def build_model(args):
if args.model == "VGG19":
model = VGG19(num_classes=7, input_shape=(48, 48, 3), dropout=0.5)
else:
model = build_resnet(args.model, input_shape=(48, 48, 3), classes=7)
return model
def __init__(self, args, arch, loss, pretrained_weights=None, state=None, cuda=True, fp16=False, distributed=False):
super(ModelAndLoss, self).__init__()
self.arch = arch
self.mask = None
print("=> creating model '{}'".format(arch))
model = models.build_resnet(arch[0], arch[1])
if pretrained_weights is not None:
print("=> using pre-trained model from a file '{}'".format(arch))
model.load_state_dict(pretrained_weights)
if cuda:
model = model.cuda()
if fp16:
model = network_to_half(model)
if distributed:
model = DDP(model)
if not state is None:
model.load_state_dict(state)
# define loss function (criterion) and optimizer
criterion = loss()
if cuda:
criterion = criterion.cuda()
self.model = model
self.loss = criterion
def build_model(args): # hàm truyền vào các tham số để xây model
if args.model == "VGG19":
model = VGG19(num_classes=7, input_shape=(48, 48, 3), dropout=0.5)
else:
model = build_resnet(args.model, input_shape=(48, 48, 3), classes=7)
return model
def build_model(input_shape=(1, 3, 480, 640),
n_classes=4,
weights_dir=None,
pretrained_weight_fname=None,
cuda=True):
backbone = build_resnet(input_shape=input_shape)
model = fcn(input_shape, backbone)
if cuda:
backbone = backbone.cuda()
model = model.cuda()
if pretrained_weight_fname is not None and weights_dir is not None:
restore_weights(model, weights_dir, pretrained_weight_fname)
return model, backbone
def tf_infer(img, save_ckpt=True, restore_from_tfckpt=False, ckpt_path=None):
pred_tensor = build_resnet(img_batch=img,
scope=MODEL_NAME,
is_training=False,
freeze_norm=True,
num_cls=1000)
if restore_from_tfckpt:
print("restore weights from tf_CKPT")
assert not ckpt_path is None, "ckpt_path is None, Erro"
restore_op = tf.train.Saver()
else:
print('restore weights from MxnetWeights')
restore_op = create_resotre_op(MODEL_NAME, Mxnet_Weights_PATH)
if DEBUG:
from resnet_utils import debug_dict
print(debug_dict)
assert len(debug_dict) >= 3, "debug_dict size erro, len is :{}".format(
len(debug_dict))
if save_ckpt:
save_dir = '../tf_ckpts'
if not os.path.exists(save_dir):
os.mkdir(save_dir)
saver = tf.train.Saver(max_to_keep=30)
save_ckpt = os.path.join(save_dir, '%s.ckpt' % MODEL_NAME)
with tf.Session() as sess:
if restore_from_tfckpt:
restore_op.restore(sess, ckpt_path)
else:
sess.run(restore_op)
if DEBUG:
name_val = {}
for name in debug_dict.keys():
name_val[name] = sess.run(debug_dict[name])
pred = sess.run(pred_tensor)
if save_ckpt:
saver.save(sess, save_ckpt)
return pred
def __init__(self, model_dict):
super(ImageClassifier, self).__init__()
model_type = model_dict['model_type']
self.backbone = build_resnet(model_dict['backbone'], model_type)
self.neck = build_neck()
self.head = build_head(model_dict['head'])