data_dir, args.filename_list, transformations, bin_width_degrees)
elif args.dataset == 'AFW':
pose_dataset = datasets.AFW(args.data_dir, args.filename_list, transformations, bin_width_degrees)
else:
print 'Error: not a valid dataset name'
sys.exit()
train_loader = torch.utils.data.DataLoader(dataset=pose_dataset,
batch_size=batch_size,
shuffle=True,
def test():
args = parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_id
mae_yaw = 0
mae_pitch = 0
mae_roll = 0
images = tf.placeholder(dtype=tf.float32,
shape=[None, 224, 224, 3],
name='image_batch')
is_training = tf.placeholder(tf.bool)
num_bins = 66
keep_prob = 0
with slim.arg_scope(nets.resnet_v1.resnet_arg_scope):
net, endpoints = nets.resnet_v1.resnet_v1_50(images,
num_classes=None,
is_training=is_training)
with tf.variable_scope('Logits'):
net = tf.squeeze(net, axis=[1, 2])
net = slim.dropout(net, keep_prob, scope='scope')
yaw = slim.fully_connected(net,
num_outputs=num_bins,
activation_fn=None,
scope='fc_yaw')
pitch = slim.fully_connected(net,
num_outputs=num_bins,
activation_fn=None,
scope='fc_pitch')
roll = slim.fully_connected(net,
num_outputs=num_bins,
activation_fn=None,
scope='fc_roll')
yaw_predicted = tf.nn.softmax(yaw)
pitch_predicted = tf.nn.softmax(pitch)
roll_predicted = tf.nn.softmax(roll)
idx_tensor = [idx for idx in range(66)]
idx_tensor = tf.convert_to_tensor(idx_tensor, tf.float32)
yaw_predicted = tf.reduce_sum(yaw_predicted * idx_tensor, 1) * 3 - 99
pitch_predicted = tf.reduce_sum(pitch_predicted * idx_tensor, 1) * 3 - 99
roll_predicted = tf.reduce_sum(roll_predicted * idx_tensor, 1) * 3 - 99
config = tf.ConfigProto(log_device_placement=False,
allow_soft_placement=True)
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
saver = tf.train.Saver()
ckpt = tf.train.latest_checkpoint(args.checkpoint_dir)
saver.restore(sess, ckpt)
test_data = datasets.AFLW2000()
for i in range(test_data.length):
img = test_data[i][0]
img = img.resize((224, 224), Image.BILINEAR)
precesssed_img = datasets.nomalizing(img, [0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
b_img = tf.expand_dims(precesssed_img, 0)
b_img = b_img.eval()
feed_dict = {images: b_img, is_training: False}
pre_yaw, pre_pitch, pre_roll = sess.run(
[yaw_predicted, pitch_predicted, roll_predicted],
feed_dict=feed_dict)
cont_labels = test_data[i][2]
yaw_label = cont_labels[0]
pitch_label = cont_labels[1]
roll_label = cont_labels[2]
mae_yaw = mae_yaw + abs(pre_yaw - yaw_label)
mae_pitch = mae_pitch + abs(pre_pitch - pitch_label)
mae_roll = mae_roll + abs(pre_roll - roll_label)
mean_yaw = mae_yaw / test_data.length
mean_pitch = mae_pitch / test_data.length
mean_roll = mae_roll / test_data.length
return mean_yaw, mean_pitch, mean_roll
if args.dataset == 'Pose_300W_LP':
pose_dataset = datasets.Pose_300W_LP(args.data_dir, args.filename_list,
transformations)
elif args.dataset == 'Pose_300W_LP_random_ds':
pose_dataset = datasets.Pose_300W_LP_random_ds(args.data_dir,
args.filename_list,
transformations)
elif args.dataset == 'Synhead':
pose_dataset = datasets.Synhead(args.data_dir, args.filename_list,
transformations)
elif args.dataset == 'AFLW2000':
pose_dataset = datasets.AFLW2000(args.data_dir, args.filename_list,
transformations)
elif args.dataset == 'BIWI':
pose_dataset = datasets.BIWI(args.data_dir, args.filename_list,
transformations)
elif args.dataset == 'AFLW':
pose_dataset = datasets.AFLW(args.data_dir, args.filename_list,
transformations)
elif args.dataset == 'AFLW_aug':
pose_dataset = datasets.AFLW_aug(args.data_dir, args.filename_list,
transformations)
elif args.dataset == 'AFW':
pose_dataset = datasets.AFW(args.data_dir, args.filename_list,
transforms.CenterCrop(224), transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
data_dir = os.path.join('Data\\RAW', args.data_dir)
vdata_dir = os.path.join('Data\\RAW', args.vdata_dir)
if args.dataset == 'Pose_300W_Training':
pose_dataset = datasets.Pose_300W_Training(data_dir, args.filename_list, transformations, max_angle=max_angle, bin_angle=angles_per_bin)
elif args.dataset == 'Pose_300W_PNP_Training':
pose_dataset = datasets.Pose_300W_PNP_Training(data_dir, args.filename_list, transformations, max_angle=max_angle, bin_angle=angles_per_bin)
else:
print('Error: not a valid dataset name')
sys.exit()
if args.vdataset == 'AFLW2000':
validate_pose_dataset = datasets.AFLW2000(vdata_dir, args.vfilename_list, vtransformations)
else:
print('Error: not a valid validation dataset name')
sys.exit()
train_loader = torch.utils.data.DataLoader(dataset=pose_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=2)
validation_loader = torch.utils.data.DataLoader(dataset=validate_pose_dataset,
batch_size=1,
num_workers=2)
model.cuda(gpu)
criterion = nn.CrossEntropyLoss().cuda(gpu)
if args.dataset == 'Pose_300W_LP':
train_filename_list = os.path.join(args.filename_list,
'train_filename_all.npy')
val_filename_list = os.path.join(args.filename_list,
'val_filename_all.npy')
pose_dataset_train = datasets.Pose_300W_LP(args.data_dir, num_bins,
train_filename_list,
transformations, args.debug)
pose_dataset_val = datasets.Pose_300W_LP(args.data_dir, num_bins,
val_filename_list,
transformations, args.debug)
elif args.dataset == 'AFLW2000':
test_filename_list = os.path.join(args.filename_list,
'test_filename.npy')
pose_dataset = datasets.AFLW2000(args.data_dir, num_bins,
args.filename_list, transformations)
else:
print('Error: not a valid dataset name')
sys.exit()
train_loader = torch.utils.data.DataLoader(dataset=pose_dataset_train,
batch_size=batch_size,
shuffle=True,
num_workers=4)
val_loader = torch.utils.data.DataLoader(dataset=pose_dataset_val,
batch_size=batch_size,
shuffle=True,
num_workers=4)
## Handling Class Imbalance
yaw_l = []
yaw = Dense(units=BIN_NUM, name='yaw')(feature)
pitch = Dense(units=BIN_NUM, name='pitch')(feature)
roll = Dense(units=BIN_NUM, name='roll')(feature)
model = Model(inputs=inputs, outputs=[yaw, pitch, roll])
model.compile(optimizer='adam',
loss={
'yaw': loss_angle,
'pitch': loss_angle,
'roll': loss_angle,
})
# model.load_weights(BIWI_PRETRAINED)
aflw2000 = datasets.AFLW2000(AFLW2000_DATA_DIR,
'/filename_list.txt',
batch_size=BATCH_SIZE,
input_size=INPUT_SIZE)
model.summary()
model.fit(x=aflw2000.data_generator(),
epochs=EPOCHS,
steps_per_epoch=aflw2000.train_num // BATCH_SIZE,
max_queue_size=10,
workers=1,
verbose=1)
model.save(MODELS_DIR + "hopenet2.h5")
