def test_multiple_setups(n_epochs, batches, optimizer_types, learning_rates):
# Chain transformations using Compose
data_transform = transforms.Compose(
[Rescale(250), RandomCrop(224),
Normalize(), ToTensor()])
transformed_dataset = FacialKeyPointsDataset(
csv_file='data/training_frames_keypoints.csv',
root_dir='data/training/',
transform=data_transform)
neural_net = NaimishNet()
neural_net.to(device)
# Use a for loop to test and see how the variables influence the model accuracy
for batch in batches:
data_loader = DataLoader(transformed_dataset,
batch_size=batch,
shuffle=True,
num_workers=4)
for optimizer in optimizer_types:
for lr in learning_rates:
print("Batch: ", batch)
print("Optimizer: ", optimizer)
print("Learning rate: ", lr)
print("NNet: ", "NaimishNet")
train_net(neural_net, n_epochs, nn.MSELoss(), data_loader,
optimizer, lr)
def train_and_save(n_epoch, batch_size, optimizer, learning_rate, path):
data_transform = transforms.Compose(
[Rescale(250), RandomCrop(224),
Normalize(), ToTensor()])
transformed_dataset = FacialKeypointsDataset(
csv_file='data/training_frames_keypoints.csv',
root_dir='data/training/',
transform=data_transform)
data_loader = DataLoader(transformed_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
net = KurbakovNet()
train_net(net, n_epoch, nn.MSELoss(), data_loader, optimizer,
learning_rate)
torch.save(net.state_dict(), path)
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_num
# gpus 0, 1
device0 = torch.device('cuda:0')
device1 = torch.device('cuda:1')
if not torch.cuda.is_available():
raise Exception("No GPU found")
else:
print("===> GPU on")
# Normalization method
mean = get_mean(args.norm_value)
std = get_std(args.norm_value)
norm_method = Normalize(mean, std)
# path setting
train_path = join(args.root_path, args.train_data)
pretrain_path = join(args.root_path, args.pretrain_path)
log_path = join(args.root_path, args.log_path)
print("===> Loading datasets")
train_data = DatasetFromFolder(train_path, args.resize_w, args.resize_h,
args.crop_size, args.fps, args.train_frames,
args.horizontal_flip, args.norm_value,
norm_method, args.num_of_vid,
args.train_frames)
train_data_loader = DataLoader(dataset=train_data,
num_workers=4,
batch_size=args.batch_size,
