Training loop stuff

source

add_callbacks

 add_callbacks (func)

source

Loss

 Loss (train=True, valid=True)

Initialize self. See help(type(self)) for accurate signature.

source

MultiClassAccuracy

 MultiClassAccuracy (train=True, valid=True)

Initialize self. See help(type(self)) for accurate signature.

source

Metric

 Metric (train=True, valid=True)

Initialize self. See help(type(self)) for accurate signature.

source

DictLoggerCallback

 DictLoggerCallback (metrics=None, history=None)

Initialize self. See help(type(self)) for accurate signature.

source

Learner

 Learner (dataloaders, model, loss_func, optimizer, callbacks=[])

Initialize self. See help(type(self)) for accurate signature.

source

one_hot_encode_batch

 one_hot_encode_batch (y, n_classes)

source

metrics_last_pretty

 metrics_last_pretty (metrics, metrics_dict)

Return an str with the last values of the metrics, extended to match the length of the metric names, min 10 characters

source

metrics_names_pretty

 metrics_names_pretty (metrics)

Return an str with the names of the metrics, extended to at least 8 characters

source

ProgressBarCallback

 ProgressBarCallback (metrics=['loss'], plot=True,
                      plot_train_skip_ylim=10, plot_smooth_training=0)

Initialize self. See help(type(self)) for accurate signature.

Type Default Details
metrics list [‘loss’] metrics to display, must be in the `learner.history`` dict
plot bool True plot the metrics
plot_train_skip_ylim int 10 skip the first N training metrics when calculating the ylim
plot_smooth_training int 0 smooth the training metrics with a moving average of N
Returns None 
from tidygrad.utils.data import DataLoader, DataLoaders
from tidygrad.utils.datasets import MNIST, mnist_batch_tfm
from tidygrad.func import sigmoid, relu, BCE_loss, CrossEntropy_loss
from tidygrad.optim import Adam
from functools import partial
BS = 1024

mnist_train = DataLoader(MNIST(fashion=True), batch_size=BS, shuffle=True, batch_tfms=[mnist_batch_tfm])
mnist_test = DataLoader(MNIST(train=False, fashion=True), batch_size=BS, shuffle=False, batch_tfms=[mnist_batch_tfm])
INT_DIM = 128

w1 = Tensor(np.random.randn(784, INT_DIM) * 0.1, "w1", requires_grad=True)
b1 = Tensor(np.ones((1, INT_DIM)) * 0.1, "b1", requires_grad=True)
w2 = Tensor(np.random.randn(INT_DIM, 10) * 0.1, "w2", requires_grad=True)
# b2 = Tensor(np.zeros((1, 10)), "b2")

def linear_model(inputs, params, debug=list()):
    inputs.data = inputs.data.reshape(inputs.data.shape[0], -1)
    x = inputs
    w1, b1, w2 = params
    # print("model inputs", x)

    z1 = relu(x.mmul(w1, "z1") + b1)
    z2 = z1.mmul(w2, "z2")

    # print("model outputs", z2)

    return z2

MM_func = partial(linear_model, params=[w1, b1, w2])
optimizer = Adam([w1, b1, w2], lr=0.005)

loss_f = lambda preds, targets: CrossEntropy_loss(preds, one_hot_encode_batch(targets.data, n_classes=10)).mean()
# loss_f = lambda preds, targets: CrossEntropy_loss(preds, one_hot_encode_batch(targets.data, 10))

student = Learner(
    dataloaders=DataLoaders(mnist_train, mnist_test),
    model=MM_func,
    loss_func=loss_f,
    optimizer=optimizer,
    callbacks=[DictLoggerCallback(), ProgressBarCallback(metrics=[
        "loss",
        "accuracy",
    ], plot_train_skip_ylim=15, plot_smooth_training=5)],
)
student.fit(epochs=5)

Ep  | loss       accuracy   | val_loss   val_accuracy
0   | 0.044432   0.844727   | 0.047987   0.829536    
1   | 0.041537   0.853516   | 0.044185   0.839193    
2   | 0.032763   0.871094   | 0.041888   0.851020    
3   | 0.035643   0.875000   | 0.038516   0.860352    
4   | 0.031717   0.891602   | 0.037972   0.864475    

/home/xl0/work/projects/grads/tidygrad/tidygrad/optim.py:54: RuntimeWarning: underflow encountered in square
  self.vs[i] = self.beta2 * self.vs[i] + (1 - self.beta2) * p.grad**2
/home/xl0/work/projects/grads/tidygrad/tidygrad/optim.py:54: RuntimeWarning: underflow encountered in multiply
  self.vs[i] = self.beta2 * self.vs[i] + (1 - self.beta2) * p.grad**2