def load_data(small_data=False, verbose=False):
"""
Output: A dictionary with the following keys and value shapes
if small_data is True:
X_train (100, 32, 32, 4)
y_train (100, 32, 32, 3)
X_val (10, 32, 32, 4)
y_val (10, 32, 32, 3)
else (small_data is False):
X_train (49000, 32, 32, 4)
y_train (49000, 32, 32, 3)
X_val (1000, 32, 32, 4)
y_val (1000, 32, 32, 3)
X_test (1000, 32, 32, 4)
y_test (1000, 32, 32, 3)
"""
print 'Loading data'
data = get_CIFAR10_data(mode=1)
if small_data:
num_train = 100
data = {
'X_train': data['X_train'][:num_train],
'y_train': data['y_train'][:num_train],
'X_val': data['X_val'][:(num_train / 10)],
'y_val': data['y_val'][:(num_train / 10)],
}
if verbose:
for k, v in data.iteritems():
print '%s: ' % k, v.shape
return data
def test_two_layer_net_solver():
model = TwoLayerNet()
solver = None
data = get_CIFAR10_data()
solver = Solver(model, data, update_rule='sgd', optim_config={'learning_rate': 1e-3}, lr_decay=0.95)
solver.train()
def test_two_layer_net_solver():
from code.solver import Solver
from code.data_utils import get_CIFAR10_data
from code.classifiers.fc_net import TwoLayerNet
model = TwoLayerNet()
solver = None
data = get_CIFAR10_data()
solver = Solver(model,
data,
update_rule='sgd',
optim_config={'learning_rate': 1e-3},
lr_decay=0.95)
solver.train()
def test_sgd():
from code.data_utils import get_CIFAR10_data
from code.classifiers.linear_classifier import Softmax
softmax = Softmax()
data = get_CIFAR10_data()
X_train, y_train = data['X_train'], data['y_train']
X_val, y_val = data['X_val'], data['y_val']
np.random.seed(1)
loss_hist = softmax.train(X_train.reshape((X_train.shape[0], -1)),
y_train,
learning_rate=1e-7,
reg=5e4,
num_iters=1500,
verbose=False)
assert loss_hist[
-1] < 2.2, 'Problem with softmax train sgd, loss: %s' % loss_hist[-1]
y_val_pred = softmax.predict(X_val.reshape((X_val.shape[0], -1)))
val_acc = np.mean(y_val == y_val_pred)
assert val_acc > 0.28, 'Problem with softmax train sgd, val accuracy %f' % val_acc
# All gradients above this will be clipped
GRAD_CLIP = 100
# How often should we check the output?
PRINT_FREQ = 1000
# Number of epochs to train the net
NUM_EPOCHS = 50
# Batch Size
BATCH_SIZE = 20 #128
NUM_LABELS = 256
# Load data
data = get_CIFAR10_data(mode=1)
"""
data is a dictionary with the following keys and value shapes:
X_val: (1000, 4, 32, 32)
X_train: (49000, 4, 32, 32)
X_test: (1000, 4, 32, 32)
y_val: (1000, 3, 32, 32)
y_train: (49000, 3, 32, 32)
y_test: (1000, 3, 32, 32)
"""
# Work with a small subset of data
num_train = 100
small_data = {
'X_train': data['X_train'][:num_train],
'y_train': data['y_train'][:num_train],