def conv_bn_conv_bn_pool2x2(inp_layer, conv_filters, conv_shapes, res_shape,
training_name):
assert conv_shapes[0][1] == conv_shapes[0][2]
pad1 = conv_shapes[0][1] // 2
conv1 = layers.Conv((conv_filters[0], ) + conv_shapes[0], {
'stride': 1,
'pad': pad1
}, inp_layer)
conv1 = layers.SpatialBatchnorm((conv_filters[0], ) + res_shape,
training_name, conv1)
conv1 = layers.Relu(conv1)
conv1 = layers.Dropout(0.6, training_name, conv1)
assert conv_shapes[1][0] == conv_shapes[1][1]
pad2 = conv_shapes[1][1] // 2
conv2 = layers.Conv((conv_filters[1], conv_filters[0]) + conv_shapes[1], {
'stride': 1,
'pad': pad2
}, conv1)
conv2 = layers.SpatialBatchnorm((conv_filters[1], ) + res_shape,
training_name, conv2)
conv2 = layers.Relu(conv2)
conv2 = layers.Dropout(0.6, training_name, conv2)
pool = layers.MaxPool((2, 2), 2, conv2)
return pool
def __init__(self, D_in, outspec, ks=io_size, st=io_stride):
""" outspec should be an Ordered Dict """
nn.Module.__init__(self)
pd = layers.pad_size(ks, "same")
self.output_layers = []
for (name,d_out) in outspec.items():
setattr(self, name, layers.Conv(D_in, d_out, ks, st, pd, bias=True))
self.output_layers.append(name)
def main():
input_X = [[1, 2, 3], [4, 5, 6], [7, 8, 9]]
W = [[1, 1, 1], [0, 0, 0], [1, 1, 1]]
b = [0]
input_X = np.array(input_X).reshape((1, 3, 3, 1))
W = np.array(W).reshape((1, 3, 3, 1))
b = np.array(b).reshape((1, 1, 1, 1))
conv = layers.Conv(num_input_channels=1,
filter_size=3,
padding='valid',
stride=2,
num_filters=3,
initializer='xavier_normal')
pool = layers.Pool(2, 1)
output_Z = conv.forward(input_X)
output_A = pool.forward(input_X)
print("### test convolutional layer ###")
print("input_X = ", input_X)
print("output_Z = ", output_Z)
print("### test pooling layer ###")
print("output_A = ", output_A)
x_train = mf.reshape_x(x_train)
x_test = mf.reshape_x(x_test)
# train shape of (60000, 1, 28, 28) test shape (10000, 1, 28, 28)
x_train = x_train.reshape(x_train.shape[0], 1, 28, 28).astype('float32')
x_test = x_test.reshape(x_test.shape[0], 1, 28, 28).astype('float32')
# min max scale from 0-255 to 0-1 scale
x_train = (x_train - np.min(x_train)) / (np.max(x_train) - np.min(x_train))
x_test = (x_test - np.min(x_test)) / (np.max(x_test) - np.min(x_test))
x_dims = (1, 28, 28)
num_classes = 10
class_names = np.unique(y_train)
# Conv layers with x dims 2 filters with kernel 3x3 stride of 1 and no padding
conv1 = layers.Conv(x_dims,
n_filter=2,
h_filter=3,
w_filter=3,
stride=1,
padding=0)
# activation for layer 1 'sigmoid'
sig = mf.sigmoid()
# MaxPool layer 2x2 stride of 1
pool1 = layers.Maxpool(conv1.out_dim, size=2, stride=2)
# Conv layer with 2 filters kernel size of 3x3 stride of 1 and no padding
conv2 = layers.Conv(pool1.out_dim,
n_filter=2,
h_filter=3,
w_filter=3,
stride=1,
padding=0)
# activation for layer 2 rectified linear
relu = mf.ReLU()
from net import *
from cs231n.solver import *
import layers
if __name__ == "__main__":
# Instantiation example
i1 = layers.Input()
c1 = layers.Conv((8, 3, 3, 3), {'stride': 1, 'pad': 1}, i1)
flat = 8 * 28 * 28
s1 = layers.Reshape((flat, ), c1)
a1 = layers.Affine((flat, 10), s1)
l1 = layers.SoftmaxLoss('y', 'loss', a1)
try:
layers.load_network('network')
except IOError:
pass
model = NeuralNetwork(i1, l1, 'loss', layers.params, layers.grads)
data = {
'X_train': np.ones((2**10, 3, 28, 28)) * 0.1,
'y_train': np.ones(2**10, dtype=np.int) * 2,
'X_val': np.ones((2**3, 3, 28, 28)) * 0.1,
'y_val': np.ones(2**3, dtype=np.int) * 2
}
solver = Solver(model,
data,
update_rule='sgd',
optim_config={
'learning_rate': 1e-3,
},
