def run_experiment(constraint_adj=False): X,output,cost,P= build_network(8,3,constraint_adj) parameters = P.values() grads = T.grad(cost,wrt=parameters) train = theano.function( inputs=[X], outputs=cost, updates=updates.adadelta(parameters,grads) ) test = theano.function( inputs=[X], outputs=output, ) data = np.eye(8,dtype=np.int8) # data = np.vstack((data,)) print "Training..." for _ in xrange(100000): np.random.shuffle(data) train(data) hidden_activations = theano.function( inputs=[X], outputs=T.nnet.sigmoid(T.dot(X,P.W_input_hidden)+P.b_hidden) ) #print_arr(test(np.eye(8,dtype=np.int32))) #print_arr(1/(1 + np.exp(-parameters[0].get_value())),1) return hinton(hidden_activations(np.eye(8,dtype=np.int8)))
def make_train(input_size,output_size,mem_size,mem_width,hidden_size=100):
P = Parameters()
# Build controller. ctrl is a network that takes an external and read input
# and returns the output of the network and its hidden layer
ctrl = controller.build(P,input_size,output_size,mem_size,mem_width,hidden_size)
# Build model that predicts output sequence given input sequence
predict = model.build(P,mem_size,mem_width,hidden_size,ctrl)
input_seq = T.matrix('input_sequence')
output_seq = T.matrix('output_sequence')
[M,weights,output_seq_pred] = predict(input_seq)
# Setup for adadelta updates
cross_entropy = T.sum(T.nnet.binary_crossentropy(5e-6 + (1 - 2*5e-6)*output_seq_pred,output_seq),axis=1)
params = P.values()
l2 = T.sum(0)
for p in params:
l2 = l2 + (p ** 2).sum()
cost = T.sum(cross_entropy) + 1e-3*l2
# clip gradients
grads = [ T.clip(g,-100,100) for g in T.grad(cost,wrt=params) ]
train = theano.function(
inputs=[input_seq,output_seq],
outputs=cost,
updates=updates.adadelta(params,grads)
)
return P,train
def make_train(input_size, output_size, mem_size, mem_width, hidden_sizes=[100]):
P = Parameters()
ctrl = controller.build(P, input_size, output_size,
mem_size, mem_width, hidden_sizes)
predict = model.build(P, mem_size, mem_width, hidden_sizes[-1], ctrl)
input_seq = T.matrix('input_sequence')
output_seq = T.matrix('output_sequence')
seqs = predict(input_seq)
output_seq_pred = seqs[-1]
cross_entropy = T.sum(T.nnet.binary_crossentropy(
5e-6 + (1 - 2 * 5e-6) * output_seq_pred, output_seq), axis=1)
params = P.values()
l2 = T.sum(0)
for p in params:
l2 = l2 + (p ** 2).sum()
cost = T.sum(cross_entropy) + 1e-3 * l2
grads = [T.clip(g, -100, 100) for g in T.grad(cost, wrt=params)]
train = theano.function(
inputs=[input_seq, output_seq],
outputs=cost,
updates=updates.adadelta(params, grads)
)
return P, train
def make_train(input_size,
output_size,
mem_size,
mem_width,
hidden_sizes=[100]):
P = Parameters()
ctrl = controller.build(P, input_size, output_size, mem_size, mem_width,
hidden_sizes)
predict = model.build(P, mem_size, mem_width, hidden_sizes[-1], ctrl)
input_seq = T.matrix('input_sequence')
output_seq = T.matrix('output_sequence')
seqs = predict(input_seq)
output_seq_pred = seqs[-1]
cross_entropy = T.sum(T.nnet.binary_crossentropy(
5e-6 + (1 - 2 * 5e-6) * output_seq_pred, output_seq),
axis=1)
cost = T.sum(cross_entropy) # + 1e-3 * l2
params = P.values()
grads = [T.clip(g, -100, 100) for g in T.grad(cost, wrt=params)]
train = theano.function(inputs=[input_seq, output_seq],
outputs=T.sum(cross_entropy),
updates=updates.adadelta(params, grads))
return P, train
def make_functions(inputs,outputs,params,grads,lr): shapes = [ p.get_value().shape for p in params ] acc_grads = [ theano.shared(np.zeros(s,dtype=np.float32)) for s in shapes ] count = theano.shared(np.float32(0)) acc_update = [ (a,a+g) for a,g in zip(acc_grads,grads) ] + [ (count,count + 1.) ] # deltas = acc_grads deltas = [ ag / count for ag in acc_grads ] grads_norms = [ T.sqrt(T.sum(g**2)) for g in deltas ] deltas = [ T.switch(T.gt(n,1.),1.*g/n,g) for n,g in zip(grads_norms,deltas) ] # param_update = [ (p, p - lr * g) for p,g in zip(params,deltas) ] param_update = updates.adadelta(params,deltas,learning_rate=lr) # ,learning_rate=lr,rho=np.float32(0.95) clear_update = [ (a,np.zeros(s,dtype=np.float32)) for a,s in zip(acc_grads,shapes) ] + [ (count,0) ] acc = theano.function( inputs = inputs, outputs = [outputs,output_ans[ans_lbl]], updates = acc_update, on_unused_input='warn', # mode=theano.compile.MonitorMode(post_func=detect_nan) ) update = theano.function( inputs=[lr], updates = param_update + clear_update, outputs = [ T.sqrt(T.sum(T.sqr(w))) for w in deltas ], on_unused_input='warn', # mode=theano.compile.MonitorMode(post_func=detect_nan) ) return acc,update
def make_train(input_size, output_size, mem_size, mem_width, hidden_size=100):
P = Parameters()
# Build controller. ctrl is a network that takes an external and read input
# and returns the output of the network and its hidden layer
ctrl = controller.build(P, input_size, output_size, mem_size, mem_width,
hidden_size)
# Build model that predicts output sequence given input sequence
predict = model.build(P, mem_size, mem_width, hidden_size, ctrl)
input_seq = T.matrix('input_sequence')
output_seq = T.matrix('output_sequence')
[M, weights, output_seq_pred] = predict(input_seq)
# Setup for adadelta updates
cross_entropy = T.sum(T.nnet.binary_crossentropy(
5e-6 + (1 - 2 * 5e-6) * output_seq_pred, output_seq),
axis=1)
params = P.values()
l2 = T.sum(0)
for p in params:
l2 = l2 + (p**2).sum()
cost = T.sum(cross_entropy) + 1e-3 * l2
# clip gradients
grads = [T.clip(g, -100, 100) for g in T.grad(cost, wrt=params)]
train = theano.function(inputs=[input_seq, output_seq],
outputs=cost,
updates=updates.adadelta(params, grads))
return P, train
def build_model(hidden_size, predict_only=False):
X = T.matrix('X')
Y = T.ivector('Y')
#* (0.001 * U.initial_weights(2,hidden_size) + np.array([[0,0,1,1],[1,1,0,0]])))
W_input_hidden = U.create_shared(U.initial_weights(2, hidden_size))
b_hidden = U.create_shared(U.initial_weights(hidden_size))
W_hidden_predict = U.create_shared(U.initial_weights(hidden_size, 2))
b_predict = U.create_shared(U.initial_weights(2))
params = [W_input_hidden, b_hidden, W_hidden_predict, b_predict]
hidden_lin = T.dot(X, W_input_hidden) + b_hidden
hidden = T.nnet.sigmoid(hidden_lin)
predict = T.nnet.softmax(T.dot(hidden, W_hidden_predict) + b_predict)
cost = -T.mean(T.log(
predict[T.arange(Y.shape[0]), Y])) + 1e-3 * adjacency_constraint(
hidden_lin) # + 1e-4 * sum(T.sum(p**2) for p in params)
accuracy = T.mean(T.eq(T.argmax(predict, axis=1), Y))
grad = T.grad(cost, params)
train = theano.function(
inputs=[X, Y],
#updates = updates.momentum(params,grad,0.9999,0.1) if not predict_only else None,
#updates = updates.momentum(params,grad,0.999,0.0005),
updates=updates.adadelta(params, grad),
outputs=[accuracy, W_input_hidden, b_hidden, (hidden > 0.5)])
predict = theano.function(inputs=[X], outputs=predict[:, 0])
i = T.iscalar('i')
hidden_p = theano.function(inputs=[X, i], outputs=hidden[:, i])
return train, predict, hidden_p, params
def turing_updates(cost , lr) :
params = P.values()
#whether add P weight decay
l2 = T.sum(0)
for p in params:
l2 = l2 + (p ** 2).sum()
all_cost = cost + 1e-3 * l2
clipper = updates.clip(5.)
g = T.grad(all_cost, wrt=params)
grads = clipper(g)
# grads = [T.clip(g, -5, 5) for g in T.grad(all_cost, wrt=params)]
# return updates.rmsprop(params, grads, learning_rate=lr)
return updates.adadelta(params, grads, learning_rate=lr)
def make_train_functions():
P = Parameters()
X = T.bvector('X')
Y = T.ivector('Y')
aux = {}
predict = model.build(
P,
input_size=128,
embedding_size=64,
controller_size=256,
stack_size=256,
output_size=128,
)
output = predict(X,aux=aux)
error = - T.log(output[T.arange(Y.shape[0]),((128+1 + Y)%(128+1))])
error = error[-(Y.shape[0]/2):]
parameters = P.values()
gradients = T.grad(T.sum(error),wrt=parameters)
shapes = [ p.get_value().shape for p in parameters ]
count = theano.shared(np.float32(0))
acc_grads = [
theano.shared(np.zeros(s,dtype=np.float32))
for s in shapes
]
acc_update = [ (a,a+g) for a,g in zip(acc_grads,gradients) ] +\
[ (count,count + np.float32(1)) ]
acc_clear = [ (a,np.float32(0) * a) for a in acc_grads ] +\
[ (count,np.int32(0)) ]
avg_grads = [ (g / count) for g in acc_grads ]
avg_grads = [ clip(g,1) for g in acc_grads ]
acc = theano.function(
inputs=[X,Y],
outputs=T.mean(error),
updates = acc_update,
)
update = theano.function(
inputs=[],
updates=updates.adadelta(parameters,avg_grads,learning_rate=1e-8) + acc_clear
)
test = theano.function(
inputs=[X],
outputs=T.argmax(output,axis=1)[-(X.shape[0]/2):],
)
return acc,update,test
def make_train_functions():
P = Parameters()
X = T.bvector('X')
Y = T.ivector('Y')
aux = {}
predict = model.build(
P,
input_size=128,
embedding_size=64,
controller_size=256,
stack_size=256,
output_size=128,
)
output = predict(X, aux=aux)
error = -T.log(output[T.arange(Y.shape[0]), ((128 + 1 + Y) % (128 + 1))])
error = error[-(Y.shape[0] / 2):]
parameters = P.values()
gradients = T.grad(T.sum(error), wrt=parameters)
shapes = [p.get_value().shape for p in parameters]
count = theano.shared(np.float32(0))
acc_grads = [theano.shared(np.zeros(s, dtype=np.float32)) for s in shapes]
acc_update = [ (a,a+g) for a,g in zip(acc_grads,gradients) ] +\
[ (count,count + np.float32(1)) ]
acc_clear = [ (a,np.float32(0) * a) for a in acc_grads ] +\
[ (count,np.int32(0)) ]
avg_grads = [(g / count) for g in acc_grads]
avg_grads = [clip(g, 1) for g in acc_grads]
acc = theano.function(
inputs=[X, Y],
outputs=T.mean(error),
updates=acc_update,
)
update = theano.function(
inputs=[],
updates=updates.adadelta(parameters, avg_grads, learning_rate=1e-8) +
acc_clear)
test = theano.function(
inputs=[X],
outputs=T.argmax(output, axis=1)[-(X.shape[0] / 2):],
)
return acc, update, test
def train_model(docs,
wordvec_size,
hidden_size,
error_threshold,
update_mu=1e-3,
update_eps=0.95):
X, parameters, hidden, hidden1_reproduction, input_reproduction, unrolled = build_network(
wordvec_size, hidden_size)
#hidden, hidden_rep, input_rep, unrlld = f(docs)
error = build_error(X, hidden, hidden1_reproduction, input_reproduction)
cost = error # + 1e-6*sum( T.sum(abs(p)) for p in parameters )
gradients = T.grad(cost, wrt=parameters)
eps = T.dscalar('eps')
mu = T.dscalar('mu')
train = theano.function(inputs=[X, eps, mu],
updates=updates.adadelta(parameters, gradients, mu,
eps),
outputs=error)
error = 10
count = 0
for i in range(10):
start_time = time.time()
error = 0
for doc in docs:
error += train(doc, update_mu, update_eps)
if count % 1 == 0:
print "iter=%d" % count, time.time() - start_time, error / len(
docs)
count += 1
f = theano.function(
inputs=[X],
outputs=[hidden, hidden1_reproduction, input_reproduction, unrolled])
print "Finish count=%d error=%f" % (count, error)
return f
def make_functions(inputs, outputs, params, grads, lr):
shapes = [p.get_value().shape for p in params]
acc_grads = [theano.shared(np.zeros(s, dtype=np.float32)) for s in shapes]
count = theano.shared(np.float32(0))
acc_update = [(a, a + g)
for a, g in zip(acc_grads, grads)] + [(count, count + 1.)]
# deltas = acc_grads
deltas = [ag / count for ag in acc_grads]
grads_norms = [T.sqrt(T.sum(g**2)) for g in deltas]
deltas = [
T.switch(T.gt(n, 1.), 1. * g / n, g)
for n, g in zip(grads_norms, deltas)
]
# param_update = [ (p, p - lr * g) for p,g in zip(params,deltas) ]
param_update = updates.adadelta(
params, deltas,
learning_rate=lr) # ,learning_rate=lr,rho=np.float32(0.95)
clear_update = [(a, np.zeros(s, dtype=np.float32))
for a, s in zip(acc_grads, shapes)] + [(count, 0)]
acc = theano.function(
inputs=inputs,
outputs=[outputs, output_ans[ans_lbl]],
updates=acc_update,
on_unused_input='warn',
# mode=theano.compile.MonitorMode(post_func=detect_nan)
)
update = theano.function(
inputs=[lr],
updates=param_update + clear_update,
outputs=[T.sqrt(T.sum(T.sqr(w))) for w in deltas],
on_unused_input='warn',
# mode=theano.compile.MonitorMode(post_func=detect_nan)
)
return acc, update
def build_model(hidden_size,predict_only=False):
X = T.matrix('X')
Y = T.ivector('Y')
#* (0.001 * U.initial_weights(2,hidden_size) + np.array([[0,0,1,1],[1,1,0,0]])))
W_input_hidden = U.create_shared(U.initial_weights(2,hidden_size))
b_hidden = U.create_shared(U.initial_weights(hidden_size))
W_hidden_predict = U.create_shared(U.initial_weights(hidden_size,2))
b_predict = U.create_shared(U.initial_weights(2))
params = [W_input_hidden,b_hidden,W_hidden_predict,b_predict]
hidden_lin = T.dot(X,W_input_hidden) + b_hidden
hidden = T.nnet.sigmoid(hidden_lin)
predict = T.nnet.softmax(T.dot(hidden,W_hidden_predict) + b_predict)
cost = -T.mean(T.log(predict[T.arange(Y.shape[0]),Y])) + 1e-3*adjacency_constraint(hidden_lin)# + 1e-4 * sum(T.sum(p**2) for p in params)
accuracy = T.mean(T.eq(T.argmax(predict,axis=1),Y))
grad = T.grad(cost,params)
train = theano.function(
inputs = [X,Y],
#updates = updates.momentum(params,grad,0.9999,0.1) if not predict_only else None,
#updates = updates.momentum(params,grad,0.999,0.0005),
updates = updates.adadelta(params,grad),
outputs = [accuracy,W_input_hidden,b_hidden,(hidden>0.5)]
)
predict = theano.function(
inputs = [X],
outputs = predict[:,0]
)
i = T.iscalar('i')
hidden_p = theano.function(
inputs = [X,i],
outputs = hidden[:,i]
)
return train,predict,hidden_p,params
def run_experiment(constraint_adj=False):
X, output, cost, P = build_network(8, 3, constraint_adj)
parameters = P.values()
grads = T.grad(cost, wrt=parameters)
train = theano.function(inputs=[X],
outputs=cost,
updates=updates.adadelta(parameters, grads))
test = theano.function(
inputs=[X],
outputs=output,
)
data = np.eye(8, dtype=np.int8)
# data = np.vstack((data,))
print "Training..."
for _ in xrange(100000):
np.random.shuffle(data)
train(data)
hidden_activations = theano.function(
inputs=[X],
outputs=T.nnet.sigmoid(T.dot(X, P.W_input_hidden) + P.b_hidden))
#print_arr(test(np.eye(8,dtype=np.int32)))
#print_arr(1/(1 + np.exp(-parameters[0].get_value())),1)
return hinton(hidden_activations(np.eye(8, dtype=np.int8)))
X,parameters,hidden,hidden1_reproduction,input_reproduction,unrolled = build_network(8,64)
f = theano.function(
inputs = [X],
outputs = [hidden,hidden1_reproduction,input_reproduction,unrolled]
)
error = build_error(X,hidden,hidden1_reproduction,input_reproduction)
cost = error # + 1e-6*sum( T.sum(abs(p)) for p in parameters )
gradients = T.grad(cost,wrt=parameters)
eps = T.dscalar('eps')
mu = T.dscalar('mu')
train = theano.function(
inputs = [X,eps,mu],
updates = updates.adadelta(parameters,gradients,mu,eps),
outputs = error
)
#example = np.vstack((np.eye(8),np.eye(8)))
example = np.eye(8)
error = 10
lr = 0.0001
t = 0
while error > 0.0001:
np.random.shuffle(example)
#error = train(example,lr,min(1 - 3.0/(t+5),0.999))
error = train(example,1e-6,0.95)
#error = train(example,lr,0)
print error
t += 1
predict = T.nnet.softmax(T.dot(hidden, W_hidden_output) + b_output)
return X, predict
def label_seq(string):
idxs = font.indexify(string)
result = np.ones((len(idxs) * 2 + 1, ), dtype=np.int32) * -1
result[np.arange(len(idxs)) * 2 + 1] = idxs
print result
return result
if __name__ == "__main__":
P = Parameters()
X = T.matrix('X')
Y = T.ivector('Y')
X, predict = build_model(P, X, 10, 10, 10)
cost = ctc.cost(predict, Y)
params = P.values()
grad = T.grad(cost, wrt=params)
train = theano.function(inputs=[X, Y],
outputs=cost,
updates=updates.adadelta(params, grad))
for _ in xrange(10):
print train(
np.eye(10, dtype=np.float32)[::-1], np.arange(10, dtype=np.int32))
predict = T.nnet.softmax(T.dot(hidden, W_hidden_output) + b_output)
return X, predict
def label_seq(string):
idxs = font.indexify(string)
result = np.ones((len(idxs) * 2 + 1,), dtype=np.int32) * -1
result[np.arange(len(idxs)) * 2 + 1] = idxs
print result
return result
if __name__ == "__main__":
P = Parameters()
X = T.matrix('X')
Y = T.ivector('Y')
X, predict = build_model(P, X, 10, 10, 10)
cost = ctc.cost(predict, Y)
params = P.values()
grad = T.grad(cost, wrt=params)
train = theano.function(
inputs=[X, Y],
outputs=cost,
updates=updates.adadelta(params, grad)
)
for _ in xrange(10):
print train(np.eye(10, dtype=np.float32)[::-1], np.arange(10, dtype=np.int32))
