def train_model():
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# creates the dataloader for pytorch
batch_size = 100
train_loader, validation_loader, test_loader = fashion_mnist_dataloaders(
batch_size, device, "2d")
net = ConvForcedQuantization()
net = net.to(device)
num_epochs = 200
net.train_model(train_loader, validation_loader, test_loader, num_epochs,
"Fashion-Conv-Forced-Quantization")
def train_model(real_model_folder):
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
batch_size = 100
real_model_param_path = os.path.join(model_path, real_model_folder,
"PiReal.param.pickle")
real_param = torch.load(real_model_param_path, map_location="cpu")
logit_net = PiLogitSign(real_param, DiscreteWeights.TERNARY)
logit_net = logit_net.to(device)
train_loader, validation_loader, test_loader = fashion_mnist_dataloaders(
batch_size, device, "flat")
logit_net.train_model(train_loader, validation_loader, test_loader, 500,
"Fashion-Pi-Sign-500e-Ternary", True)
def train_model():
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# basic dataset holder
# creates the dataloader for pytorch
batch_size = 100
train_loader, validation_loader, test_loader = fashion_mnist_dataloaders(
batch_size, device, "flat")
net = PiReal()
net = net.to(device)
num_epochs = 500
# will save metrics and model to disk
return net.train_model(train_loader,
validation_loader,
test_loader,
num_epochs,
model_name="Fashion-pi-real-500epoch")