def train_item(args, model, sentence):
loss = None
seq = [
model.wlookup[int(model.w2i.get(entry, 0))]
for entry in sentence.preprocessed_sentence
]
if len(seq) > 0:
encoded_sequence = encode_sequence(model, seq, model.sentence_rnn)
last_output = encoded_sequence[-1]
global_max = max_pooling(encoded_sequence)
global_min = average_pooling(encoded_sequence)
context = dy.concatenate([last_output, global_max, global_min])
y_pred = dy.logistic((model.mlp_w * context) + model.mlp_b)
if sentence.permissions[args.permission_type]:
loss = dy.binary_log_loss(y_pred, dy.scalarInput(1))
else:
loss = dy.binary_log_loss(y_pred, dy.scalarInput(0))
loss.backward()
model.trainer.update()
loss_val = loss.scalar_value()
dy.renew_cg()
return loss_val
return 0
def scorer(self, q_d_hists, q_idf, bm25_score, overlap_features, p):
"""
Makes all the calculations and returns a relevance score
"""
idf_vec = dy.inputVector(q_idf)
bm25_score = dy.scalarInput(bm25_score)
overlap_features = dy.inputVector(overlap_features)
# Pass each query term representation through the MLP
term_scores = []
for hist in q_d_hists:
q_d_hist = dy.reshape(dy.inputVector(hist), (1, len(hist)))
hidd_out = dy.rectify(q_d_hist * self.W_1 + self.b_1)
for i in range(0, self.mlp_layers):
hidd_out = dy.rectify(hidd_out * self.W_n[i] + self.b_n[i])
term_scores.append(hidd_out * self.W_last + self.b_last)
# Term Gating
gating_weights = idf_vec * self.w_g
bm25_feature = bm25_score * self.W_bm25 + self.b_bm25
drop_out = dy.scalarInput(1)
drop_num = (np.random.rand(1) < p)/p #p= probability of keeping a unit active
drop_out.set(drop_num)
bm25_feature *= drop_out
drmm_score = dy.transpose(dy.concatenate(term_scores)) * dy.reshape(gating_weights, (len(q_idf), 1)) #basic MLPs output
doc_score = dy.transpose(dy.concatenate([drmm_score, overlap_features])) * self.W_scores + self.b_scores #extra features layer
return doc_score
def __train(model, data):
tagged_loss = 0
untagged_loss = 0
for index, sentence_report in enumerate(data):
for phrase in sentence_report.all_phrases:
loss = None
encoded_phrase = __encode_sequence(model, phrase)
if model.options.external_info != "no_info":
encoded_phrase = dy.concatenate(
[encoded_phrase, model.doclookup[sentence_report.app_id]])
y_pred = dy.logistic((model.mlp_w * encoded_phrase) + model.mlp_b)
if sentence_report.mark:
loss = dy.binary_log_loss(y_pred, dy.scalarInput(1))
else:
loss = dy.binary_log_loss(y_pred, dy.scalarInput(0))
if sentence_report.mark:
tagged_loss += loss.scalar_value() / (index + 1)
else:
untagged_loss += loss.scalar_value() / (index + 1)
loss.backward()
model.trainer.update()
dy.renew_cg()
def perceptron_loss(scores, reference):
if use_cost_augmented:
predictions = hamming_augmented_decode(scores, reference)
else:
predictions = [np.argmax(score.npvalue()) for score in scores]
margin = dy.scalarInput(-2)
if predictions != reference:
reference_score = calc_sequence_score(scores, reference)
prediction_score = calc_sequence_score(scores, predictions)
if use_cost_augmented:
# One could actually get the hamming augmented value during decoding, but we didn't do it here for
# demonstration purpose.
hamming = dy.scalarInput(hamming_cost(predictions, reference))
loss = prediction_score + hamming - reference_score
else:
loss = prediction_score - reference_score
if use_hinge:
loss = dy.emax([dy.scalarInput(0), loss - margin])
return loss
else:
return dy.scalarInput(0)
def perceptron_loss(scores, reference):
if use_cost_augmented:
predictions = hamming_augmented_decode(scores, reference)
else:
predictions = [np.argmax(score.npvalue()) for score in scores]
margin = dy.scalarInput(-2)
if predictions != reference:
reference_score = calc_sequence_score(scores, reference)
prediction_score = calc_sequence_score(scores, predictions)
if use_cost_augmented:
# One could actually get the hamming augmented value during decoding, but we didn't do it here for
# demonstration purpose.
hamming = dy.scalarInput(hamming_cost(predictions, reference))
loss = prediction_score + hamming - reference_score
else:
loss = prediction_score - reference_score
if use_hinge:
loss = dy.emax([dy.scalarInput(0), loss - margin])
return loss
else:
return dy.scalarInput(0)
def __train(self, data):
def encode_sequence(seq):
rnn_forward = self.phrase_rnn[0].initial_state()
for entry in seq:
vec = self.wlookup[int(self.w2i.get(entry, 0))]
rnn_forward = rnn_forward.add_input(vec)
return rnn_forward.output()
tagged_loss = 0
untagged_loss = 0
for index, sentence_report in enumerate(data):
for phrase in sentence_report.all_phrases:
loss = None
encoded_phrase = encode_sequence(phrase)
y_pred = dy.logistic((self.mlp_w*encoded_phrase) + self.mlp_b)
if sentence_report.mark:
loss = dy.binary_log_loss(y_pred, dy.scalarInput(1))
else:
loss = dy.binary_log_loss(y_pred, dy.scalarInput(0))
if index % 1000 == 0:
print("Description : {}".format(index+1))
print("Marked {} Prediction Result {} : ".format(sentence_report.mark, y_pred.scalar_value()))
print("Tagged loss {} Untagged Loss {} Total loss {}".format(tagged_loss, untagged_loss, tagged_loss+untagged_loss))
if sentence_report.mark:
tagged_loss += loss.scalar_value()/(index+1)
else:
untagged_loss += loss.scalar_value()/(index+1)
loss.backward()
self.trainer.update()
dy.renew_cg()
def learn(self, characters, target_mgc, guided_att=True):
num_mgc = target_mgc.shape[0]
# print num_mgc
dy.renew_cg()
output_mgc, output_stop, output_attention = self._predict(
characters, target_mgc)
losses = []
index = 0
for mgc, real_mgc in zip(output_mgc, target_mgc):
t_mgc = dy.inputVector(real_mgc)
# losses.append(self._compute_binary_divergence(mgc, t_mgc) )
losses.append(dy.l1_distance(mgc, t_mgc))
if index % 3 == 0:
# attention loss
if guided_att:
att = output_attention[index / 3]
losses.append(
self._compute_guided_attention(att, index / 3,
len(characters) + 2,
num_mgc / 3))
# EOS loss
stop = output_stop[index / 3]
if index >= num_mgc - 6:
losses.append(dy.l1_distance(stop, dy.scalarInput(-0.8)))
else:
losses.append(dy.l1_distance(stop, dy.scalarInput(0.8)))
index += 1
loss = dy.esum(losses)
loss_val = loss.value() / num_mgc
loss.backward()
self.trainer.update()
return loss_val
def beam_search(self, char_seq, truth = None, mu =0.):
start_agenda = Agenda(self.options['beam_size'])
init_state = self.params['lstm'].initial_state().add_input(self.param_exprs['<bos>'])
init_y = dy.tanh(self.param_exprs['pW'] * init_state.output() + self.param_exprs['pb'])
init_score = dy.scalarInput(0.)
start_agenda.push(Sentence(score=init_score.scalar_value(),score_expr=init_score,LSTMState =init_state, y= init_y , prevState = None, wlen=None))
agenda = [start_agenda]
for idx, _ in enumerate(char_seq,1): # from left to right, character by character
now = Agenda(self.options['beam_size'])
for wlen in xrange(1,min(idx,self.options['max_word_len'])+1): # generate candidate word vectors
word = self.word_repr(char_seq[idx-wlen:idx])
word_score = dy.dot_product(word,self.param_exprs['U'])
for sent in agenda[idx-wlen]: # join segmentation
if truth is not None:
margin = dy.scalarInput(mu*wlen if truth[idx-1]!=wlen else 0.)
score = margin + sent.score_expr + dy.dot_product(sent.y, word) + word_score
else:
score = sent.score_expr + dy.dot_product(sent.y, word) + word_score
if now.happy_with(score.scalar_value()):
new_state = sent.LSTMState.add_input(word)
new_y = dy.tanh(self.param_exprs['pW'] * new_state.output() + self.param_exprs['pb'])
now.push(Sentence(score=score.scalar_value(),score_expr=score,LSTMState=new_state,y=new_y, prevState=sent, wlen=wlen))
agenda.append(now)
if truth is not None:
return agenda[-1].max().score_expr
return agenda
def beam_train_max_margin_with_answer_guidence(self, init_state, gold_ans):
# perform two beam search; one for prediction and the other for state action suff
# max reward y = argmax(r(y)) with the help of gold_ans
# max y' = argmax f(x,y) - R(y')
# loss = max(f(x,y') - f(x,y) + R(y) - R(y') , 0)
#end_state_list = self.beam_predict(init_state)
end_state_list = self.beam_predict_max_violation(
init_state, gold_ans) # have to use this to make it work....
reward_list = [x.reward(gold_ans) for x in end_state_list]
violation_list = [
s.path_score_expression.value() - reward
for s, reward in zip(end_state_list, reward_list)
]
best_score_state_idx = violation_list.index(max(
violation_list)) # find the best scoring seq with minimal reward
best_score_state = end_state_list[best_score_state_idx]
best_score_state_reward = reward_list[best_score_state_idx]
loss_value = 0
if self.only_one_best:
best_states = self.beam_find_actions_with_answer_guidence(
init_state, gold_ans)
if best_states == []:
return 0, []
best_reward_state = best_states[0]
#print ("debug: found best_reward_state: qid =", best_reward_state.qinfo.seq_qid, best_reward_state)
best_reward_state_reward = best_reward_state.reward(gold_ans)
#print ("debug: best_reward_state_reward =", best_reward_state_reward)
loss = dt.rectify(best_score_state.path_score_expression -
best_reward_state.path_score_expression +
dt.scalarInput(best_reward_state_reward -
best_score_state_reward))
else:
best_states = self.beam_find_actions_with_answer_guidence(
init_state, gold_ans)
best_states_rewards = [s.reward(gold_ans) for s in best_states]
max_reward = max(best_states_rewards)
best_states = [
s for s, r in zip(best_states, best_states_rewards)
if r == max_reward
]
loss = dt.average([
dt.rectify(best_score_state.path_score_expression -
best_reward_state.path_score_expression +
dt.scalarInput(max_reward -
best_score_state_reward))
for best_reward_state in best_states
])
loss_value = loss.value()
loss.backward()
self.neural_model.learner.update()
#print ("debug: beam_train_max_margin_with_answer_guidence done. loss_value =", loss_value)
return loss_value, best_states
def training_session(self, sentence, print_logger, pool):
lstm_output = self.network.get_lstm_output(sentence)
length = len(sentence)
raw_exprs = self.network.edge_eval.get_complete_raw_exprs(lstm_output)
yield raw_exprs
scores = self.network.edge_eval.raw_exprs_to_scores(raw_exprs, length)
exprs = self.network.edge_eval.raw_exprs_to_exprs(raw_exprs, length)
gold = [entry.parent_id for entry in sentence]
heads_future = pool.apply_async(self.decoder,
(scores, gold if self.options.cost_augment else None))
yield None
heads = heads_future.get()
if self.labelsFlag:
edges = [(head, sentence[modifier].relation, modifier)
for modifier, head in enumerate(gold[1:], 1)]
label_exprs = list(self.network.get_label_scores(lstm_output, edges))
yield label_exprs
label_loss = self.label_decoder(edges, label_exprs, self.statistics.labels, True)
else:
label_loss = dn.scalarInput(0.0)
yield []
head_exprs = [(exprs[h][i] - exprs[g][i] + 1)
for i, (h, g) in enumerate(zip(heads, gold)) if
h != g]
print_logger.correct_edge += len(sentence) - len(head_exprs)
print_logger.total_edge += len(sentence)
head_loss = dn.esum(head_exprs) if head_exprs else dn.scalarInput(0.0)
yield label_loss + head_loss
def train(self, indices, gold_arcs, gold_labels, pos_indices=None):
total_arc_loss = 0
total_label_loss = 0
start = time.time()
for i in range(len(indices)):
states = self.states(
indices[i],
pos_indices[i] if pos_indices is not None else None)
arc_scores = self.score_arcs(states, value=False)
label_scores = self.score_labels(states, gold_arcs[i], value=False)
arc_loss = self.arc_loss(gold_arcs[i], arc_scores)
label_loss = self.label_loss(gold_labels[i], label_scores)
if len(arc_loss) > 0:
arc_loss = dynet.esum(arc_loss)
else:
arc_loss = dynet.scalarInput(0)
if len(label_loss) > 0:
label_loss = dynet.esum(label_loss)
else:
label_loss = dynet.scalarInput(0)
loss = dynet.esum([arc_loss, label_loss])
arc_loss = arc_loss.value()
label_loss = label_loss.value()
total_arc_loss += arc_loss
total_label_loss += label_loss
loss.backward()
self.trainer.update()
dynet.renew_cg()
print(time.time() - start)
return total_arc_loss, total_label_loss
def copy_src_probs_pick(token_type, token_literal):
if token_type not in copy_atts:
return dy.scalarInput(0.0)
selected_indexes = copy_history[token_type][token_literal]
if len(selected_indexes) == 0:
return dy.scalarInput(0.0)
probs = copy_src_probs(token_type)
return dy.sum_elems(dy.select_rows(probs, selected_indexes))
def decomp_attend(self, vecsA, vecsB):
# Fq^T Fc -> need to expedite using native matrix/tensor multiplication
Fq = vecsA # the original word vector, not yet passing a NN as in Eq.1, # need a function F
Fc = vecsB # need a function F
expE = []
for fq in Fq:
row = []
for fc in Fc:
row.append(dt.exp(dt.dot_product(fq, fc)))
expE.append(row)
#print ("debug: expE", expE[0][0].value())
invSumExpEi = []
for i in xrange(len(Fq)):
invSumExpEi.append(dt.pow(dt.esum(expE[i]), dt.scalarInput(-1)))
invSumExpEj = []
for j in xrange(len(Fc)):
invSumExpEj.append(
dt.pow(dt.esum([expE[i][j] for i in xrange(len(Fq))]),
dt.scalarInput(-1)))
beta = []
for i in xrange(len(Fq)):
s = dt.esum([Fc[j] * expE[i][j] for j in xrange(len(Fc))])
beta.append(s * invSumExpEi[i])
#print("debug: beta", beta[0].value())
alpha = []
for j in xrange(len(Fc)):
s = dt.esum([Fc[j] * expE[i][j] for i in xrange(len(Fq))])
alpha.append(s * invSumExpEj[j])
#print("debug: alpha", alpha[0].value())
# Compare
v1i = [
dt.logistic(dt.concatenate([Fq[i], beta[i]]))
for i in xrange(len(Fq))
] # need a function G
v2j = [
dt.logistic(dt.concatenate([Fc[j], alpha[j]]))
for j in xrange(len(Fc))
] # need a function G
#print ("debug: v1i", v1i[0].value())
#print ("debug: v2j", v2j[0].value())
# Aggregate
v1 = dt.esum(v1i)
v2 = dt.esum(v2j)
#print ("debug: v1.value()", v1.value())
#print ("debug: v2.value()", v2.value())
#colScore = dt.logistic(dt.dot_product(self.SelHW, dt.concatenate([v1,v2])))
return dt.dot_product(v1, v2)
def other_loss_function(self, pred, gold, word_id, lexicon):
if len(lexicon[word_id]) > 0:
n_labels = len(lexicon[word_id])
return dynet.scalarInput(1 / n_labels) * dynet.esum(
[-dynet.log(pred[k]) for k in sorted(lexicon[word_id])])
else:
n_labels = len(lexicon["LEX_POS"])
return dynet.scalarInput(1 / n_labels) * dynet.esum(
[-dynet.log(pred[k]) for k in sorted(lexicon["LEX_POS"])])
def _attend(self, query, mask=None):
# query ((H), B)
# mask ((T, 1), B)
query = unsqueeze(query, 0)
# ((1, H), B) * ((H, T), B) -> ((1, T), B) -> ((T, 1), B)
attn_scores = dy.cdiv(dy.transpose(query * self.context), dy.scalarInput(self.scale))
if mask is not None:
attn_scores = dy.cmult(attn_scores, mask[0]) + (mask[1] * dy.scalarInput(-1e9))
return dy.softmax(attn_scores) # ((T, 1), B)
def _make_input(self, seq, lang_id, runtime):
x_list = []
encoder_states_list = [None]
lang_emb = self.lang_embeddings[lang_id]
# add the root
x_list.append(self.padd_embeddings[0])
for entry in seq:
word = entry.word
# prepare lexical embeddings
char_emb, encoder_states = self.character_network.compute_embeddings(
word, runtime=runtime, language_embeddings=lang_emb)
encoder_states_list.append(encoder_states)
word = word.lower()
if word in self.encodings.word2int:
holistic_emb = self.holistic_embeddings[
self.encodings.word2int[word]]
else:
holistic_emb = self.holistic_embeddings[
self.encodings.word2int['<UNK>']]
# dropout lexical embeddings
if runtime:
w_emb = char_emb + holistic_emb
else:
p1 = random.random()
p2 = random.random()
m1 = 1
m2 = 1
if p1 < self.config.input_dropout_prob:
m1 = 0
if p2 < self.config.input_dropout_prob:
m2 = 0
scale = 1.0
if m1 + m2 > 0:
scale = float(2) / (m1 + m2)
m1 = dy.scalarInput(m1)
m2 = dy.scalarInput(m2)
scale = dy.scalarInput(scale)
w_emb = (char_emb * m1 + holistic_emb * m2) * scale
x_list.append(dy.concatenate([w_emb, lang_emb]))
# close sequence
x_list.append(self.padd_embeddings[1])
encoder_states_list.append(None)
return x_list, encoder_states_list
def greedy_search(self, char_seq, truth = None, mu =0.):
init_state = self.params['lstm'].initial_state().add_input(self.param_exprs['<bos>'])
init_y = dy.tanh(self.param_exprs['pW'] * init_state.output() + self.param_exprs['pb'])
init_score = dy.scalarInput(0.)
init_sentence = Sentence(score=init_score.scalar_value(),score_expr=init_score,LSTMState =init_state, y= init_y , prevState = None, wlen=None, golden=True)
if truth is not None:
cembs = [ dy.dropout(dy.lookup(self.params['embed'],char),self.options['dropout_rate']) for char in char_seq ]
else:
cembs = [dy.lookup(self.params['embed'],char) for char in char_seq ]
#cembs = [ dy.dropout(dy.lookup(self.params['embed'],char),self.options['dropout_rate']) for char in char_seq ]
start_agenda = init_sentence
agenda = [start_agenda]
for idx, _ in enumerate(char_seq,1): # from left to right, character by character
now = None
for wlen in range(1,min(idx,self.options['max_word_len'])+1): # generate word candidate vectors
# join segmentation sent + word
word = self.word_repr(char_seq[idx-wlen:idx], cembs[idx-wlen:idx])
sent = agenda[idx-wlen]
if truth is not None:
word = dy.dropout(word,self.options['dropout_rate'])
word_score = dy.dot_product(word,self.param_exprs['U'])
if truth is not None:
golden = sent.golden and truth[idx-1]==wlen
margin = dy.scalarInput(mu*wlen if truth[idx-1]!=wlen else 0.)
score = margin + sent.score_expr + dy.dot_product(sent.y, word) + word_score
else:
golden = False
score = sent.score_expr + dy.dot_product(sent.y, word) + word_score
good = (now is None or now.score < score.scalar_value())
if golden or good:
new_state = sent.LSTMState.add_input(word)
new_y = dy.tanh(self.param_exprs['pW'] * new_state.output() + self.param_exprs['pb'])
new_sent = Sentence(score=score.scalar_value(),score_expr=score,LSTMState=new_state,y=new_y, prevState=sent, wlen=wlen, golden=golden)
if good:
now = new_sent
if golden:
golden_sent = new_sent
agenda.append(now)
if truth is not None and truth[idx-1]>0 and (not now.golden):
return (now.score_expr - golden_sent.score_expr)
if truth is not None:
return (now.score_expr - golden_sent.score_expr)
return agenda
def _attend(self, query, mask=None):
# query ((H), B)
# mask ((T, 1), B)
query = unsqueeze(query, 0)
# ((1, H), B) * ((H, T), B) -> ((1, T), B) -> ((T, 1), B)
attn_scores = dy.cdiv(dy.transpose(query * self.context),
dy.scalarInput(self.scale))
if mask is not None:
attn_scores = dy.cmult(attn_scores,
mask[0]) + (mask[1] * dy.scalarInput(-1e9))
return dy.softmax(attn_scores) # ((T, 1), B)
def get_last_layer_context_representations(
self, sentence, context_representations_for_crf_loss,
context_representations_for_md_loss):
last_layer_context_representations = context_representations_for_crf_loss
if self.parameters['active_models'] in [1, 2, 3]:
if self.parameters['active_models'] == 1 and \
self.parameters['integration_mode'] != 0:
assert False, "integration_mode should be set to zero when active_models == 1"
if self.parameters['debug'] == 1:
print(("str_words", sentence["str_words"]))
morph_analysis_representations, morph_analysis_scores = \
self.get_morph_analysis_representations_and_scores(sentence,
context_representations_for_md_loss)
selected_morph_analysis_representations = \
self.disambiguate_morph_analyzes(morph_analysis_scores)
if 'golden_morph_analysis_indices' in list(sentence.keys()):
md_loss = dynet.esum([
dynet.pickneglogsoftmax(morph_analysis_scores_for_word,
golden_idx)
for golden_idx, morph_analysis_scores_for_word in zip(
sentence['golden_morph_analysis_indices'],
morph_analysis_scores)
])
else:
md_loss = dynet.scalarInput(0)
if self.parameters['integration_mode'] == 2:
# on the other hand, we can implement two layer of contexts, which we use the
# first for morphological disambiguation and then concatenate the predicted/computed/
# selected morphological analysis representation to use for calculating tag_scores
last_layer_context_representations = \
[dynet.concatenate([context,
morph_analysis_representations[word_pos]
[selected_morph_analysis_representation_pos]])
for word_pos, (selected_morph_analysis_representation_pos, context) in
enumerate(
zip(selected_morph_analysis_representations, context_representations_for_crf_loss))]
if md_loss.value() > 1000:
logging.error("BEEP")
else:
# only the plain old NER model
# we must decide whether we should implement the morphological embeddings scheme here.
md_loss = dynet.scalarInput(0)
selected_morph_analysis_representations = None
last_layer_context_representations = context_representations_for_crf_loss
assert last_layer_context_representations is not None
return last_layer_context_representations, md_loss, selected_morph_analysis_representations
def get_factor_expressions(fws,
bws,
tfemb,
tfdict,
valid_fes,
sentence,
spaths_x=None,
cpaths_x=None):
factexprs = {}
sentlen = len(fws)
sortedtfd = sorted(list(tfdict.keys()))
targetspan = (sortedtfd[0], sortedtfd[-1])
for j in range(sentlen):
istart = 0
if USE_SPAN_CLIP and j > ALLOWED_SPANLEN:
istart = max(0, j - ALLOWED_SPANLEN)
for i in range(istart, j + 1):
spanlen = dy.scalarInput(j - i + 1)
logspanlen = dy.scalarInput(math.log(j - i + 1))
spanwidth = sp_x[SpanWidth.howlongisspan(i, j)]
spanpos = ap_x[ArgPosition.whereisarg((i, j), targetspan)]
fbemb_ij_basic = dy.concatenate([
fws[i][j], bws[i][j], tfemb, spanlen, logspanlen, spanwidth,
spanpos
])
if USE_DEPS:
outs = oh_s[OutHeads.getnumouts(i, j, sentence.outheads)]
shp = spaths_x[sentence.shortest_paths[(i, j, targetspan[0])]]
fbemb_ij = dy.concatenate([fbemb_ij_basic, outs, shp])
elif USE_CONSTITS:
isconstit = dy.scalarInput((i, j) in sentence.constitspans)
lca = ct_x[sentence.lca[(i, j)][1]]
phrp = cpaths_x[sentence.cpaths[(i, j, targetspan[0])]]
fbemb_ij = dy.concatenate(
[fbemb_ij_basic, isconstit, lca, phrp])
else:
fbemb_ij = fbemb_ij_basic
for y in valid_fes:
fctr = Factor(i, j, y)
if USE_HIER and y in feparents:
fefixed = dy.esum([fe_x[y]] +
[fe_x[par] for par in feparents[y]])
else:
fefixed = fe_x[y]
fbemb_ijy = dy.concatenate([fefixed, fbemb_ij])
factexprs[fctr] = w_f * dy.rectify(w_z * fbemb_ijy + b_z) + b_f
return factexprs
def beam_search(self, char_seq, truth=None, mu=0.):
start_agenda = Agenda(self.options['beam_size'])
init_state = self.params['lstm'].initial_state().add_input(
self.param_exprs['<bos>'])
init_y = dy.tanh(self.param_exprs['pW'] * init_state.output() +
self.param_exprs['pb'])
init_score = dy.scalarInput(0.)
start_agenda.push(
Sentence(score=init_score.scalar_value(),
score_expr=init_score,
LSTMState=init_state,
y=init_y,
prevState=None,
wlen=None))
agenda = [start_agenda]
for idx, _ in enumerate(
char_seq, 1): # from left to right, character by character
now = Agenda(self.options['beam_size'])
for wlen in xrange(1,
min(idx, self.options['max_word_len']) +
1): # generate candidate word vectors
word = self.word_repr(char_seq[idx - wlen:idx])
word_score = dy.dot_product(word, self.param_exprs['U'])
for sent in agenda[idx - wlen]: # join segmentation
if truth is not None:
margin = dy.scalarInput(
mu * wlen if truth[idx - 1] != wlen else 0.)
score = margin + sent.score_expr + dy.dot_product(
sent.y, word) + word_score
else:
score = sent.score_expr + dy.dot_product(
sent.y, word) + word_score
if now.happy_with(score.scalar_value()):
new_state = sent.LSTMState.add_input(word)
new_y = dy.tanh(self.param_exprs['pW'] *
new_state.output() +
self.param_exprs['pb'])
now.push(
Sentence(score=score.scalar_value(),
score_expr=score,
LSTMState=new_state,
y=new_y,
prevState=sent,
wlen=wlen))
agenda.append(now)
if truth is not None:
return agenda[-1].max().score_expr
return agenda
def calc_loss(batch_distances, batch_scores, lamb):
batch_losses = [
lamb * dy.scalarInput(d) - (batch_scores[0] - s)
for s, d in zip(batch_scores[1:], batch_distances[1:])
]
losses_pos = [
l if l.npvalue() >= 0 else dy.scalarInput(0) for l in batch_losses
]
if len(losses_pos) == 0:
return 0
return dy.esum(losses_pos)
def rule_loss_collector(
beam_item # type: BeamItem
):
if beam_item.sync_rule is None:
yield dn.scalarInput(0.0)
raise StopIteration
node_info = beam_item.node_info_ref()
if node_info.early_updated:
yield dn.scalarInput(0.0)
raise StopIteration
if self.options.use_graph_embedding:
# calculate rule loss
correspondents = node_info.correspondents
pred_expr = correspondents[beam_item.sync_rule]
gold_expr = correspondents[node_info.gold_rule]
loss = (pred_expr - gold_expr) if pred_expr is not gold_expr else dn.scalarInput(0.0)
else:
loss = dn.scalarInput(0.0)
# add rule correctness statistics
print_logger.total_count += 1
if beam_item.sync_rule == node_info.gold_rule:
print_logger.correct_count += 1
# calculate edge loss
gold_item = node_info.gold_item
if gold_item is not beam_item:
predict_edge_scores = edge_feature_to_scores(
beam_item.own_features - gold_item.own_features,
return_expr=True) # ignore common features
gold_edge_scores = edge_feature_to_scores(
gold_item.own_features - beam_item.own_features, True)
print_logger.total_gold_score += gold_item.score
print_logger.total_predict_score += beam_item.score
struct_loss = predict_edge_scores - gold_edge_scores
loss += struct_loss
yield loss
# loss of children node
if beam_item.left is not None:
for i in rule_loss_collector(beam_item.left):
yield i
if beam_item.right is not None:
for i in rule_loss_collector(beam_item.right):
yield i
node_info.early_updated = 1
def greedy_train_max_sumlogllh(self, init_state, gold_actions):
total_obj = dt.scalarInput(0)
cur_state = init_state
res = 0
idx = 0
while True:
if cur_state.is_end():
break
action_list = list(cur_state.get_action_set())
new_expression_list, meta_info_list = cur_state.get_next_score_expressions(
action_list)
prob_list = dt.softmax(new_expression_list)
gold_action = gold_actions[idx]
action_idx = action_list.index(gold_action)
total_obj += -(dt.log(prob_list[action_idx]))
cur_state = cur_state.get_new_state_after_action(
gold_action, meta_info_list[action_idx])
idx += 1
#print (cur_state)
res = total_obj.scalar_value()
total_obj.backward()
self.neural_model.learner.update()
return res
def learn(self, seq):
output, proj_x3 = self._predict(seq, runtime=False)
# arcs
for iSrc in range(len(seq)):
for iDst in range(len(seq)):
if iDst > iSrc:
o = output[iSrc][iDst] # the softmax portion
t = get_link(seq, iSrc, iDst)
# if t==1:
# self.losses.append(-dy.log(dy.pick(o, t)))
self.losses.append(dy.binary_log_loss(
o, dy.scalarInput(t)))
# labels
gs_chains, labels = self._get_gs_chains(seq)
for chain, label in zip(gs_chains, labels):
label_rnn = self.label_decoder.initial_state()
for index in chain:
label_rnn = label_rnn.add_input(proj_x3[index])
label_softmax = dy.softmax(
self.label_w.expr(update=True) * label_rnn.output() +
self.label_b.expr(update=True))
self.losses.append(-dy.log(
dy.pick(label_softmax, self.encodings.label2int[label])))
def decode(self, states, y, encoded_input, train=False):
def sample(probs):
return np.argmax(probs)
s = self.decoder_rnn.initial_state()
start_encoded = self.l2e["sep"].encode("<s>", "sep")
out = []
loss = dy.scalarInput(0.)
#s = s.add_input(states[-1]) #s.add_input(dy.concatenate([start_encoded, states[-1]]))
s = s.add_input(dy.concatenate([start_encoded, states[-1]]))
generated_string = []
for char in y:
true_char_encoded = self.l2e["l"].encode(char, "l")
scores = self.predict_letter(s.output(), y)
generated_string.append(scores)
weighted_states = self.attend(s.output(), states, encoded_input)
#s = s.add_input(weighted_states) #s.add_input(dy.concatenate([true_char_encoded, weighted_states]))
s = s.add_input(
dy.concatenate([true_char_encoded, weighted_states]))
if char in self.C2I:
loss += dy.pickneglogsoftmax(scores, self.C2I[char])
return loss, generated_string
def get_loss_classification(self, inputs, seq):
"""
Computes classification loss for this sequence based on input vectors.
"""
W_ent1 = dy.parameter(self.l2ent1)
W_ent1b = dy.parameter(self.l2ent1b)
W_ent2 = dy.parameter(self.l2ent2)
W_ent2b = dy.parameter(self.l2ent2b)
def ff(h):
return W_ent2 * dy.tanh(W_ent1 * h + W_ent1b) + W_ent2b
boundaries = get_boundaries(seq.l_seq) if self.__is_training else \
get_boundaries(seq.bio_pred)
losses = []
if not self.__is_training: seq.ent_pred = []
for (s, t, entity) in boundaries:
h = bilstm_single(inputs[s:t + 1], self.elstm1, self.elstm2)
g = ff(h)
if self.__is_training:
gold = self.e_enc[entity]
losses.append(dy.pickneglogsoftmax(g, gold))
else:
seq.ent_pred.append(self.e_dec[np.argmax(g.npvalue())])
if self.entemb_path:
string = stringfy(seq.w_seq, s, t)
with open(self.entemb_path, 'a') as inf:
inf.write(string + '\t')
for val in g.vec_value():
inf.write(str(val) + ' ')
inf.write('\n')
classification_loss = dy.esum(losses) if losses else dy.scalarInput(0.)
return classification_loss
def get_loss_boundary(self, inputs, seq):
"""
Computes boundary loss for this sequence based on input vectors.
"""
W_bio1 = dy.parameter(self.l2bio1)
W_bio1b = dy.parameter(self.l2bio1b)
W_bio2 = dy.parameter(self.l2bio2)
W_bio2b = dy.parameter(self.l2bio2b)
def ff(h):
return W_bio2 * dy.tanh(W_bio1 * h + W_bio1b) + W_bio2b
gs = [ff(h) for h in inputs] # Inputs now 3 dimensional ("BIO scores")
if self.loss == "global":
boundary_loss = self.get_loss_boundary_global(gs, seq)
elif self.loss == "local":
boundary_loss = self.get_loss_boundary_local(gs, seq)
else:
sys.exit("Unknown loss \"{0}\"".format(self.loss))
losses = []
if not self.__is_training: seq.bio_pred = []
for i, g in enumerate(gs):
if self.__is_training:
gold = self.__BIO_ENC[seq.l_seq[i][0]]
losses.append(dy.pickneglogsoftmax(g, gold))
else:
seq.bio_pred.append(self.__BIO_DEC[np.argmax(g.npvalue())])
boundary_loss = dy.esum(losses) if losses else dy.scalarInput(0.)
return boundary_loss
def get_loss_boundary_global(self, score_vecs, seq):
start_b = dy.parameter(self.start_bias)
T = dy.parameter(self.trans_mat)
end_b = dy.parameter(self.end_bias)
if not self.__is_training:
seq.bio_pred = viterbi(start_b, T, end_b, score_vecs, self.valid)
return dy.scalarInput(0.)
pi = [[None for _ in xrange(3)] for _ in xrange(len(score_vecs))]
for y in xrange(3):
pi[0][y] = score_vecs[0][y] + start_b[y]
for i in xrange(1, len(pi)):
for y in xrange(3):
pi[i][y] = dy.logsumexp([
pi[i - 1][y_prev] + T[y_prev][y] + score_vecs[i][y]
for y_prev in xrange(3)
])
normalizer = dy.logsumexp([pi[-1][y] + end_b[y] for y in xrange(3)])
gold_score = score_crf(start_b, T, end_b, score_vecs,
[self.__BIO_ENC[l[0]] for l in seq.l_seq])
return normalizer - gold_score
def beam_train_max_margin(self, init_state, gold_ans):
#still did not use the gold sequence but use the min risk training
#max reward y = argmax(r(y))
#max y' = argmax f(x,y) - R(y')
# loss = max(f(x,y') - f(x,y) + R(y) - R(y') , 0)
end_state_list = self.beam_predict(init_state)
reward_list = [x.reward(gold_ans) for x in end_state_list]
violation_list = [
s.score - reward for s, reward in zip(end_state_list, reward_list)
]
best_score_state_idx = violation_list.index(max(
violation_list)) # find the best scoring seq with minimal reward
best_reward_state_idx = reward_list.index(
max(reward_list)) # find seq with the max reward in beam
best_score_state = end_state_list[best_score_state_idx]
best_reward_state = end_state_list[best_reward_state_idx]
best_score_state_reward = reward_list[best_score_state_idx]
best_reward_state_reward = reward_list[best_reward_state_idx]
loss = dt.rectify(best_score_state.path_score_expression -
best_reward_state.path_score_expression +
dt.scalarInput(best_reward_state_reward -
best_score_state_reward))
loss_value = loss.value()
loss.backward()
self.neural_model.learner.update()
return loss_value
def train(self, mini_batch, num_train, k):
words, pos_tags, chars, langs, signs, masks = mini_batch
# Getting the last hidden layer from BiLSTM.
rnn_out = self.rnn_mlp(mini_batch, True)
h_out = rnn_out[-1]
t_out_d = dy.reshape(h_out, (h_out.dim()[0][0], h_out.dim()[1]))
t_out = dy.transpose(t_out_d)
# Calculating the kq values for NCE.
kq = dy.scalarInput(float(k) / num_train)
lkq = dy.log(kq)
loss_values = []
for i in range(len(langs)):
for j in range(i + 1, len(langs)):
if (langs[i] != langs[j]) and (signs[i] == 1 or signs[j] == 1):
lu = -dy.squared_distance(t_out[i], t_out[j])
denom = dy.log(dy.exp(lu) + kq)
if signs[i] == signs[j]: # both one
nom = lu
else:
nom = lkq
loss_values.append(denom - nom)
err_value = 0
if len(loss_values) > 0:
err = dy.esum(loss_values) / len(loss_values)
err.forward()
err_value = err.value()
err.backward()
self.trainer.update()
dy.renew_cg()
return err_value
def beam_train_max_margin_with_goldactions(self, init_state, gold_actions):
#max y = gold y
#max y' = argmax f(x,y)
# loss = max(f(x,y') - f(x,y) + R(y) - R(y') , 0)
#loss
#end_state_list = self.beam_predict(init_state) # top-k argmax_y f(x,y)
end_state_list = self.beam_predict_max_violation(
init_state, gold_actions
) # top-k argmax_y f(x,y) + R(y*) - R(y) // Current implementation is the same as Hamming distance
best_score_state = end_state_list[0]
reward_list = [x.reward(gold_actions) for x in end_state_list]
best_reward_state = self.get_goldstate_with_gold_actions(
init_state, gold_actions)
best_reward = best_reward_state.reward(gold_actions)
loss = dt.rectify(best_score_state.path_score_expression -
best_reward_state.path_score_expression +
dt.scalarInput(best_reward - reward_list[0]))
loss_value = loss.value()
loss.backward()
self.neural_model.learner.update()
return loss_value
def get_constit_loss(fws, bws, goldspans):
if not USE_PTB_CONSTITS:
raise Exception("should not be using the constit loss now!",
USE_PTB_CONSTITS)
if len(goldspans) == 0:
return None, 0
losses = []
sentlen = len(fws)
for j in range(sentlen):
istart = 0
if USE_SPAN_CLIP and j > ALLOWED_SPANLEN:
istart = max(0, j - ALLOWED_SPANLEN)
for i in range(istart, j + 1):
constit_ij = w_c * dy.rectify(
w_fb * dy.concatenate([fws[i][j], bws[i][j]]) + b_fb) + b_c
logloss = dy.log_softmax(constit_ij)
isconstit = int((i, j) in goldspans)
losses.append(pick(logloss, isconstit))
ptbconstitloss = dy.scalarInput(DELTA) * -esum(losses)
numspanstagged = len(losses)
return ptbconstitloss, numspanstagged
def loss_function(recon_x, x, mu, logvar):
BCE = dy.binary_log_loss(recon_x, x) # equiv to torch.nn.functional.binary_cross_entropy(?,?, size_average=False)
# see Appendix B from VAE paper:
# Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014
# https://arxiv.org/abs/1312.6114
# 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
KLD = -0.5 * dy.sum_elems(1 + logvar - dy.pow(mu, dy.scalarInput(2)) - dy.exp(logvar))
return BCE + KLD
def _attend(self, query, mask=None):
# query ((H), B)
# mask ((T, 1), B)
projected_state = self.decoder * query # ((H,), B)
non_lin = dy.tanh(dy.colwise_add(self.context_proj, projected_state)) # ((H, T), B)
attn_scores = dy.transpose(self.v * non_lin) # ((1, H), B) * ((H, T), B) -> ((1, T), B) -> ((T, 1), B)
if mask is not None:
attn_scores = dy.cmult(attn_scores, mask[0]) + (mask[1] * dy.scalarInput(-1e9))
return dy.softmax(attn_scores) # ((T, 1), B)
def truth_score(self, word_seq):
wembs = [self.param_exprs['<bos>']]+[self.word_repr(word) for word in word_seq]
init_state = self.params['lstm'].initial_state()
hidden_states = init_state.transduce(wembs)
score = dy.scalarInput(0.)
for h, w in zip(hidden_states[:-1],wembs[1:]):
y = dy.tanh(self.param_exprs['pW'] * h + self.param_exprs['pb'])
score = score + dy.dot_product(y,w) +dy.dot_product(w,self.param_exprs['U'])
return score
def score_sentence(self, emissions, tags):
"""Get the score of a given sentence.
:param emissions: List[dy.Expression ((H,), B)]
:param tags: List[int]
Returns:
dy.Expression ((1,), B)
"""
tags = np.concatenate((np.array([self.start_idx], dtype=int), tags))
score = dy.scalarInput(0)
transitions = self.transitions
for i, e in enumerate(emissions):
# Due to Dynet being column based it is best to use the transition
# matrix so that x -> y is T[y, x].
score += dy.pick(dy.pick(transitions, tags[i + 1]), tags[i]) + dy.pick(e, tags[i + 1])
score += dy.pick(dy.pick(transitions, self.end_idx), tags[-1])
return score
def node_iteration(rel, g, node, opts, assoc_model, trainer, log_file, is_source):
"""
Perform one iteration of trying to score a node's neighbors above negative samples.
"""
# true instances likelihood
trues = targets(g, node) if is_source else sources(g, node)
side = '->' if is_source else '<-'
if len(trues) == 0: return 0.0
if opts.debug:
dy.renew_cg(immediate_compute = True, check_validity = True)
else:
dy.renew_cg()
# compute association score as dynet expression (can't do this above due to staleness)
true_scores = []
for tr in trues:
if is_source:
j_assoc_score = assoc_model.word_assoc_score(node, tr, rel)
else:
j_assoc_score = assoc_model.word_assoc_score(tr, node, rel)
if log_file is not None:
log_file.write('{} {}\tTRUE_{}\t{:.3e}\n'\
.format(node, side, tr, j_assoc_score.scalar_value()))
true_scores.append(j_assoc_score)
# false targets likelihood - negative sampling (uniform)
# collect negative samples
if opts.nll:
sample_scores = [[ts] for ts in true_scores]
else:
margins = []
neg_samples = [np.random.choice(range(N)) for _ in range(opts.neg_samp * len(trues))]
# remove source and true targets if applicable
for t in [node] + trues:
if t in neg_samples:
neg_samples.remove(t)
neg_samples.append(np.random.choice(range(N)))
for (i,ns) in enumerate(neg_samples):
# compute association score as dynet expression
if is_source:
ns_assoc_score = assoc_model.word_assoc_score(node, ns, rel)
else:
ns_assoc_score = assoc_model.word_assoc_score(ns, node, rel)
if log_file is not None:
log_file.write('{} {}\tNEG_{}\t{:.3e}\n'\
.format(node, side, ns, ns_assoc_score.scalar_value()))
corresponding_true = i // opts.neg_samp
if opts.nll:
sample_scores[corresponding_true].append(ns_assoc_score)
else:
# TODO maybe use dy.hinge()
ctt_score = true_scores[corresponding_true]
margin = ctt_score - ns_assoc_score
margins.append(dy.rectify(dy.scalarInput(1.0) - margin))
# compute overall loss
if opts.nll:
if len(sample_scores) == 0:
dy_loss = dy.scalarInput(0.0)
else:
dy_loss = dy.esum([dy.pickneglogsoftmax(dy.concatenate(scrs), 0) for scrs in sample_scores])
else:
if len(margins) == 0:
dy_loss = dy.scalarInput(0.0)
else:
dy_loss = dy.esum(margins)
sc_loss = dy_loss.scalar_value()
if log_file is not None:
log_file.write('{}\tLOSS\t{:.3e}\n'\
.format(node, sc_loss))
# backprop and recompute score
if opts.v > 1:
timeprint('overall loss for relation {}, node {} as {} = {:.6f}'\
.format(rel, node, 'source' if is_source else 'target', sc_loss))
dy_loss.backward()
trainer.update()
return sc_loss
pW1 = m.add_parameters((HIDDEN_SIZE, 2), device="GPU:1")
pb1 = m.add_parameters(HIDDEN_SIZE, device="GPU:1")
pW2 = m.add_parameters((HIDDEN_SIZE, HIDDEN_SIZE), device="GPU:0")
pb2 = m.add_parameters(HIDDEN_SIZE, device="GPU:0")
pV = m.add_parameters((1, HIDDEN_SIZE), device="CPU")
pa = m.add_parameters(1, device="CPU")
if len(sys.argv) == 2:
m.populate_from_textfile(sys.argv[1])
dy.renew_cg()
W1, b1, W2, b2, V, a = dy.parameter(pW1, pb1, pW2, pb2, pV, pa)
x = dy.vecInput(2, "GPU:1")
y = dy.scalarInput(0, "CPU")
h1 = dy.tanh((W1*x) + b1)
h1_gpu0 = dy.to_device(h1, "GPU:0")
h2 = dy.tanh((W2*h1_gpu0) + b2)
h2_cpu = dy.to_device(h2, "CPU")
if xsent:
y_pred = dy.logistic((V*h2_cpu) + a)
loss = dy.binary_log_loss(y_pred, y)
T = 1
F = 0
else:
y_pred = (V*h2_cpu) + a
loss = dy.squared_distance(y_pred, y)
T = 1
F = -1
HIDDEN_SIZE = 8
ITERATIONS = 2000
m = dy.Model()
trainer = dy.SimpleSGDTrainer(m)
W = m.add_parameters((HIDDEN_SIZE, 2))
b = m.add_parameters(HIDDEN_SIZE)
V = m.add_parameters((1, HIDDEN_SIZE))
a = m.add_parameters(1)
if len(sys.argv) == 2:
m.populate_from_textfile(sys.argv[1])
x = dy.vecInput(2)
y = dy.scalarInput(0)
h = dy.tanh((W*x) + b)
if xsent:
y_pred = dy.logistic((V*h) + a)
loss = dy.binary_log_loss(y_pred, y)
T = 1
F = 0
else:
y_pred = (V*h) + a
loss = dy.squared_distance(y_pred, y)
T = 1
F = -1
for iter in range(ITERATIONS):
mloss = 0.0
