def train(self, xy_data, rnns=None):
# This function trains one RNN for each possible output digit
self.rnns = rnns if rnns is not None else self.rnns
# xs = [np.array(self.encode_expr(self.scramble_double(x))) for x,y in xy_data]
ys = [self.encode_expr(lengthen(y, self.y_len)) for x, y in xy_data]
# for printing purposes only
dev_data = get_data('data/dev.txt')
for i, rnn_i in enumerate(self.rnns):
# where i is the index of the rnn we're using
print 'i', i
xs_i = [
np.array(self.encode_expr(self.scramble(x, i)))
for x, y in xy_data
]
ys_i = [y[i] for y in ys]
dev_xs_i = [
np.array(self.encode_expr(self.scramble(x, i)))
for x, y in dev_data
]
dev_ys_i = [
self.encode_expr(lengthen(y, self.y_len))[i]
for x, y in dev_data
]
rnn_i.grad_check(dev_xs_i[0], dev_ys_i[0])
for j in xrange(self.n_epochs):
for x, y in zip(xs_i, ys):
rnn_i.train_point_sgd(x, y[i], self.alpha)
# print 'train loss', rnn_i.compute_loss(xs_i, ys_i)
if j % 10 == 0:
print 'dev loss', rnn_i.compute_loss(dev_xs_i, dev_ys_i)
# # extra stuff to print
# for x,y in zip(xs_i,ys)[:5]:
# yhat = rnn_i.predict(x)
# print x, yhat, np.argmax(yhat)
return self.rnns
def train(self, xy_data, rnns = None):
# This function trains one RNN for each possible output digit
self.rnns = rnns if rnns is not None else self.rnns
# xs = [np.array(self.encode_expr(self.scramble_double(x))) for x,y in xy_data]
ys = [self.encode_expr(lengthen(y, self.y_len)) for x,y in xy_data]
# for printing purposes only
dev_data = get_data('data/dev.txt')
for i,rnn_i in enumerate(self.rnns):
# where i is the index of the rnn we're using
print 'i',i
xs_i = [np.array(self.encode_expr(self.scramble(x, i))) for x,y in xy_data]
ys_i = [y[i] for y in ys]
dev_xs_i = [np.array(self.encode_expr(self.scramble(x, i))) for x,y in dev_data]
dev_ys_i = [self.encode_expr(lengthen(y, self.y_len))[i] for x,y in dev_data]
rnn_i.grad_check(dev_xs_i[0], dev_ys_i[0])
for j in xrange(self.n_epochs):
for x,y in zip(xs_i, ys):
rnn_i.train_point_sgd(x, y[i], self.alpha)
# print 'train loss', rnn_i.compute_loss(xs_i, ys_i)
if j % 10 == 0:
print 'dev loss', rnn_i.compute_loss(dev_xs_i, dev_ys_i)
# # extra stuff to print
# for x,y in zip(xs_i,ys)[:5]:
# yhat = rnn_i.predict(x)
# print x, yhat, np.argmax(yhat)
return self.rnns
def learn(self, xy_data):
# This function is supposed to learn some stats from the training
# data and store it in our cache
# Structure of Cache:
# (index of x0 and x1, digit x0 of first arg, digit x1 of second arg, index of answer)
# the counter then stores how many times we see each character
# appear for this digit of the answer
self.cache = defaultdict(Counter)
for x,y in xy_data:
# we split our args by the operator
nums = self.num_split(x)
y = lengthen(y, self.y_len)
for i,(x0,x1) in enumerate(zip(nums[0], nums[1])): # for digit of x0,x1
for j,y0 in enumerate(y): # for digit of y
self.cache[(i,x0,x1,j)][y0] += 1
return self.cache
def learn(self, xy_data):
# This function is supposed to learn some stats from the training
# data and store it in our cache
# Structure of Cache:
# (index of x0 and x1, digit x0 of first arg, digit x1 of second arg, index of answer)
# the counter then stores how many times we see each character
# appear for this digit of the answer
self.cache = defaultdict(Counter)
for x, y in xy_data:
# we split our args by the operator
nums = self.num_split(x)
y = lengthen(y, self.y_len)
for i, (x0, x1) in enumerate(zip(nums[0],
nums[1])): # for digit of x0,x1
for j, y0 in enumerate(y): # for digit of y
self.cache[(i, x0, x1, j)][y0] += 1
return self.cache
def lengthen_double(self, x_string):
# format from '21 + 12' -> '021 + 012'
return ' + '.join([lengthen(s, self.x_len) for s in x_string.split(' + ')])
def lengthen_double(self, x_string):
# format from '21 + 12' -> '021 + 012'
return ' + '.join(
[lengthen(s, self.x_len) for s in x_string.split(' + ')])
def num_split(self, x_string):
# split args by operator, e.g. '21 + 12' => ['021','012']
# assuming only addition for now
return [lengthen(s, self.x_len) for s in x_string.split(' + ')]
def num_split(self, x_string):
# split args by operator, e.g. '21 + 12' => ['021','012']
# assuming only addition for now
return [lengthen(s, self.x_len) for s in x_string.split(' + ')]