print('Testing conv')
print('dx error: ', rel_error(dx_num, dx))
print('dw error: ', rel_error(dw_num, dw))
print('db error: ', rel_error(db_num, db))
# TODO: write script to check the backpropagation on the whole CNN is correct
if check_cnn_back:
model = CNN(image_size=(3, 8, 8),
channels=3,
conv_kernel=5,
pool_kernel=2,
hidden_units=10,
n_classes=10)
X = np.random.randn(2, 3, 8, 8)
y = (np.random.rand(2) * 10).astype(np.int32)
model.oracle(X, y)
loss_fn = SoftmaxCE()
for group in model.param_groups:
for name, param in group.items():
xx = param["param"]
def fx(v):
tmp = param['param']
param['param'] = v
f_x, _ = loss_fn(model.score(X), y)
param['param'] = tmp
return f_x
d_num = eval_numerical_gradient(fx, xx, verbose=False)
if check_cnn_back:
inputs = 2
input_dim = (3, 16, 16)
hidden_units = 10
num_classes = 10
num_filters = 3
filter_size = 3
pool_size = 2
np.random.seed(231)
X = np.random.randn(inputs, *input_dim)
y = np.random.randint(num_classes, size=inputs)
model = CNN(input_dim, num_filters, filter_size, pool_size, hidden_units,
num_classes)
loss, score = model.oracle(X, y)
print('loss:', loss)
a = ['w1', 'w2', 'w3']
b = ['b1', 'b2', 'b3']
for param_name in sorted(a):
f = lambda _: model.oracle(X, y)[0]
param_grad_num = eval_numerical_gradient(
f,
model.param_groups['w'][param_name]['param'],
verbose=False,
h=0.00001)
e = rel_error(param_grad_num,
model.param_groups['w'][param_name]['grad'])
print('%s relative error: %e' %
(param_name,