def init_net_without_loading_data(self, config):
"""This should be called after loading all required data."""
self.config = config
if config.is_output and (not os.path.exists(config.output_dir)):
os.makedirs(config.output_dir)
[num_total_cases, input_dim] = self.train_data.X.shape
self.num_total_cases = num_total_cases
self.input_dim = input_dim
self.num_minibatches = num_total_cases / config.minibatch_size
if self.num_minibatches < 1:
self.num_minibatches = 1
# initialize the network
self.num_layers = config.num_layers
self.layer = []
in_dim = input_dim
for i in range(self.num_layers):
layer_spec = config.layer[i]
self.layer.append(Layer(
in_dim, layer_spec.out_dim, layer_spec.act_type,
layer_spec.weight_decay, layer_spec.weight_constraint,
layer_spec.dropout))
in_dim = layer_spec.out_dim
self.output = OutputLayer(in_dim, config.output.out_dim,
config.output.output_type, config.output.weight_decay,
config.output.weight_constraint, config.output.dropout)
# if not linear output (regression) load task loss function
if not isinstance(self.output.act_type, act.LinearOutput):
if config.task_loss_file != None:
self.task_loss = self.read_loss(config.task_loss_file)
print 'Loading task loss from %s' % config.task_loss_file
else:
self.task_loss = 1 - np.eye(self.train_data.K)
print 'No task loss specified, using 0-1 loss.'
# To use multi-class hinge output, a training loss function is required
if isinstance(self.output.act_type, act.MulticlassHingeOutput):
if config.train_loss_file != None:
self.train_loss = self.read_loss(config.train_loss_file)
print 'Loading surrogate loss from %s' % config.train_loss_file
else:
self.train_loss = 1 - np.eye(self.train_data.K)
print 'No surrogate loss specified, using 0-1 loss.'
self.output.act_type.set_loss(self.train_loss)
# initialize the weights in every layer
self._init_weights(config.init_scale, config.random_seed)
def forward_time(self, ix, prev_cell):
"""Forward in time t given current input and previous hidden cell state."""
# FIXME: There maybe multiple layers here.
# Compute the hidden state
(hidden, output) = self.create_cell()
hidden.forward(self.Wz, self.Wi, self.Wf, self.Wo, self.Rz, self.Ri, self.Rf, self.Ro, self.pi, self.pf,
self.po, self.bz, self.bi, self.bf, self.bo, ix, prev_cell[0].c, prev_cell[0].h)
# Compute the output
output = OutputLayer(self.hidden_size)
output.forward(self.V, hidden.h, self.c)
return (hidden, output)
def forward_time(self, ix, prev_cell):
"""Forward in time t given current input and previous hidden cell state."""
# FIXME: There maybe multiple layers here.
# Compute the hidden state
(hidden, output) = self.create_cell()
hidden.forward(self.Wz, self.Wi, self.Wf, self.Wo, self.Rz, self.Ri,
self.Rf, self.Ro, self.pi, self.pf, self.po, self.bz,
self.bi, self.bf, self.bo, ix, prev_cell[0].c,
prev_cell[0].h)
# Compute the output
output = OutputLayer(self.hidden_size)
output.forward(self.V, hidden.h, self.c)
return (hidden, output)
class TrainingResultsTest( unittest.TestCase ):
def setUp( self ):
self.tr = TrainingResults()
from opencl import OpenCL
from layer import InputLayer, Layer, OutputLayer, ExecutionContext
self.ocl = OpenCL( pyopencl.create_some_context() )
self.i = InputLayer( 2, self.ocl )
self.h = Layer( 3, self.ocl )
self.o = OutputLayer( 1, self.ocl )
self.i.link_next( self.h )
self.h.link_next( self.o, 0, 3 )
self.nnc = ExecutionContext( self.i, self.o, allow_training = True )
self.i.set_weights( numpy.array( [ 0.1 ] * self.i.weights_count, numpy.float32 ) )
self.h.set_weights( numpy.array( [ 0.2 ] * self.h.weights_count, numpy.float32 ) )
self.o.set_weights( numpy.array( [ 0.3 ] * self.o.weights_count, numpy.float32 ) )
def assertArrayEqual( self, ar1, ar2 ):
self.assertEqual( len( ar1 ), len( ar2 ) )
for x, y in zip( numpy.array( ar1, numpy.float32 ), numpy.array( ar2, numpy.float32 ) ):
self.assertAlmostEqual( x, y, places = 5 )
def test_store( self ):
self.tr.reset()
self.assertEqual( self.tr.iterations, numpy.int32( 0 ) )
self.assertGreater( self.tr.minimal_error, numpy.float32( 1e6 ) )
self.assertIsNone( self.tr.optimal_weights )
self.assertAlmostEqual( self.tr.total_time, 0.0 )
self.assertAlmostEqual( self.tr.opencl_time, 0.0 )
self.i.set_inputs( numpy.array( [1.0, 1.0], numpy.float32 ), is_blocking = True )
self.i.process()
initial_result = self.o.get_outputs()
self.tr.store_weights( self.nnc )
self.i.set_weights( numpy.array( [ 0.4 ] * self.i.weights_count, numpy.float32 ) )
self.i.process()
self.assertNotEqual( initial_result, self.o.get_outputs() )
self.tr.apply_weights( self.nnc )
self.i.process()
self.assertArrayEqual( initial_result , self.o.get_outputs() )
def __init__(self, cfg):
self.model = cfg["model"]
self.cost = cfg["cost"]
self.struct = cfg["struct"]
self.activation = cfg["activation"]
# size is the number of weight matrices
self.size = len(cfg["struct"]) - 1
# init of layers
self.layers = []
layer_cfg = {
"model": cfg["model"],
"load_weights": cfg["load_weights"],
"activation": cfg["activation"],
"cost": cfg["cost"]
}
for l in range(self.size):
layer_cfg["shape"] = [
cfg["struct"][l], # input size
cfg["struct"][l + 1] # output size
]
layer_cfg["ID"] = l
if l == self.size - 1:
self.layers.append(OutputLayer(layer_cfg))
else:
self.layers.append(Layer(layer_cfg))
def Mininet_test():
from functions import Sigmoid, Quadratic
w_h = np.ones((2, 2)) * 0.5
w_o = np.ones((1, 3)) * 0.5
f_activation = Sigmoid
f_cost = Quadratic
layers = [
Layer(f_activation, weights=w_h),
OutputLayer(f_activation, f_cost, weights=w_o)
]
net = Network(f_activation, f_cost, layers=layers)
input = [1]
act, cost = net.train(input=input, target=input, learning_rate=1)
print(act, cost)
for l in net.layers:
print(l.delta_w)
def load(name):
with open(name + "/p.pkl", "rb") as file:
print(file)
params = pickle.load(file)
layers = []
for index in range(params["no_of_layers"] - 1):
path = name + "/l" + str(index) + ".npy"
weights = np.load(path)
layers.append(
Layer(f_activation=params["f_activation"], weights=weights))
path = name + "/l" + str(params["no_of_layers"] - 1) + ".npy"
weights = np.load(path)
layers.append(
OutputLayer(f_activation=params["f_activation"],
f_cost=params["f_cost"],
weights=weights))
return Network(f_activation=params["f_activation"],
f_cost=params["f_cost"],
layers=layers)
def init_net(self, config):
"""config is an instance of class Config"""
import os
self.config = config
if config.is_output and (not os.path.exists(config.output_dir)):
os.makedirs(config.output_dir)
self.train_data = self.read_data(config.train_data_file)
if config.is_val:
self.val_data = self.read_data(config.val_data_file)
if config.is_test:
self.test_data = self.read_data(config.test_data_file)
[num_total_cases, input_dim] = self.train_data.X.shape
self.num_total_cases = num_total_cases
self.input_dim = input_dim
self.num_minibatches = num_total_cases / config.minibatch_size
if self.num_minibatches < 1:
self.num_minibatches = 1
# initialize the network
self.num_layers = config.num_layers
self.layer = []
in_dim = input_dim
for i in range(0, self.num_layers):
self.layer.append(
Layer(in_dim, config.layer[i].out_dim,
config.layer[i].act_type))
in_dim = config.layer[i].out_dim
self.output = OutputLayer(in_dim, config.output.out_dim,
config.output.output_type)
# To use multi-class hinge output, we need to specify the loss function
if isinstance(self.output.act_type, act.MulticlassHingeOutput):
if config.loss_file != None:
self.output.act_type.set_loss(self.read_loss(config.loss_file))
else:
self.output.act_type.set_loss(1 - np.eye(self.train_data.K))
# initialize the weights in every layer
self._init_weights(config.init_scale, config.random_seed)
def forward_propagation(self, x):
"""Forward Progation of a single sample."""
tau = len(x)
prev_h = sp.zeros(self.n_hiddens)
cells = [None for i in range(tau)]
for i in range(tau):
# Compute the hidden state
time_input = x[i]
hidden = HiddenLayer()
hidden.forward(self.U, time_input, self.W, prev_h, self.b)
# Compute the output
prev_h = hidden.h
output = OutputLayer()
output.forward(self.V, hidden.h, self.c)
cells[i] = (hidden, output)
return cells
def setUpClass( self ):
from opencl import OpenCL
from layer import InputLayer, OutputLayer, ExecutionContext
self.ocl = OpenCL( pyopencl.create_some_context() )
self.i = InputLayer( 2, self.ocl )
self.o = OutputLayer( 1, self.ocl )
self.i.link_next( self.o )
self.nnc = ExecutionContext( self.i, self.o, allow_training = True )
self.i.set_weights( numpy.array( [ 0.1 ] * self.i.weights_count, numpy.float32 ) )
self.o.set_weights( numpy.array( [ 0.3 ] * self.o.weights_count, numpy.float32 ) )
self.tr = TrainingResults()
self._create_method()
def __init_layers(self, layer_spec):
self.layers = []
last_index = len(layer_spec) - 1
for i, size in enumerate(layer_spec):
if i == 0:
self.layers.append(InputLayer(size, self.activation_fn))
elif i == last_index:
self.layers.append(OutputLayer(size, self.activation_fn))
else:
self.layers.append(HiddenLayer(size, self.activation_fn))
for i in range(len(self.layers) - 1):
self.__join_layers(self.layers[i], self.layers[i+1])
def __init__(self,
architecture=[784, 100, 10],
activation='sigmoid',
learning_rate=0.1,
momentum=0.5,
weight_decay=1e-4,
dropout=0.5,
early_stopping=True,
seed=99):
"""
Neural network model initializer.
"""
# Attributes
self.architecture = architecture
self.activation = activation
self.learning_rate = learning_rate
self.momentum = momentum
self.weight_decay = weight_decay
self.dropout = dropout
self.early_stopping = early_stopping
self.seed = seed
# Turn `activation` and `learning_rate` to class instances
if not isinstance(self.activation, Activation):
self.activation = Activation(self.activation)
if not isinstance(self.learning_rate, LearningRate):
self.learning_rate = LearningRate(self.learning_rate)
# Initialize a list of layers
self.layers = []
for i, (n_in,
n_out) in enumerate(zip(architecture[:-2],
architecture[1:-1])):
l = HiddenLayer('layer{}'.format(i), n_in, n_out, self.activation,
self.learning_rate, self.momentum,
self.weight_decay, self.dropout, self.seed + i)
self.layers.append(l)
# Output layer
n_in, n_out = architecture[-2], architecture[-1]
l = OutputLayer('output_layer', n_in, n_out, self.learning_rate,
self.momentum, self.weight_decay, self.dropout,
self.seed + i + 1)
self.layers.append(l)
# Training updates
self.epoch = 0
self.training_error = []
self.validation_error = []
self.training_loss = []
self.validation_loss = []
def create_layers(f_activation, f_cost, struct):
layers = []
for input, output in zip(struct, struct[1:-1]):
layer = Layer(f_activation, size=[output, input])
layers.append(layer)
layer = OutputLayer(f_activation,
f_cost,
size=[struct[-1], struct[-2]])
layers.append(layer)
return layers
def init_net(self, config):
"""config is an instance of class Config"""
import os
self.config = config
if config.is_output and (not os.path.exists(config.output_dir)):
os.makedirs(config.output_dir)
self.train_data = self.read_data(config.train_data_file)
if config.is_val:
self.val_data = self.read_data(config.val_data_file)
if config.is_test:
self.test_data = self.read_data(config.test_data_file)
[num_total_cases, input_dim] = self.train_data.X.shape
self.num_total_cases = num_total_cases
self.input_dim = input_dim
self.num_minibatches = num_total_cases / config.minibatch_size
if self.num_minibatches < 1:
self.num_minibatches = 1
# initialize the network
self.num_layers = config.num_layers
self.layer = []
in_dim = input_dim
for i in range(0, self.num_layers):
self.layer.append(Layer(
in_dim, config.layer[i].out_dim, config.layer[i].act_type))
in_dim = config.layer[i].out_dim
self.output = OutputLayer(in_dim, config.output.out_dim,
config.output.output_type)
# To use multi-class hinge output, we need to specify the loss function
if isinstance(self.output.act_type, act.MulticlassHingeOutput):
if config.loss_file != None:
self.output.act_type.set_loss(self.read_loss(config.loss_file))
else:
self.output.act_type.set_loss(1 - np.eye(self.train_data.K))
# initialize the weights in every layer
self._init_weights(config.init_scale, config.random_seed)
def __init__(self,
model,
struct,
activation="sigmoid",
cost="quadratic",
load_weights=False,
weight_range=[-1, 1],
bias_range=[0.5, 0.95]):
self.model = model
self.struct = struct
self.activation = activation
self.cost = cost
# size is the number of weight matrices
self.size = len(self.struct) - 1
# init of layers
self.layers = []
for l in range(self.size):
if l == self.size - 1:
self.layers.append(
OutputLayer(model=self.model,
ID=l,
activation=self.activation,
load_weights=load_weights,
in_size=self.struct[l],
out_size=self.struct[l + 1],
weight_range=weight_range,
bias_range=bias_range,
cost=self.cost))
else:
self.layers.append(
Layer(model=self.model,
ID=l,
activation=self.activation,
load_weights=load_weights,
in_size=self.struct[l],
out_size=self.struct[l + 1],
weight_range=weight_range,
bias_range=bias_range))
def main():
sess = tf.Session()
image = read_image('../data/heart.jpg')
image = np.reshape(image, [1, 224, 224, 3]) # type numpy.ndarray
image.astype(np.float32)
parser = Parser('../data/alexnet.cfg')
network_builder = NetworkBuilder("test") # type: NetworkBuilder
network_builder.set_parser(parser)
network = network_builder.build() # type: Network
network.add_input_layer(InputLayer(tf.float32, [None, 224, 224, 3]))
network.add_output_layer(OutputLayer())
network.connect_each_layer()
sess.run(tf.global_variables_initializer())
fc_layer = sess.run(network.output, feed_dict={network.input: image})
def test():
data = pd.read_csv("train.csv").values
x = data[:, 1:]
t = np.identity(10, dtype=np.uint8)[data[:, 0]]
x = (x - x.min()) / x.max()
x = (x - x.mean()) / x.std()
validate = int(x.shape[0] * 0.75)
x_train, t_train = x[:validate], t[:validate]
x_test, t_test = x[validate:], t[validate:]
network = NeuralNetwork(784, error="R2error", optimizer="Gradient")
network.add(Layer(100, activation="Sigmoid"))
network.add(Layer(50, activation="Sigmoid"))
network.add(OutputLayer(10, activation="Softmax"))
network.fit(x_train, t_train, epoch_time=EPOCH_TIME, batch_size=BATCH_SIZE)
network.print_accurate(x_test, t_test)
def build_net_from_copy(self, copy):
"""
Rebuild the net from a copy made by make_copy.
"""
nnstore = copy
self.num_layers = len(nnstore.layer)
self.layer = []
for i in range(self.num_layers):
in_dim, out_dim = nnstore.layer[i].W.shape
new_layer = Layer(in_dim, out_dim, nnstore.layer[i].act_type)
new_layer.load_weight(nnstore.layer[i].W, nnstore.layer[i].b)
self.layer.append(new_layer)
in_dim, out_dim = nnstore.output.W.shape
new_layer = OutputLayer(in_dim, out_dim, nnstore.output.act_type)
new_layer.load_weight(nnstore.output.W, nnstore.output.b)
self.output = new_layer
if self.num_layers > 0:
self.input_dim = self.layer[0].W.shape[0]
else:
self.input_dim = self.output.W.shape[0]
def init_net(self, config):
"""config is an instance of class Config"""
import os
self.config = config
if config.is_output and (not os.path.exists(config.output_dir)):
os.makedirs(config.output_dir)
self.train_data = self.read_data(config.train_data_file)
if config.is_val:
self.val = self.read_data(config.val_data_file)
if config.is_test:
self.test = self.read_data(config.test_data_file)
[num_total_cases, input_dim] = self.train_data.X.shape
self.num_total_cases = num_total_cases
self.input_dim = input_dim
self.num_minibatches = num_total_cases / config.minibatch_size
if self.num_minibatches < 1:
self.num_minibatches += 1
# initialize the network
self.num_layers = config.num_layers
self.layer = []
in_dim = input_dim
for i in range(0, self.num_layers):
self.layer.append(Layer(
in_dim, config.layer[i].out_dim, config.layer[i].act_type))
in_dim = config.layer[i].out_dim
self.output = OutputLayer(in_dim, config.output.out_dim,
config.output.output_type)
# initialize the weights in every layer
self._init_weights(config.init_scale, config.random_seed)
class NN:
"""A class for general purpose neural networks, trained with
backpropagation. The type of activation functions, number of hidden layers
and number of units in each layer, the output function, and other options
during training can be configured."""
def __init__(self):
pass
def init_net(self, config):
"""config is an instance of class Config"""
import os
self.config = config
if config.is_output and (not os.path.exists(config.output_dir)):
os.makedirs(config.output_dir)
self.train_data = self.read_data(config.train_data_file)
if config.is_val:
self.val_data = self.read_data(config.val_data_file)
if config.is_test:
self.test_data = self.read_data(config.test_data_file)
[num_total_cases, input_dim] = self.train_data.X.shape
self.num_total_cases = num_total_cases
self.input_dim = input_dim
self.num_minibatches = num_total_cases / config.minibatch_size
if self.num_minibatches < 1:
self.num_minibatches = 1
# initialize the network
self.num_layers = config.num_layers
self.layer = []
in_dim = input_dim
for i in range(0, self.num_layers):
self.layer.append(Layer(
in_dim, config.layer[i].out_dim, config.layer[i].act_type))
in_dim = config.layer[i].out_dim
self.output = OutputLayer(in_dim, config.output.out_dim,
config.output.output_type)
# To use multi-class hinge output, we need to specify the loss function
if isinstance(self.output.act_type, act.MulticlassHingeOutput):
if config.loss_file != None:
self.output.act_type.set_loss(self.read_loss(config.loss_file))
else:
self.output.act_type.set_loss(1 - np.eye(self.train_data.K))
# initialize the weights in every layer
self._init_weights(config.init_scale, config.random_seed)
def _init_weights(self, init_scale, random_seed=None):
if random_seed:
np.random.seed(random_seed)
for i in range(0, self.num_layers):
self.layer[i].init_weight(init_scale)
self.output.init_weight(init_scale)
def train(self):
config = self.config
# convert t into a matrix in 1-of-K representation if it is a vector
t = self.train_data.T
if not self.config.is_regression:
T_matrix = self.output.act_type.label_vec_to_mat(t, self.train_data.K)
else:
T_matrix = t
layer_config = LayerConfig()
layer_config.learn_rate = config.learn_rate
layer_config.momentum = config.momentum
layer_config.weight_decay = config.weight_decay
nnstore = NNStore()
nnstore.init_from_net(self)
self.display_training_info(-1, 0, 0)
t_start = time.time()
for epoch in range(0, config.num_epochs):
# shuffle the dataset
idx = np.random.permutation(self.num_total_cases)
train_X = self.train_data.X[idx]
train_T = T_matrix[idx]
loss = 0
for batch in range(0, self.num_minibatches):
i_start = batch * config.minibatch_size
if not batch == self.num_minibatches - 1:
i_end = i_start + config.minibatch_size
else:
i_end = self.num_total_cases
X = train_X[i_start:i_end]
T = train_T[i_start:i_end]
Xbelow = X
# forward pass
for i in range(0, self.num_layers):
Xbelow = self.layer[i].forward(Xbelow)
self.output.forward(Xbelow)
# compute loss
loss += self.output.loss(T)
# backprop
dLdXabove = self.output.backprop(layer_config)
for i in range(self.num_layers-1, -1, -1):
dLdXabove = self.layer[i].backprop(dLdXabove, layer_config)
# statistics
avg_loss = 1.0 * loss / self.num_total_cases
if (epoch + 1) % config.epoch_to_display == 0:
self.display_training_info(epoch, avg_loss, time.time() - t_start)
t_start = time.time()
if (epoch + 1) % config.epoch_to_save == 0:
nnstore.update_from_net(self)
nnstore.write(config.output_dir + '/m' + str(epoch + 1) + '.pdata')
def display_training_info(self, epoch, loss, time):
"""Print training information. Use the config information to determine
what information to display."""
if self.config.is_val:
if self.config.is_test:
self._display_training_info(epoch, loss, time,
self.train_data.X, self.train_data.T,
val_data=self.val_data.X, val_labels=self.val_data.T,
test_data=self.test_data.X, test_labels=self.test_data.T)
else:
self._display_training_info(epoch, loss, time,
self.train_data.X, self.train_data.T,
val_data=self.val_data.X, val_labels=self.val_data.T)
else:
if self.config.is_test:
self._display_training_info(epoch, loss, time,
self.train_data.X, self.train_data.T,
test_data=self.test_data.X, test_labels=self.test_data.T)
else:
self._display_training_info(epoch, loss, time,
self.train_data.X, self.train_data.T)
def _display_training_info(self, epoch, loss, time,
train_data, train_labels, val_data=None, val_labels=None,
test_data=None, test_labels=None):
"""Print training information during training."""
print 'epoch %d, loss %.4f,' % (epoch + 1, loss),
# print loss if it is a regression problem
if self.config.is_regression:
if val_data != None and val_labels != None:
self.predict(val_data)
avg_loss = self.output.loss(val_labels) / val_labels.shape[0]
print 'val_loss %.4f,' % (avg_loss),
if test_data != None and test_labels != None:
self.predict(test_data)
avg_loss = self.output.loss(test_labels) / test_labels.shape[0]
print 'test_loss %.4f,' % (avg_loss),
else:
# print accuracy if it is a classification problem
ypred = self.predict(train_data)
acc = (ypred == train_labels.squeeze()).mean()
print 'acc %.4f,' % acc,
if val_data != None and val_labels != None:
ypred = self.predict(val_data)
acc = (ypred == val_labels.squeeze()).mean()
print 'val_acc %.4f,' % acc,
if test_data != None and test_labels != None:
ypred = self.predict(test_data)
acc = (ypred == test_labels.squeeze()).mean()
print 'test_acc %.4f,' % acc,
if self.config.display_winc:
for i in range(0, self.num_layers):
print 'winc%d %.5f,' % (i+1, np.abs(self.layer[i].Winc).max()),
print 'winc_out %.5f,' % np.abs(self.output.Winc).max(),
print 'time %.2f' % time
def _forward(self, X):
"""Do a forward pass without computing the output and predictions.
Used as a subroutine for function predict and check_grad."""
Xbelow = X
for i in range(0, self.num_layers):
Xbelow = self.layer[i].forward(Xbelow)
self.output.forward(Xbelow)
def predict(self, X):
"""Make prediction using the current network.
X: N*D data matrix
ispad: if True, X is padded by an extra dimension of constant 1's
Return an N-element vector of predicted labels.
"""
self._forward(X)
return self.output.predict()
def read_data(self, data_file_name):
"""(data_file_name) --> data
Read from the specified data file, return a data object, which is an
object with three attributes, X, T and K. X and T are the data and
target matrices respectively, and K is the dimensionality of the output.
Each of X and T is a matrix with N rows, N is the number of data
cases."""
f = open(data_file_name)
data_dict = pickle.load(f)
f.close()
data = Data()
data.X = data_dict['data']
data.T = data_dict['labels']
data.K = data_dict['K']
return data
def read_loss(self, loss_file_name):
"""(data_file_name) --> loss
Read from the specified data file, return a loss matrix.
"""
f = open(loss_file_name)
d = pickle.load(f)
f.close()
return d['loss']
def display(self):
print '%d training cases' % self.train_data.X.shape[0]
if self.config.is_val:
print '%d validation cases' % self.val_data.X.shape[0]
if self.config.is_test:
print '%d test cases' % self.test_data.X.shape[0]
print '[' + str(self.output) + ']'
for i in range(self.num_layers-1, -1, -1):
print '[' + str(self.layer[i]) + ']'
print '[input ' + str(self.input_dim) + ']'
print 'learn_rate : ' + str(self.config.learn_rate)
print 'init_scale : ' + str(self.config.init_scale)
print 'momentum : ' + str(self.config.momentum)
print 'weight_decay : ' + str(self.config.weight_decay)
print 'minibatch_size : ' + str(self.config.minibatch_size)
print 'num_epochs : ' + str(self.config.num_epochs)
print 'epoch_to_save : ' + str(self.config.epoch_to_save)
def check_grad(self):
# check the gradient of the 1st layer weights
import scipy.optimize as opt
ncases = 100
def f(w):
if self.num_layers == 0:
Wtemp = self.output.W
self.output.W = w.reshape(Wtemp.shape)
else:
Wtemp = self.layer[0].W
self.layer[0].W = w.reshape(Wtemp.shape)
self._forward(self.train_data.X[:ncases,:])
Z = self.train_data.T[:ncases]
if not self.config.is_regression:
Z = self.output.act_type.label_vec_to_mat(Z, self.train_data.K)
L = self.output.loss(Z) / Z.shape[0]
if self.num_layers == 0:
self.output.W = Wtemp
else:
self.layer[0].W = Wtemp
return L
def fgrad(w):
if self.num_layers == 0:
Wtemp = self.output.W
self.output.W = w.reshape(Wtemp.shape)
else:
Wtemp = self.layer[0].W
self.layer[0].W = w.reshape(Wtemp.shape)
self._forward(self.train_data.X[:ncases,:])
Z = self.train_data.T[:ncases]
if not self.config.is_regression:
Z = self.output.act_type.label_vec_to_mat(Z, self.train_data.K)
self.output.loss(Z)
self.output.gradient()
dLdXabove = self.output.dLdXtop[:,:-1]
for i in range(self.num_layers-1, -1, -1):
self.layer[i].gradient(dLdXabove)
dLdXabove = self.layer[i].dLdXbelow[:,:-1]
if self.num_layers == 0:
grad_w = self.output.dLdW
else:
grad_w = self.layer[0].dLdW
if self.num_layers == 0:
self.output.W = Wtemp
else:
self.layer[0].W = Wtemp
return grad_w.reshape(np.prod(grad_w.shape)) / Z.shape[0]
if self.num_layers == 0:
#W = np.random.randn(
# self.output.W.shape[0], self.output.W.shape[1]) * 1e-3
W = self.output.W
else:
#W = np.random.randn(
# self.layer[0].W.shape[0], self.layer[0].W.shape[1]) * 1e-3
W = self.layer[0].W
print "wmax: %f" % np.abs(fgrad(W.reshape(np.prod(W.shape)))).max()
print "check_grad err: %f" % opt.check_grad(
f, fgrad, W.reshape(np.prod(W.shape)))
epoch = 1
batch_size = 8
interval = 1 # 経過の表示間隔
n_sample = 7 # 誤差計測のサンプル数
n_flt = 6 # n_flt:フィルタ数
flt_h = 3 # flt_h:フィルタ高さ
flt_w = 3 # flt_w:フィルタ幅
# stride:ストライド幅, pad:パディング幅
# -- 各層の初期化 --
cl_1 = ConvLayer(img_ch, img_h, img_w, n_flt, flt_h, flt_w, 1, 1)
pl_1 = PoolingLayer(cl_1.y_ch, cl_1.y_h, cl_1.y_w, 2, 0)
n_fc_in = pl_1.y_ch * pl_1.y_h * pl_1.y_w
ml_1 = MiddleLayer(n_fc_in, 10)
ol_1 = OutputLayer(10, 10)
# -- 順伝播 --
def forward_propagation(x):
n_bt = x.shape[0]
print("forward_propagation()")
print(" x.shape", x.shape)
images = x.reshape(n_bt, img_ch, img_h, img_w)
print(" x.reshape", images.shape)
cl_1.forward(images)
pl_1.forward(cl_1.y)
print(" pl_1.y.shape", pl_1.y.shape, "バッチ数,フィルタ数,img縦,img横")
class NN:
"""A class for general purpose neural networks, trained with
backpropagation. The type of activation functions, number of hidden layers
and number of units in each layer, the output function, and other options
during training can be configured."""
def __init__(self):
pass
def init_net(self, config):
"""config is an instance of class Config"""
import os
self.config = config
if config.is_output and (not os.path.exists(config.output_dir)):
os.makedirs(config.output_dir)
self.train_data = self.read_data(config.train_data_file)
if config.is_val:
self.val = self.read_data(config.val_data_file)
if config.is_test:
self.test = self.read_data(config.test_data_file)
[num_total_cases, input_dim] = self.train_data.X.shape
self.num_total_cases = num_total_cases
self.input_dim = input_dim
self.num_minibatches = num_total_cases / config.minibatch_size
if self.num_minibatches < 1:
self.num_minibatches += 1
# initialize the network
self.num_layers = config.num_layers
self.layer = []
in_dim = input_dim
for i in range(0, self.num_layers):
self.layer.append(Layer(
in_dim, config.layer[i].out_dim, config.layer[i].act_type))
in_dim = config.layer[i].out_dim
self.output = OutputLayer(in_dim, config.output.out_dim,
config.output.output_type)
# initialize the weights in every layer
self._init_weights(config.init_scale, config.random_seed)
def _init_weights(self, init_scale, random_seed=None):
if random_seed:
np.random.seed(random_seed)
for i in range(0, self.num_layers):
self.layer[i].init_weight(init_scale)
self.output.init_weight(init_scale)
def train(self):
config = self.config
layer_config = LayerConfig()
layer_config.learn_rate = config.learn_rate
layer_config.momentum = config.momentum
layer_config.weight_decay = config.weight_decay
nnstore = NNStore()
nnstore.init_from_net(self)
for epoch in range(0, config.num_epochs):
# shuffle the data cases
idx = np.random.permutation(self.num_total_cases)
train_X = self.train_data.X[idx]
train_T = self.train_data.T[idx]
loss = 0
for batch in range(0, self.num_minibatches):
i_start = batch * config.minibatch_size
if not batch == self.num_minibatches - 1:
i_end = i_start + config.minibatch_size
else:
i_end = self.num_total_cases
X = train_X[i_start:i_end]
T = train_T[i_start:i_end]
Xbelow = X
# forward pass
for i in range(0, self.num_layers):
Xbelow = self.layer[i].forward(Xbelow)
self.output.forward(Xbelow)
# compute loss
loss += self.output.loss(T)
# backprop
dLdXabove = self.output.backprop(layer_config)
for i in range(self.num_layers-1, -1, -1):
dLdXabove = self.layer[i].backprop(dLdXabove, layer_config)
# statistics
avg_loss = 1.0 * loss / self.num_total_cases
if (epoch + 1) % config.epoch_to_display == 0:
print 'epoch ' + str(epoch + 1) + ', loss = ' + str(avg_loss)
if (epoch + 1) % config.epoch_to_save == 0:
nnstore.update_from_net(self)
nnstore.write(config.output_dir + '/m' + str(epoch + 1) + '.pdata')
def read_data(self, data_file_name):
"""(data_file_name) --> data
Read from the specified data file, return a data object, which is an
object with three attributes, X, T and K. X and T are the data and
target matrix respectively, and K is the dimensionality of the output.
Each of X and T is a matrix with N rows, N is the number of data
cases"""
f = open(data_file_name)
data_dict = pickle.load(f)
f.close()
X = data_dict['data']
t = data_dict['labels']
K = data_dict['K']
if len(t.shape) == 1 or t.shape[0] == 1 or t.shape[1] == 1:
T = util.vec_to_mat(t, K)
else:
T = t
data = Data()
data.X = X
data.T = T
data.K = K
return data
def save_net(self, model_file_name):
"""Save the current neural net to a file."""
pass
def display(self):
print '[' + str(self.output) + ']'
for i in range(self.num_layers-1, -1, -1):
print '[' + str(self.layer[i]) + ']'
print '[input ' + str(self.input_dim) + ']'
print 'learn_rate : ' + str(self.config.learn_rate)
print 'init_scale : ' + str(self.config.init_scale)
print 'momentum : ' + str(self.config.momentum)
print 'weight_decay : ' + str(self.config.weight_decay)
def create_cell(self):
# FIXME: Should create cells base on network structure
hidden = HiddenLayer(self.hidden_size)
output = OutputLayer(self.hidden_size)
return (hidden, output)
class NN:
"""A class for general purpose neural networks, trained with
backpropagation. The type of activation functions, number of hidden layers
and number of units in each layer, the output function, and other options
during training can be configured."""
def __init__(self):
self.task_loss_fn = None
def load_train_data(self, data):
self.train_data = data
self.train_data.X = gnp.garray(data.X)
def load_val_data(self, data):
self.val_data = data
self.val_data.X = gnp.garray(data.X)
def load_test_data(self, data):
self.test_data = data
self.test_data.X = gnp.garray(data.X)
def init_net_without_loading_data(self, config):
"""This should be called after loading all required data."""
self.config = config
if config.is_output and (not os.path.exists(config.output_dir)):
os.makedirs(config.output_dir)
[num_total_cases, input_dim] = self.train_data.X.shape
self.num_total_cases = num_total_cases
self.input_dim = input_dim
self.num_minibatches = num_total_cases / config.minibatch_size
if self.num_minibatches < 1:
self.num_minibatches = 1
# initialize the network
self.num_layers = config.num_layers
self.layer = []
in_dim = input_dim
for i in range(self.num_layers):
layer_spec = config.layer[i]
self.layer.append(Layer(
in_dim, layer_spec.out_dim, layer_spec.act_type,
layer_spec.weight_decay, layer_spec.weight_constraint,
layer_spec.dropout))
in_dim = layer_spec.out_dim
self.output = OutputLayer(in_dim, config.output.out_dim,
config.output.output_type, config.output.weight_decay,
config.output.weight_constraint, config.output.dropout)
# if not linear output (regression) load task loss function
if not isinstance(self.output.act_type, act.LinearOutput):
if config.task_loss_file != None:
self.task_loss = self.read_loss(config.task_loss_file)
print 'Loading task loss from %s' % config.task_loss_file
else:
self.task_loss = 1 - np.eye(self.train_data.K)
print 'No task loss specified, using 0-1 loss.'
# To use multi-class hinge output, a training loss function is required
if isinstance(self.output.act_type, act.MulticlassHingeOutput):
if config.train_loss_file != None:
self.train_loss = self.read_loss(config.train_loss_file)
print 'Loading surrogate loss from %s' % config.train_loss_file
else:
self.train_loss = 1 - np.eye(self.train_data.K)
print 'No surrogate loss specified, using 0-1 loss.'
self.output.act_type.set_loss(self.train_loss)
# initialize the weights in every layer
self._init_weights(config.init_scale, config.random_seed)
def init_net(self, config):
"""config is an instance of class Config"""
self.train_data = self.read_data(config.train_data_file)
print 'Loading training data from %s' % config.train_data_file
if config.is_val:
self.val_data = self.read_data(config.val_data_file)
print 'Loading validation data from %s' % config.val_data_file
if config.is_test:
self.test_data = self.read_data(config.test_data_file)
print 'Loading test data from %s' % config.test_data_file
self.init_net_without_loading_data(config)
def load_net(self, model_file):
"""Load a saved model from a specified file."""
nnstore = NNStore()
nnstore.load(model_file)
self.build_net_from_copy(nnstore)
def make_copy(self):
"""
Make a CPU copy of the net. This copy can be used to recover the net.
"""
nnstore = NNStore()
nnstore.init_from_net(self)
return nnstore
def build_net_from_copy(self, copy):
"""
Rebuild the net from a copy made by make_copy.
"""
nnstore = copy
self.num_layers = len(nnstore.layer)
self.layer = []
for i in range(self.num_layers):
in_dim, out_dim = nnstore.layer[i].W.shape
new_layer = Layer(in_dim, out_dim, nnstore.layer[i].act_type)
new_layer.load_weight(nnstore.layer[i].W, nnstore.layer[i].b)
self.layer.append(new_layer)
in_dim, out_dim = nnstore.output.W.shape
new_layer = OutputLayer(in_dim, out_dim, nnstore.output.act_type)
new_layer.load_weight(nnstore.output.W, nnstore.output.b)
self.output = new_layer
if self.num_layers > 0:
self.input_dim = self.layer[0].W.shape[0]
else:
self.input_dim = self.output.W.shape[0]
def _init_weights(self, init_scale, random_seed=None):
if random_seed:
np.random.seed(random_seed)
for i in range(0, self.num_layers):
self.layer[i].init_weight(init_scale)
self.output.init_weight(init_scale)
def set_task_loss(self, task_loss_fn):
"""Set the task loss function to be user defined task loss.
task_loss_fn should have a signature like this:
task_loss_fn(OutputType, Y, Z, A)
"""
self.task_loss_fn = task_loss_fn
def _compute_loss(self, X, T, batch_size=1000):
n_total = X.shape[0]
n_batches = n_total / batch_size
loss = 0
for i in range(n_batches):
gnp.free_reuse_cache()
i_start = i * batch_size
if i < n_batches - 1:
i_end = i_start + batch_size
else:
i_end = n_total
Xbatch = X[i_start:i_end]
Tbatch = T[i_start:i_end]
self._forward(Xbatch)
loss += self.output.loss(Tbatch)
return loss / n_total
def train(self):
config = self.config
# convert t into a matrix in 1-of-K representation if it is a vector
t = self.train_data.T
T_matrix = self.output.act_type.label_vec_to_mat(t, self.train_data.K)
layer_config = LayerConfig()
layer_config.learn_rate = config.learn_rate
layer_config.momentum = config.init_momentum
layer_config.weight_decay = config.weight_decay
nnstore = NNStore()
nnstore.init_from_net(self)
best_net = NNStore()
best_net.init_from_net(self)
train_acc, val_acc, test_acc = self.display_training_info(
-1,
self._compute_loss(
self.train_data.X, T_matrix, config.minibatch_size),
0)
acc_rec = np.zeros((config.num_epochs / config.epoch_to_display + 1, 4))
acc_rec[0, 0] = 0
acc_rec[0, 1] = train_acc
if config.is_val:
acc_rec[0, 2] = val_acc
if config.is_test:
acc_rec[0, 3] = test_acc
t_start = time.time()
best_acc = val_acc
if self.config.is_test:
best_test_acc = test_acc
best_epoch = -1
for epoch in range(0, config.num_epochs):
gnp.free_reuse_cache()
# decrease learning rate over time
layer_config.learn_rate = config.learn_rate / \
(epoch / config.lr_drop_rate + 1)
# TODO [dirty] special for Lnsvm
if isinstance(self.output.act_type, act.LnsvmVariantOutput):
#self.output.act_type.n = 3.0 - (3.0 - 0.5) / 50 * epoch
self.output.act_type.n = 0.5
if self.output.act_type.n < 0.5:
self.output.act_type.n = 0.5
if (epoch + 1) % config.epoch_to_display == 0:
print 'n %.4f' % self.output.act_type.n,
if epoch >= config.switch_epoch:
layer_config.momentum = config.final_momentum
# shuffle the dataset
idx = np.random.permutation(self.num_total_cases)
#idx = np.arange(self.num_total_cases)
train_X = self.train_data.X[idx]
train_T = T_matrix[idx]
if config.input_noise > 0:
train_X = train_X * (gnp.rand(train_X.shape) > config.input_noise)
# train_X = train_X + gnp.randn(train_X.shape) * config.input_noise
loss = 0
for batch in range(0, self.num_minibatches):
i_start = batch * config.minibatch_size
if not batch == self.num_minibatches - 1:
i_end = i_start + config.minibatch_size
else:
i_end = self.num_total_cases
X = train_X[i_start:i_end]
T = train_T[i_start:i_end]
# forward pass
self._forward(X)
# compute loss
loss += self.output.loss(T)
if self.output.Y.isnan().any():
import ipdb
ipdb.set_trace()
print 'batch #%d <-- nan' % batch
# backprop
dLdXabove = self.output.backprop(layer_config)
for i in range(self.num_layers-1, -1, -1):
dLdXabove = self.layer[i].backprop(dLdXabove, layer_config)
# statistics
avg_loss = 1.0 * loss / self.num_total_cases
if (epoch + 1) % config.epoch_to_display == 0:
train_acc, val_acc, test_acc = self.display_training_info(
epoch, avg_loss, time.time() - t_start)
if val_acc == None:
val_acc = train_acc
if (config.show_task_loss and val_acc < best_acc) or \
(not config.show_task_loss and val_acc > best_acc):
best_acc = val_acc
best_net.update_from_net(self)
if config.is_test:
best_test_acc = test_acc
best_epoch = epoch
t_start = time.time()
acc_rec[(epoch + 1) / config.epoch_to_display, 0] = epoch + 1
acc_rec[(epoch + 1) / config.epoch_to_display, 1] = train_acc
if config.is_val:
acc_rec[(epoch + 1) / config.epoch_to_display, 2] = val_acc
if config.is_test:
acc_rec[(epoch + 1) / config.epoch_to_display, 3] = test_acc
if (epoch + 1) % config.epoch_to_save == 0:
nnstore.update_from_net(self)
nnstore.write(config.output_dir + '/m' + str(epoch + 1) + '.pdata')
print '----------------------------------------------------------------'
if config.show_task_loss:
s = 'loss'
else:
s = 'acc'
if config.is_val:
print 'Best val_%s %.4f' % (s, best_acc),
else:
print 'Best train_%s %.4f' % (s, best_acc),
if config.is_test:
print '--> test_%s %.4f' % (s, best_test_acc),
print 'at epoch %d' % (best_epoch + 1)
if config.is_output:
f = open('%s/acc_rec.pdata' % config.output_dir, 'w')
pickle.dump(acc_rec, f, -1)
f.close()
self.write_config('%s/cfg.txt' % config.output_dir)
# save the best net
fname = config.output_dir + '/best_net.pdata'
print 'Saving the best model to ' + fname
best_net.write(fname)
if config.is_test:
return (best_acc, best_test_acc)
else:
return (best_acc)
def display_training_info(self, epoch, loss, time):
"""Print training information. Use the config information to determine
what information to display.
Return a 3-tuple (train acc, val acc, test acc)
val acc and test acc will be 0 if no validation/test data are given
"""
if self.config.is_val:
if self.config.is_test:
return self._display_training_info(epoch, loss, time,
self.train_data.X, self.train_data.T,
val_data=self.val_data.X, val_labels=self.val_data.T,
test_data=self.test_data.X, test_labels=self.test_data.T)
else:
return self._display_training_info(epoch, loss, time,
self.train_data.X, self.train_data.T,
val_data=self.val_data.X, val_labels=self.val_data.T)
else:
if self.config.is_test:
return self._display_training_info(epoch, loss, time,
self.train_data.X, self.train_data.T,
test_data=self.test_data.X, test_labels=self.test_data.T)
else:
return self._display_training_info(epoch, loss, time,
self.train_data.X, self.train_data.T)
def _display_training_info(self, epoch, loss, time,
train_data, train_labels, val_data=None, val_labels=None,
test_data=None, test_labels=None):
"""Print training information during training."""
print 'epoch %d, surrogate loss %.4f,' % (epoch + 1, loss),
train_acc = 0
val_acc = None
test_acc = None
acc = 0
# print loss if it is a regression problem
if self.config.is_regression:
# TODO [Dirty code]
#self.predict(train_data)
#avg_loss = self.output.task_loss(train_labels, self.task_loss_fn)
avg_loss = np.sqrt(self._compute_loss(train_data, train_labels) * 2)
print 'train_loss %.4f,' % avg_loss,
if val_data != None and val_labels != None:
#self.predict(val_data)
#avg_loss = self.output.task_loss(val_labels, self.task_loss_fn)
avg_loss = np.sqrt(self._compute_loss(val_data, val_labels) * 2)
print 'val_loss %.4f,' % (avg_loss),
val_acc = avg_loss
if test_data != None and test_labels != None:
#self.predict(test_data)
#avg_loss = self.output.task_loss(test_labels, self.task_loss_fn)
avg_loss = np.sqrt(self._compute_loss(test_data, test_labels) * 2)
print 'test_loss %.4f,' % (avg_loss),
test_acc = avg_loss
else:
# print accuracy if it is a classification problem
ypred = self.predict(train_data)
if self.config.show_accuracy:
acc = (ypred == train_labels.squeeze()).mean()
print 'acc %.4f,' % acc,
if self.config.show_task_loss:
acc = self.task_loss[ypred, train_labels].mean()
print 'loss %.4f,' % acc,
train_acc = acc
if val_data != None and val_labels != None:
ypred = self.predict(val_data)
if self.config.show_accuracy:
acc = (ypred == val_labels.squeeze()).mean()
print 'val_acc %.4f,' % acc,
if self.config.show_task_loss:
acc = self.task_loss[ypred, val_labels].mean()
print 'val_loss %.4f,' % acc,
val_acc = acc
if test_data != None and test_labels != None:
ypred = self.predict(test_data)
if self.config.show_accuracy:
acc = (ypred == test_labels.squeeze()).mean()
print 'test_acc %.4f,' % acc,
if self.config.show_task_loss:
acc = self.task_loss[ypred, test_labels].mean()
print 'test_loss %.4f,' % acc,
test_acc = acc
if self.config.display_winc:
self.display_winc()
print 'time %.2f' % time
return (train_acc, val_acc, test_acc)
def display_winc(self):
"""Display scale of weight updates. This can be used by external
applications."""
for i in range(0, self.num_layers):
print 'winc%d %.5f,' % (i+1, gnp.abs(self.layer[i].Winc).max()),
print 'winc_out %.5f,' % gnp.abs(self.output.Winc).max(),
def _forward(self, X):
"""Do a forward pass without computing the output and predictions.
Used as a subroutine for function predict and check_grad."""
Xbelow = X
for i in range(self.num_layers):
Xbelow = self.layer[i].forward(Xbelow)
self.output.forward(Xbelow)
def predict(self, X):
"""Make prediction using the current network.
X: N*D data matrix
Return an N-element vector of predicted labels.
"""
self._forward(X)
return self.output.predict()
def forward(self, X):
"""Compute the activation for each class.
X: N*D data matrix
Return a N*D activation matrix A.
"""
self._forward(X)
return self.output.A
def _backprop(self, config):
"""Backpropagate through the net from the output layer. This will be
used as an external interface for semi-supervised application, and the
backprop starts from the `update_weights` method of the output layer,
rather than the `backprop` method."""
dLdXabove = self.output.update_weights(config)
for i in range(self.num_layers-1, -1, -1):
dLdXabove = self.layer[i].backprop(dLdXabove, config)
def eval_task_loss(self, X, z, loss):
"""Evaluate the performance of the net using task specific loss.
Classification problems only.
X: N*D data matrix
z: N-d ground truth matrix.
loss: K*K matrix, K is the number of classes.
Return the average loss over all datacases.
"""
y = self.predict(X)
return loss[z, y].mean()
def read_data(self, data_file_name):
"""(data_file_name) --> data
Read from the specified data file, return a data object, which is an
object with three attributes, X, T and K. X and T are the data and
target matrices respectively, and K is the dimensionality of the output.
Each of X and T is a matrix with N rows, N is the number of data
cases."""
f = open(data_file_name)
data_dict = pickle.load(f)
f.close()
data = Data()
data.X = gnp.garray(data_dict['data'])
#data.T = data_dict['labels'].astype(np.float)
data.T = data_dict['labels']
data.K = data_dict['K']
return data
def read_loss(self, loss_file_name):
"""(data_file_name) --> loss
Read from the specified data file, return a loss matrix.
"""
f = open(loss_file_name)
d = pickle.load(f)
f.close()
return d
def write_config(self, filename):
f = open(filename, 'w')
f.write('%d training cases\n' % self.train_data.X.shape[0])
if self.config.is_val:
f.write('%d validation cases\n' % self.val_data.X.shape[0])
if self.config.is_test:
f.write('%d test cases\n' % self.test_data.X.shape[0])
f.write('[' + str(self.output) + ']\n')
for i in range(self.num_layers-1, -1, -1):
f.write('[' + str(self.layer[i]) + ']\n')
f.write('[input ' + str(self.input_dim) + ']\n')
f.write('learn_rate : ' + str(self.config.learn_rate) + '\n')
f.write('init_scale : ' + str(self.config.init_scale) + '\n')
f.write('init_momentum : ' + str(self.config.init_momentum) + '\n')
f.write('switch_epoch : ' + str(self.config.switch_epoch) + '\n')
f.write('final_momentum : ' + str(self.config.final_momentum) + '\n')
f.write('weight_decay : ' + str(self.config.weight_decay) + '\n')
f.write('minibatch_size : ' + str(self.config.minibatch_size) + '\n')
f.write('num_epochs : ' + str(self.config.num_epochs) + '\n')
f.write('epoch_to_save : ' + str(self.config.epoch_to_save) + '\n')
f.close()
def display_structure(self):
print '[' + str(self.output) + ']'
for i in range(self.num_layers-1, -1, -1):
print '[' + str(self.layer[i]) + ']'
print '[input ' + str(self.input_dim) + ']'
def display(self):
print '%d training cases' % self.train_data.X.shape[0]
if self.config.is_val:
print '%d validation cases' % self.val_data.X.shape[0]
if self.config.is_test:
print '%d test cases' % self.test_data.X.shape[0]
self.display_structure()
print 'learn_rate : ' + str(self.config.learn_rate)
print 'init_scale : ' + str(self.config.init_scale)
print 'init_momentum : ' + str(self.config.init_momentum)
print 'switch_epoch : ' + str(self.config.switch_epoch)
print 'final_momentum : ' + str(self.config.final_momentum)
print 'weight_decay : ' + str(self.config.weight_decay)
print 'minibatch_size : ' + str(self.config.minibatch_size)
print 'num_epochs : ' + str(self.config.num_epochs)
print 'epoch_to_save : ' + str(self.config.epoch_to_save)
if self.config.is_output:
print 'output_dir : ' + self.config.output_dir
def check_grad(self):
# check the gradient of the 1st layer weights
import scipy.optimize as opt
ncases = 100
def f(w):
if self.num_layers == 0:
Wtemp = self.output.W
self.output.W = gnp.garray(w.reshape(Wtemp.shape))
else:
Wtemp = self.layer[0].W
self.layer[0].W = gnp.garray(w.reshape(Wtemp.shape))
self._forward(self.train_data.X[:ncases,:])
Z = self.train_data.T[:ncases]
Z = self.output.act_type.label_vec_to_mat(Z, self.train_data.K)
L = self.output.loss(Z) / Z.shape[0]
if self.num_layers == 0:
self.output.W = Wtemp
else:
self.layer[0].W = Wtemp
return L
def fgrad(w):
if self.num_layers == 0:
Wtemp = self.output.W
self.output.W = gnp.garray(w.reshape(Wtemp.shape))
else:
Wtemp = self.layer[0].W
self.layer[0].W = gnp.garray(w.reshape(Wtemp.shape))
self._forward(self.train_data.X[:ncases,:])
Z = self.train_data.T[:ncases]
Z = self.output.act_type.label_vec_to_mat(Z, self.train_data.K)
self.output.loss(Z)
self.output.gradient()
dLdXabove = self.output.dLdXtop
for i in range(self.num_layers-1, -1, -1):
self.layer[i].gradient(dLdXabove)
dLdXabove = self.layer[i].dLdXbelow
if self.num_layers == 0:
grad_w = self.output.dLdW
else:
grad_w = self.layer[0].dLdW
if self.num_layers == 0:
self.output.W = Wtemp
else:
self.layer[0].W = Wtemp
return grad_w.reshape(np.prod(grad_w.shape)).asarray() / Z.shape[0]
if self.num_layers == 0:
W = self.output.W
else:
W = self.layer[0].W
W = W.asarray()
def finite_diff_grad(f, x0):
eps = 1e-8
approx = np.zeros(len(x0))
for i in xrange(len(x0)):
x0plus = x0.copy()
x0minus = x0.copy()
x0plus[i] += eps
x0minus[i] -= eps
approx[i] = (f(x0plus) - f(x0minus)) / (2 * eps)
return approx
net_grad = fgrad(W.reshape(W.size))
fd_grad = finite_diff_grad(f, W.reshape(W.size))
print "wmax: %f" % np.abs(net_grad).max()
print "finite difference grad scale: %f" % np.abs(fd_grad).max()
print "check_grad err: %f" % np.sqrt(((fd_grad - net_grad)**2).sum())
class TrainingMethodTest( unittest.TestCase ):
@classmethod
def setUpClass( self ):
from opencl import OpenCL
from layer import InputLayer, OutputLayer, ExecutionContext
self.ocl = OpenCL( pyopencl.create_some_context() )
self.i = InputLayer( 2, self.ocl )
self.o = OutputLayer( 1, self.ocl )
self.i.link_next( self.o )
self.nnc = ExecutionContext( self.i, self.o, allow_training = True )
self.i.set_weights( numpy.array( [ 0.1 ] * self.i.weights_count, numpy.float32 ) )
self.o.set_weights( numpy.array( [ 0.3 ] * self.o.weights_count, numpy.float32 ) )
self.tr = TrainingResults()
self._create_method()
@classmethod
def _create_method( self ):
pass
def assertArrayEqual( self, ar1, ar2 ):
self.assertEqual( len( ar1 ), len( ar2 ) )
for x, y in zip( numpy.array( ar1, numpy.float32 ), numpy.array( ar2, numpy.float32 ) ):
self.assertAlmostEqual( x, y, places = 5 )
def test_create( self ):
self.setUpClass()
if not getattr( self, 'method', None ):
return
self.assertAlmostEqual( self.method.n, 0.5 )
self.assertAlmostEqual( self.method.alpha, 0.2 )
self.assertAlmostEqual( self.method.kw, 1.03 )
self.assertAlmostEqual( self.method.pd, 0.7 )
self.assertAlmostEqual( self.method.pi, 1.02 )
self.assertAlmostEqual( self.method.last_error, 0.0 )
self.assertEqual( self.method.offline, False )
def test_randomize_weights( self ):
if not getattr( self, 'method', None ):
return
self.i.set_weights( numpy.array( [ 0.1 ] * self.i.weights_count, numpy.float32 ) )
self.assertTrue( all( map( lambda x: abs( x - 0.1 ) < 0.0001, self.i.get_weights() ) ) )
self.method.randomize_weights( self.nnc )
w1 = self.i.get_weights()
self.assertFalse( all( map( lambda x: abs( x - 0.1 ) < 0.0001, self.i.get_weights() ) ) )
self.method.randomize_weights( self.nnc )
self.assertFalse( all( map( lambda x: abs( x - 0.1 ) < 0.0001, self.i.get_weights() ) ) )
self.assertNotAlmostEqual( ( w1 - self.i.get_weights() ).sum(), 0.0 )
def test_adjust_weights( self ):
if not getattr( self, 'method', None ):
return
self.method.last_error = numpy.float32( 1.0 )
self.method.n = numpy.float32( 0.5 )
self.method.kw = numpy.float32( 1.03 )
self.method.pd = numpy.float32( 0.5 )
self.method.pi = numpy.float32( 1.5 )
self.method.adjust_training_parameters( 1.2 )
self.assertAlmostEqual( self.method.n, 0.25 )
self.assertAlmostEqual( self.method.last_error, 1.2 )
self.method.adjust_training_parameters( 1.0 )
self.assertAlmostEqual( self.method.n, 0.375 )
self.assertAlmostEqual( self.method.last_error, 1.0 )
self.method.adjust_training_parameters( 1.0 )
self.assertAlmostEqual( self.method.n, 0.5625 )
self.assertAlmostEqual( self.method.last_error, 1.0 )
def test_prepare_training( self ):
if not getattr( self, 'method', None ):
return
self.method.prepare_training( self.nnc )
self.assertIsInstance( self.method._weights_delta_buf, pyopencl.Buffer )
class NN:
"""A class for general purpose neural networks, trained with
backpropagation. The type of activation functions, number of hidden layers
and number of units in each layer, the output function, and other options
during training can be configured."""
def __init__(self):
pass
def init_net(self, config):
"""config is an instance of class Config"""
import os
self.config = config
if config.is_output and (not os.path.exists(config.output_dir)):
os.makedirs(config.output_dir)
self.train_data = self.read_data(config.train_data_file)
if config.is_val:
self.val_data = self.read_data(config.val_data_file)
if config.is_test:
self.test_data = self.read_data(config.test_data_file)
[num_total_cases, input_dim] = self.train_data.X.shape
self.num_total_cases = num_total_cases
self.input_dim = input_dim
self.num_minibatches = num_total_cases / config.minibatch_size
if self.num_minibatches < 1:
self.num_minibatches = 1
# initialize the network
self.num_layers = config.num_layers
self.layer = []
in_dim = input_dim
for i in range(0, self.num_layers):
self.layer.append(
Layer(in_dim, config.layer[i].out_dim,
config.layer[i].act_type))
in_dim = config.layer[i].out_dim
self.output = OutputLayer(in_dim, config.output.out_dim,
config.output.output_type)
# To use multi-class hinge output, we need to specify the loss function
if isinstance(self.output.act_type, act.MulticlassHingeOutput):
if config.loss_file != None:
self.output.act_type.set_loss(self.read_loss(config.loss_file))
else:
self.output.act_type.set_loss(1 - np.eye(self.train_data.K))
# initialize the weights in every layer
self._init_weights(config.init_scale, config.random_seed)
def _init_weights(self, init_scale, random_seed=None):
if random_seed:
np.random.seed(random_seed)
for i in range(0, self.num_layers):
self.layer[i].init_weight(init_scale)
self.output.init_weight(init_scale)
def train(self):
config = self.config
# convert t into a matrix in 1-of-K representation if it is a vector
t = self.train_data.T
if not self.config.is_regression:
T_matrix = self.output.act_type.label_vec_to_mat(
t, self.train_data.K)
else:
T_matrix = t
layer_config = LayerConfig()
layer_config.learn_rate = config.learn_rate
layer_config.momentum = config.momentum
layer_config.weight_decay = config.weight_decay
nnstore = NNStore()
nnstore.init_from_net(self)
self.display_training_info(-1, 0, 0)
t_start = time.time()
for epoch in range(0, config.num_epochs):
# shuffle the dataset
idx = np.random.permutation(self.num_total_cases)
train_X = self.train_data.X[idx]
train_T = T_matrix[idx]
loss = 0
for batch in range(0, self.num_minibatches):
i_start = batch * config.minibatch_size
if not batch == self.num_minibatches - 1:
i_end = i_start + config.minibatch_size
else:
i_end = self.num_total_cases
X = train_X[i_start:i_end]
T = train_T[i_start:i_end]
Xbelow = X
# forward pass
for i in range(0, self.num_layers):
Xbelow = self.layer[i].forward(Xbelow)
self.output.forward(Xbelow)
# compute loss
loss += self.output.loss(T)
# backprop
dLdXabove = self.output.backprop(layer_config)
for i in range(self.num_layers - 1, -1, -1):
dLdXabove = self.layer[i].backprop(dLdXabove, layer_config)
# statistics
avg_loss = 1.0 * loss / self.num_total_cases
if (epoch + 1) % config.epoch_to_display == 0:
self.display_training_info(epoch, avg_loss,
time.time() - t_start)
t_start = time.time()
if (epoch + 1) % config.epoch_to_save == 0:
nnstore.update_from_net(self)
nnstore.write(config.output_dir + '/m' + str(epoch + 1) +
'.pdata')
def display_training_info(self, epoch, loss, time):
"""Print training information. Use the config information to determine
what information to display."""
if self.config.is_val:
if self.config.is_test:
self._display_training_info(epoch,
loss,
time,
self.train_data.X,
self.train_data.T,
val_data=self.val_data.X,
val_labels=self.val_data.T,
test_data=self.test_data.X,
test_labels=self.test_data.T)
else:
self._display_training_info(epoch,
loss,
time,
self.train_data.X,
self.train_data.T,
val_data=self.val_data.X,
val_labels=self.val_data.T)
else:
if self.config.is_test:
self._display_training_info(epoch,
loss,
time,
self.train_data.X,
self.train_data.T,
test_data=self.test_data.X,
test_labels=self.test_data.T)
else:
self._display_training_info(epoch, loss, time,
self.train_data.X,
self.train_data.T)
def _display_training_info(self,
epoch,
loss,
time,
train_data,
train_labels,
val_data=None,
val_labels=None,
test_data=None,
test_labels=None):
"""Print training information during training."""
print 'epoch %d, loss %.4f,' % (epoch + 1, loss),
# print loss if it is a regression problem
if self.config.is_regression:
if val_data != None and val_labels != None:
self.predict(val_data)
avg_loss = self.output.loss(val_labels) / val_labels.shape[0]
print 'val_loss %.4f,' % (avg_loss),
if test_data != None and test_labels != None:
self.predict(test_data)
avg_loss = self.output.loss(test_labels) / test_labels.shape[0]
print 'test_loss %.4f,' % (avg_loss),
else:
# print accuracy if it is a classification problem
ypred = self.predict(train_data)
acc = (ypred == train_labels.squeeze()).mean()
print 'acc %.4f,' % acc,
if val_data != None and val_labels != None:
ypred = self.predict(val_data)
acc = (ypred == val_labels.squeeze()).mean()
print 'val_acc %.4f,' % acc,
if test_data != None and test_labels != None:
ypred = self.predict(test_data)
acc = (ypred == test_labels.squeeze()).mean()
print 'test_acc %.4f,' % acc,
if self.config.display_winc:
for i in range(0, self.num_layers):
print 'winc%d %.5f,' % (i + 1, np.abs(
self.layer[i].Winc).max()),
print 'winc_out %.5f,' % np.abs(self.output.Winc).max(),
print 'time %.2f' % time
def _forward(self, X):
"""Do a forward pass without computing the output and predictions.
Used as a subroutine for function predict and check_grad."""
Xbelow = X
for i in range(0, self.num_layers):
Xbelow = self.layer[i].forward(Xbelow)
self.output.forward(Xbelow)
def predict(self, X):
"""Make prediction using the current network.
X: N*D data matrix
ispad: if True, X is padded by an extra dimension of constant 1's
Return an N-element vector of predicted labels.
"""
self._forward(X)
return self.output.predict()
def read_data(self, data_file_name):
"""(data_file_name) --> data
Read from the specified data file, return a data object, which is an
object with three attributes, X, T and K. X and T are the data and
target matrices respectively, and K is the dimensionality of the output.
Each of X and T is a matrix with N rows, N is the number of data
cases."""
f = open(data_file_name)
data_dict = pickle.load(f)
f.close()
data = Data()
data.X = data_dict['data']
data.T = data_dict['labels']
data.K = data_dict['K']
return data
def read_loss(self, loss_file_name):
"""(data_file_name) --> loss
Read from the specified data file, return a loss matrix.
"""
f = open(loss_file_name)
d = pickle.load(f)
f.close()
return d['loss']
def display(self):
print '%d training cases' % self.train_data.X.shape[0]
if self.config.is_val:
print '%d validation cases' % self.val_data.X.shape[0]
if self.config.is_test:
print '%d test cases' % self.test_data.X.shape[0]
print '[' + str(self.output) + ']'
for i in range(self.num_layers - 1, -1, -1):
print '[' + str(self.layer[i]) + ']'
print '[input ' + str(self.input_dim) + ']'
print 'learn_rate : ' + str(self.config.learn_rate)
print 'init_scale : ' + str(self.config.init_scale)
print 'momentum : ' + str(self.config.momentum)
print 'weight_decay : ' + str(self.config.weight_decay)
print 'minibatch_size : ' + str(self.config.minibatch_size)
print 'num_epochs : ' + str(self.config.num_epochs)
print 'epoch_to_save : ' + str(self.config.epoch_to_save)
def check_grad(self):
# check the gradient of the 1st layer weights
import scipy.optimize as opt
ncases = 100
def f(w):
if self.num_layers == 0:
Wtemp = self.output.W
self.output.W = w.reshape(Wtemp.shape)
else:
Wtemp = self.layer[0].W
self.layer[0].W = w.reshape(Wtemp.shape)
self._forward(self.train_data.X[:ncases, :])
Z = self.train_data.T[:ncases]
if not self.config.is_regression:
Z = self.output.act_type.label_vec_to_mat(Z, self.train_data.K)
L = self.output.loss(Z) / Z.shape[0]
if self.num_layers == 0:
self.output.W = Wtemp
else:
self.layer[0].W = Wtemp
return L
def fgrad(w):
if self.num_layers == 0:
Wtemp = self.output.W
self.output.W = w.reshape(Wtemp.shape)
else:
Wtemp = self.layer[0].W
self.layer[0].W = w.reshape(Wtemp.shape)
self._forward(self.train_data.X[:ncases, :])
Z = self.train_data.T[:ncases]
if not self.config.is_regression:
Z = self.output.act_type.label_vec_to_mat(Z, self.train_data.K)
self.output.loss(Z)
self.output.gradient()
dLdXabove = self.output.dLdXtop[:, :-1]
for i in range(self.num_layers - 1, -1, -1):
self.layer[i].gradient(dLdXabove)
dLdXabove = self.layer[i].dLdXbelow[:, :-1]
if self.num_layers == 0:
grad_w = self.output.dLdW
else:
grad_w = self.layer[0].dLdW
if self.num_layers == 0:
self.output.W = Wtemp
else:
self.layer[0].W = Wtemp
return grad_w.reshape(np.prod(grad_w.shape)) / Z.shape[0]
if self.num_layers == 0:
#W = np.random.randn(
# self.output.W.shape[0], self.output.W.shape[1]) * 1e-3
W = self.output.W
else:
#W = np.random.randn(
# self.layer[0].W.shape[0], self.layer[0].W.shape[1]) * 1e-3
W = self.layer[0].W
print "wmax: %f" % np.abs(fgrad(W.reshape(np.prod(W.shape)))).max()
print "check_grad err: %f" % opt.check_grad(
f, fgrad, W.reshape(np.prod(W.shape)))