Exemple #1
0
def _4e():
    net = hw.DigitsConvNet()
    train, val = utils.torch_digits()
    n_epochs = 30
    optimizer = optim.SGD(net.parameters(), lr=0.005)
    loss_func = nn.CrossEntropyLoss()

    train_el, val_el = hw.fit_and_validate(net, optimizer, loss_func, train, val, n_epochs, batch_size=16)

    _4_plotter(train_el, val_el, 'SGD, lr = 0.005, batch = 16, epochs = 30')
Exemple #2
0
def _4d():
    net = hw.DigitsConvNet()
    train, val = utils.torch_digits()
    n_epochs = 30
    optimizer = optim.SGD(net.parameters(), lr=0.005)
    scheduler = optim.lr_scheduler.ExponentialLR(optimizer, 0.95)

    loss_func = nn.CrossEntropyLoss()

    train_el, val_el = hw.fit_and_validate(net, optimizer, loss_func, train, val, n_epochs, sch=scheduler)
    torch.save(net.cpu().state_dict(), './conv.pb')

    _4_plotter(train_el, val_el, 'SGD, lr = 0.005, gamma = 0.95, batch = 1, epochs = 30')
Exemple #3
0
def _4f():
    net = hw.DigitsConvNet()
    train, val = utils.torch_digits()
    n_epochs = 30
    optimizer = optim.SGD(net.parameters(), lr=0.005)
    scheduler = optim.lr_scheduler.ExponentialLR(optimizer, 0.95)
    loss_func = nn.CrossEntropyLoss()

    train_el, val_el = hw.fit_and_validate(net, optimizer, loss_func, train, val, n_epochs, batch_size=16, sch=scheduler)
    digits, labels, val_X, val_Y = utils.torch_digits(mode=True)

    intermediate = net.intermediate(digits)

    utils.plot_PCA(intermediate.cpu().detach().numpy(), labels.cpu().detach().numpy())
Exemple #4
0
def _4g():
    net = hw.DigitsConvNet()
    train, val = utils.torch_digits()
    n_epochs = 30
    optimizer = optim.SGD(net.parameters(), lr=0.005)
    scheduler = optim.lr_scheduler.ExponentialLR(optimizer, 0.95)
    loss_func = nn.CrossEntropyLoss()

    train_el, val_el = hw.fit_and_validate(net, optimizer, loss_func, train, val, n_epochs, batch_size=16, sch=scheduler)

    digits, labels, val_X, val_Y = utils.torch_digits(mode=True)

    labels = labels.cpu().detach().numpy()
    val_Y = val_Y.cpu().detach().numpy()
    intermediate_train = net.intermediate(digits).cpu().detach().numpy()
    intermediate_val = net.intermediate(val_X).cpu().detach().numpy()

    knn = KNeighborsClassifier(n_neighbors=5, algorithm='kd_tree')
    knn.fit(intermediate_train, labels)
    accuracy = knn.score(intermediate_val, val_Y) 

    print(accuracy)