def load_data(data_dir, verbose=False):
dataset = data_utils.get_CIFAR10_data(data_dir)
if verbose:
for k, v in dataset.items():
print("%s : %s " % (k, v.shape))
return dataset
def load_data(data_dir, verbose=False):
if verbose is True:
print("Loading data from %s" % data_dir)
dataset = data_utils.get_CIFAR10_data(data_dir)
if verbose:
for k, v in dataset.items():
print("%s : %s " % (k, v.shape))
return dataset
def load_data(self, data_dir: str) -> None:
self.dataset = data_utils.get_CIFAR10_data(data_dir)
self.train_data: Dict[str, Any] = {
'X_train': self.dataset['X_train'][:self.num_train],
'y_train': self.dataset['y_train'][:self.num_train],
'X_val': self.dataset['X_val'][:self.num_train],
'y_val': self.dataset['y_val'][:self.num_train]
}
if self.verbose:
for k, v in self.dataset.items():
print("%s : %s " % (k, v.shape))
def overfit():
# Data
dataset = data_utils.get_CIFAR10_data('datasets/cifar-10-batches-py')
# Hyperparameters
# for now we just some random params, not found by search
reg = 1e-2
weight_scale = 2e-3
learning_rate = 1e-3
# Training parameters
num_epochs = 40
#train_sizes = [50, 100, 150, 200]
train_sizes = [200, 400, 800, 1000, 1500]
solv_dict = {}
for size in train_sizes:
overfit_data = {
'X_train': dataset['X_train'][:size],
'y_train': dataset['y_train'][:size],
'X_val': dataset['X_val'][:size],
'y_val': dataset['y_val'][:size]
}
model = convnet.ConvNetLayer(hidden_dims=[256],
num_filters=[16],
filter_size=5,
reg=reg,
weight_scale=weight_scale)
solv = solver.Solver(model,
overfit_data,
num_epochs=num_epochs,
optim_config={'learning_rate': learning_rate})
print("Overfitting on %d examples in %d epochs using the following network" % (size, num_epochs))
print(model)
solv.train()
dkey = 'size_%d' % size
solv_dict[dkey] = solv
# Check that we can actually overfit
# Plot the results
fig, ax = vis_solver.get_train_fig()
vis_solver.plot_solver_compare(ax, solv_dict)
plt.show()