def test_model_acc():
correct = 0
for test_batch_in, test_batch_out in test_batches:
test_output = my_model(Variable(test_batch_in)).data
correct += np.sum(np.argmax(test_output, axis=1) == np.argmax(test_batch_out, axis=1))
my_acc = correct / len(test_dataset)
return my_acc
def __init__(self, num_input, num_output):
super().__init__()
self.num_input = num_input
self.num_output = num_output
self.weights = Variable(
np.random.normal(0, 0.05, (self.num_input, self.num_output)))
self.biases = variable.zeros(self.num_output, np.float32)
self.register_variables(('weight', self.weights),
('bias', self.biases))
def __init__(self,
input_channels,
output_channels,
kernel_size,
stride=1,
padding=0,
add_bias=True):
super().__init__()
if isinstance(kernel_size, int):
kernel_size = (kernel_size, kernel_size)
if isinstance(stride, int):
stride = (stride, stride)
if isinstance(padding, int):
padding = (padding, padding)
self.input_channels = input_channels
self.output_channels = output_channels
self.kernel_size = kernel_size
self.padding = padding
self.stride = stride
self.weights = Variable(
np.random.normal(0, 0.05,
(self.output_channels, self.input_channels,
self.kernel_size[0], self.kernel_size[1])))
self.add_bias = add_bias
if self.add_bias:
self.biases = variable.zeros(output_channels)
self.register_variables(('weight', self.weights),
('bias', self.biases))
else:
self.register_variables(('weight', self.weights))
self.img2col = Img2Col(self.kernel_size, self.stride)
for test_batch_in, test_batch_out in test_batches:
test_output = my_model(Variable(test_batch_in)).data
correct += np.sum(np.argmax(test_output, axis=1) == np.argmax(test_batch_out, axis=1))
my_acc = correct / len(test_dataset)
return my_acc
finished = False
for it in range(iterations):
if finished:
break
train_batches.shuffle()
start = time()
for i_b, (batch_in, batch_out) in enumerate(train_batches):
model_input = Variable(batch_in)
good_output = Variable(batch_out)
model_output = my_model(model_input)
err = loss(good_output, model_output)
optimizer.zero_grad()
err.backward()
optimizer.step()
if i_b % 100 == 0:
print(i_b)
print("epoch time: {:.2f} seconds".format(time() - start))
acc = test_model_acc()
print("model accuracy: {}".format(acc))
class Conv2d(Module):
def __init__(self,
input_channels,
output_channels,
kernel_size,
stride=1,
padding=0,
add_bias=True):
super().__init__()
if isinstance(kernel_size, int):
kernel_size = (kernel_size, kernel_size)
if isinstance(stride, int):
stride = (stride, stride)
if isinstance(padding, int):
padding = (padding, padding)
self.input_channels = input_channels
self.output_channels = output_channels
self.kernel_size = kernel_size
self.padding = padding
self.stride = stride
self.weights = Variable(
np.random.normal(0, 0.05,
(self.output_channels, self.input_channels,
self.kernel_size[0], self.kernel_size[1])))
self.add_bias = add_bias
if self.add_bias:
self.biases = variable.zeros(output_channels)
self.register_variables(('weight', self.weights),
('bias', self.biases))
else:
self.register_variables(('weight', self.weights))
self.img2col = Img2Col(self.kernel_size, self.stride)
# https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/making_faster.html
def forward(self, input_variable: Variable) -> Variable:
"""
Performs 2D convolution on input_variable.
Args:
input_variable (np.array): input image allowed shapes:
[N, C, H, W], [C, H, W]
N - batches,
C - channels,
H - height,
W - width
"""
img2col = self.img2col(input_variable)
unformed_res = self.weights.reshape(self.weights.shape[0],
-1) @ img2col
img_w = input_variable.shape[-1]
img_h = input_variable.shape[-2]
# new image width
new_w = (img_w - self.kernel_size[0]) // self.stride[0] + 1
# new image height
new_h = (img_h - self.kernel_size[1]) // self.stride[1] + 1
batch_input = len(input_variable.shape) == 4
if batch_input:
output_shape = (input_variable.shape[0], self.output_channels,
new_h, new_w)
else:
output_shape = (self.output_channels, new_h, new_w)
if self.add_bias:
unformed_res = (unformed_res.swap_axes(-1, -2) +
self.biases).swap_axes(-1, -2)
return unformed_res.reshape(*output_shape)