def evaluate_lenet5(learning_rate=0.02,
n_epochs=4,
L2_weight=1e-5,
extra_size=4,
emb_size=300,
batch_size=100,
filter_size=[3, 3],
maxSentLen=40,
hidden_size=[300, 300],
max_term_len=4,
p_mode='conc'):
model_options = locals().copy()
print "model options", model_options
seed = 1234
np.random.seed(seed)
rng = np.random.RandomState(
seed) #random seed, control the model generates the same results
all_sentences_l, all_masks_l, all_sentences_r, all_masks_r, all_word1, all_word2, all_word1_mask, all_word2_mask, all_labels, all_extra, word2id = load_wordnet_hyper_vs_all_with_words(
maxlen=maxSentLen, wordlen=max_term_len
) #minlen, include one label, at least one word in the sentence
# test_sents_l, test_masks_l, test_sents_r, test_masks_r, test_labels, word2id =load_ACE05_dataset(maxSentLen, word2id)
test_sents_l, test_masks_l, test_sents_r, test_masks_r, test_word1, test_word2, test_word1_mask, test_word2_mask, test_labels, test_extra, word2id = load_task_hyper_vs_all_with_words(
LenciBenotto_file, maxSentLen, word2id, wordlen=max_term_len)
store_word2id(word2id, root_dic + 'LenciBenotto_word2id.pkl')
# exit(0)
total_size = len(all_sentences_l)
hold_test_size = 10000
train_size = total_size - hold_test_size
train_sents_l = np.asarray(all_sentences_l[:train_size], dtype='int32')
# dev_sents_l=np.asarray(all_sentences_l[1], dtype='int32')
# test_sents_l=np.asarray(all_sentences_l[-test_size:], dtype='int32')
test_sents_l = np.asarray(test_sents_l, dtype='int32')
train_masks_l = np.asarray(all_masks_l[:train_size],
dtype=theano.config.floatX)
# dev_masks_l=np.asarray(all_masks_l[1], dtype=theano.config.floatX)
# test_masks_l=np.asarray(all_masks_l[-test_size:], dtype=theano.config.floatX)
test_masks_l = np.asarray(test_masks_l, dtype=theano.config.floatX)
train_sents_r = np.asarray(all_sentences_r[:train_size], dtype='int32')
# dev_sents_r=np.asarray(all_sentences_r[1] , dtype='int32')
# test_sents_r=np.asarray(all_sentences_r[-test_size:], dtype='int32')
test_sents_r = np.asarray(test_sents_r, dtype='int32')
train_masks_r = np.asarray(all_masks_r[:train_size],
dtype=theano.config.floatX)
# dev_masks_r=np.asarray(all_masks_r[1], dtype=theano.config.floatX)
# test_masks_r=np.asarray(all_masks_r[-test_size:], dtype=theano.config.floatX)
test_masks_r = np.asarray(test_masks_r, dtype=theano.config.floatX)
train_word1 = np.asarray(all_word1[:train_size], dtype='int32')
train_word2 = np.asarray(all_word2[:train_size], dtype='int32')
test_word1 = np.asarray(test_word1, dtype='int32')
test_word2 = np.asarray(test_word2, dtype='int32')
train_word1_mask = np.asarray(all_word1_mask[:train_size],
dtype=theano.config.floatX)
train_word2_mask = np.asarray(all_word2_mask[:train_size],
dtype=theano.config.floatX)
test_word1_mask = np.asarray(test_word1_mask, dtype=theano.config.floatX)
test_word2_mask = np.asarray(test_word2_mask, dtype=theano.config.floatX)
train_labels_store = np.asarray(all_labels[:train_size], dtype='int32')
# dev_labels_store=np.asarray(all_labels[1], dtype='int32')
# test_labels_store=np.asarray(all_labels[-test_size:], dtype='int32')
test_labels_store = np.asarray(test_labels, dtype='int32')
train_extra = np.asarray(all_extra[:train_size],
dtype=theano.config.floatX)
test_extra = np.asarray(test_extra, dtype=theano.config.floatX)
# train_size=len(train_labels_store)
# dev_size=len(dev_labels_store)
test_size = len(test_labels_store)
print 'train size: ', train_size, ' test size: ', test_size
vocab_size = len(word2id) + 1
rand_values = rng.normal(
0.0, 0.01,
(vocab_size, emb_size)) #generate a matrix by Gaussian distribution
#here, we leave code for loading word2vec to initialize words
rand_values[0] = np.array(np.zeros(emb_size), dtype=theano.config.floatX)
id2word = {y: x for x, y in word2id.iteritems()}
word2vec = load_word2vec()
rand_values = load_word2vec_to_init(rand_values, id2word, word2vec)
init_embeddings = theano.shared(
value=np.array(rand_values, dtype=theano.config.floatX), borrow=True
) #wrap up the python variable "rand_values" into theano variable
store_model_to_file(root_dic + 'LenciBenotto_best_para_init_embeddings',
[init_embeddings])
#now, start to build the input form of the model
sents_ids_l = T.imatrix()
sents_mask_l = T.fmatrix()
sents_ids_r = T.imatrix()
sents_mask_r = T.fmatrix()
word1_ids = T.imatrix()
word2_ids = T.imatrix()
word1_mask = T.fmatrix()
word2_mask = T.fmatrix()
extra = T.fvector()
labels = T.ivector()
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
def embed_input(emb_matrix, sent_ids):
return emb_matrix[sent_ids.flatten()].reshape(
(batch_size, maxSentLen, emb_size)).dimshuffle(0, 2, 1)
embed_input_l = embed_input(
init_embeddings, sents_ids_l
) #embeddings[sents_ids_l.flatten()].reshape((batch_size,maxSentLen, emb_size)).dimshuffle(0,2,1) #the input format can be adapted into CNN or GRU or LSTM
embed_input_r = embed_input(
init_embeddings, sents_ids_r
) #embeddings[sents_ids_r.flatten()].reshape((batch_size,maxSentLen, emb_size)).dimshuffle(0,2,1)
embed_word1 = init_embeddings[word1_ids.flatten()].reshape(
(batch_size, word1_ids.shape[1], emb_size))
embed_word2 = init_embeddings[word2_ids.flatten()].reshape(
(batch_size, word2_ids.shape[1], emb_size))
word1_embedding = T.sum(embed_word1 * word1_mask.dimshuffle(0, 1, 'x'),
axis=1)
word2_embedding = T.sum(embed_word2 * word2_mask.dimshuffle(0, 1, 'x'),
axis=1)
'''create_AttentiveConv_params '''
conv_W, conv_b = create_conv_para(rng,
filter_shape=(hidden_size[1], 1,
emb_size, filter_size[0]))
conv_W_context, conv_b_context = create_conv_para(
rng, filter_shape=(hidden_size[1], 1, emb_size, 1))
NN_para = [conv_W, conv_b, conv_W_context]
'''
attentive convolution function
'''
term_vs_term_layer = Conv_for_Pair(
rng,
origin_input_tensor3=embed_word1.dimshuffle(0, 2, 1),
origin_input_tensor3_r=embed_word2.dimshuffle(0, 2, 1),
input_tensor3=embed_word1.dimshuffle(0, 2, 1),
input_tensor3_r=embed_word2.dimshuffle(0, 2, 1),
mask_matrix=word1_mask,
mask_matrix_r=word2_mask,
image_shape=(batch_size, 1, emb_size, max_term_len),
image_shape_r=(batch_size, 1, emb_size, max_term_len),
filter_shape=(hidden_size[1], 1, emb_size, filter_size[0]),
filter_shape_context=(hidden_size[1], 1, emb_size, 1),
W=conv_W,
b=conv_b,
W_context=conv_W_context,
b_context=conv_b_context)
tt_embeddings_l = term_vs_term_layer.attentive_maxpool_vec_l
tt_embeddings_r = term_vs_term_layer.attentive_maxpool_vec_r
p_ww = T.concatenate([
tt_embeddings_l, tt_embeddings_r, tt_embeddings_l * tt_embeddings_r,
tt_embeddings_l - tt_embeddings_r
],
axis=1)
term_vs_def_layer = Conv_for_Pair(
rng,
origin_input_tensor3=embed_word1.dimshuffle(0, 2, 1),
origin_input_tensor3_r=embed_input_r,
input_tensor3=embed_word1.dimshuffle(0, 2, 1),
input_tensor3_r=embed_input_r,
mask_matrix=word1_mask,
mask_matrix_r=sents_mask_r,
image_shape=(batch_size, 1, emb_size, max_term_len),
image_shape_r=(batch_size, 1, emb_size, maxSentLen),
filter_shape=(hidden_size[1], 1, emb_size, filter_size[0]),
filter_shape_context=(hidden_size[1], 1, emb_size, 1),
W=conv_W,
b=conv_b,
W_context=conv_W_context,
b_context=conv_b_context)
td_embeddings_l = term_vs_def_layer.attentive_maxpool_vec_l
td_embeddings_r = term_vs_def_layer.attentive_maxpool_vec_r
p_wd = T.concatenate([
td_embeddings_l, td_embeddings_r, td_embeddings_l * td_embeddings_r,
td_embeddings_l - td_embeddings_r
],
axis=1)
def_vs_term_layer = Conv_for_Pair(
rng,
origin_input_tensor3=embed_input_l,
origin_input_tensor3_r=embed_word2.dimshuffle(0, 2, 1),
input_tensor3=embed_input_l,
input_tensor3_r=embed_word2.dimshuffle(0, 2, 1),
mask_matrix=sents_mask_l,
mask_matrix_r=word2_mask,
image_shape=(batch_size, 1, emb_size, maxSentLen),
image_shape_r=(batch_size, 1, emb_size, max_term_len),
filter_shape=(hidden_size[1], 1, emb_size, filter_size[0]),
filter_shape_context=(hidden_size[1], 1, emb_size, 1),
W=conv_W,
b=conv_b,
W_context=conv_W_context,
b_context=conv_b_context)
dt_embeddings_l = def_vs_term_layer.attentive_maxpool_vec_l
dt_embeddings_r = def_vs_term_layer.attentive_maxpool_vec_r
p_dw = T.concatenate([
dt_embeddings_l, dt_embeddings_r, dt_embeddings_l * dt_embeddings_r,
dt_embeddings_l - dt_embeddings_r
],
axis=1)
def_vs_def_layer = Conv_for_Pair(
rng,
origin_input_tensor3=embed_input_l,
origin_input_tensor3_r=embed_input_r,
input_tensor3=embed_input_l,
input_tensor3_r=embed_input_r,
mask_matrix=sents_mask_l,
mask_matrix_r=sents_mask_r,
image_shape=(batch_size, 1, emb_size, maxSentLen),
image_shape_r=(batch_size, 1, emb_size, maxSentLen),
filter_shape=(hidden_size[1], 1, emb_size, filter_size[0]),
filter_shape_context=(hidden_size[1], 1, emb_size, 1),
W=conv_W,
b=conv_b,
W_context=conv_W_context,
b_context=conv_b_context)
dd_embeddings_l = def_vs_def_layer.attentive_maxpool_vec_l
dd_embeddings_r = def_vs_def_layer.attentive_maxpool_vec_r
p_dd = T.concatenate([
dd_embeddings_l, dd_embeddings_r, dd_embeddings_l * dd_embeddings_r,
dd_embeddings_l - dd_embeddings_r
],
axis=1)
if p_mode == 'conc':
p = T.concatenate([p_ww, p_wd, p_dw, p_dd], axis=1)
p_len = 4 * 4 * hidden_size[1]
else:
p = T.max(T.concatenate([
p_ww.dimshuffle('x', 0, 1),
p_wd.dimshuffle('x', 0, 1),
p_dw.dimshuffle('x', 0, 1),
p_dd.dimshuffle('x', 0, 1)
],
axis=0),
axis=0)
p_len = 4 * hidden_size[1]
# HL_input = T.concatenate([p,cosine_matrix1_matrix2_rowwise(word1_embedding,word2_embedding).dimshuffle(0,'x'),extra.dimshuffle(0,'x')],axis=1)
# HL_input_size=p_len+1+1
#
# HL_layer_1=HiddenLayer(rng, input=HL_input, n_in=HL_input_size, n_out=hidden_size[1], activation=T.tanh)
"form input to LR classifier"
LR_input = T.concatenate([
p,
cosine_matrix1_matrix2_rowwise(word1_embedding,
word2_embedding).dimshuffle(0, 'x'),
extra.dimshuffle(0, 'x')
],
axis=1)
LR_input_size = p_len + 1 + 1
# LR_input = HL_layer_1.output
# LR_input_size = hidden_size[1]
U_a = create_ensemble_para(
rng, 2, LR_input_size) # the weight matrix hidden_size*2
LR_b = theano.shared(value=np.zeros((2, ), dtype=theano.config.floatX),
name='LR_b',
borrow=True) #bias for each target class
LR_para = [U_a, LR_b]
layer_LR = LogisticRegression(
rng,
input=LR_input,
n_in=LR_input_size,
n_out=2,
W=U_a,
b=LR_b,
bias=0.25
) #basically it is a multiplication between weight matrix and input feature vector
loss = layer_LR.negative_log_likelihood(
labels
) #for classification task, we usually used negative log likelihood as loss, the lower the better.
# L2_reg = (conv_W**2).sum()+(conv_W_context**2).sum()+(U_a**2).sum()
params = NN_para + LR_para #[init_embeddings]
cost = loss #+L2_weight*L2_reg
updates = Gradient_Cost_Para(cost, params, learning_rate)
train_model = theano.function([
sents_ids_l, sents_mask_l, sents_ids_r, sents_mask_r, word1_ids,
word2_ids, word1_mask, word2_mask, extra, labels
],
cost,
updates=updates,
allow_input_downcast=True,
on_unused_input='ignore')
test_model = theano.function([
sents_ids_l, sents_mask_l, sents_ids_r, sents_mask_r, word1_ids,
word2_ids, word1_mask, word2_mask, extra, labels
], [layer_LR.errors(labels), layer_LR.y_pred, layer_LR.prop_for_posi],
allow_input_downcast=True,
on_unused_input='ignore')
###############
# TRAIN MODEL #
###############
print '... training'
# early-stopping parameters
patience = 50000000000 # look as this many examples regardless
start_time = time.time()
mid_time = start_time
past_time = mid_time
epoch = 0
done_looping = False
n_train_batches = train_size / batch_size
train_batch_start = list(
np.arange(n_train_batches) * batch_size) + [train_size - batch_size]
# n_dev_batches=dev_size/batch_size
# dev_batch_start=list(np.arange(n_dev_batches)*batch_size)+[dev_size-batch_size]
n_test_batches = test_size / batch_size
n_test_remain = test_size % batch_size
if n_test_remain != 0:
test_batch_start = list(
np.arange(n_test_batches) * batch_size) + [test_size - batch_size]
else:
test_batch_start = list(np.arange(n_test_batches) * batch_size)
# max_acc_dev=0.0
max_ap_test = 0.0
max_ap_topk_test = 0.0
max_f1 = 0.0
cost_i = 0.0
train_indices = range(train_size)
while epoch < n_epochs:
epoch = epoch + 1
random.Random(100).shuffle(
train_indices
) #shuffle training set for each new epoch, is supposed to promote performance, but not garrenteed
iter_accu = 0
for batch_id in train_batch_start: #for each batch
# iter means how many batches have been run, taking into loop
iter = (epoch - 1) * n_train_batches + iter_accu + 1
iter_accu += 1
train_id_batch = train_indices[batch_id:batch_id + batch_size]
cost_i += train_model(
train_sents_l[train_id_batch], train_masks_l[train_id_batch],
train_sents_r[train_id_batch], train_masks_r[train_id_batch],
train_word1[train_id_batch], train_word2[train_id_batch],
train_word1_mask[train_id_batch],
train_word2_mask[train_id_batch], train_extra[train_id_batch],
train_labels_store[train_id_batch])
#after each 1000 batches, we test the performance of the model on all test data
if iter % 100 == 0:
print 'Epoch ', epoch, 'iter ' + str(
iter) + ' average cost: ' + str(cost_i / iter), 'uses ', (
time.time() - past_time) / 60.0, 'min'
past_time = time.time()
pred_labels = []
probs = []
gold_labels = []
error_sum = 0.0
for idd, test_batch_id in enumerate(
test_batch_start): # for each test batch
error_i, pred_i, prob_i = test_model(
test_sents_l[test_batch_id:test_batch_id + batch_size],
test_masks_l[test_batch_id:test_batch_id + batch_size],
test_sents_r[test_batch_id:test_batch_id + batch_size],
test_masks_r[test_batch_id:test_batch_id + batch_size],
test_word1[test_batch_id:test_batch_id + batch_size],
test_word2[test_batch_id:test_batch_id + batch_size],
test_word1_mask[test_batch_id:test_batch_id +
batch_size],
test_word2_mask[test_batch_id:test_batch_id +
batch_size],
test_extra[test_batch_id:test_batch_id + batch_size],
test_labels_store[test_batch_id:test_batch_id +
batch_size])
error_sum += error_i
pred_labels += list(pred_i)
probs += list(prob_i)
if n_test_remain != 0:
probs = probs[:(len(test_batch_start) - 1) *
batch_size] + probs[-n_test_remain:]
assert len(test_labels) == len(probs)
# test_acc=1.0-error_sum/(len(test_batch_start))
test_ap = apk(test_labels, probs, k=len(test_labels))
test_ap_top100 = apk(test_labels, probs, k=100)
if test_ap > max_ap_test:
max_ap_test = test_ap
store_model_to_file(
root_dic + 'LenciBenotto_best_para_' +
str(max_ap_test), params)
if test_ap_top100 > max_ap_topk_test:
max_ap_topk_test = test_ap_top100
print '\t\tcurrent ap:', test_ap, ' ; ', '\t\tmax_ap: ', max_ap_test, 'ap@100: ', test_ap_top100, '\tmax_ap@100:', max_ap_topk_test
print 'Epoch ', epoch, 'uses ', (time.time() - mid_time) / 60.0, 'min'
mid_time = time.time()
#print 'Batch_size: ', update_freq
end_time = time.time()
print >> sys.stderr, ('The code for file ' + os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
def evaluate_lenet5(learning_rate=0.02, n_epochs=4, L2_weight=1e-5, extra_size=4, emb_size=300, batch_size=100, filter_size=[3,3], maxSentLen=40, hidden_size=[300,300], max_term_len=4, p_mode = 'conc'):
model_options = locals().copy()
print "model options", model_options
seed=1234
np.random.seed(seed)
rng = np.random.RandomState(seed) #random seed, control the model generates the same results
# all_sentences_l, all_masks_l, all_sentences_r, all_masks_r, all_word1,all_word2,all_word1_mask,all_word2_mask,all_labels, all_extra, word2id =load_wordnet_hyper_vs_all_with_words(maxlen=maxSentLen, wordlen=max_term_len) #minlen, include one label, at least one word in the sentence
# test_sents_l, test_masks_l, test_sents_r, test_masks_r, test_labels, word2id =load_ACE05_dataset(maxSentLen, word2id)
word2id = load_word2id(root_dic+'LenciBenotto_word2id.pkl')
test_sents_l, test_masks_l, test_sents_r, test_masks_r, test_word1,test_word2,test_word1_mask,test_word2_mask,test_labels, test_extra, word2id, group_size_list = load_task_hyper_vs_all_with_allDefComb(LenciBenotto_file,maxSentLen, word2id, wordlen=max_term_len)
test_sents_l=np.asarray(test_sents_l, dtype='int32')
test_masks_l=np.asarray(test_masks_l, dtype=theano.config.floatX)
test_sents_r=np.asarray(test_sents_r, dtype='int32')
test_masks_r=np.asarray(test_masks_r, dtype=theano.config.floatX)
test_word1=np.asarray(test_word1, dtype='int32')
test_word2=np.asarray(test_word2, dtype='int32')
test_word1_mask=np.asarray(test_word1_mask, dtype=theano.config.floatX)
test_word2_mask=np.asarray(test_word2_mask, dtype=theano.config.floatX)
test_labels_store=np.asarray(test_labels, dtype='int32')
test_extra=np.asarray(test_extra, dtype=theano.config.floatX)
# train_size=len(train_labels_store)
# dev_size=len(dev_labels_store)
test_size=len(test_sents_l)
print ' test size: ', test_size
vocab_size=len(word2id)+1
rand_values=rng.normal(0.0, 0.01, (vocab_size, emb_size)) #generate a matrix by Gaussian distribution
rand_values[0]=np.array(np.zeros(emb_size),dtype=theano.config.floatX)
id2word = {y:x for x,y in word2id.iteritems()}
word2vec=load_word2vec()
rand_values=load_word2vec_to_init(rand_values, id2word, word2vec)
init_embeddings=theano.shared(value=np.array(rand_values,dtype=theano.config.floatX), borrow=True) #wrap up the python variable "rand_values" into theano variable
# load_model_from_file(root_dic+'Weeds_best_para_init_embeddings', [init_embeddings])
#now, start to build the input form of the model
sents_ids_l=T.imatrix()
sents_mask_l=T.fmatrix()
sents_ids_r=T.imatrix()
sents_mask_r=T.fmatrix()
word1_ids = T.imatrix()
word2_ids = T.imatrix()
word1_mask = T.fmatrix()
word2_mask = T.fmatrix()
extra = T.fvector()
labels=T.ivector()
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
def embed_input(emb_matrix, sent_ids):
return emb_matrix[sent_ids.flatten()].reshape((batch_size,maxSentLen, emb_size)).dimshuffle(0,2,1)
embed_input_l=embed_input(init_embeddings, sents_ids_l)#embeddings[sents_ids_l.flatten()].reshape((batch_size,maxSentLen, emb_size)).dimshuffle(0,2,1) #the input format can be adapted into CNN or GRU or LSTM
embed_input_r=embed_input(init_embeddings, sents_ids_r)#embeddings[sents_ids_r.flatten()].reshape((batch_size,maxSentLen, emb_size)).dimshuffle(0,2,1)
embed_word1 = init_embeddings[word1_ids.flatten()].reshape((batch_size,word1_ids.shape[1], emb_size))
embed_word2 = init_embeddings[word2_ids.flatten()].reshape((batch_size,word2_ids.shape[1], emb_size))
word1_embedding = T.sum(embed_word1*word1_mask.dimshuffle(0,1,'x'), axis=1)
word2_embedding = T.sum(embed_word2*word2_mask.dimshuffle(0,1,'x'), axis=1)
'''create_AttentiveConv_params '''
conv_W, conv_b=create_conv_para(rng, filter_shape=(hidden_size[1], 1, emb_size, filter_size[0]))
conv_W_context, conv_b_context=create_conv_para(rng, filter_shape=(hidden_size[1], 1, emb_size, 1))
NN_para=[conv_W, conv_b,conv_W_context]
'''
attentive convolution function
'''
term_vs_term_layer = Conv_for_Pair(rng,
origin_input_tensor3=embed_word1.dimshuffle(0,2,1),
origin_input_tensor3_r = embed_word2.dimshuffle(0,2,1),
input_tensor3=embed_word1.dimshuffle(0,2,1),
input_tensor3_r = embed_word2.dimshuffle(0,2,1),
mask_matrix = word1_mask,
mask_matrix_r = word2_mask,
image_shape=(batch_size, 1, emb_size, max_term_len),
image_shape_r = (batch_size, 1, emb_size, max_term_len),
filter_shape=(hidden_size[1], 1, emb_size, filter_size[0]),
filter_shape_context=(hidden_size[1], 1,emb_size, 1),
W=conv_W, b=conv_b,
W_context=conv_W_context, b_context=conv_b_context)
tt_embeddings_l = term_vs_term_layer.attentive_maxpool_vec_l
tt_embeddings_r = term_vs_term_layer.attentive_maxpool_vec_r
p_ww = T.concatenate([tt_embeddings_l,tt_embeddings_r,tt_embeddings_l*tt_embeddings_r,tt_embeddings_l-tt_embeddings_r], axis=1)
term_vs_def_layer = Conv_for_Pair(rng,
origin_input_tensor3=embed_word1.dimshuffle(0,2,1),
origin_input_tensor3_r = embed_input_r,
input_tensor3=embed_word1.dimshuffle(0,2,1),
input_tensor3_r = embed_input_r,
mask_matrix = word1_mask,
mask_matrix_r = sents_mask_r,
image_shape=(batch_size, 1, emb_size, max_term_len),
image_shape_r = (batch_size, 1, emb_size, maxSentLen),
filter_shape=(hidden_size[1], 1, emb_size, filter_size[0]),
filter_shape_context=(hidden_size[1], 1,emb_size, 1),
W=conv_W, b=conv_b,
W_context=conv_W_context, b_context=conv_b_context)
td_embeddings_l = term_vs_def_layer.attentive_maxpool_vec_l
td_embeddings_r = term_vs_def_layer.attentive_maxpool_vec_r
p_wd = T.concatenate([td_embeddings_l,td_embeddings_r,td_embeddings_l*td_embeddings_r,td_embeddings_l-td_embeddings_r], axis=1)
def_vs_term_layer = Conv_for_Pair(rng,
origin_input_tensor3=embed_input_l,
origin_input_tensor3_r = embed_word2.dimshuffle(0,2,1),
input_tensor3=embed_input_l,
input_tensor3_r = embed_word2.dimshuffle(0,2,1),
mask_matrix = sents_mask_l,
mask_matrix_r = word2_mask,
image_shape=(batch_size, 1, emb_size, maxSentLen),
image_shape_r = (batch_size, 1, emb_size, max_term_len),
filter_shape=(hidden_size[1], 1, emb_size, filter_size[0]),
filter_shape_context=(hidden_size[1], 1,emb_size, 1),
W=conv_W, b=conv_b,
W_context=conv_W_context, b_context=conv_b_context)
dt_embeddings_l = def_vs_term_layer.attentive_maxpool_vec_l
dt_embeddings_r = def_vs_term_layer.attentive_maxpool_vec_r
p_dw = T.concatenate([dt_embeddings_l,dt_embeddings_r,dt_embeddings_l*dt_embeddings_r,dt_embeddings_l-dt_embeddings_r], axis=1)
def_vs_def_layer = Conv_for_Pair(rng,
origin_input_tensor3=embed_input_l,
origin_input_tensor3_r = embed_input_r,
input_tensor3=embed_input_l,
input_tensor3_r = embed_input_r,
mask_matrix = sents_mask_l,
mask_matrix_r = sents_mask_r,
image_shape=(batch_size, 1, emb_size, maxSentLen),
image_shape_r = (batch_size, 1, emb_size, maxSentLen),
filter_shape=(hidden_size[1], 1, emb_size, filter_size[0]),
filter_shape_context=(hidden_size[1], 1,emb_size, 1),
W=conv_W, b=conv_b,
W_context=conv_W_context, b_context=conv_b_context)
dd_embeddings_l = def_vs_def_layer.attentive_maxpool_vec_l
dd_embeddings_r = def_vs_def_layer.attentive_maxpool_vec_r
p_dd = T.concatenate([dd_embeddings_l,dd_embeddings_r,dd_embeddings_l*dd_embeddings_r,dd_embeddings_l-dd_embeddings_r], axis=1)
if p_mode == 'conc':
p=T.concatenate([p_ww, p_wd, p_dw, p_dd], axis=1)
p_len = 4*4*hidden_size[1]
else:
p = T.max(T.concatenate([p_ww.dimshuffle('x',0,1),p_wd.dimshuffle('x',0,1),p_dw.dimshuffle('x',0,1),p_dd.dimshuffle('x',0,1)],axis=0), axis=0)
p_len =4*hidden_size[1]
# HL_input = T.concatenate([p,cosine_matrix1_matrix2_rowwise(word1_embedding,word2_embedding).dimshuffle(0,'x'),extra.dimshuffle(0,'x')],axis=1)
# HL_input_size=p_len+1+1
#
# HL_layer_1=HiddenLayer(rng, input=HL_input, n_in=HL_input_size, n_out=hidden_size[1], activation=T.tanh)
"form input to LR classifier"
LR_input = T.concatenate([p,cosine_matrix1_matrix2_rowwise(word1_embedding,word2_embedding).dimshuffle(0,'x'),extra.dimshuffle(0,'x')],axis=1)
LR_input_size=p_len+1+1
# LR_input = HL_layer_1.output
# LR_input_size = hidden_size[1]
U_a = create_ensemble_para(rng, 2, LR_input_size) # the weight matrix hidden_size*2
LR_b = theano.shared(value=np.zeros((2,),dtype=theano.config.floatX),name='LR_b', borrow=True) #bias for each target class
LR_para=[U_a, LR_b]
layer_LR=LogisticRegression(rng, input=LR_input, n_in=LR_input_size, n_out=2, W=U_a, b=LR_b, bias=0.25) #basically it is a multiplication between weight matrix and input feature vector
loss=layer_LR.negative_log_likelihood(labels) #for classification task, we usually used negative log likelihood as loss, the lower the better.
# L2_reg = (conv_W**2).sum()+(conv_W_context**2).sum()+(U_a**2).sum()
params = NN_para+LR_para #[init_embeddings]
# load_model_from_file('/save/wenpeng/datasets/HypeNet/HyperDef_label_meta_best_para_0.938730853392', params)
load_model_from_file(root_dic+'LenciBenotto_best_para_0.557286573332', params)
'''
0.552587544259; current ap: 0.574037513126 ap@100 0.918481316424
0.557286573332; current ap: 0.576498645289 ap@100 0.909032657538
'''
test_model = theano.function([sents_ids_l, sents_mask_l, sents_ids_r, sents_mask_r, word1_ids,word2_ids,word1_mask,word2_mask,extra], [layer_LR.y_pred,layer_LR.prop_for_posi], allow_input_downcast=True, on_unused_input='ignore')
###############
# TRAIN MODEL #
###############
print '... training'
# early-stopping parameters
patience = 50000000000 # look as this many examples regardless
start_time = time.time()
mid_time = start_time
past_time= mid_time
epoch = 0
done_looping = False
n_test_batches=test_size/batch_size
n_test_remain = test_size%batch_size
if n_test_remain!=0:
test_batch_start=list(np.arange(n_test_batches)*batch_size)+[test_size-batch_size]
else:
test_batch_start=list(np.arange(n_test_batches)*batch_size)
# max_acc_dev=0.0
max_ap_test=0.0
max_ap_topk_test=0.0
max_f1=0.0
pred_labels =[]
probs = []
gold_labels =[]
error_sum=0.0
for idd, test_batch_id in enumerate(test_batch_start): # for each test batch
pred_i, prob_i=test_model(
test_sents_l[test_batch_id:test_batch_id+batch_size],
test_masks_l[test_batch_id:test_batch_id+batch_size],
test_sents_r[test_batch_id:test_batch_id+batch_size],
test_masks_r[test_batch_id:test_batch_id+batch_size],
test_word1[test_batch_id:test_batch_id+batch_size],
test_word2[test_batch_id:test_batch_id+batch_size],
test_word1_mask[test_batch_id:test_batch_id+batch_size],
test_word2_mask[test_batch_id:test_batch_id+batch_size],
test_extra[test_batch_id:test_batch_id+batch_size])
# error_sum+=error_i
pred_labels+=list(pred_i)
probs+=list(prob_i)
print len(test_sents_l), len(probs)
if n_test_remain !=0:
probs = probs[:(len(test_batch_start)-1)*batch_size]+probs[-n_test_remain:]
print len(test_sents_l), len(probs)
assert len(test_sents_l) == len(probs)
assert sum(group_size_list) == len(probs)
#max prob in group
max_probs = []
prior_size = 0
for i in range(len(group_size_list)):
sub_probs = probs[prior_size:prior_size+group_size_list[i]]
prior_size += group_size_list[i]
max_probs.append(max(sub_probs))
print len(group_size_list),len(max_probs),len(test_labels)
assert len(test_labels) == len(max_probs)
# test_acc=1.0-error_sum/(len(test_batch_start))
test_ap = apk(test_labels, max_probs, k=len(test_labels))
test_ap_top100 = apk(test_labels, max_probs, k=100)
# if test_ap > max_ap_test:
# max_ap_test=test_ap
# store_model_to_file('/save/wenpeng/datasets/EVALution/HyperDef_label_4ways_conc_test_on_EVA_allDefComb_best_para_'+str(max_ap_test), params)
# if test_ap_top100 > max_ap_topk_test:
# max_ap_topk_test=test_ap_top100
print '\t\tcurrent ap:', test_ap,'ap@100', test_ap_top100
def evaluate_lenet5(learning_rate=0.1, n_epochs=4, L2_weight=0.001, emb_size=70, batch_size=50, filter_size=3, maxSentLen=50, nn='CNN'):
hidden_size=emb_size
model_options = locals().copy()
print "model options", model_options
rng = np.random.RandomState(1234) #random seed, control the model generates the same results
all_sentences_l, all_masks_l, all_sentences_r, all_masks_r,all_labels, word2id =load_SNLI_dataset(maxlen=maxSentLen) #minlen, include one label, at least one word in the sentence
train_sents_l=np.asarray(all_sentences_l[0], dtype='int32')
dev_sents_l=np.asarray(all_sentences_l[1], dtype='int32')
test_sents_l=np.asarray(all_sentences_l[2], dtype='int32')
train_masks_l=np.asarray(all_masks_l[0], dtype=theano.config.floatX)
dev_masks_l=np.asarray(all_masks_l[1], dtype=theano.config.floatX)
test_masks_l=np.asarray(all_masks_l[2], dtype=theano.config.floatX)
train_sents_r=np.asarray(all_sentences_r[0], dtype='int32')
dev_sents_r=np.asarray(all_sentences_r[1] , dtype='int32')
test_sents_r=np.asarray(all_sentences_r[2] , dtype='int32')
train_masks_r=np.asarray(all_masks_r[0], dtype=theano.config.floatX)
dev_masks_r=np.asarray(all_masks_r[1], dtype=theano.config.floatX)
test_masks_r=np.asarray(all_masks_r[2], dtype=theano.config.floatX)
train_labels_store=np.asarray(all_labels[0], dtype='int32')
dev_labels_store=np.asarray(all_labels[1], dtype='int32')
test_labels_store=np.asarray(all_labels[2], dtype='int32')
train_size=len(train_labels_store)
dev_size=len(dev_labels_store)
test_size=len(test_labels_store)
vocab_size=len(word2id)+1
rand_values=rng.normal(0.0, 0.01, (vocab_size, emb_size)) #generate a matrix by Gaussian distribution
#here, we leave code for loading word2vec to initialize words
# rand_values[0]=np.array(np.zeros(emb_size),dtype=theano.config.floatX)
# id2word = {y:x for x,y in word2id.iteritems()}
# word2vec=load_word2vec()
# rand_values=load_word2vec_to_init(rand_values, id2word, word2vec)
embeddings=theano.shared(value=np.array(rand_values,dtype=theano.config.floatX), borrow=True) #wrap up the python variable "rand_values" into theano variable
#now, start to build the input form of the model
sents_ids_l=T.imatrix()
sents_mask_l=T.fmatrix()
sents_ids_r=T.imatrix()
sents_mask_r=T.fmatrix()
labels=T.ivector()
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
common_input_l=embeddings[sents_ids_l.flatten()].reshape((batch_size,maxSentLen, emb_size)) #the input format can be adapted into CNN or GRU or LSTM
common_input_r=embeddings[sents_ids_r.flatten()].reshape((batch_size,maxSentLen, emb_size))
#conv
if nn=='CNN':
conv_W, conv_b=create_conv_para(rng, filter_shape=(hidden_size, 1, emb_size, filter_size))
conv_W_into_matrix=conv_W.reshape((conv_W.shape[0], conv_W.shape[2]*conv_W.shape[3]))
NN_para=[conv_W, conv_b]
conv_input_l = common_input_l.dimshuffle((0,'x', 2,1)) #(batch_size, 1, emb_size, maxsenlen)
conv_model_l = Conv_with_input_para(rng, input=conv_input_l,
image_shape=(batch_size, 1, emb_size, maxSentLen),
filter_shape=(hidden_size, 1, emb_size, filter_size), W=conv_W, b=conv_b)
conv_output_l=conv_model_l.narrow_conv_out #(batch, 1, hidden_size, maxsenlen-filter_size+1)
conv_output_into_tensor3_l=conv_output_l.reshape((batch_size, hidden_size, maxSentLen-filter_size+1))
mask_for_conv_output_l=T.repeat(sents_mask_l[:,filter_size-1:].reshape((batch_size, 1, maxSentLen-filter_size+1)), hidden_size, axis=1) #(batch_size, emb_size, maxSentLen-filter_size+1)
mask_for_conv_output_l=(1.0-mask_for_conv_output_l)*(mask_for_conv_output_l-10)
masked_conv_output_l=conv_output_into_tensor3_l+mask_for_conv_output_l #mutiple mask with the conv_out to set the features by UNK to zero
sent_embeddings_l=T.max(masked_conv_output_l, axis=2) #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size
conv_input_r = common_input_r.dimshuffle((0,'x', 2,1)) #(batch_size, 1, emb_size, maxsenlen)
conv_model_r = Conv_with_input_para(rng, input=conv_input_r,
image_shape=(batch_size, 1, emb_size, maxSentLen),
filter_shape=(hidden_size, 1, emb_size, filter_size), W=conv_W, b=conv_b)
conv_output_r=conv_model_r.narrow_conv_out #(batch, 1, hidden_size, maxsenlen-filter_size+1)
conv_output_into_tensor3_r=conv_output_r.reshape((batch_size, hidden_size, maxSentLen-filter_size+1))
mask_for_conv_output_r=T.repeat(sents_mask_r[:,filter_size-1:].reshape((batch_size, 1, maxSentLen-filter_size+1)), hidden_size, axis=1) #(batch_size, emb_size, maxSentLen-filter_size+1)
mask_for_conv_output_r=(1.0-mask_for_conv_output_r)*(mask_for_conv_output_r-10)
masked_conv_output_r=conv_output_into_tensor3_r+mask_for_conv_output_r #mutiple mask with the conv_out to set the features by UNK to zero
sent_embeddings_r=T.max(masked_conv_output_r, axis=2) #(batch_size, hidden_size) # each sentence then have an embedding of length hidden_size
#GRU
if nn=='GRU':
U1, W1, b1=create_GRU_para(rng, emb_size, hidden_size)
NN_para=[U1, W1, b1] #U1 includes 3 matrices, W1 also includes 3 matrices b1 is bias
gru_input_l = common_input_l.dimshuffle((0,2,1)) #gru requires input (batch_size, emb_size, maxSentLen)
gru_layer_l=GRU_Batch_Tensor_Input_with_Mask(gru_input_l, sents_mask_l, hidden_size, U1, W1, b1)
sent_embeddings_l=gru_layer_l.output_sent_rep # (batch_size, hidden_size)
gru_input_r = common_input_r.dimshuffle((0,2,1)) #gru requires input (batch_size, emb_size, maxSentLen)
gru_layer_r=GRU_Batch_Tensor_Input_with_Mask(gru_input_r, sents_mask_r, hidden_size, U1, W1, b1)
sent_embeddings_r=gru_layer_r.output_sent_rep # (batch_size, hidden_size)
#LSTM
if nn=='LSTM':
LSTM_para_dict=create_LSTM_para(rng, emb_size, hidden_size)
NN_para=LSTM_para_dict.values() # .values returns a list of parameters
lstm_input_l = common_input_l.dimshuffle((0,2,1)) #LSTM has the same inpur format with GRU
lstm_layer_l=LSTM_Batch_Tensor_Input_with_Mask(lstm_input_l, sents_mask_l, hidden_size, LSTM_para_dict)
sent_embeddings_l=lstm_layer_l.output_sent_rep # (batch_size, hidden_size)
lstm_input_r = common_input_r.dimshuffle((0,2,1)) #LSTM has the same inpur format with GRU
lstm_layer_r=LSTM_Batch_Tensor_Input_with_Mask(lstm_input_r, sents_mask_r, hidden_size, LSTM_para_dict)
sent_embeddings_r=lstm_layer_r.output_sent_rep # (batch_size, hidden_size)
HL_layer_1_input = T.concatenate([sent_embeddings_l,sent_embeddings_r, sent_embeddings_l*sent_embeddings_r, cosine_matrix1_matrix2_rowwise(sent_embeddings_l,sent_embeddings_r).dimshuffle(0,'x')],axis=1)
HL_layer_1_input_size = hidden_size*3+1
HL_layer_1=HiddenLayer(rng, input=HL_layer_1_input, n_in=HL_layer_1_input_size, n_out=hidden_size, activation=T.tanh)
HL_layer_2=HiddenLayer(rng, input=HL_layer_1.output, n_in=hidden_size, n_out=hidden_size, activation=T.tanh)
#classification layer, it is just mapping from a feature vector of size "hidden_size" to a vector of only two values: positive, negative
LR_input_size=HL_layer_1_input_size+2*hidden_size
U_a = create_ensemble_para(rng, 3, LR_input_size) # the weight matrix hidden_size*2
LR_b = theano.shared(value=np.zeros((3,),dtype=theano.config.floatX),name='LR_b', borrow=True) #bias for each target class
LR_para=[U_a, LR_b]
LR_input=T.concatenate([HL_layer_1_input, HL_layer_1.output, HL_layer_2.output],axis=1)
layer_LR=LogisticRegression(rng, input=T.tanh(LR_input), n_in=LR_input_size, n_out=3, W=U_a, b=LR_b) #basically it is a multiplication between weight matrix and input feature vector
loss=layer_LR.negative_log_likelihood(labels) #for classification task, we usually used negative log likelihood as loss, the lower the better.
params = [embeddings]+NN_para+LR_para+HL_layer_1.params+HL_layer_2.params # put all model parameters together
# L2_reg =L2norm_paraList([embeddings,conv_W, U_a])
# diversify_reg= Diversify_Reg(U_a.T)+Diversify_Reg(conv_W_into_matrix)
cost=loss#+Div_reg*diversify_reg#+L2_weight*L2_reg
grads = T.grad(cost, params) # create a list of gradients for all model parameters
accumulator=[]
for para_i in params:
eps_p=np.zeros_like(para_i.get_value(borrow=True),dtype=theano.config.floatX)
accumulator.append(theano.shared(eps_p, borrow=True))
updates = []
for param_i, grad_i, acc_i in zip(params, grads, accumulator):
acc = acc_i + T.sqr(grad_i)
updates.append((param_i, param_i - learning_rate * grad_i / (T.sqrt(acc)+1e-8))) #1e-8 is add to get rid of zero division
updates.append((acc_i, acc))
#train_model = theano.function([sents_id_matrix, sents_mask, labels], cost, updates=updates, on_unused_input='ignore')
train_model = theano.function([sents_ids_l, sents_mask_l, sents_ids_r, sents_mask_r, labels], cost, updates=updates, allow_input_downcast=True, on_unused_input='ignore')
dev_model = theano.function([sents_ids_l, sents_mask_l, sents_ids_r, sents_mask_r, labels], layer_LR.errors(labels), allow_input_downcast=True, on_unused_input='ignore')
test_model = theano.function([sents_ids_l, sents_mask_l, sents_ids_r, sents_mask_r, labels], layer_LR.errors(labels), allow_input_downcast=True, on_unused_input='ignore')
###############
# TRAIN MODEL #
###############
print '... training'
# early-stopping parameters
patience = 50000000000 # look as this many examples regardless
start_time = time.time()
mid_time = start_time
past_time= mid_time
epoch = 0
done_looping = False
n_train_batches=train_size/batch_size
train_batch_start=list(np.arange(n_train_batches)*batch_size)+[train_size-batch_size]
n_dev_batches=dev_size/batch_size
dev_batch_start=list(np.arange(n_dev_batches)*batch_size)+[dev_size-batch_size]
n_test_batches=test_size/batch_size
test_batch_start=list(np.arange(n_test_batches)*batch_size)+[test_size-batch_size]
max_acc_dev=0.0
max_acc_test=0.0
while epoch < n_epochs:
epoch = epoch + 1
train_indices = range(train_size)
random.Random(200).shuffle(train_indices) #shuffle training set for each new epoch, is supposed to promote performance, but not garrenteed
iter_accu=0
cost_i=0.0
for batch_id in train_batch_start: #for each batch
# iter means how many batches have been run, taking into loop
iter = (epoch - 1) * n_train_batches + iter_accu +1
iter_accu+=1
train_id_batch = train_indices[batch_id:batch_id+batch_size]
cost_i+= train_model(
train_sents_l[train_id_batch],
train_masks_l[train_id_batch],
train_sents_r[train_id_batch],
train_masks_r[train_id_batch],
train_labels_store[train_id_batch])
#after each 1000 batches, we test the performance of the model on all test data
if iter%500==0:
print 'Epoch ', epoch, 'iter '+str(iter)+' average cost: '+str(cost_i/iter), 'uses ', (time.time()-past_time)/60.0, 'min'
past_time = time.time()
# if epoch >=3 and iter >= len(train_batch_start)*2.0/3 and iter%500==0:
# print 'Epoch ', epoch, 'iter '+str(iter)+' average cost: '+str(cost_i/iter), 'uses ', (time.time()-past_time)/60.0, 'min'
# past_time = time.time()
error_sum=0.0
for dev_batch_id in dev_batch_start: # for each test batch
error_i=dev_model(
dev_sents_l[dev_batch_id:dev_batch_id+batch_size],
dev_masks_l[dev_batch_id:dev_batch_id+batch_size],
dev_sents_r[dev_batch_id:dev_batch_id+batch_size],
dev_masks_r[dev_batch_id:dev_batch_id+batch_size],
dev_labels_store[dev_batch_id:dev_batch_id+batch_size]
)
error_sum+=error_i
dev_accuracy=1.0-error_sum/(len(dev_batch_start))
if dev_accuracy > max_acc_dev:
max_acc_dev=dev_accuracy
print 'current dev_accuracy:', dev_accuracy, '\t\t\t\t\tmax max_acc_dev:', max_acc_dev
#best dev model, do test
error_sum=0.0
for test_batch_id in test_batch_start: # for each test batch
error_i=test_model(
test_sents_l[test_batch_id:test_batch_id+batch_size],
test_masks_l[test_batch_id:test_batch_id+batch_size],
test_sents_r[test_batch_id:test_batch_id+batch_size],
test_masks_r[test_batch_id:test_batch_id+batch_size],
test_labels_store[test_batch_id:test_batch_id+batch_size]
)
error_sum+=error_i
test_accuracy=1.0-error_sum/(len(test_batch_start))
if test_accuracy > max_acc_test:
max_acc_test=test_accuracy
print '\t\tcurrent testbacc:', test_accuracy, '\t\t\t\t\tmax_acc_test:', max_acc_test
else:
print 'current dev_accuracy:', dev_accuracy, '\t\t\t\t\tmax max_acc_dev:', max_acc_dev
print 'Epoch ', epoch, 'uses ', (time.time()-mid_time)/60.0, 'min'
mid_time = time.time()
#print 'Batch_size: ', update_freq
end_time = time.time()
print >> sys.stderr, ('The code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
return max_acc_test
