def run_experiment(self, dataset, word_embedding, exp_name):
# load parameters
num_maps_word = self.options["num_maps_word"]
drop_rate_word = self.options["drop_rate_word"]
drop_rate_sentence = self.options["drop_rate_sentence"]
word_window = self.options["word_window"]
word_dim = self.options["word_dim"]
k_max_word = self.options["k_max_word"]
batch_size = self.options["batch_size"]
rho = self.options["rho"]
epsilon = self.options["epsilon"]
norm_lim = self.options["norm_lim"]
max_iteration = self.options["max_iteration"]
k_portion = self.options["k_portion"]
sentence_len = len(dataset[0][0][0][0])
# compute the sentence flags
train_flags, test_flags = construct_sentence_flag(dataset)
train_k_value = construct_dynamic_k(train_flags, k_portion)
test_k_value = construct_dynamic_k(test_flags, k_portion)
train_flags = theano.shared(value=np.asarray(train_flags, dtype=theano.config.floatX), borrow=True)
test_flags = theano.shared(value=np.asarray(test_flags, dtype=theano.config.floatX), borrow=True)
train_k = theano.shared(value=np.asarray(train_k_value, dtype=theano.config.floatX), borrow=True)
test_k = theano.shared(value=np.asarray(test_k_value, dtype=theano.config.floatX), borrow=True)
# define the parameters
x = T.tensor3("x")
y = T.ivector("y")
sen_flags = T.matrix("flag")
sen_k = T.matrix("sen_k")
rng = np.random.RandomState(1234)
words = theano.shared(value=np.asarray(word_embedding,
dtype=theano.config.floatX),
name="embedding", borrow=True)
zero_vector_tensor = T.vector()
zero_vec = np.zeros(word_dim, dtype=theano.config.floatX)
set_zero = theano.function([zero_vector_tensor], updates=[(words, T.set_subtensor(words[0,:], zero_vector_tensor))])
x_emb = words[T.cast(x.flatten(), dtype="int32")].reshape((x.shape[0]*x.shape[1], 1, x.shape[2], words.shape[1]))
dropout_x_emb = nn.dropout_from_layer(rng, x_emb, drop_rate_word)
# compute convolution on words layer
word_filter_shape = (num_maps_word, 1, word_window, word_dim)
word_pool_size = (sentence_len - word_window + 1, 1)
dropout_word_conv = nn.ConvPoolLayer(rng,
input=dropout_x_emb,
input_shape=None,
filter_shape=word_filter_shape,
pool_size=word_pool_size,
activation=Tanh,
k=k_max_word)
sent_vec_dim = num_maps_word*k_max_word
dropout_sent_vec = dropout_word_conv.output.reshape((x.shape[0] * x.shape[1], sent_vec_dim))
word_conv = nn.ConvPoolLayer(rng,
input=dropout_x_emb*(1 - drop_rate_word),
input_shape=None,
filter_shape=word_filter_shape,
pool_size=word_pool_size,
activation=Tanh,
k=k_max_word,
W=dropout_word_conv.W,
b=dropout_word_conv.b)
sent_vec = word_conv.output.reshape((x.shape[0] * x.shape[1], sent_vec_dim))
# construct sentence level classifier
n_in = sent_vec_dim
n_out = 1
sen_W_values = np.zeros((n_in, n_out), dtype=theano.config.floatX)
sen_W = theano.shared(value=sen_W_values, borrow=True, name="logis_W")
sen_b_value = nn.as_floatX(0.0)
sen_b = theano.shared(value=sen_b_value, borrow=True, name="logis_b")
drop_sent_prob = T.nnet.sigmoid(T.dot(dropout_sent_vec, sen_W) + sen_b)
sent_prob = T.nnet.sigmoid(T.dot(sent_vec, sen_W*(1-drop_rate_sentence)) + sen_b)
# reform the sent vec to doc level
drop_sent_prob = drop_sent_prob.reshape((x.shape[0], x.shape[1]))
sent_prob = sent_prob.reshape((x.shape[0], x.shape[1]))
"""
# the pos probability bag label is the avg of the probs
drop_doc_prob = T.sum(drop_sent_prob * sen_flags, axis=1) / T.sum(sen_flags, axis=1)
doc_prob = T.sum(sent_prob * sen_flags, axis=1) / T.sum(sen_flags, axis=1)
"""
# using the dynamic top k max probability as bag level probability
# compute the dynamic K for each documents
drop_doc_prob = T.sum(T.sort(drop_sent_prob, axis=1) * sen_k, axis=1) / T.sum(sen_k, axis=1)
doc_prob = T.sum(T.sort(sent_prob, axis=1) * sen_k, axis=1) / T.sum(sen_k, axis=1)
drop_doc_prob = T.clip(drop_doc_prob, nn.as_floatX(1e-7), nn.as_floatX(1 - 1e-7 ))
doc_prob = T.clip(doc_prob, nn.as_floatX(1e-7), nn.as_floatX(1 - 1e-7 ))
doc_preds = doc_prob > 0.5
# instance level cost
drop_sent_cost = T.sum(T.maximum(0.0, nn.as_floatX(.5) - T.sgn(drop_sent_prob.reshape((x.shape[0]*x.shape[1], n_out)) - nn.as_floatX(0.6)) * T.dot(dropout_sent_vec, sen_W)) * sen_flags.reshape((x.shape[0]*x.shape[1], n_out))) / T.sum(sen_flags)
# we need that the most positive instance at least 0.7 in pos bags
# and at most 0.1 in neg bags
# we want the number of positive instance should at least ...
# and non of the positive instances in the negative bags
# compute the number of positive instance
positive_count = T.sum((drop_sent_prob * sen_flags) > 0.5, axis=1)
pos_cost = T.maximum(nn.as_floatX(0.0), nn.as_floatX(2) - positive_count)
neg_cost = T.maximum(nn.as_floatX(0.0), positive_count)
penal_cost = T.mean(pos_cost * y + neg_cost * (nn.as_floatX(1.0) - y))
# add the sentence similarity constrains
sen_sen = T.dot(dropout_sent_vec, dropout_sent_vec.T)
sen_sqr = T.sum(dropout_sent_vec ** 2, axis=1)
sen_sqr_left = sen_sqr.dimshuffle(0, 'x')
sen_sqr_right = sen_sqr.dimshuffle('x', 0)
sen_sim_matrix = sen_sqr_left - 2 * sen_sqr + sen_sqr_right
sen_sim_matrix = T.exp(-1 * sen_sim_matrix)
sen_sim_prob = drop_sent_prob.reshape((x.shape[0]*x.shape[1], 1)) - drop_sent_prob.flatten()
sen_sim_prob = sen_sim_prob ** 2
sen_sim_flag = T.dot(sen_flags.reshape((x.shape[0]*x.shape[1],1)), sen_flags.reshape((1,x.shape[0]*x.shape[1])))
sen_sim_cost = T.sum(sen_sim_matrix * sen_sim_prob * sen_sim_flag) / T.sum(sen_sim_flag)
# bag level cost
drop_bag_cost = T.mean(-y * T.log(drop_doc_prob) * nn.as_floatX(0.6) - (1 - y) * T.log(1 - drop_doc_prob) * nn.as_floatX(0.4))
#drop_cost = drop_bag_cost * nn.as_floatX(3.0) + drop_sent_cost + nn.as_floatX(2.0) * penal_cost
drop_cost = drop_bag_cost * nn.as_floatX(0.6) + drop_sent_cost * nn.as_floatX(0.1) + penal_cost * nn.as_floatX(0.5) + sen_sim_cost * nn.as_floatX(0.0001)
# collect parameters
self.params.append(words)
self.params += dropout_word_conv.params
self.params.append(sen_W)
self.params.append(sen_b)
grad_updates = nn.sgd_updates_adadelta(self.params,
drop_cost,
rho,
epsilon,
norm_lim)
# construct the dataset
# random the
train_x, train_y = nn.shared_dataset(dataset[0])
test_x, test_y = nn.shared_dataset(dataset[1])
test_cpu_y = dataset[1][1]
n_train_batches = int(np.ceil(1.0 * len(dataset[0][0]) / batch_size))
n_test_batches = int(np.ceil(1.0 * len(dataset[1][0]) / batch_size))
# construt the model
index = T.iscalar()
train_func = theano.function([index], [drop_cost, drop_bag_cost, drop_sent_cost, penal_cost, sen_sim_cost], updates=grad_updates,
givens={
x: train_x[index*batch_size:(index+1)*batch_size],
y: train_y[index*batch_size:(index+1)*batch_size],
sen_flags: train_flags[index*batch_size:(index+1)*batch_size],
sen_k: train_k[index*batch_size:(index+1)*batch_size]
})
test_func = theano.function([index], doc_preds,
givens={
x:test_x[index*batch_size:(index+1)*batch_size],
sen_k:test_k[index*batch_size:(index+1)*batch_size]
})
get_train_sent_prob = theano.function([index], sent_prob,
givens={
x:train_x[index*batch_size:(index+1)*batch_size]
})
get_test_sent_prob = theano.function([index], sent_prob,
givens={
x:test_x[index*batch_size:(index+1)*batch_size]
})
epoch = 0
best_score = 0
log_file = open("./log/%s.log" % exp_name, 'w')
while epoch <= max_iteration:
start_time = timeit.default_timer()
epoch += 1
costs = []
for minibatch_index in np.random.permutation(range(n_train_batches)):
cost_epoch = train_func(minibatch_index)
costs.append(cost_epoch)
set_zero(zero_vec)
total_train_cost, train_bag_cost, train_sent_cost, train_penal_cost, train_sim_cost = zip(*costs)
print "Iteration %d, total_cost %f bag_cost %f sent_cost %f penal_cost %f sim cost %f\n" % (epoch, np.mean(total_train_cost), np.mean(train_bag_cost), np.mean(train_sent_cost), np.mean(train_penal_cost), np.mean(train_sim_cost))
if epoch % 1 == 0:
test_preds = []
for i in xrange(n_test_batches):
test_y_pred = test_func(i)
test_preds.append(test_y_pred)
test_preds = np.concatenate(test_preds)
test_score = 1 - np.mean(np.not_equal(test_cpu_y, test_preds))
precision, recall, beta, support = precision_recall_fscore_support(test_cpu_y, test_preds, pos_label=1)
if beta[1] > best_score or epoch % 5 == 0:
best_score = beta[1]
# save the sentence vectors
train_sens = [get_train_sent_prob(i) for i in range(n_train_batches)]
test_sens = [get_test_sent_prob(i) for i in range(n_test_batches)]
train_sens = np.concatenate(train_sens, axis=0)
test_sens = np.concatenate(test_sens, axis=0)
out_train_sent_file = "./results/%s_train_sent_%d.vec" % (exp_name, epoch)
out_test_sent_file = "./results/%s_test_sent_%d.vec" % (exp_name, epoch)
with open(out_test_sent_file, 'w') as test_f, open(out_train_sent_file, 'w') as train_f:
cPickle.dump(train_sens, train_f)
cPickle.dump(test_sens, test_f)
print "Get best performace at %d iteration %f" % (epoch, test_score)
log_file.write("Get best performance at %d iteration %f \n" % (epoch, test_score))
end_time = timeit.default_timer()
print "Iteration %d , precision, recall, f1" % epoch, precision, recall, beta
log_file.write("Iteration %d, neg precision %f, pos precision %f, neg recall %f pos recall %f , neg f1 %f, pos f1 %f, total_cost %f bag_cost %f sent_cost %f penal_cost %f\n" % (epoch, precision[0], precision[1], recall[0], recall[1], beta[0], beta[1], np.mean(total_train_cost), np.mean(train_bag_cost), np.mean(train_sent_cost), np.mean(train_penal_cost)))
print "Using time %f m" % ((end_time -start_time)/60.)
log_file.write("Uing time %f m\n" % ((end_time - start_time)/60.))
end_time = timeit.default_timer()
print "Iteration %d Using time %f m" % ( epoch, (end_time -start_time)/60.)
log_file.write("Uing time %f m\n" % ((end_time - start_time)/60.))
log_file.flush()
log_file.close()
def run_experiment(self, dataset, word_embedding, exp_name):
# load parameters
num_maps_word = self.options["num_maps_word"]
drop_rate_word = self.options["drop_rate_word"]
drop_rate_sentence = self.options["drop_rate_sentence"]
word_window = self.options["word_window"]
word_dim = self.options["word_dim"]
k_max_word = self.options["k_max_word"]
batch_size = self.options["batch_size"]
rho = self.options["rho"]
epsilon = self.options["epsilon"]
norm_lim = self.options["norm_lim"]
max_iteration = self.options["max_iteration"]
sentence_len = len(dataset[0][0][0][0])
# compute the sentence flags
train_flags, test_flags = construct_sentence_flag(dataset)
train_flags = theano.shared(value=np.asarray(train_flags, dtype=theano.config.floatX), borrow=True)
test_flags = theano.shared(value=np.asarray(test_flags, dtype=theano.config.floatX), borrow=True)
# define the parameters
x = T.tensor3("x")
y = T.ivector("y")
sen_flags = T.matrix("flag")
rng = np.random.RandomState(1234)
words = theano.shared(value=np.asarray(word_embedding,
dtype=theano.config.floatX),
name="embedding", borrow=True)
zero_vector_tensor = T.vector()
zero_vec = np.zeros(word_dim, dtype=theano.config.floatX)
set_zero = theano.function([zero_vector_tensor], updates=[(words, T.set_subtensor(words[0,:], zero_vector_tensor))])
x_emb = words[T.cast(x.flatten(), dtype="int32")].reshape((x.shape[0]*x.shape[1], 1, x.shape[2], words.shape[1]))
dropout_x_emb = nn.dropout_from_layer(rng, x_emb, drop_rate_word)
# compute convolution on words layer
word_filter_shape = (num_maps_word, 1, word_window, word_dim)
word_pool_size = (sentence_len - word_window + 1, 1)
dropout_word_conv = nn.ConvPoolLayer(rng,
input=dropout_x_emb,
input_shape=None,
filter_shape=word_filter_shape,
pool_size=word_pool_size,
activation=Tanh,
k=k_max_word)
sent_vec_dim = num_maps_word*k_max_word
dropout_sent_vec = dropout_word_conv.output.reshape((x.shape[0] * x.shape[1], sent_vec_dim))
word_conv = nn.ConvPoolLayer(rng,
input=dropout_x_emb*(1 - drop_rate_word),
input_shape=None,
filter_shape=word_filter_shape,
pool_size=word_pool_size,
activation=Tanh,
k=k_max_word,
W=dropout_word_conv.W,
b=dropout_word_conv.b)
sent_vec = word_conv.output.reshape((x.shape[0] * x.shape[1], sent_vec_dim))
# construct sentence level classifier
n_in = sent_vec_dim
n_out = 1
sen_W_values = np.zeros((n_in, n_out), dtype=theano.config.floatX)
sen_W = theano.shared(value=sen_W_values, borrow=True, name="logis_W")
sen_b_value = nn.as_floatX(0.0)
sen_b = theano.shared(value=sen_b_value, borrow=True, name="logis_b")
drop_sent_prob = T.nnet.sigmoid(T.dot(dropout_sent_vec, sen_W) + sen_b)
sent_prob = T.nnet.sigmoid(T.dot(sent_vec, sen_W*(1-drop_rate_sentence)) + sen_b)
# reform the sent vec to doc level
drop_sent_prob = drop_sent_prob.reshape((x.shape[0], x.shape[1]))
sent_prob = sent_prob.reshape((x.shape[0], x.shape[1]))
# the pos probability bag label is the avg of the probs
drop_doc_prob = T.sum(drop_sent_prob * sen_flags, axis=1) / T.sum(sen_flags, axis=1)
doc_prob = T.sum(sent_prob * sen_flags, axis=1) / T.sum(sen_flags, axis=1)
drop_doc_prob = T.clip(drop_doc_prob, nn.as_floatX(1e-7), nn.as_floatX(1 - 1e-7 ))
doc_prob = T.clip(doc_prob, nn.as_floatX(1e-7), nn.as_floatX(1 - 1e-7 ))
"""
# the pos probability bag label equals to 1 - all negative
drop_doc_prob = T.prod(drop_sent_prob, axis=1)
drop_doc_prob = T.set_subtensor(drop_doc_prob[:,1], 1 - drop_doc_prob[:,0])
doc_prob = T.prod(sent_prob, axis=1)
doc_prob = T.set_subtensor(doc_prob[:,1], 1 - doc_prob[:,0])
# the pos probability bag label is the most positive probability
drop_doc_prob = T.max(drop_sent_prob, axis=1)
drop_doc_prob = T.clip(drop_doc_prob, nn.as_floatX(1e-7), nn.as_floatX(1 - 1e-7 ))
doc_prob = T.max(sent_prob, axis=1)
doc_prob = T.clip(doc_prob, nn.as_floatX(1e-7), nn.as_floatX(1 - 1e-7 ))
"""
doc_preds = doc_prob > 0.5
# instance level cost
drop_sent_cost = T.sum(T.maximum(0.0, nn.as_floatX(.5) - T.sgn(drop_sent_prob.reshape((x.shape[0]*x.shape[1], n_out)) - nn.as_floatX(0.6)) * T.dot(dropout_sent_vec, sen_W)) * sen_flags.reshape((x.shape[0]*x.shape[1], n_out))) / T.sum(sen_flags)
# we need that the most positive instance at least 0.7 in pos bags
# and at most 0.1 in neg bags
# we want the number of positive instance should at least ...
# and non of the positive instances in the negative bags
# compute the number of positive instance
positive_count = T.sum((drop_sent_prob * sen_flags) > 0.5, axis=1)
pos_cost = T.maximum(nn.as_floatX(0.0), nn.as_floatX(2) - positive_count)
neg_cost = T.maximum(nn.as_floatX(0.0), positive_count)
"""
most_positive_prob = T.max(drop_sent_prob, axis=1)
pos_cost = T.maximum(0.0, nn.as_floatX(0.6) - most_positive_prob)
neg_cost = T.maximum(0.0, most_positive_prob - nn.as_floatX(0.05))
"""
penal_cost = T.mean(pos_cost * y + neg_cost * (nn.as_floatX(1.0) - y))
# add the sentence similarity constrains
sen_sen = T.dot(dropout_sent_vec, dropout_sent_vec.T)
sen_sqr = T.sum(dropout_sent_vec ** 2, axis=1)
sen_sqr_left = sen_sqr.dimshuffle(0, 'x')
sen_sqr_right = sen_sqr.dimshuffle('x', 0)
sen_sim_matrix = sen_sqr_left - 2 * sen_sqr + sen_sqr_right
sen_sim_matrix = T.exp(-1 * sen_sim_matrix)
sen_sim_prob = drop_sent_prob.reshape((x.shape[0]*x.shape[1], 1)) - drop_sent_prob.flatten()
sen_sim_prob = sen_sim_prob ** 2
sen_sim_flag = T.dot(sen_flags.reshape((x.shape[0]*x.shape[1],1)), sen_flags.reshape((1,x.shape[0]*x.shape[1])))
sen_sim_cost = T.sum(sen_sim_matrix * sen_sim_prob * sen_sim_flag) / T.sum(sen_sim_flag)
# bag level cost
drop_bag_cost = T.mean(-y * T.log(drop_doc_prob) * nn.as_floatX(0.6) - (1 - y) * T.log(1 - drop_doc_prob) * nn.as_floatX(0.4))
#drop_cost = drop_bag_cost * nn.as_floatX(3.0) + drop_sent_cost + nn.as_floatX(2.0) * penal_cost
drop_cost = drop_bag_cost * nn.as_floatX(0.6) + drop_sent_cost * nn.as_floatX(0.1) + penal_cost * nn.as_floatX(0.5) + sen_sim_cost * nn.as_floatX(0.0001)
# collect parameters
self.params.append(words)
self.params += dropout_word_conv.params
self.params.append(sen_W)
self.params.append(sen_b)
grad_updates = nn.sgd_updates_adadelta(self.params,
drop_cost,
rho,
epsilon,
norm_lim)
# construct the dataset
# random the
train_x, train_y = nn.shared_dataset(dataset[0])
test_x, test_y = nn.shared_dataset(dataset[1])
test_cpu_y = dataset[1][1]
n_train_batches = int(np.ceil(1.0 * len(dataset[0][0]) / batch_size))
n_test_batches = int(np.ceil(1.0 * len(dataset[1][0]) / batch_size))
# construt the model
index = T.iscalar()
train_func = theano.function([index], [drop_cost, drop_bag_cost, drop_sent_cost, penal_cost, sen_sim_cost], updates=grad_updates,
givens={
x: train_x[index*batch_size:(index+1)*batch_size],
y: train_y[index*batch_size:(index+1)*batch_size],
sen_flags: train_flags[index*batch_size:(index+1)*batch_size]
})
test_func = theano.function([index], doc_preds,
givens={
x:test_x[index*batch_size:(index+1)*batch_size],
sen_flags: test_flags[index*batch_size:(index+1)*batch_size]
})
get_train_sent_prob = theano.function([index], sent_prob,
givens={
x:train_x[index*batch_size:(index+1)*batch_size]
})
get_test_sent_prob = theano.function([index], sent_prob,
givens={
x:test_x[index*batch_size:(index+1)*batch_size]
})
epoch = 0
best_score = 0
log_file = open("./log/%s.log" % exp_name, 'w')
while epoch <= max_iteration:
start_time = timeit.default_timer()
epoch += 1
costs = []
for minibatch_index in np.random.permutation(range(n_train_batches)):
cost_epoch = train_func(minibatch_index)
costs.append(cost_epoch)
set_zero(zero_vec)
total_train_cost, train_bag_cost, train_sent_cost, train_penal_cost, train_sim_cost = zip(*costs)
print "Iteration %d, total_cost %f bag_cost %f sent_cost %f penal_cost %f sim cost %f\n" % (epoch, np.mean(total_train_cost), np.mean(train_bag_cost), np.mean(train_sent_cost), np.mean(train_penal_cost), np.mean(train_sim_cost))
if epoch % 1 == 0:
test_preds = []
for i in xrange(n_test_batches):
test_y_pred = test_func(i)
test_preds.append(test_y_pred)
test_preds = np.concatenate(test_preds)
test_score = 1 - np.mean(np.not_equal(test_cpu_y, test_preds))
precision, recall, beta, support = precision_recall_fscore_support(test_cpu_y, test_preds, pos_label=1)
if beta[1] > best_score or epoch % 5 == 0:
best_score = beta[1]
# save the sentence vectors
train_sens = [get_train_sent_prob(i) for i in range(n_train_batches)]
test_sens = [get_test_sent_prob(i) for i in range(n_test_batches)]
train_sens = np.concatenate(train_sens, axis=0)
test_sens = np.concatenate(test_sens, axis=0)
out_train_sent_file = "./results/%s_train_sent_%d.vec" % (exp_name, epoch)
out_test_sent_file = "./results/%s_test_sent_%d.vec" % (exp_name, epoch)
with open(out_test_sent_file, 'w') as test_f, open(out_train_sent_file, 'w') as train_f:
cPickle.dump(train_sens, train_f)
cPickle.dump(test_sens, test_f)
print "Get best performace at %d iteration %f" % (epoch, test_score)
log_file.write("Get best performance at %d iteration %f \n" % (epoch, test_score))
end_time = timeit.default_timer()
print "Iteration %d , precision, recall, f1" % epoch, precision, recall, beta
log_file.write("Iteration %d, neg precision %f, pos precision %f, neg recall %f pos recall %f , neg f1 %f, pos f1 %f, total_cost %f bag_cost %f sent_cost %f penal_cost %f\n" % (epoch, precision[0], precision[1], recall[0], recall[1], beta[0], beta[1], np.mean(total_train_cost), np.mean(train_bag_cost), np.mean(train_sent_cost), np.mean(train_penal_cost)))
print "Using time %f m" % ((end_time -start_time)/60.)
log_file.write("Uing time %f m\n" % ((end_time - start_time)/60.))
end_time = timeit.default_timer()
print "Iteration %d Using time %f m" % ( epoch, (end_time -start_time)/60.)
log_file.write("Uing time %f m\n" % ((end_time - start_time)/60.))
log_file.flush()
log_file.close()
def run_experiment(self, dataset, word_embedding, exp_name):
# load parameters
num_maps_word = self.options["num_maps_word"]
drop_rate_word = self.options["drop_rate_word"]
word_window = self.options["word_window"]
word_dim = self.options["word_dim"]
k_max_word = self.options["k_max_word"]
num_maps_sentence = self.options["num_maps_sentence"]
drop_rate_sentence = self.options["drop_rate_sentence"]
sentence_window = self.options["sentence_window"]
k_max_sentence = self.options["k_max_sentence"]
batch_size = self.options["batch_size"]
rho = self.options["rho"]
epsilon = self.options["epsilon"]
norm_lim = self.options["norm_lim"]
max_iteration = self.options["max_iteration"]
sentence_len = len(dataset[0][0][0][0])
sentence_num = len(dataset[0][0][0])
# define the parameters
x = T.tensor3("x")
y = T.ivector("y")
rng = np.random.RandomState(1234)
words = theano.shared(value=np.asarray(word_embedding,
dtype=theano.config.floatX),
name="embedding",
borrow=True)
zero_vector_tensor = T.vector()
zero_vec = np.zeros(word_dim, dtype=theano.config.floatX)
set_zero = theano.function(
[zero_vector_tensor],
updates=[(words, T.set_subtensor(words[0, :],
zero_vector_tensor))])
x_emb = words[T.cast(x.flatten(), dtype="int32")].reshape(
(x.shape[0] * x.shape[1], 1, x.shape[2], words.shape[1]))
dropout_x_emb = nn.dropout_from_layer(rng, x_emb, drop_rate_word)
# compute convolution on words layer
word_filter_shape = (num_maps_word, 1, word_window, word_dim)
word_pool_size = (sentence_len - word_window + 1, 1)
dropout_word_conv = nn.ConvPoolLayer(rng,
input=dropout_x_emb,
input_shape=None,
filter_shape=word_filter_shape,
pool_size=word_pool_size,
activation=Tanh,
k=k_max_word)
sent_vec_dim = num_maps_word * k_max_word
dropout_sent_vec = dropout_word_conv.output.reshape(
(x.shape[0], 1, x.shape[1], sent_vec_dim))
dropout_sent_vec = nn.dropout_from_layer(rng, dropout_sent_vec,
drop_rate_sentence)
word_conv = nn.ConvPoolLayer(rng,
input=dropout_x_emb *
(1 - drop_rate_word),
input_shape=None,
filter_shape=word_filter_shape,
pool_size=word_pool_size,
activation=Tanh,
k=k_max_word,
W=dropout_word_conv.W,
b=dropout_word_conv.b)
sent_vec = word_conv.output.reshape(
(x.shape[0], 1, x.shape[1], sent_vec_dim))
# construct the convolution layer on sentences
sent_filter_shape = (num_maps_sentence, 1, sentence_window,
sent_vec_dim)
sent_pool_size = (sentence_num - sentence_window + 1, 1)
dropout_sent_conv = nn.ConvPoolLayer(rng,
input=dropout_sent_vec,
input_shape=None,
filter_shape=sent_filter_shape,
pool_size=sent_pool_size,
activation=Tanh,
k=k_max_sentence)
sent_conv = nn.ConvPoolLayer(rng,
input=sent_vec * (1 - drop_rate_sentence),
input_shape=None,
filter_shape=sent_filter_shape,
pool_size=sent_pool_size,
activation=Tanh,
k=k_max_sentence,
W=dropout_sent_conv.W,
b=dropout_sent_conv.b)
dropout_doc_vec = dropout_sent_conv.output.flatten(2)
doc_vec = sent_conv.output.flatten(2)
doc_vec_dim = num_maps_sentence * k_max_sentence
# construct classifier
dropout_logistic_layer = nn.LogisticRegressionLayer(
input=dropout_doc_vec, n_in=doc_vec_dim, n_out=2)
logistic_layer = nn.LogisticRegressionLayer(input=doc_vec,
n_in=doc_vec_dim,
n_out=2,
W=dropout_logistic_layer.W,
b=dropout_logistic_layer.b)
dropout_cost = dropout_logistic_layer.negative_log_likelihood(y)
cost = logistic_layer.negative_log_likelihood(y)
preds = logistic_layer.y_pred
errors = logistic_layer.errors(y)
# collect parameters
self.params.append(words)
self.params += dropout_word_conv.params
self.params += dropout_sent_conv.params
self.params += dropout_logistic_layer.params
grad_updates = nn.sgd_updates_adadelta(self.params, dropout_cost, rho,
epsilon, norm_lim)
# construct the dataset
train_x, train_y = nn.shared_dataset(dataset[0])
test_x, test_y = nn.shared_dataset(dataset[1])
test_cpu_y = dataset[1][1]
n_train_batches = int(np.ceil(1.0 * len(dataset[0][0]) / batch_size))
n_test_batches = int(np.ceil(1.0 * len(dataset[1][0]) / batch_size))
# construt the model
index = T.iscalar()
train_func = theano.function(
[index],
dropout_cost,
updates=grad_updates,
givens={
x: train_x[index * batch_size:(index + 1) * batch_size],
y: train_y[index * batch_size:(index + 1) * batch_size]
})
test_func = theano.function(
[index],
preds,
givens={x: test_x[index * batch_size:(index + 1) * batch_size]})
get_train_sentvec = theano.function(
[index],
sent_vec,
givens={x: train_x[index * batch_size:(index + 1) * batch_size]})
get_test_sentvec = theano.function(
[index],
sent_vec,
givens={x: test_x[index * batch_size:(index + 1) * batch_size]})
epoch = 0
best_score = 0
raw_train_x = dataset[0][0]
raw_test_x = dataset[1][0]
# get the sentence number for each document
number_train_sens = []
number_test_sens = []
for doc in raw_train_x:
sen_num = 0
for sen in doc:
if np.any(sen):
sen_num += 1
number_train_sens.append(sen_num)
for doc in raw_test_x:
sen_num = 0
for sen in doc:
if np.any(sen):
sen_num += 1
number_test_sens.append(sen_num)
log_file = open("./log/%s.log" % exp_name, 'w')
while epoch <= max_iteration:
start_time = timeit.default_timer()
epoch += 1
costs = []
for minibatch_index in np.random.permutation(
range(n_train_batches)):
cost_epoch = train_func(minibatch_index)
costs.append(cost_epoch)
set_zero(zero_vec)
if epoch % 5 == 0:
test_preds = []
for i in xrange(n_test_batches):
test_y_pred = test_func(i)
test_preds.append(test_y_pred)
test_preds = np.concatenate(test_preds)
test_score = 1 - np.mean(np.not_equal(test_cpu_y, test_preds))
precision, recall, beta, support = precision_recall_fscore_support(
test_cpu_y, test_preds, pos_label=1)
if test_score > best_score:
best_score = test_score
# save the sentence vectors
train_sens = [
get_train_sentvec(i) for i in range(n_train_batches)
]
test_sens = [
get_test_sentvec(i) for i in range(n_test_batches)
]
train_sens = np.concatenate(train_sens, axis=0)
test_sens = np.concatenate(test_sens, axis=0)
out_train_sent_file = "./results/%s_train_sent.vec" % exp_name
out_test_sent_file = "./results/%s_test_sent.vec" % exp_name
with open(out_train_sent_file,
'w') as train_f, open(out_test_sent_file,
'w') as test_f:
for i in range(len(train_sens)):
tr_doc_vect = train_sens[i][
0][:number_train_sens[i]]
train_f.write(
json.dumps(tr_doc_vect.tolist()) + "\n")
for i in range(len(test_sens)):
te_doc_vect = test_sens[i][0][:number_test_sens[i]]
test_f.write(
json.dumps(te_doc_vect.tolist()) + "\n")
print "Get best performace at %d iteration" % epoch
log_file.write("Get best performance at %d iteration\n" %
epoch)
end_time = timeit.default_timer()
print "Iteration %d , precision, recall, support" % epoch, precision, recall, support
log_file.write(
"Iteration %d, neg precision %f, pos precision %f, neg recall %f pos recall %f \n"
%
(epoch, precision[0], precision[1], recall[0], recall[1]))
print "Using time %f m" % ((end_time - start_time) / 60.)
log_file.write("Uing time %f m\n" %
((end_time - start_time) / 60.))
end_time = timeit.default_timer()
print "Iteration %d Using time %f m" % (epoch,
(end_time - start_time) /
60.)
log_file.write("Uing time %f m\n" %
((end_time - start_time) / 60.))
log_file.flush()
log_file.close()
def run_experiment(self, dataset, word_embedding, exp_name):
# load parameters
num_maps_word = self.options["num_maps_word"]
drop_rate_word = self.options["drop_rate_word"]
word_window = self.options["word_window"]
word_dim = self.options["word_dim"]
k_max_word = self.options["k_max_word"]
num_maps_sentence = self.options["num_maps_sentence"]
drop_rate_sentence = self.options["drop_rate_sentence"]
sentence_window = self.options["sentence_window"]
k_max_sentence = self.options["k_max_sentence"]
batch_size = self.options["batch_size"]
rho = self.options["rho"]
epsilon = self.options["epsilon"]
norm_lim = self.options["norm_lim"]
max_iteration = self.options["max_iteration"]
sentence_len = len(dataset[0][0][0][0])
sentence_num = len(dataset[0][0][0])
# define the parameters
x = T.tensor3("x")
y = T.ivector("y")
rng = np.random.RandomState(1234)
words = theano.shared(value=np.asarray(word_embedding,
dtype=theano.config.floatX),
name="embedding", borrow=True)
zero_vector_tensor = T.vector()
zero_vec = np.zeros(word_dim, dtype=theano.config.floatX)
set_zero = theano.function([zero_vector_tensor], updates=[(words, T.set_subtensor(words[0,:], zero_vector_tensor))])
x_emb = words[T.cast(x.flatten(), dtype="int32")].reshape((x.shape[0]*x.shape[1], 1, x.shape[2], words.shape[1]))
dropout_x_emb = nn.dropout_from_layer(rng, x_emb, drop_rate_word)
# compute convolution on words layer
word_filter_shape = (num_maps_word, 1, word_window, word_dim)
word_pool_size = (sentence_len - word_window + 1, 1)
dropout_word_conv = nn.ConvPoolLayer(rng,
input=dropout_x_emb,
input_shape=None,
filter_shape=word_filter_shape,
pool_size=word_pool_size,
activation=Tanh,
k=k_max_word)
sent_vec_dim = num_maps_word*k_max_word
dropout_sent_vec = dropout_word_conv.output.reshape((x.shape[0], 1, x.shape[1], sent_vec_dim))
dropout_sent_vec = nn.dropout_from_layer(rng, dropout_sent_vec, drop_rate_sentence)
word_conv = nn.ConvPoolLayer(rng,
input=dropout_x_emb*(1 - drop_rate_word),
input_shape=None,
filter_shape=word_filter_shape,
pool_size=word_pool_size,
activation=Tanh,
k=k_max_word,
W=dropout_word_conv.W,
b=dropout_word_conv.b)
sent_vec = word_conv.output.reshape((x.shape[0], 1, x.shape[1], sent_vec_dim))
# construct the convolution layer on sentences
sent_filter_shape = (num_maps_sentence, 1, sentence_window, sent_vec_dim)
sent_pool_size = (sentence_num - sentence_window + 1, 1)
dropout_sent_conv = nn.ConvPoolLayer(rng,
input=dropout_sent_vec,
input_shape=None,
filter_shape=sent_filter_shape,
pool_size=sent_pool_size,
activation=Tanh,
k=k_max_sentence)
sent_conv = nn.ConvPoolLayer(rng,
input=sent_vec*(1 - drop_rate_sentence),
input_shape=None,
filter_shape=sent_filter_shape,
pool_size=sent_pool_size,
activation=Tanh,
k=k_max_sentence,
W=dropout_sent_conv.W,
b=dropout_sent_conv.b)
dropout_doc_vec = dropout_sent_conv.output.flatten(2)
doc_vec = sent_conv.output.flatten(2)
doc_vec_dim = num_maps_sentence * k_max_sentence
# construct classifier
dropout_logistic_layer = nn.LogisticRegressionLayer(
input=dropout_doc_vec,
n_in=doc_vec_dim,
n_out=2)
logistic_layer = nn.LogisticRegressionLayer(
input=doc_vec,
n_in=doc_vec_dim,
n_out=2,
W=dropout_logistic_layer.W,
b=dropout_logistic_layer.b)
dropout_cost = dropout_logistic_layer.negative_log_likelihood(y)
cost = logistic_layer.negative_log_likelihood(y)
preds = logistic_layer.y_pred
errors = logistic_layer.errors(y)
# collect parameters
self.params.append(words)
self.params += dropout_word_conv.params
self.params += dropout_sent_conv.params
self.params += dropout_logistic_layer.params
grad_updates = nn.sgd_updates_adadelta(self.params,
dropout_cost,
rho,
epsilon,
norm_lim)
# construct the dataset
train_x, train_y = nn.shared_dataset(dataset[0])
test_x, test_y = nn.shared_dataset(dataset[1])
test_cpu_y = dataset[1][1]
n_train_batches = int(np.ceil(1.0 * len(dataset[0][0]) / batch_size))
n_test_batches = int(np.ceil(1.0 * len(dataset[1][0]) / batch_size))
# construt the model
index = T.iscalar()
train_func = theano.function([index], dropout_cost, updates=grad_updates,
givens={
x: train_x[index*batch_size:(index+1)*batch_size],
y: train_y[index*batch_size:(index+1)*batch_size]
})
test_func = theano.function([index], preds,
givens={
x:test_x[index*batch_size:(index+1)*batch_size]
})
get_train_sentvec = theano.function([index], sent_vec,
givens={
x:train_x[index*batch_size:(index+1)*batch_size]
})
get_test_sentvec = theano.function([index], sent_vec,
givens={
x:test_x[index*batch_size:(index+1)*batch_size]
})
epoch = 0
best_score = 0
raw_train_x = dataset[0][0]
raw_test_x = dataset[1][0]
# get the sentence number for each document
number_train_sens = []
number_test_sens = []
for doc in raw_train_x:
sen_num = 0
for sen in doc:
if np.any(sen):
sen_num += 1
number_train_sens.append(sen_num)
for doc in raw_test_x:
sen_num = 0
for sen in doc:
if np.any(sen):
sen_num += 1
number_test_sens.append(sen_num)
log_file = open("./log/%s.log" % exp_name, 'w')
while epoch <= max_iteration:
start_time = timeit.default_timer()
epoch += 1
costs = []
for minibatch_index in np.random.permutation(range(n_train_batches)):
cost_epoch = train_func(minibatch_index)
costs.append(cost_epoch)
set_zero(zero_vec)
if epoch % 5 == 0:
test_preds = []
for i in xrange(n_test_batches):
test_y_pred = test_func(i)
test_preds.append(test_y_pred)
test_preds = np.concatenate(test_preds)
test_score = 1 - np.mean(np.not_equal(test_cpu_y, test_preds))
precision, recall, beta, support = precision_recall_fscore_support(test_cpu_y, test_preds, pos_label=1)
if test_score > best_score:
best_score = test_score
# save the sentence vectors
train_sens = [get_train_sentvec(i) for i in range(n_train_batches)]
test_sens = [get_test_sentvec(i) for i in range(n_test_batches)]
train_sens = np.concatenate(train_sens, axis=0)
test_sens = np.concatenate(test_sens, axis=0)
out_train_sent_file = "./results/%s_train_sent.vec" % exp_name
out_test_sent_file = "./results/%s_test_sent.vec" % exp_name
with open(out_train_sent_file, 'w') as train_f, open(out_test_sent_file, 'w') as test_f:
for i in range(len(train_sens)):
tr_doc_vect = train_sens[i][0][:number_train_sens[i]]
train_f.write(json.dumps(tr_doc_vect.tolist()) + "\n")
for i in range(len(test_sens)):
te_doc_vect = test_sens[i][0][:number_test_sens[i]]
test_f.write(json.dumps(te_doc_vect.tolist()) + "\n")
print "Get best performace at %d iteration" % epoch
log_file.write("Get best performance at %d iteration\n" % epoch)
end_time = timeit.default_timer()
print "Iteration %d , precision, recall, support" % epoch, precision, recall, support
log_file.write("Iteration %d, neg precision %f, pos precision %f, neg recall %f pos recall %f \n" % (epoch, precision[0], precision[1], recall[0], recall[1]))
print "Using time %f m" % ((end_time -start_time)/60.)
log_file.write("Uing time %f m\n" % ((end_time - start_time)/60.))
end_time = timeit.default_timer()
print "Iteration %d Using time %f m" % ( epoch, (end_time -start_time)/60.)
log_file.write("Uing time %f m\n" % ((end_time - start_time)/60.))
log_file.flush()
log_file.close()
def run_cnn(exp_name,
dataset, embedding,
log_fn, perf_fn,
emb_dm=100,
batch_size=100,
filter_hs=[1, 2, 3],
hidden_units=[200, 100, 11],
dropout_rate=0.5,
shuffle_batch=True,
n_epochs=300,
lr_decay=0.95,
activation=ReLU,
sqr_norm_lim=9,
non_static=True,
print_freq=5):
"""
Train and Evaluate CNN event encoder model
:dataset: list containing three elements[(train_x, train_y),
(valid_x, valid_y), (test_x, test_y)]
:embedding: word embedding with shape (|V| * emb_dm)
:filter_hs: filter height for each paralle cnn layer
:dropout_rate: dropout rate for full connected layers
:n_epochs: the max number of iterations
"""
start_time = timeit.default_timer()
rng = np.random.RandomState(1234)
input_height = len(dataset[0][0][0][0]) # number of words in the sentences
num_sens = len(dataset[0][0][0]) # number of sentences
print "--input height ", input_height
input_width = emb_dm
num_maps = hidden_units[0]
###################
# start snippet 1 #
###################
print "start to construct the model ...."
x = T.tensor3("x")
y = T.ivector("y")
words = shared(value=np.asarray(embedding,
dtype=theano.config.floatX),
name="embedding", borrow=True)
# define function to keep padding vector as zero
zero_vector_tensor = T.vector()
zero_vec = np.zeros(input_width, dtype=theano.config.floatX)
set_zero = function([zero_vector_tensor],
updates=[(words, T.set_subtensor(words[0,:], zero_vector_tensor))])
# the input for the sentence level conv layers
layer0_input = words[T.cast(x.flatten(), dtype="int32")].reshape((
x.shape[0] * x.shape[1], 1, x.shape[2], emb_dm
))
conv_layers = []
filter_shape = (num_maps, 1, filter_hs[0], emb_dm)
pool_size = (input_height - filter_hs[0] + 1, 1)
conv_layer = nn.ConvPoolLayer(rng, input=layer0_input,
input_shape=None,
filter_shape=filter_shape,
pool_size=pool_size, activation=activation)
sen_vecs = conv_layer.output.reshape((x.shape[0] * x.shape[1], num_maps))
conv_layers.append(conv_layer)
# compute preactivation for each sentences
layer_sizes = zip(hidden_units, hidden_units[1:])
full_layer_input = sen_vecs
dropout_input = sen_vecs
hidden_outs = []
drophidden_outs = []
hidden_layers = []
dropout_layers = []
droprate = 0.5
for lay_size in layer_sizes[:-1]:
U_value = np.random.random(lay_size).astype(theano.config.floatX)
b_value = np.zeros((lay_size[1],), dtype=theano.config.floatX)
U = theano.shared(U_value, borrow=True, name="U")
b = theano.shared(b_value, borrow=True, name="b")
hiddenLayer = nn.HiddenLayer(rng, full_layer_input, lay_size[0], lay_size[1], ReLU, U * (1 - droprate), b)
dropHiddenLayer = nn.DropoutHiddenLayer(rng, dropout_input, lay_size[0], lay_size[1], ReLU, droprate, U, b)
hidden_layers.append(hiddenLayer)
dropout_layers.append(dropHiddenLayer)
hidden_out = hiddenLayer.output
drophidden_out = dropHiddenLayer.output
hidden_outs.append(hidden_out)
drophidden_outs.append(drophidden_out)
full_layer_input = hidden_out
dropout_input = drophidden_out
# get the max value for each class
n_in, n_out = layer_sizes[-1]
W_value = np.random.random((n_in, n_out)).astype(theano.config.floatX)
b_value = np.zeros((n_out,), dtype=theano.config.floatX)
W = theano.shared(W_value, borrow=True, name="logis_W")
b = theano.shared(b_value, borrow=True, name="logis_b")
full_act = T.dot(hidden_outs[-1], W*(1 - droprate)) + b
dropout_act = nn.dropout_from_layer(rng, T.dot(drophidden_outs[-1], W) + b, droprate)
# compute the probability
sen_full_probs = T.nnet.softmax(full_act)
sen_dropout_probs = T.nnet.softmax(dropout_act)
# compute the sentence similarity
sen_sen = T.dot(sen_vecs, sen_vecs.T)
sen_sqr = T.sum(sen_vecs ** 2, axis=1)
sen_sqr_left = sen_sqr.dimshuffle(0, 'x')
sen_sqr_right = sen_sqr.dimshuffle('x', 0)
sen_smi_matrix = sen_sqr_left - 2 * sen_sen + sen_sqr_right
sen_smi_matrix = T.exp(-1 * sen_smi_matrix)
# compute the delta between sentence probabilities
prob_prob_full = T.dot(sen_full_probs, sen_full_probs.T)
prob_sqr_full = T.sum(sen_full_probs ** 2, axis=1)
prob_sqr_left_full = prob_sqr_full.dimshuffle(0, 'x')
prob_sqr_right_full = prob_sqr_full.dimshuffle('x', 0)
prob_delta_full = prob_sqr_left_full - 2 * prob_prob_full + prob_sqr_right_full
sen_cost_full = T.sum(sen_smi_matrix * prob_delta_full)
prob_prob_drop = T.dot(sen_dropout_probs, sen_dropout_probs.T)
prob_sqr_drop = T.sum(sen_dropout_probs ** 2, axis=1)
prob_sqr_left_drop = prob_sqr_drop.dimshuffle(0, 'x')
prob_sqr_right_drop = prob_sqr_drop.dimshuffle('x', 0)
prob_delta_drop = prob_sqr_left_drop - 2 * prob_prob_drop + prob_sqr_right_drop
sen_cost_drop = T.sum(sen_smi_matrix * prob_delta_drop)
# transform the sen probs to doc probs
# by using average probs
doc_full_probs = sen_full_probs.reshape((x.shape[0], x.shape[1], n_out))
doc_full_probs = T.mean(doc_full_probs, axis=1)
doc_dropout_probs = sen_dropout_probs.reshape((x.shape[0], x.shape[1], n_out))
doc_dropout_probs = T.mean(doc_dropout_probs, axis=1)
doc_full_y_pred = T.argmax(doc_full_probs, axis=1)
doc_dropout_y_pred = T.argmax(doc_dropout_probs, axis=1)
full_negative_likelihood = T.sum(-T.log(doc_full_probs)[T.arange(y.shape[0]), y])
dropout_negative_likelihood = T.sum(-T.log(doc_dropout_probs)[T.arange(y.shape[0]), y])
full_errors = T.mean(T.neq(doc_full_y_pred, y))
gamma = 2
full_cost = full_negative_likelihood + gamma * sen_cost_full
dropout_cost = dropout_negative_likelihood + gamma * sen_cost_drop
params = []
for conv_layer in conv_layers:
params += conv_layer.params
for dropout_layer in dropout_layers:
params += dropout_layer.params
params.append(W)
params.append(b)
if non_static:
params.append(words)
grad_updates = sgd_updates_adadelta(params,
dropout_cost,
lr_decay,
1e-6,
sqr_norm_lim)
#####################
# Construct Dataset #
#####################
print "Copy data to GPU and constrct train/valid/test func"
np.random.seed(1234)
train_x, train_y = shared_dataset(dataset[0])
valid_x, valid_y = shared_dataset(dataset[1])
test_x, test_y = shared_dataset(dataset[2])
n_train_batches = int(np.ceil(1.0 * len(dataset[0][0]) / batch_size))
n_valid_batches = int(np.ceil(1.0 * len(dataset[1][0]) / batch_size))
n_test_batches = int(np.ceil(1.0 * len(dataset[2][0]) / batch_size))
#####################
# Train model func #
#####################
index = T.iscalar()
train_func = function([index], full_cost, updates=grad_updates,
givens={
x: train_x[index*batch_size:(index+1)*batch_size],
y: train_y[index*batch_size:(index+1)*batch_size]
})
train_error = function([index], full_errors,
givens={
x: train_x[index*batch_size:(index+1)*batch_size],
y: train_y[index*batch_size:(index+1)*batch_size]
})
valid_train_func = function([index], [full_negative_likelihood, sen_cost_full], updates=grad_updates,
givens={
x: valid_x[index*batch_size:(index+1)*batch_size],
y: valid_y[index*batch_size:(index+1)*batch_size]
})
test_pred = function([index], doc_full_y_pred,
givens={
x:test_x[index*batch_size:(index+1)*batch_size],
})
# apply early stop strategy
patience = 100
patience_increase = 2
improvement_threshold = 1.005
n_valid = len(dataset[1][0])
n_test = len(dataset[2][0])
epoch = 0
best_params = None
best_validation_score = 0.
test_perf = 0
done_loop = False
log_file = open(log_fn, 'w')
print "Start to train the model....."
cpu_trn_y = np.asarray(dataset[0][1])
cpu_val_y = np.asarray(dataset[1][1])
cpu_tst_y = np.asarray(dataset[2][1])
def compute_score(true_list, pred_list):
mat = np.equal(true_list, pred_list)
score = np.mean(mat)
return score
best_test_score = 0.
while (epoch < n_epochs) and not done_loop:
start_time = timeit.default_timer()
epoch += 1
costs = []
for minibatch_index in np.random.permutation(range(n_train_batches)):
cost_epoch = train_func(minibatch_index)
costs.append(cost_epoch)
set_zero(zero_vec)
# do validatiovalidn
valid_cost, valid_sen_cost = zip(*[valid_train_func(i) for i in np.random.permutation(xrange(n_valid_batches))])
if epoch % print_freq == 0:
# do test
test_preds = np.concatenate([test_pred(i) for i in xrange(n_test_batches)])
test_score = compute_score(cpu_tst_y, test_preds)
with open(os.path.join(perf_fn, "%s_%d.pred" % (exp_name, epoch)), 'w') as epf:
for p in test_preds:
epf.write("%d\n" % int(p))
message = "Epoch %d test perf %f train cost %f, valid_sen_cost %f, valid_doc_cost %f" % (epoch, test_score, np.mean(costs), np.mean(valid_sen_cost), np.mean(valid_cost))
print message
log_file.write(message + "\n")
log_file.flush()
"""
# store the best model
if (test_score > best_test_score) or (epoch % 25 == 0):
best_test_score = test_score
# save the model
model_name = "%s_%d.model" % (exp_name, epoch)
with open(model_name, 'wb') as bm:
for p in params:
cPickle.dump(p.get_value(), bm)
"""
end_time = timeit.default_timer()
print "Finish one iteration using %f m" % ((end_time - start_time)/60.)
log_file.flush()
log_file.close()
def run_experiment(self, dataset, word_embedding, exp_name):
# load parameters
num_maps_word = self.options["num_maps_word"]
drop_rate_word = self.options["drop_rate_word"]
drop_rate_sentence = self.options["drop_rate_sentence"]
word_window = self.options["word_window"]
word_dim = self.options["word_dim"]
k_max_word = self.options["k_max_word"]
batch_size = self.options["batch_size"]
rho = self.options["rho"]
epsilon = self.options["epsilon"]
norm_lim = self.options["norm_lim"]
max_iteration = self.options["max_iteration"]
k = self.options["k_max"]
sentence_len = len(dataset[0][0][0][0])
sentence_num = len(dataset[0][0][0])
# define the parameters
x = T.tensor3("x")
y = T.ivector("y")
rng = np.random.RandomState(1234)
words = theano.shared(value=np.asarray(word_embedding,
dtype=theano.config.floatX),
name="embedding", borrow=True)
zero_vector_tensor = T.vector()
zero_vec = np.zeros(word_dim, dtype=theano.config.floatX)
set_zero = theano.function([zero_vector_tensor], updates=[(words, T.set_subtensor(words[0,:], zero_vector_tensor))])
x_emb = words[T.cast(x.flatten(), dtype="int32")].reshape((x.shape[0]*x.shape[1], 1, x.shape[2], words.shape[1]))
dropout_x_emb = nn.dropout_from_layer(rng, x_emb, drop_rate_word)
# compute convolution on words layer
word_filter_shape = (num_maps_word, 1, word_window, word_dim)
word_pool_size = (sentence_len - word_window + 1, 1)
dropout_word_conv = nn.ConvPoolLayer(rng,
input=dropout_x_emb,
input_shape=None,
filter_shape=word_filter_shape,
pool_size=word_pool_size,
activation=Tanh,
k=k_max_word)
sent_vec_dim = num_maps_word*k_max_word
dropout_sent_vec = dropout_word_conv.output.reshape((x.shape[0] * x.shape[1], sent_vec_dim))
word_conv = nn.ConvPoolLayer(rng,
input=dropout_x_emb*(1 - drop_rate_word),
input_shape=None,
filter_shape=word_filter_shape,
pool_size=word_pool_size,
activation=Tanh,
k=k_max_word,
W=dropout_word_conv.W,
b=dropout_word_conv.b)
sent_vec = word_conv.output.reshape((x.shape[0] * x.shape[1], sent_vec_dim))
theta_value = np.random.random((sent_vec_dim,1))
theta = shared(value=np.asarray(theta_value, dtype=theano.config.floatX), name="theta", borrow=True)
weighted_drop_sent_vec, weighted_sen_score = keep_max(dropout_sent_vec.reshape((x.shape[0], 1, x.shape[1], sent_vec_dim)), theta, k)
drop_doc_vec = T.sum(weighted_drop_sent_vec, axis=2).flatten(2)
weighted_sent_vec, sen_score = keep_max(sent_vec.reshape((x.shape[0], 1, x.shape[1], sent_vec_dim)), theta, k)
doc_vec = T.sum(weighted_sent_vec, axis=2).flatten(2)
# we need to constrain the number of positive sentences in positive
# collect parameters
self.params.append(words)
self.params += dropout_word_conv.params
self.params.append(sen_W)
self.params.append(sen_b)
grad_updates = nn.sgd_updates_adadelta(self.params,
drop_cost,
rho,
epsilon,
norm_lim)
# construct the dataset
train_x, train_y = nn.shared_dataset(dataset[0])
test_x, test_y = nn.shared_dataset(dataset[1])
test_cpu_y = dataset[1][1]
n_train_batches = int(np.ceil(1.0 * len(dataset[0][0]) / batch_size))
n_test_batches = int(np.ceil(1.0 * len(dataset[1][0]) / batch_size))
# construt the model
index = T.iscalar()
train_func = theano.function([index], [drop_cost, drop_bag_cost, drop_sent_cost, penal_cost], updates=grad_updates,
givens={
x: train_x[index*batch_size:(index+1)*batch_size],
y: train_y[index*batch_size:(index+1)*batch_size]
})
test_func = theano.function([index], doc_preds,
givens={
x:test_x[index*batch_size:(index+1)*batch_size]
})
get_train_sent_prob = theano.function([index], sent_prob,
givens={
x:train_x[index*batch_size:(index+1)*batch_size]
})
get_test_sent_prob = theano.function([index], sent_prob,
givens={
x:test_x[index*batch_size:(index+1)*batch_size]
})
epoch = 0
best_score = 0
raw_train_x = dataset[0][0]
raw_test_x = dataset[1][0]
# get the sentence number for each document
number_train_sens = []
number_test_sens = []
log_file = open("./log/%s.log" % exp_name, 'w')
while epoch <= max_iteration:
start_time = timeit.default_timer()
epoch += 1
costs = []
for minibatch_index in np.random.permutation(range(n_train_batches)):
cost_epoch = train_func(minibatch_index)
costs.append(cost_epoch)
set_zero(zero_vec)
total_train_cost, train_bag_cost, train_sent_cost, train_penal_cost = zip(*costs)
print "Iteration %d, total_cost %f bag_cost %f sent_cost %f penal_cost %f\n" % (epoch, np.mean(total_train_cost), np.mean(train_bag_cost), np.mean(train_sent_cost), np.mean(train_penal_cost))
if epoch % 5 == 0:
test_preds = []
for i in xrange(n_test_batches):
test_y_pred = test_func(i)
test_preds.append(test_y_pred)
test_preds = np.concatenate(test_preds)
test_score = 1 - np.mean(np.not_equal(test_cpu_y, test_preds))
precision, recall, beta, support = precision_recall_fscore_support(test_cpu_y, test_preds, pos_label=1)
if test_score > best_score:
best_score = test_score
# save the sentence vectors
train_sens = [get_train_sent_prob(i) for i in range(n_train_batches)]
test_sens = [get_test_sent_prob(i) for i in range(n_test_batches)]
train_sens = np.concatenate(train_sens, axis=0)
test_sens = np.concatenate(test_sens, axis=0)
out_train_sent_file = "./results/%s_train_sent.vec" % exp_name
out_test_sent_file = "./results/%s_test_sent.vec" % exp_name
with open(out_train_sent_file, 'w') as train_f, open(out_test_sent_file, 'w') as test_f:
cPickle.dump(train_sens, train_f)
cPickle.dump(test_sens, test_f)
print "Get best performace at %d iteration %f" % (epoch, test_score)
log_file.write("Get best performance at %d iteration %f \n" % (epoch, test_score))
end_time = timeit.default_timer()
print "Iteration %d , precision, recall, support" % epoch, precision, recall, support
log_file.write("Iteration %d, neg precision %f, pos precision %f, neg recall %f pos recall %f , total_cost %f bag_cost %f sent_cost %f penal_cost %f\n" % (epoch, precision[0], precision[1], recall[0], recall[1], np.mean(total_train_cost), np.mean(train_bag_cost), np.mean(train_sent_cost), np.mean(train_penal_cost)))
print "Using time %f m" % ((end_time -start_time)/60.)
log_file.write("Uing time %f m\n" % ((end_time - start_time)/60.))
end_time = timeit.default_timer()
print "Iteration %d Using time %f m" % ( epoch, (end_time -start_time)/60.)
log_file.write("Uing time %f m\n" % ((end_time - start_time)/60.))
log_file.flush()
log_file.close()