def Encoder(img_size, in_channel, conv_channel, filter_size, latent_dim,
dense_size, bn):
inner_conv_channel = conv_channel // 2
if img_size % 4 != 0:
print("WARNING: image size mod 4 != 0, may produce bug.")
# total input number of the input of the last conv layer, new image size = old / 2 / 2
flatten_img_size = inner_conv_channel * img_size / 4 * img_size / 4
# explain: first two layer's padding = 2, because we set W/S = W/S + floor((-F+2P)/S+1), S=2,F=5,so P=2
if VERBOSE:
print(img_size, in_channel, conv_channel, filter_size, latent_dim, bn)
model = nn.Sequential(
layers.ConvLayer(in_channel,
conv_channel,
filter_size,
stride=2,
padding=2,
bn=bn),
layers.ConvLayer(conv_channel,
inner_conv_channel,
filter_size,
stride=2,
padding=2,
bn=bn),
layers.ConvLayer(inner_conv_channel,
inner_conv_channel,
filter_size,
stride=1,
padding=2,
bn=bn), layers.Flatten(),
layers.Dense(flatten_img_size, dense_size),
layers.Dense(dense_size, latent_dim))
model = model.to(device=device, dtype=dtype)
model = torch.nn.DataParallel(model, device_ids=GPU_IDs)
return model
def create_network(available_actions_count):
# Create the input variables
s1 = T.tensor4("State")
a = T.vector("Action", dtype="int32")
q2 = T.vector("Q2")
r = T.vector("Reward")
isterminal = T.vector("IsTerminal", dtype="int8")
# Create the input layer of the network.
inputLayer = s1
new_w = resolution[0]
new_h = resolution[1]
# Add 2 convolutional layers with ReLu activation
# filter_shape = [num_filters, num_input_feature_maps, filter_height, filter_width]
input_shape_1 = [batch_size, 1, resolution[0], resolution[1]]
filter_shape_1 = [8, 1, 6, 6]
layer1 = layers.ConvLayer(input=inputLayer, filter_shape=filter_shape_1, input_shape=input_shape_1, pool_size=None)
new_w = (new_w - filter_shape_1[2] + 1)/1 # No pooling
new_h = (new_h - filter_shape_1[3] + 1)/1 # No pooling
input_shape_2 = [batch_size, 8, new_w, new_h]
filter_shape_2 = [8, 8, 3, 3]
layer2 = layers.ConvLayer(input=layer1.out, filter_shape=filter_shape_2, input_shape=input_shape_2, pool_size=None)
# Add a single fully-connected layer.
new_w = (new_w - filter_shape_2[2] + 1)/1 # No pooling
new_h = (new_h - filter_shape_2[3] + 1)/1 # No pooling
layer3 = layers.FCLayer(input=layer2.out.flatten(2), fan_in=filter_shape_2[0]*new_w*new_h, num_hidden=128)
# Add the output layer (also fully-connected).
# (no nonlinearity as it is for approximating an arbitrary real function)
layer4 = layers.FCLayer(input=layer3.out, fan_in=128, num_hidden=available_actions_count, activation=None)
layer4_out = layer4.out
q = layer4_out
# target differs from q only for the selected action. The following means:
# target_Q(s,a) = r + gamma * max Q(s2,_) if isterminal else r
target_q = T.set_subtensor(q[T.arange(q.shape[0]), a], r + discount_factor * (1 - isterminal) * q2)
loss = squared_error(q, target_q).mean()
# Update the parameters according to the computed gradient using RMSProp.
params = layer4.params + layer3.params + layer2.params + layer1.params
updates = rmsprop(loss, params, learning_rate)
# Compile the theano functions
print("Compiling the network ...")
function_learn = theano.function([s1, q2, a, r, isterminal], loss, updates=updates, name="learn_fn")
function_get_q_values = theano.function([s1], q, name="eval_fn")
function_get_best_action = theano.function([s1], T.argmax(q, axis=1), name="test_fn")
print("Network compiled.")
def simple_get_best_action(state):
state = np.expand_dims(state, axis=0)
state = np.expand_dims(state, axis=0)
state = np.repeat(state, batch_size, axis=0)
return function_get_best_action(state)
# Returns Theano objects for the net and functions.
return params, function_learn, function_get_q_values, simple_get_best_action
def Decoder(s_dim, z_dim, img_size, img_channel, conv_channel, filter_size,
dense_size, bn):
# TODO
# essentially the mirror version of Encoder
inner_conv_channel = conv_channel // 2
back_img_size = img_size // 4
flatten_img_size = inner_conv_channel * back_img_size * back_img_size
input_dim = s_dim + z_dim
pad = int(
np.floor(filter_size /
2)) # chose pad this way to fullfill floor((-F+2P)/1+1)==0
model = nn.Sequential(
layers.Dense(input_dim, dense_size),
layers.Dense(dense_size,
inner_conv_channel * back_img_size * back_img_size),
layers.Reshape((-1, inner_conv_channel, back_img_size, back_img_size)),
layers.ConvLayer(inner_conv_channel,
inner_conv_channel,
filter_size,
stride=1,
padding=pad,
bn=bn,
upsampling=True),
layers.ConvLayer(inner_conv_channel,
conv_channel,
filter_size,
stride=1,
padding=pad,
bn=bn,
upsampling=True),
layers.ConvLayer(conv_channel,
img_channel,
filter_size,
stride=1,
padding=pad,
bn=bn,
upsampling=False),
)
model = model.to(device=device, dtype=dtype)
model = torch.nn.DataParallel(model, device_ids=GPU_IDs)
return model
def CNN():
net = n.NeuralNetwork([
l.InputLayer(height=28, width=28),
l.ConvLayer(10, kernel=3, init_val=o.randSc, activator=o.sigmoid),
l.MaxLayer(pool=2),
l.FCLayer(height=10, init_val=o.randSc, activator=o.softmax)
], o.crossentropy)
optimizer = o.SGD(0.1)
num_epochs = 2
batch_size = 50
num_class = 2
return net, optimizer, num_epochs, batch_size, num_class
def __init__(self, in_channels=3):
super(Encoder, self).__init__()
layers = [nn.InstanceNorm2d(in_channels, affine=True)]
layers.append(torch.nn.ReflectionPad2d(15))
layer_spec = [[3, 32, 3, 1], [32, 32, 3, 2], [32, 64, 3, 2],
[64, 128, 3, 2], [128, 256, 3, 2]]
pad = True
for spec in layer_spec:
in_channels, out_channels, kernel_size, stride = spec
layers.append(
L.ConvLayer(in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
pad='VALID'))
self.layers = nn.Sequential(*layers)
def Loaded_nn(name='save(cnn(3_5_2)(0_1_2_8))'):
batch_of_weights = load_weights(name)
ccn_pack, fcl_pack = batch_of_weights
cnn_b, cnn_w = ccn_pack
fcl_b, fcl_w = fcl_pack
cnn_w = map(float, np.delete(cnn_w.split('\n'), -1))
cnn_b = map(float, np.delete(cnn_b.split('\n'), -1))
fcl_w = map(float, np.delete(fcl_w.split(', '), -1))
fcl_b = map(float, np.delete(fcl_b.split(', '), -1))
'''
fcl_w = map(float,np.delete(fcl_w.split('\n'),-1))
fcl_b = map(float,np.delete(fcl_b.split('\n'),-1))
'''
weights = [cnn_w, cnn_b, fcl_w, fcl_b]
con_layer = l.ConvLayer(3,
kernel=5,
init_val=o.randSc,
activator=o.sigmoid)
fc_layer = l.FCLayer(height=10, init_val=o.randSc, activator=o.softmax)
net = NeuralNetwork([
l.InputLayer(height=28, width=28), con_layer,
l.MaxLayer(pool=2), fc_layer
], o.crossentropy)
con_layer.w = np.reshape(weights[0], (np.shape(con_layer.w)))
con_layer.b = np.reshape(weights[1], (np.shape(con_layer.b)))
fc_layer.w = np.reshape(weights[2], (np.shape(fc_layer.w)))
fc_layer.b = np.reshape(weights[3], (np.shape(fc_layer.b)))
return net
def CPRS(network_list):
#print(train_data[0].shape)
#Xdim = ( Xnch, Xrow, Xcol )
ds1 = (2, 2)
#st1 = ( 2, 2 )
input = {}
weight_dim = {}
output = {}
output_num = {}
#入力パラメータ
#input_dim = train_data[0].shape
input_dim = (params_dict["InputChannel"], params_dict["InputWidth"],
params_dict["InputHeight"])
#weight_dim[0] = ( 16, input_dim[0], 5, 5 )
#output_num[0] = 100
#output_num[1] = 10
reshape_flag = True
#affine_cnt = 0
layers = []
gamma = {}
beta = {}
#for i in range( len( network_list ) ):
for i in range(params_dict['LayerNum']):
if network_list[i] == "BatchNorm":
input[i] = output[i - 1]
output[i] = input[i]
gamma[i] = theano.shared(
np.ones(input[i], dtype=theano.config.floatX))
beta[i] = theano.shared(
np.zeros(input[i], dtype=theano.config.floatX))
layers.append(convnet.BatchNormLayer(input[i], gamma[i], beta[i]))
if network_list[i] == "Convolution":
input[i] = input_dim if i == 0 else output[i - 1]
weight_dim[i] = (params_dict['Channel' + str(i + 1)], input[i][0],
params_dict['Width' + str(i + 1)],
params_dict['Height' + str(i + 1)])
output[i] = (weight_dim[i][0], input[i][1] - weight_dim[i][2] + 1,
input[i][2] - weight_dim[i][3] + 1)
layers.append(convnet.ConvLayer(input[i], weight_dim[i]))
#elif network_list[i] == "Pool":
elif network_list[i] == "Pooling":
input[i] = output[i - 1]
if input[i][1] % 2 == 0:
#input[i][1] = input[i][1] + 1
#input[i][2] = input[i][2] + 1
#print(input[i][1])
#print(input[i][2])
output[i] = (input[i][0], input[i][1] / 2, input[i][2] / 2)
else:
output[i] = (input[i][0], (input[i][1] / 2) + 1,
(input[i][2] / 2) + 1)
layers.append(
convnet.PoolLayer(input[i], output[i], ds1, bias=True))
elif network_list[i] == "Relu":
input[i] = output[i - 1]
output[i] = input[i]
layers.append(convnet.ReluLayer())
elif network_list[i] == "Affine":
if i == 0:
input[i] = (params_dict["InputChannel"],
params_dict["InputWidth"],
params_dict["InputHeight"])
else:
input[i] = output[i - 1]
if reshape_flag == True: input[i] = int(np.prod(input[i]))
#output[i] = output_num[affine_cnt]
output[i] = params_dict['Output' + str(i + 1)]
layers.append(
convnet.AffineLayer(input[i], output[i], T4toMat=reshape_flag))
reshape_flag = False
#affine_cnt += 1
elif network_list[i] == "Softmax":
layers.append(convnet.SoftmaxLayer())
cnn = CNN(layers)
return cnn
def __init__(self, rng, batch_size, input):
conv1_1 = layers.ConvLayer(rng,
input=input,
image_shape=(batch_size, 3, 480, 320),
filter_shape=(32, 3, 3, 3),
bias=True,
padding='valid')
conv1_2 = layers.ConvLayer(rng,
input=K.spatial_2d_padding(conv1_1.output),
image_shape=(batch_size, 32, 480, 320),
filter_shape=(64, 32, 3, 3),
padding='valid')
maxp1 = layers.MaxPooling(input=K.spatial_2d_padding(conv1_2.output),
poolsize=(2, 2))
conv2_1 = layers.ConvLayer(rng,
input=maxp1.output,
image_shape=(batch_size, 64, 240, 160),
filter_shape=(64, 64, 3, 3),
padding='valid')
conv2_2 = layers.ConvLayer(rng,
input=K.spatial_2d_padding(conv2_1.output),
image_shape=(batch_size, 64, 240, 160),
filter_shape=(64, 64, 3, 3),
padding='valid')
# maxp2 = layers.MaxPooling(
# input=K.spatial_2d_padding(conv2_2.output),
# poolsize=(2, 2)
# )
#
# conv3_1 = layers.ConvLayer(
# rng,
# input=maxp2.output,
# image_shape=(batch_size, 10, 120, 80),
# filter_shape=(10, 10, 3, 3),
# padding='valid'
# )
#
# conv3_2 = layers.ConvLayer(
# rng,
# input=K.spatial_2d_padding(conv3_1.output),
# image_shape=(batch_size, 10, 120, 80),
# filter_shape=(10, 10, 3, 3),
# padding='valid'
# )
#
# maxp3 = layers.MaxPooling(
# input=conv3_2.output,
# poolsize=(2, 2)
# )
#
# conv4_1 = layers.ConvLayer(
# rng,
# input=maxp3.output,
# image_shape=(batch_size, 10, 60, 40),
# filter_shape=(10, 10, 3, 3),
# padding='same'
# )
#
# conv4_2 = layers.ConvLayer(
# rng,
# input=conv4_1.output,
# image_shape=(batch_size, 10, 60, 40),
# filter_shape=(10, 10, 3, 3),
# padding='same'
# )
#
# maxp4 = layers.MaxPooling(
# input=conv4_2.output,
# poolsize=(2, 2)
# )
#
# conv5_1 = layers.ConvLayer(
# rng,
# input=maxp4.output,
# image_shape=(batch_size, 10, 30, 20),
# filter_shape=(10, 10, 3, 3),
# padding='same'
# )
#
# conv5_2 = layers.ConvLayer(
# rng,
# input=conv5_1.output,
# image_shape=(batch_size, 10, 30, 20),
# filter_shape=(10, 10, 3, 3),
# padding='same'
# )
#
# maxp5 = layers.MaxPooling(
# input=conv5_2.output,
# poolsize=(2, 2)
# )
#
# flat1_input = maxp5.output.flatten(2)
# flat1_shape = 10*15*10
#
# flat1 = layers.HiddenLayer(
# rng=rng,
# input=flat1_input,
# n_in=flat1_shape,
# n_out=flat1_shape,
# activation=T.nnet.relu
# )
#
# flat2 = layers.HiddenLayer(
# rng=rng,
# input=flat1.output,
# n_in=flat1_shape,
# n_out=flat1_shape,
# activation=T.nnet.relu
# )
#
# remax5_input = flat2.output.reshape((batch_size,10,15,10))
#
# remax5 = layers.ReverseMaxPooling(
# input=remax5_input,
# mask=maxp5.mask,
# poolsize=(2, 2)
# )
#
# deconv5_1 = layers.DeConvLayer(
# rng,
# #input=remax5.output,
# input=conv5_2.output,
# W=conv5_2.W,
# image_shape=(batch_size, 10, 30, 20),
# filter_shape=(10, 10, 3, 3),
# padding='same'
# )
#
# deconv5_2 = layers.DeConvLayer(
# rng,
# input=deconv5_1.output,
# W=conv5_1.W,
# image_shape=(batch_size, 10, 30, 20),
# filter_shape=(10, 10, 3, 3),
# padding='same'
# )
#
# remax4 = layers.ReverseMaxPooling(
# input=deconv5_2.output,
# mask=maxp4.mask,
# poolsize=(2, 2)
# )
#
# deconv4_1 = layers.DeConvLayer(
# rng,
# input=remax4.output,
# W=conv4_2.W,
# image_shape=(batch_size, 10, 60, 40),
# filter_shape=(10, 10, 3, 3),
# padding='same'
# )
#
# deconv4_2 = layers.DeConvLayer(
# rng,
# input=deconv4_1.output,
# W=conv4_1.W,
# image_shape=(batch_size, 10, 60, 40),
# filter_shape=(10, 10, 3, 3),
# padding='same'
# )
#
# remax3 = layers.ReverseMaxPooling(
# input=deconv4_2.output,
# mask=maxp3.mask,
# poolsize=(2, 2)
# )
#
# deconv3_1 = layers.DeConvLayer(
# rng,
# #input=remax3.output,
# input=conv3_2.output,
# W=conv3_2.W,
# image_shape=(batch_size, 10, 118, 78),
# filter_shape=(10, 10, 3, 3),
# padding='same'
# )
#
# deconv3_2 = layers.DeConvLayer(
# rng,
# input=K.spatial_2d_padding(deconv3_1.output),
# W=conv3_1.W,
# image_shape=(batch_size, 10, 120, 80),
# filter_shape=(10, 10, 3, 3),
# padding='same'
# )
#
# remax2 = layers.ReverseMaxPooling(
# input=deconv3_2.output,
# mask=maxp2.mask,
# poolsize=(2, 2)
# )
#
deconv2_1 = layers.DeConvLayer(
rng,
#input=remax2.output,
input=K.spatial_2d_padding(conv2_2.output),
W=conv2_2.W,
image_shape=(batch_size, 64, 240, 160),
filter_shape=(64, 64, 3, 3),
padding='same')
deconv2_2 = layers.DeConvLayer(rng,
input=deconv2_1.output,
W=conv2_1.W,
image_shape=(batch_size, 64, 240, 160),
filter_shape=(64, 64, 3, 3),
padding='same')
remax1 = layers.ReverseMaxPooling(input=deconv2_2.output,
mask=maxp1.mask,
poolsize=(2, 2))
deconv1_1 = layers.DeConvLayer(
rng,
input=remax1.output,
#input=K.spatial_2d_padding(conv1_2.output),
W=conv1_2.W,
image_shape=(batch_size, 64, 480, 320),
filter_shape=(32, 64, 3, 3),
padding='same',
)
deconv1_2 = layers.ConvLayer(rng,
input=deconv1_1.output,
image_shape=(batch_size, 32, 480, 320),
filter_shape=(1, 32, 3, 3),
bias=True,
padding='same',
activation=T.nnet.sigmoid)
self.y_pred = T.clip(deconv1_2.output, 0.001, 0.999)
# self.params=conv1_1.params+conv1_2.params+conv2_1.params+conv2_2.params+\
# conv3_1.params+conv3_2.params+conv4_1.params+conv4_2.params+conv5_1.params+conv5_2.params+\
# deconv5_1.params+deconv5_2.params+deconv4_1.params+deconv4_2.params+deconv3_1.params+\
# deconv3_2.params+deconv2_1.params+deconv2_2.params+deconv1_1.params+deconv1_2.params
self.params = conv1_1.params + conv1_2.params + conv2_1.params + conv2_2.params + deconv1_2.params
self.input = input
from activation_functions import Sigmoid, Tanh, Linear
from loss_functions import SoftMaxCrossEntropyLoss
from mnist_data import x_train_2D as X
from mnist_data import y_train_reformatted as y
from mnist_data import x_test_2D as x_test
from mnist_data import y_test_reformatted as y_test
# ===================================================================================
nn = NN.NN(loss_func=SoftMaxCrossEntropyLoss())
convLayer = layers.ConvLayer(
n_channels=9,
kernel_size=9,
weight_init='glorot',
activation_func=Tanh(),
flatten=True,
dropout=0.8,
)
# pool = layers.PoolingLayer(pool_size=7, flatten=True)
dense = layers.DenseLayer(10, Linear(), weight_initialisation='glorot')
nn.add_layer(convLayer)
# nn.add_layer(pool)
nn.add_layer(dense)
optimizer = optimizers.MomentumSGD(learning_rate=0.01, momentum=0.90)
trainer = NN.Trainer(nn, optimizer)
# statistic
batch_size = 50
def __init__(self):
self.convlayer = ly.ConvLayer()
self.relu = ly.ReLU()
