def summary(self, input_shape, batch_size=1, device='cpu'):
print("[%s] Network summary..." % (self.__class__.__name__))
torchsummary.summary(self,
input_size=input_shape,
batch_size=batch_size,
device=device)
input = torch.randn([1, *input_shape], dtype=torch.float)
counter = add_flops_counting_methods(self)
counter.eval().start_flops_count()
counter(input)
print('Flops: {}'.format(
flops_to_string(counter.compute_average_flops_cost())))
print(
'----------------------------------------------------------------')
args.lr = args.learning_rate
others = args.weight_decay * 0.01
if not os.path.isdir(train_config['save_dir']):
os.mkdir(train_config['save_dir'])
model_name = train_config["Model"]
if data_config["dataset_name"].startswith('cifar'):
model = models.cifar.__dict__[model_name](
num_classes=data_config["num_classes"])
else:
model = models.imagenet.__dict__[model_name](
num_classes=data_config["num_classes"], pretrained=True)
model_eval = add_flops_counting_methods(model)
model_eval.eval().start_flops_count()
batch = torch.FloatTensor(1, 3, data_config["w"], data_config["h"])
out = model_eval(batch)
N_flop = model.compute_average_flops_cost()
total_parameters = netParams(model)
print("num_classes: {}".format(data_config["num_classes"]))
print("total_parameters: {}".format(total_parameters))
if use_cuda:
if num_gpu > 1:
model = torch.nn.DataParallel(model)
print("make DataParallel")
model = model.cuda()
start_epoch = 0
Max_val_acc1 = 0.0
#reporting results in fps:
total=0
for x in range(0,200):
image = torch.randn(1, 3, 360, 640).cuda()
with torch.no_grad():
a = time.perf_counter()
output = net.forward(image)
torch.cuda.synchronize()
b = time.perf_counter()
total+=b-a
print(str(200/total))
batch = torch.FloatTensor(1, 3, 512, 1024).cuda()
model = add_flops_counting_methods(net)
model.eval().start_flops_count()
_ = model(batch)
print('Flops: {}'.format(flops_to_string(model.compute_average_flops_cost())))
print('Params: ' + get_model_parameters_number(model))
composed_transforms_ts = transforms.Compose([
augment.RandomHorizontalFlip(),
augment.CenterCrop((512,1024)),
augment.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),#Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),#augment. Normalize_cityscapes(mean=(72.39, 82.91, 73.16)),#123.15, 115.90, 103.06 72.39, 82.91, 73.16
augment.ToTensor()])