def load_shit():
text_test = './../texts/melville.txt'
char_map_obj = Character_Map(text_test,'mapping.dat',overwrite=True, break_line=None)
unique_char = char_map_obj.unique_char
char_map_obj.k_map()
x, y, shared_x, shared_y = char_map_obj.gen_x_and_y(filename=None)
# print(shared_x, shared_y.get_value().shape[0])
nh = 100
nx = len(char_map_obj.unique_char)
ny = nx
trainer = RNN(nh,nx,ny)
trainer.load_param('param_6-10_17:52/param_epoch199.dat')
f = trainer.compile_gen_sentence()
for xi in x[100:150]:
y_guess = f(xi[0])
y_argmax = [np.argmax(y) for y in y_guess]
char_y = [unique_char[int(yi)] for yi in y_argmax]
print(char_y)
print(''.join(char_y))
return trainer, char_map_obj, x, y, shared_x, shared_y
def main_test():
rng = np.random.RandomState(1234)
filename = './../texts/melville.txt'
foo = Character_Map(filename,'mapping.dat',overwrite=True)
# print(len(foo.mapping))
map_matrix = foo.k_map()
train, valid, test = foo.gen_train_valid_test(filename=None)
# print(train[1].get_value().dtype)
# print(train[1].get_value()[:10].shape)
x = T.tensor3('x')
y = T.imatrix('y')
# x = T.matrix('x')
# rnnlayer = RNNlayer(x,77,77)
# f = theano.function(inputs=[x], outputs=rnnlayer.output)
# foo = f(train[0].get_value()[:10])
# print(foo.shape)
rnn = RNN(x,[77],rng=rng) #the number of unique characters in Moby Dick
ftest = theano.function(inputs=[x], outputs=rnn.p_y_given_x)
# print(ftest(train[0].get_value()[:10]).shape)
print("Compiling training and testing functions...")
t0 = time.time()
ftrain = theano.function(inputs=[x,y],outputs=rnn.neg_log_likelihood(y))
# ftest = theano.function(inputs=[x,y], outputs=rnn.error(y))
# ftest1 = theano.function(inputs=[x,y],outputs=[rnn.y_pred, y])
print("Completed compiling functions. Took {:.2f} seconds".format(time.time() - t0))
for i in xrange(2):
print(ftrain(train[0].get_value()[i*10:(i+1)*10], train[1].get_value()[i*10:(i+1)*10]))
def train_NN(mu, n_epoch, mini_batch): """ Train the neural net """ text_test = './../texts/melville.txt' char_map_obj = Character_Map(text_test,'mapping.dat',overwrite=True, break_line=None) char_map_obj.k_map() x, y, shared_x, shared_y = char_map_obj.gen_x_and_y(filename=None) nh = 100 nx = len(char_map_obj.unique_char) ny = nx trainer = RNNClass(nh,nx,ny) trainer.train_index((shared_x,shared_y),mu,n_epoch,mini_batch)
# compile a theano function that takes the initial x value
# and returns y vectors for each of the subsequent positions.
# """
# x = T.vector('x')
# y = self.gen_random_sentence(x)
# f = theano.function([x],y)
# return f
if __name__ == '__main__':
text_test = './../texts/melville.txt'
char_map_obj = Character_Map(text_test,'mapping.dat',overwrite=True, break_line=None)
char_map_obj.k_map()
x, y, shared_x, shared_y = char_map_obj.gen_x_and_y(filename=None)
# print(shared_x, shared_y.get_value().shape[0])
nh = 100
nx = len(char_map_obj.unique_char)
ny = nx
trainer = RNNClass(nh,nx,ny)
# jobs = []
# for i in xrange(2):
# p = multiprocessing.Process(target=trainer.train, args=((shared_x,shared_y),0.03,1000,10,))
# jobs.append(p)
# p.start()
# trainer.load_param('param_epoch95.dat')
trainer.train_index(training_data=(shared_x,shared_y),
def test_train_RNN(**kwargs):
"""
kwargs
"""
filename = kwargs.get('filename','./../texts/melville.txt')
n_hidden = kwargs.get('n_hidden',77)
n_epochs = kwargs.get('n_epochs',100)
minibatch_size = kwargs.get('minibatch_size',100)
lr = kwargs.get('lr',0.01)
charmap = Character_Map(filename,'mapping.dat',overwrite=True)
charmap.k_map()
train, valid, test = charmap.gen_train_valid_test(filename=None)
train_set_x, train_set_y = train
valid_set_x, valid_set_y = valid
test_set_x, test_set_y = test
n_train_batches = train_set_x.get_value(borrow=True).shape[0] // minibatch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] // minibatch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] // minibatch_size
# print(n_train_batches, n_valid_batches, n_test_batches)
print("Train size: {}, valid size {}, test size {}".format(train_set_x.get_value(borrow=True).shape[0],
valid_set_x.get_value(borrow=True).shape[0],
test_set_x.get_value(borrow=True).shape[0]))
index = T.lscalar()
x = T.tensor3('x')
y = T.imatrix('y')
rng = np.random.RandomState(1234)
rnn = RNN(x,[n_hidden]) #i need to change this to take into account different in and out sizes.
cost = rnn.neg_log_likelihood(y)
print("Compiling training, testing and validating functions...")
t0 = time.time()
test_model = theano.function(
inputs=[index],
outputs=rnn.error(y),
givens={
x: test_set_x[index * minibatch_size:(index + 1) * minibatch_size],
y: test_set_y[index * minibatch_size:(index + 1) * minibatch_size]
}
)
valid_model = theano.function(
inputs=[index],
outputs=rnn.error(y),
givens={
x: valid_set_x[index * minibatch_size:(index + 1) * minibatch_size],
y: valid_set_y[index * minibatch_size:(index + 1) * minibatch_size]
}
)
gparams = [T.grad(cost, param) for param in rnn.params]
updates = [
(param, param-lr*gparam) for param, gparam in zip(rnn.params,gparams)
]
train_model = theano.function(
inputs = [index],
outputs = cost,
updates = updates,
givens = {
x: train_set_x[index * minibatch_size:(index + 1) * minibatch_size],
y: train_set_y[index * minibatch_size:(index + 1) * minibatch_size]
}
)
print("Completed compiling functions. Took {:.2f} seconds".format(time.time() - t0))
print("Starting training...")
valid_freq = 4
best_valid = np.inf
for epoch in xrange(n_epochs):
for minibatch_index in xrange(n_train_batches-1):
mean_cost = train_model(minibatch_index)
iteration_number = epoch*n_train_batches + minibatch_index
if iteration_number % valid_freq == 0:
valid_losses = np.array([valid_model(i) for i in xrange(n_valid_batches)])
# print(valid_losses)
mean_valid = np.mean(valid_losses)
print("Minibatch number: {}\nEpoch number: {}\nValidation Error {}".format(minibatch_index,epoch,mean_valid))
if mean_valid < best_valid:
best_valid = mean_valid
print("Best Validation so far: {}".format(best_valid))
else:
print("Number of iterations: {}, cost {}".format(iteration_number,mean_cost))
print("Done optimizing")
def load_dataset(filename):
foo = Character_Map(filename,'mapping.dat',overwrite=True)
# print(len(foo.mapping))
map_matrix = foo.k_map()
return foo.gen_train_valid_test(filename=None)
