first:
I didn't change any parameters in this code:
args.original_lr = 1e-6
args.lr = 1e-7
args.batch_size = 1
args.momentum = 0.95 # 0.95
args.decay = 51e-4 # 51e-4
args.start_epoch = 0
args.epochs = 400
args.steps = [-1,1,100,150]
args.scales = [1,1,1,1]
args.workers = 4
second:
I train from the vgg16 weights, and the inital code did something wrong with python 3.7, so I changed it as followed:
self.frontend = make_layers(self.frontend_feat)
self.backend = make_layers(self.backend_feat,in_channels = 512,dilation = True)
self.output_layer = nn.Conv2d(64, 1, kernel_size=1)
if not load_weights:
mod = models.vgg16(pretrained = True)
bb_dict = mod.state_dict()
self._initialize_weights()
pretrained_dict = self.frontend.state_dict()
for keys, values in pretrained_dict.items():
pretrained_dict[keys] = bb_dict['features.'+ keys]
last:
the avg loss in the train can reach around 300, it had a long way from the best MAE, so how can I do to reach the best MAE in ShanghaiTech partA
first:
I didn't change any parameters in this code:
args.original_lr = 1e-6
args.lr = 1e-7
args.batch_size = 1
args.momentum = 0.95 # 0.95
args.decay = 51e-4 # 51e-4
args.start_epoch = 0
args.epochs = 400
args.steps = [-1,1,100,150]
args.scales = [1,1,1,1]
args.workers = 4
second:
I train from the vgg16 weights, and the inital code did something wrong with python 3.7, so I changed it as followed:
self.frontend = make_layers(self.frontend_feat)
self.backend = make_layers(self.backend_feat,in_channels = 512,dilation = True)
self.output_layer = nn.Conv2d(64, 1, kernel_size=1)
if not load_weights:
mod = models.vgg16(pretrained = True)
bb_dict = mod.state_dict()
self._initialize_weights()
pretrained_dict = self.frontend.state_dict()
for keys, values in pretrained_dict.items():
pretrained_dict[keys] = bb_dict['features.'+ keys]
last:
the avg loss in the train can reach around 300, it had a long way from the best MAE, so how can I do to reach the best MAE in ShanghaiTech partA