def encode(input_, lengths):
forward, backward = lstm(input_)
states = dy.concatenate_cols(forward)
final_states_forward = dy.pick_batch(states, lengths, dim=1)
states = dy.concatenate_cols(backward)
final_states_backward = dy.pick_batch(states, lengths, dim=1)
return dy.concatenate([final_states_forward, final_states_backward])
def generate_output(self,
translator,
initial_state,
src_length=None,
forced_trg_ids=None):
# Output variables
score = []
word_ids = []
attentions = []
logsoftmaxes = []
states = []
masks = []
# Search Variables
done = None
current_state = initial_state
for length in range(self.max_len):
prev_word = word_ids[length - 1] if length > 0 else None
current_output = translator.generate_one_step(
prev_word, current_state)
current_state = current_output.state
if forced_trg_ids is None:
word_id = np.argmax(current_output.logsoftmax.npvalue(),
axis=0)
if len(word_id.shape) == 2:
word_id = word_id[0]
else:
if batchers.is_batched(forced_trg_ids):
word_id = [
forced_trg_ids[i][length]
for i in range(len(forced_trg_ids))
]
else:
word_id = [forced_trg_ids[length]]
logsoft = dy.pick_batch(current_output.logsoftmax, word_id)
if done is not None:
word_id = [
word_id[i] if not done[i] else Vocab.ES
for i in range(len(done))
]
# masking for logsoftmax
mask = [1 if not done[i] else 0 for i in range(len(done))]
logsoft = dy.cmult(logsoft, dy.inputTensor(mask, batched=True))
masks.append(mask)
# Packing outputs
score.append(logsoft.npvalue())
word_ids.append(word_id)
attentions.append(current_output.attention)
logsoftmaxes.append(
dy.pick_batch(current_output.logsoftmax, word_id))
states.append(translator.get_nobp_state(current_state))
# Check if we are done.
done = [x == Vocab.ES for x in word_id]
if all(done):
break
masks.insert(0, [1 for _ in range(len(done))])
words = np.stack(word_ids, axis=1)
score = np.sum(score, axis=0)
return [
SearchOutput(words, attentions, score, logsoftmaxes, states, masks)
]
def predict_chunks_by_tokens(self, w_t, chunk_batch):
ender = [self.lattice_vocab.chunk_end.i] * self.BATCH_SIZE
lps = []
state = self.lattice_rnn.initial_state(dropout=self.DROPOUT)
cs = [[self.lattice_vocab.chunk_start.i] * self.BATCH_SIZE
] + chunk_batch
cum_lp = dynet.scalarInput(0.0, device=self.args.param_device)
for i, (cc, nc) in enumerate(zip(cs, cs[1:])):
if self.args.concat_context_vector:
x_t = dynet.pick_batch(self.vocab_R, cc)
state.add_input(x_t)
else:
if i == 0:
state.add_input(self.project_main_to_lattice_init_R * w_t)
else:
x_t = dynet.pick_batch(self.vocab_R, cc)
state.add_input(x_t)
y_t = state.output()
y_t = dynet.to_device(y_t, self.args.param_device)
if self.DROPOUT:
y_t = dynet.cmult(y_t, self.dropout_mask_lattice_y_t)
if self.args.concat_context_vector:
y_t = dynet.concatenate([y_t, w_t])
r_t = dynet.affine_transform([
self.vocab_bias, self.vocab_R,
dynet.tanh(
dynet.affine_transform(
[self.lattice_bias, self.lattice_R, y_t]))
])
if i > 0:
lps.append(cum_lp + -dynet.pickneglogsoftmax_batch(r_t, ender))
cum_lp = cum_lp + -dynet.pickneglogsoftmax_batch(r_t, nc)
lps.append(cum_lp)
return lps
def cross_entropy_loss(self, score, next_word, cur_word): if self.__ls: log_prob = dy.log_softmax(score) if self.__lm is None: loss = - dy.pick_batch(log_prob, next_word) * (1 - self.__ls_eps) - \ dy.mean_elems(log_prob) * self.__ls_eps else: loss = - dy.pick_batch(log_prob, next_word) * (1 - self.__ls_eps) - \ dy.dot_product(self.__lm.next_expr(cur_word), log_prob) * self.__ls_eps else: loss = dy.pickneglogsoftmax(score, next_word) return loss
def predict_logprobs(self, X, Y):
"""
Returns the log probabilities of the predictions for this model (batched version).
Returns a matrix of log probabilities.
@param X: the input indexes from which to predict
@param Y: a list of references indexes for which to extract the prob.
@return the matrix of predicted logprobabilities for each of the provided ref y in Y
as a numpy array
"""
assert (len(X) == len(Y))
assert (all([len(x) == len(y) for x, y in zip(X, Y)]))
nlines = len(X)
X = zip(*X) #transposes the batch
Y = zip(*Y) #transposes the batch
if self.tied:
dy.renew_cg()
state = self.rnn.initial_state()
E = dy.parameter(self.embedding_matrix)
preds = []
lookups = [dy.pick_batch(E, xcolumn) for xcolumn in X]
outputs = state.transduce(lookups)
ypred_batch = [
dy.pickneglogsoftmax_batch(E * lstm_out, y)
for lstm_out, y in zip(outputs, Y)
]
dy.forward(ypred_batch)
if nlines > 1:
preds = [(-col.npvalue()).tolist()[0] for col in ypred_batch]
else:
preds = [(-col.npvalue()).tolist() for col in ypred_batch]
return list(zip(*preds)) #final back transposition
else:
dy.renew_cg()
state = self.rnn.initial_state()
O = dy.parameter(self.output_weights)
E = dy.parameter(self.embedding_matrix)
preds = []
lookups = [dy.pick_batch(E, xcolumn) for xcolumn in X]
outputs = state.transduce(lookups)
ypred_batch = [
dy.pickneglogsoftmax_batch(O * lstm_out, y)
for lstm_out, y in zip(outputs, Y)
]
dy.forward(ypred_batch)
preds = [(-col.npvalue()).tolist()[0] for col in ypred_batch]
if nlines > 1:
preds = [(-col.npvalue()).tolist()[0] for col in ypred_batch]
else:
preds = [(-col.npvalue()).tolist() for col in ypred_batch]
return list(zip(*preds)) #final back transposition
def cross_entropy_loss(self, s, nw, cw):
"""Calculates the cross-entropy
"""
if self.ls:
log_prob = dy.log_softmax(s)
if self.lm is None:
loss = - dy.pick_batch(log_prob, nw) * (1 - self.ls_eps) - \
dy.mean_elems(log_prob) * self.ls_eps
else:
loss = - dy.pick_batch(log_prob, nw) * (1 - self.ls_eps) - \
dy.dot_product(self.lm_e, log_prob) * self.ls_eps
else:
loss = dy.pickneglogsoftmax_batch(s, nw)
return loss
def compress_chunk(self, chunks, masks=None):
compression_batch_size = len(chunks[0])
# token_embeddings = [dynet.reshape(dynet.select_cols(self.vocab_lookup, tokens), (self.args.dim,), compression_batch_size)
# token_embeddings = [dynet.reshape(dynet.transpose(dynet.select_rows(self.vocab_R, tokens)), (self.args.dim,), compression_batch_size)
token_embeddings = [
dynet.pick_batch(self.vocab_R, tokens) for tokens in chunks
]
fwd_state = self.lattice_fwd_comp_rnn.initial_state(
mb_size=compression_batch_size, dropout=self.DROPOUT)
bwd_state = self.lattice_bwd_comp_rnn.initial_state(
mb_size=compression_batch_size, dropout=self.DROPOUT)
if masks is None:
fwd_emb = fwd_state.transduce(token_embeddings)[-1]
bwd_emb = bwd_state.transduce(list(reversed(token_embeddings)))[-1]
else:
masks = [
dynet.inputTensor(
mask, batched=True, device=self.args.param_device)
if min(mask) == 0 else None for mask in masks
]
fwd_emb = fwd_state.transduce(token_embeddings, masks)[-1]
bwd_emb = bwd_state.transduce(reversed(token_embeddings),
reversed(masks))[-1]
emb = dynet.concatenate([fwd_emb, bwd_emb])
emb = dynet.to_device(emb, self.args.param_device)
return emb
def get_chunk_embedding(self, chunks, masks=None):
if masks is None:
merged_chunks = [
self.lattice_vocab.pp(chunk)
for chunk in map(list, zip(*chunks))
]
else:
merged_chunks = [
self.lattice_vocab.masked_pp(chunk, mask)
for chunk, mask in zip(map(list, zip(
*chunks)), map(list, zip(*masks)))
]
chunk_emb_is = [
self.chunk_vocab[chunk].i if chunk in self.chunk_vocab.strings else
self.chunk_vocab['<chunk_unk>'].i for chunk in merged_chunks
]
# fixed_embs = dynet.reshape(dynet.transpose(dynet.select_rows(self.chunk_vocab_R, chunk_emb_is)), (self.args.dim,), len(chunk_emb_is))
# fixed_embs = dynet.reshape(dynet.select_cols(self.chunk_vocab_lookup, chunk_emb_is), (self.args.dim,), len(chunk_emb_is))
fixed_embs = dynet.pick_batch(self.chunk_vocab_R, chunk_emb_is)
if self.args.no_dynamic_embs:
return fixed_embs
else:
dynamic_embs = self.compress_chunk(chunks, masks)
full_embs = dynet.concatenate([fixed_embs, dynamic_embs])
return full_embs
def on_calc_additional_loss(self, reward):
if not self.learn_segmentation:
return None
ret = LossBuilder()
if self.length_prior_alpha > 0:
reward += self.segment_length_prior * self.length_prior_alpha
reward = dy.cdiv(reward - dy.mean_batches(reward),
dy.std_batches(reward))
# Baseline Loss
if self.use_baseline:
baseline_loss = []
for i, baseline in enumerate(self.bs):
baseline_loss.append(dy.squared_distance(reward, baseline))
ret.add_loss("Baseline", dy.esum(baseline_loss))
# Reinforce Loss
lmbd = self.lmbd.get_value(self.warmup_counter)
if lmbd > 0.0:
reinforce_loss = []
# Calculating the loss of the baseline and reinforce
for i in range(len(self.segment_decisions)):
ll = dy.pick_batch(self.segment_logsoftmaxes[i],
self.segment_decisions[i])
if self.use_baseline:
r_i = reward - self.bs[i]
else:
r_i = reward
reinforce_loss.append(dy.logistic(r_i) * ll)
ret.add_loss("Reinforce", -dy.esum(reinforce_loss) * lmbd)
# Total Loss
return ret
def sample_one(
self,
translator: 'xnmt.models.translators.AutoRegressiveTranslator',
initial_state: decoders.AutoRegressiveDecoderState,
forced_trg_ids: Optional[Sequence[numbers.Integral]] = None
) -> SearchOutput:
# Search variables
current_words = None
current_state = initial_state
done = None
# Outputs
logsofts = []
samples = []
states = []
attentions = []
masks = []
# Sample to the max length
for length in range(self.max_len):
translator_output = translator.generate_one_step(
current_words, current_state)
if forced_trg_ids is None:
sample = translator_output.logsoftmax.tensor_value(
).categorical_sample_log_prob().as_numpy()
if len(sample.shape) == 2:
sample = sample[0]
else:
sample = [
forced_trg[length]
if forced_trg.sent_len() > length else Vocab.ES
for forced_trg in forced_trg_ids
]
logsoft = dy.pick_batch(translator_output.logsoftmax, sample)
if done is not None:
sample = [
sample[i] if not done[i] else Vocab.ES
for i in range(len(done))
]
# masking for logsoftmax
mask = [1 if not done[i] else 0 for i in range(len(done))]
logsoft = dy.cmult(logsoft, dy.inputTensor(mask, batched=True))
masks.append(mask)
# Appending output
logsofts.append(logsoft)
samples.append(sample)
states.append(translator.get_nobp_state(translator_output.state))
attentions.append(translator_output.attention)
# Next time step
current_words = sample
current_state = translator_output.state
# Check done
done = [x == Vocab.ES for x in sample]
# Check if we are done.
if all(done):
break
# Packing output
scores = dy.esum(logsofts).npvalue()
masks.insert(0, [1 for _ in range(len(done))])
samples = np.stack(samples, axis=1)
return SearchOutput(samples, attentions, scores, logsofts, states,
masks)
def calc_loss(
self, x: dy.Expression,
y: Union[numbers.Integral,
List[numbers.Integral]]) -> dy.Expression:
if self.can_loss_be_derived_from_scores():
scores = self.calc_scores(x)
# single mode
if not batchers.is_batched(y):
loss = dy.pickneglogsoftmax(scores, y)
# minibatch mode
else:
loss = dy.pickneglogsoftmax_batch(scores, y)
else:
log_prob = self.calc_log_probs(x)
if not batchers.is_batched(y):
loss = -dy.pick(log_prob, y)
else:
loss = -dy.pick_batch(log_prob, y)
if self.label_smoothing > 0:
ls_loss = -dy.mean_elems(log_prob)
loss = ((1 - self.label_smoothing) *
loss) + (self.label_smoothing * ls_loss)
return loss
def decode_loss(self, src_encodings, tgt_seqs):
"""
:param tgt_seqs: (tgt_heads, tgt_labels): list (length=batch_size) of (src_len)
"""
# todo(NOTE): Sentences should start with empty token (as root of dependency tree)!
tgt_heads, tgt_labels = tgt_seqs
src_len = len(tgt_heads[0])
batch_size = len(tgt_heads)
np_tgt_heads = np.array(tgt_heads).flatten() # (src_len * batch_size)
np_tgt_labels = np.array(tgt_labels).flatten()
s_arc, s_label = self.cal_scores(src_encodings) # (src_len, src_len, bs), ([(src_len, src_len, bs)])
s_arc_value = s_arc.npvalue()
s_arc_choice = np.argmax(s_arc_value, axis=0).transpose().flatten() # (src_len * batch_size)
s_pick_labels = [dy.pick_batch(dy.reshape(score, (src_len,), batch_size=src_len * batch_size), s_arc_choice)
for score in s_label]
s_argmax_labels = dy.concatenate(s_pick_labels, d=0) # n_labels, src_len * batch_size
reshape_s_arc = dy.reshape(s_arc, (src_len,), batch_size=src_len * batch_size)
arc_loss = dy.pickneglogsoftmax_batch(reshape_s_arc, np_tgt_heads)
label_loss = dy.pickneglogsoftmax_batch(s_argmax_labels, np_tgt_labels)
loss = dy.sum_batches(arc_loss + label_loss) / batch_size
return loss
def calc_loss(self, policy_reward, only_final_reward=True):
loss = losses.FactoredLossExpr()
## Calculate baseline
pred_reward, baseline_loss = self.calc_baseline_loss(policy_reward, only_final_reward)
if only_final_reward:
rewards = [policy_reward - pw_i for pw_i in pred_reward]
else:
rewards = [pr_i - pw_i for pr_i, pw_i in zip(policy_reward, pred_reward)]
loss.add_loss("rl_baseline", baseline_loss)
## Z-Normalization
rewards = dy.concatenate(rewards, d=0)
if self.z_normalization:
rewards_value = rewards.value()
rewards_mean = np.mean(rewards_value)
rewards_std = np.std(rewards_value) + 1e-10
rewards = (rewards - rewards_mean) / rewards_std
## Calculate Confidence Penalty
if self.confidence_penalty:
cp_loss = self.confidence_penalty.calc_loss(self.policy_lls)
loss.add_loss("rl_confpen", cp_loss)
## Calculate Reinforce Loss
reinf_loss = []
# Loop through all action in one sequence
for i, (policy, action) in enumerate(zip(self.policy_lls, self.actions)):
# Main Reinforce calculation
reward = dy.pick(rewards, i)
ll = dy.pick_batch(policy, action)
if self.valid_pos is not None:
ll = dy.pick_batch_elems(ll, self.valid_pos[i])
reward = dy.pick_batch_elems(reward, self.valid_pos[i])
reinf_loss.append(dy.sum_batches(ll * reward))
loss.add_loss("rl_reinf", -self.weight * dy.esum(reinf_loss))
## the composed losses
return loss
def cross_entropy_loss(self, scores, next_words):
if self.label_smoothing:
log_softmax = dy.log_softmax(scores)
return -dy.pick_batch(log_softmax, next_words) * (1 - self.label_smoothing) \
- dy.mean_elems(log_softmax) * self.label_smoothing
else:
return dy.pickneglogsoftmax_batch(scores, next_words)
def calc_loss(self, mlp_dec_state, ref_action):
"""
Label Smoothing is implemented with reference to Section 7 of the paper
"Rethinking the Inception Architecture for Computer Vision"
(https://arxiv.org/pdf/1512.00567.pdf)
"""
scores = self.get_scores(mlp_dec_state)
if self.label_smoothing == 0.0:
# single mode
if not xnmt.batcher.is_batched(ref_action):
return dy.pickneglogsoftmax(scores, ref_action)
# minibatch mode
else:
return dy.pickneglogsoftmax_batch(scores, ref_action)
else:
log_prob = dy.log_softmax(scores)
if not xnmt.batcher.is_batched(ref_action):
pre_loss = -dy.pick(log_prob, ref_action)
else:
pre_loss = -dy.pick_batch(log_prob, ref_action)
ls_loss = -dy.mean_elems(log_prob)
loss = ((1 - self.label_smoothing) *
pre_loss) + (self.label_smoothing * ls_loss)
return loss
def calc_loss(
self, x: dy.Expression,
y: Union[numbers.Integral,
List[numbers.Integral]]) -> dy.Expression:
scores = self.calc_scores(x)
if self.label_smoothing == 0.0:
# single mode
if not batchers.is_batched(y):
loss = dy.pickneglogsoftmax(scores, y)
# minibatch mode
else:
loss = dy.pickneglogsoftmax_batch(scores, y)
else:
log_prob = dy.log_softmax(scores)
if not batchers.is_batched(y):
pre_loss = -dy.pick(log_prob, y)
else:
pre_loss = -dy.pick_batch(log_prob, y)
ls_loss = -dy.mean_elems(log_prob)
loss = ((1 - self.label_smoothing) *
pre_loss) + (self.label_smoothing * ls_loss)
return loss
def embed(self, x):
if self.train and self.word_dropout > 0.0 and self.word_id_mask is None:
batch_size = x.batch_size() if xnmt.batcher.is_batched(x) else 1
self.word_id_mask = [set(np.random.choice(self.vocab_size, int(self.vocab_size * self.word_dropout), replace=False)) for _ in range(batch_size)]
emb_e = dy.parameter(self.embeddings)
# single mode
if not xnmt.batcher.is_batched(x):
if self.train and self.word_id_mask and x in self.word_id_mask[0]:
ret = dy.zeros((self.emb_dim,))
else:
ret = dy.pick(emb_e, index=x)
if self.fix_norm is not None:
ret = dy.cdiv(ret, dy.l2_norm(ret))
if self.fix_norm != 1:
ret *= self.fix_norm
# minibatch mode
else:
ret = dy.pick_batch(emb_e, x)
if self.fix_norm is not None:
ret = dy.cdiv(ret, dy.l2_norm(ret))
if self.fix_norm != 1:
ret *= self.fix_norm
if self.train and self.word_id_mask and any(x[i] in self.word_id_mask[i] for i in range(x.batch_size())):
dropout_mask = dy.inputTensor(np.transpose([[0.0]*self.emb_dim if x[i] in self.word_id_mask[i] else [1.0]*self.emb_dim for i in range(x.batch_size())]), batched=True)
ret = dy.cmult(ret, dropout_mask)
if self.train and self.weight_noise > 0.0:
ret = dy.noise(ret, self.weight_noise)
return ret
def decode_loss(self, src_encodings, tgt_seqs):
"""
:param tgt_seqs: (tgt_heads, tgt_labels): list (length=batch_size) of (src_len)
"""
# todo(NOTE): Sentences should start with empty token (as root of dependency tree)!
tgt_heads, tgt_labels = tgt_seqs
src_len = len(tgt_heads[0])
batch_size = len(tgt_heads)
np_tgt_heads = np.array(tgt_heads).flatten() # (src_len * batch_size)
np_tgt_labels = np.array(tgt_labels).flatten()
s_arc, s_label = self.cal_scores(src_encodings) # (src_len, src_len, bs), ([(src_len, src_len, bs)])
s_arc_value = s_arc.npvalue()
s_arc_choice = np.argmax(s_arc_value, axis=0).transpose().flatten() # (src_len * batch_size)
s_pick_labels = [dy.pick_batch(dy.reshape(score, (src_len,), batch_size=src_len * batch_size), s_arc_choice)
for score in s_label]
s_argmax_labels = dy.concatenate(s_pick_labels, d=0) # n_labels, src_len * batch_size
reshape_s_arc = dy.reshape(s_arc, (src_len,), batch_size=src_len * batch_size)
arc_loss = dy.pickneglogsoftmax_batch(reshape_s_arc, np_tgt_heads)
label_loss = dy.pickneglogsoftmax_batch(s_argmax_labels, np_tgt_labels)
loss = dy.sum_batches(arc_loss + label_loss) / batch_size
return loss
def __call__(self, translator, dec_state, src, trg):
# TODO: apply trg.mask ?
samples = []
logsofts = []
self.bs = []
done = [False for _ in range(len(trg))]
for _ in range(self.sample_length):
dec_state.context = translator.attender.calc_context(dec_state.rnn_state.output())
if self.use_baseline:
h_t = dy.tanh(translator.decoder.context_projector(dy.concatenate([dec_state.rnn_state.output(), dec_state.context])))
self.bs.append(self.baseline(dy.nobackprop(h_t)))
logsoft = dy.log_softmax(translator.decoder.get_scores(dec_state))
sample = logsoft.tensor_value().categorical_sample_log_prob().as_numpy()[0]
# Keep track of previously sampled EOS
sample = [sample_i if not done_i else Vocab.ES for sample_i, done_i in zip(sample, done)]
# Appending and feeding in the decoder
logsoft = dy.pick_batch(logsoft, sample)
logsofts.append(logsoft)
samples.append(sample)
dec_state = translator.decoder.add_input(dec_state, translator.trg_embedder.embed(xnmt.batcher.mark_as_batch(sample)))
# Check if we are done.
done = list(six.moves.map(lambda x: x == Vocab.ES, sample))
if all(done):
break
samples = np.stack(samples, axis=1).tolist()
self.eval_score = []
for trg_i, sample_i in zip(trg, samples):
# Removing EOS
try:
idx = sample_i.index(Vocab.ES)
sample_i = sample_i[:idx]
except ValueError:
pass
try:
idx = trg_i.words.index(Vocab.ES)
trg_i.words = trg_i.words[:idx]
except ValueError:
pass
# Calculate the evaluation score
score = 0 if not len(sample_i) else self.evaluation_metric.evaluate_fast(trg_i.words, sample_i)
self.eval_score.append(score)
self.true_score = dy.inputTensor(self.eval_score, batched=True)
loss = LossBuilder()
if self.use_baseline:
for i, (score, _) in enumerate(zip(self.bs, logsofts)):
logsofts[i] = dy.cmult(logsofts[i], score - self.true_score)
loss.add_loss("Reinforce", dy.sum_elems(dy.esum(logsofts)))
else:
loss.add_loss("Reinforce", dy.sum_elems(dy.cmult(-self.true_score, dy.esum(logsofts))))
if self.use_baseline:
baseline_loss = []
for bs in self.bs:
baseline_loss.append(dy.squared_distance(self.true_score, bs))
loss.add_loss("Baseline", dy.sum_elems(dy.esum(baseline_loss)))
return loss
def score_one_sequence(self, tag_scores, tags, batch_size):
''' tags: list of tag ids at each time step '''
# print tags, batch_size
# print batch_size
# print "scoring one sentence"
tags = [[self.start_id] * batch_size
] + tags # len(tag_scores) = len(tags) - 1
score = dy.inputTensor(np.zeros(batch_size), batched=True)
# tag_scores = dy.concatenate_cols(tag_scores) # tot_tags, sent_len, batch_size
# print "tag dim: ", tag_scores.dim()
for i in range(len(tags) - 1):
score += dy.pick_batch(dy.lookup_batch(self.transition_matrix, tags[i + 1]), tags[i]) \
+ dy.pick_batch(tag_scores[i], tags[i + 1])
score += dy.pick_batch(
dy.lookup_batch(self.transition_matrix,
[self.end_id] * batch_size), tags[-1])
return score
def predict_logprobs(self,X,Y,structural=True,hidden_out=False):
"""
Returns the log probabilities of the predictions for this model (batched version).
@param X: the input indexes from which to predict (each xdatum is expected to be an iterable of integers)
@param Y: a list of references indexes for which to extract the prob
@param structural: switches between structural and lexical logprob evaluation
@param hidden_out: outputs an additional list of hidden dimension vectors
@return the list of predicted logprobabilities for each of the provided ref y in Y
"""
assert(len(X) == len(Y))
assert(all(len(x) == self.input_length for x in X))
if structural:
dy.renew_cg()
W = dy.parameter(self.hidden_weights)
E = dy.parameter(self.input_embeddings)
A = dy.parameter(self.action_weights)
batched_X = zip(*X) #transposes the X matrix
embeddings = [dy.pick_batch(E, xcolumn) for xcolumn in batched_X]
xdense = dy.concatenate(embeddings)
preds = dy.pickneglogsoftmax_batch(A * dy.tanh( W * xdense ),Y).value()
return [-ypred for ypred in preds]
else:#lexical
if self.tied:
dy.renew_cg()
W = dy.parameter(self.hidden_weights)
E = dy.parameter(self.input_embeddings)
batched_X = zip(*X) #transposes the X matrix
embeddings = [dy.pick_batch(E, xcolumn) for xcolumn in batched_X]
xdense = dy.concatenate(embeddings)
preds = dy.pickneglogsoftmax_batch(E * dy.tanh( W * xdense ),Y).value()
return [-ypred for ypred in preds]
else:
dy.renew_cg()
O = dy.parameter(self.output_embeddings)
W = dy.parameter(self.hidden_weights)
E = dy.parameter(self.input_embeddings)
batched_X = zip(*X) #transposes the X matrix
embeddings = [dy.pick_batch(E, xcolumn) for xcolumn in batched_X]
xdense = dy.concatenate(embeddings)
preds = dy.pickneglogsoftmax_batch(O * dy.tanh( W * xdense ),Y).value()
return [-ypred for ypred in preds]
def _embed_word(self, word, is_batched):
if is_batched:
embedding = dy.pick_batch(
self.embeddings,
word) if self.is_dense else self.embeddings.batch(word)
else:
embedding = dy.pick(
self.embeddings,
index=word) if self.is_dense else self.embeddings[word]
return embedding
def on_calc_additional_loss(self, translator_loss):
if not self.learn_segmentation or self.segment_decisions is None:
return None
reward = -translator_loss["mle"]
if not self.log_reward:
reward = dy.exp(reward)
reward = dy.nobackprop(reward)
# Make sure that reward is not scalar, but rather based on the each batch item
assert reward.dim()[1] == len(self.src_sent)
# Mask
enc_mask = self.enc_mask.get_active_one_mask().transpose() if self.enc_mask is not None else None
# Compose the lose
ret = LossBuilder()
## Length prior
alpha = self.length_prior_alpha.value() if self.length_prior_alpha is not None else 0
if alpha > 0:
reward += self.segment_length_prior * alpha
# reward z-score normalization
if self.z_normalization:
reward = dy.cdiv(reward-dy.mean_batches(reward), dy.std_batches(reward) + EPS)
## Baseline Loss
if self.use_baseline:
baseline_loss = []
for i, baseline in enumerate(self.bs):
loss = dy.squared_distance(reward, baseline)
if enc_mask is not None:
loss = dy.cmult(dy.inputTensor(enc_mask[i], batched=True), loss)
baseline_loss.append(loss)
ret.add_loss("Baseline", dy.esum(baseline_loss))
if self.print_sample:
print(dy.exp(self.segment_logsoftmaxes[i]).npvalue().transpose()[0])
## Reinforce Loss
lmbd = self.lmbd.value()
if lmbd > 0.0:
reinforce_loss = []
# Calculating the loss of the baseline and reinforce
for i in range(len(self.segment_decisions)):
ll = dy.pick_batch(self.segment_logsoftmaxes[i], self.segment_decisions[i])
if self.use_baseline:
r_i = reward - dy.nobackprop(self.bs[i])
else:
r_i = reward
if enc_mask is not None:
ll = dy.cmult(dy.inputTensor(enc_mask[i], batched=True), ll)
reinforce_loss.append(r_i * -ll)
loss = dy.esum(reinforce_loss) * lmbd
ret.add_loss("Reinforce", loss)
if self.confidence_penalty:
ls_loss = self.confidence_penalty(self.segment_logsoftmaxes, enc_mask)
ret.add_loss("Confidence Penalty", ls_loss)
# Total Loss
return ret
def next(self, w, c, test=True, state=None):
e = dy.pick_batch(self.E, w)
if not test:
e = dy.dropout_dim(e, 0, self.wdr)
# Run LSTM
if state is None:
self.ds = self.ds.add_input(e)
next_state = self.ds
else:
next_state = state.add_input(e)
h = next_state.output()
return h, e, next_state
def next(self, word_idx, context, train, cur_state=None):
embs = dy.pick_batch(self.E, word_idx)
if train:
embs = dy.dropout_dim(embs, 0, self.word_dropout)
x = dy.concatenate([embs, context])
if cur_state is None:
self.dec_state = self.dec_state.add_input(x)
next_state = self.dec_state
else:
next_state = cur_state.add_input(x)
hidden = next_state.output()
return hidden, embs, next_state
def learn(self, batch_size):
if self.prioritized:
if not self.memory.is_full(): return -np.inf
indices, exps, weights = self.memory.sample(batch_size, self.beta)
else:
exps = self.memory.sample(batch_size)
obss, actions, rewards, obs_nexts, dones = self._process(exps)
dy.renew_cg()
target_network = self.target_network if self.use_double_dqn else self.network
if self.dueling:
target_values, v = target_network(obs_nexts, batched=True)
target_values = target_values.npvalue() + v.npvalue()
else:
target_values = target_network(obs_nexts, batched=True)
target_values = target_values.npvalue()
target_values = np.max(target_values, axis=0)
target_values = rewards + self.reward_decay * (target_values *
(1 - dones))
dy.renew_cg()
if self.dueling:
all_values_expr, v = self.network(obss, batched=True)
else:
all_values_expr = self.network(obss, batched=True)
picked_values = dy.pick_batch(all_values_expr, actions)
diff = (picked_values + v if self.dueling else
picked_values) - dy.inputTensor(target_values, batched=True)
if self.prioritized:
self.memory.update(indices, np.transpose(np.abs(diff.npvalue())))
losses = dy.pow(diff, dy.constant(1, 2))
if self.prioritized:
losses = dy.cmult(losses, dy.inputTensor(weights, batched=True))
loss = dy.sum_batches(losses)
loss_value = loss.npvalue()
loss.backward()
self.trainer.update()
self.epsilon = max(self.epsilon - self.epsilon_decrease,
self.epsilon_lower)
if self.prioritized:
self.beta = min(self.beta + self.beta_increase, 1.)
self.learn_step += 1
if self.use_double_dqn and self.learn_step % self.n_replace_target == 0:
self.target_network.update(self.network)
return loss_value
def next(self, w, c, test=True, state=None):
if isinstance(w, dy.Expression):
e = w
else:
e = dy.pick_batch(self.E, w)
if not test:
e = dy.dropout_dim(e, 0, self.wdr)
x = dy.concatenate([e, c])
# Run LSTM
if state is None:
self.ds = self.ds.add_input(x)
next_state = self.ds
else:
next_state = state.add_input(x)
h = next_state.output()
return h, e, next_state
def embed_sentence(self, ws, pwords, ts, chars, is_train):
cembed = [dy.lookup_batch(self.clookup, c) for c in chars]
char_fwd, char_bckd = self.char_lstm.builder_layers[0][0].initial_state().transduce(cembed)[-1], \
self.char_lstm.builder_layers[0][1].initial_state().transduce(reversed(cembed))[-1]
crnn = dy.reshape(dy.concatenate_cols([char_fwd, char_bckd]), (self.options.we, ws.shape[0] * ws.shape[1]))
cnn_reps = [list() for _ in range(len(ws))]
for i in range(ws.shape[0]):
cnn_reps[i] = dy.pick_batch(crnn, [i * ws.shape[1] + j for j in range(ws.shape[1])], 1)
wembed = [dy.lookup_batch(self.wlookup, ws[i]) + dy.lookup_batch(self.elookup, pwords[i]) + cnn_reps[i] for i in range(len(ws))]
posembed = [dy.lookup_batch(self.tlookup, ts[i]) for i in range(len(ts))]
if (not is_train) or self.options.dropout == 0:
return [dy.concatenate([wembed[i], posembed[i]]) for i in range(len(ts))]
else:
emb_masks = self.generate_emb_mask(ws.shape[0], ws.shape[1])
return [dy.concatenate([dy.cmult(w, wm), dy.cmult(pos, posm)]) for w, pos, (wm, posm) in
zip(wembed, posembed, emb_masks)]
def process_batch(self, batch, training=False):
self.TRAINING_ITER = training
self.DROPOUT = self.args.dropout if (
self.TRAINING_ITER and self.args.dropout > 0) else None
self.BATCH_SIZE = len(batch)
sents, masks = self.vocab.batchify(batch)
self.instantiate_parameters()
init_state = self.rnn.initial_state(mb_size=self.BATCH_SIZE,
dropout=self.DROPOUT)
# embeddings = [dynet.reshape(dynet.select_cols(self.vocab_lookup, toks), (self.args.dim,), self.BATCH_SIZE)
# embeddings = [dynet.reshape(dynet.transpose(dynet.select_rows(self.vocab_R, toks)), (self.args.dim*2,), self.BATCH_SIZE)
embeddings = [dynet.pick_batch(self.vocab_R, toks) for toks in sents]
outputs = init_state.transduce(embeddings)
outputs = [
dynet.to_device(out, self.args.param_device) for out in outputs
]
if self.DROPOUT:
y_ts = [dynet.cmult(y_t, self.dropout_mask_y_t) for y_t in outputs]
else:
y_ts = outputs
r_ts = [
dynet.affine_transform([
self.vocab_bias, self.vocab_R,
dynet.tanh(dynet.affine_transform([self.bias, self.R, y_t]))
]) for y_t in y_ts
]
errs = [
dynet.pickneglogsoftmax_batch(r_t, toks)
for r_t, toks in zip(r_ts, sents[1:])
]
for tok_i, (err, mask) in enumerate(zip(errs, masks[1:])):
if min(mask) == 0:
errs[tok_i] = err * dynet.inputTensor(
mask, batched=True, device=self.args.param_device)
err = dynet.esum(errs)
char_count = [1 + len(self.vocab.pp(sent[1:-1])) for sent in batch]
word_count = [len(sent[1:]) for sent in batch]
# word_count = [2+self.vocab.pp(sent[1:-1]).count(' ') for sent in batch]
return {"loss": err, "charcount": char_count, "wordcount": word_count}
def compute_loss(self, words, extwords, tags, true_arcs, true_labels):
arc_logits, rel_logits = self.forward(words, extwords, tags, True)
seq_len = len(true_arcs)
targets_1D = dynet_flatten_numpy(true_arcs)
flat_arc_logits = dy.reshape(arc_logits, (seq_len, ), seq_len)
losses = dy.pickneglogsoftmax_batch(flat_arc_logits, targets_1D)
arc_loss = dy.sum_batches(losses)
flat_rel_logits = dy.reshape(rel_logits, (seq_len, self.rel_size),
seq_len)
partial_rel_logits = dy.pick_batch(flat_rel_logits, targets_1D)
targets_rel1D = dynet_flatten_numpy(true_labels)
losses = dy.pickneglogsoftmax_batch(partial_rel_logits, targets_rel1D)
rel_loss = dy.sum_batches(losses)
loss = arc_loss + rel_loss
return loss
def decode_loss(self, src_encodings, masks, tgt_seqs, sents_len):
"""
:param tgt_seqs: (tgt_heads, tgt_labels): list (length=batch_size) of (src_len)
"""
tgt_heads, tgt_labels = tgt_seqs
src_len = len(tgt_heads[0])
batch_size = len(tgt_heads)
np_tgt_heads = np.array(tgt_heads).flatten() # (src_len * batch_size)
np_tgt_labels = np.array(tgt_labels).flatten()
np_masks = np.array(masks).transpose().flatten()
masks_expr = dy.inputVector(np_masks)
masks_expr = dy.reshape(masks_expr, (1, ),
batch_size=src_len * batch_size)
s_arc, s_label = self.cal_scores(
src_encodings
) # (src_len, src_len, bs), ([(src_len, src_len, bs)])
s_arc = dy.select_cols(s_arc, range(1, src_len + 1))
s_label = [
dy.select_cols(label, range(1, src_len + 1)) for label in s_label
]
s_arc_value = s_arc.npvalue()
s_arc_choice = np.argmax(
s_arc_value,
axis=0).transpose().flatten() # (src_len * batch_size)
s_pick_labels = [
dy.pick_batch(
dy.reshape(score, (src_len + 1, ),
batch_size=src_len * batch_size), s_arc_choice)
for score in s_label
]
s_argmax_labels = dy.concatenate(s_pick_labels,
d=0) # n_labels, src_len * batch_size
reshape_s_arc = dy.reshape(s_arc, (src_len + 1, ),
batch_size=src_len * batch_size)
arc_loss = dy.pickneglogsoftmax_batch(reshape_s_arc, np_tgt_heads)
arc_loss = arc_loss * masks_expr
label_loss = dy.pickneglogsoftmax_batch(s_argmax_labels, np_tgt_labels)
label_loss = label_loss * masks_expr
loss = dy.sum_batches(arc_loss + label_loss)
return loss
def learn(self, batch_size):
if self.prioritized:
if not self.memory.is_full(): return -np.inf
indices, exps, weights = self.memory.sample(batch_size, self.beta)
else:
exps = self.memory.sample(batch_size)
obss, actions, rewards, obs_nexts, dones = self._process(exps)
dy.renew_cg()
target_network = self.target_network if self.use_double_dqn else self.network
if self.dueling:
target_values, v = target_network(obs_nexts, batched=True)
target_values = target_values.npvalue() + v.npvalue()
else:
target_values = target_network(obs_nexts, batched=True)
target_values = target_values.npvalue()
target_values = np.max(target_values, axis=0)
target_values = rewards + self.reward_decay * (target_values * (1 - dones))
dy.renew_cg()
if self.dueling:
all_values_expr, v = self.network(obss, batched=True)
else:
all_values_expr = self.network(obss, batched=True)
picked_values = dy.pick_batch(all_values_expr, actions)
diff = (picked_values + v if self.dueling else picked_values) - dy.inputTensor(target_values, batched=True)
if self.prioritized:
self.memory.update(indices, np.transpose(np.abs(diff.npvalue())))
losses = dy.pow(diff, dy.constant(1, 2))
if self.prioritized:
losses = dy.cmult(losses, dy.inputTensor(weights, batched=True))
loss = dy.sum_batches(losses)
loss_value = loss.npvalue()
loss.backward()
self.trainer.update()
self.epsilon = max(self.epsilon - self.epsilon_decrease, self.epsilon_lower)
if self.prioritized:
self.beta = min(self.beta + self.beta_increase, 1.)
self.learn_step += 1
if self.use_double_dqn and self.learn_step % self.n_replace_target == 0:
self.target_network.update(self.network)
return loss_value
# regular lookup
a = lp[1].npvalue()
b = lp[2].npvalue()
c = lp[3].npvalue()
# batch lookup instead of single elements.
# two ways of doing this.
abc1 = dy.lookup_batch(lp, [1,2,3])
print(abc1.npvalue())
abc2 = lp.batch([1,2,3])
print(abc2.npvalue())
print(np.hstack([a,b,c]))
# use pick and pickneglogsoftmax in batch mode
# (must be used in conjunction with lookup_batch):
print("\nPick")
W = dy.parameter( m.add_parameters((5, 10)) )
h = W * lp.batch([1,2,3])
print(h.npvalue())
print(dy.pick_batch(h,[1,2,3]).npvalue())
print(dy.pick(W*lp[1],1).value(), dy.pick(W*lp[2],2).value(), dy.pick(W*lp[3],3).value())
# using pickneglogsoftmax_batch
print("\nPick neg log softmax")
print((-dy.log(dy.softmax(h))).npvalue())
print(dy.pickneglogsoftmax_batch(h,[1,2,3]).npvalue())