def decode(model, dec_lstm, vectors, output):
output = [EOS] + list(output) + [EOS]
output = [char2int[c] for c in output]
w = pc.parameter(model["decoder_w"])
b = pc.parameter(model["decoder_b"])
s = dec_lstm.initial_state().add_input(pc.vecInput(STATE_SIZE * 2))
loss = []
for char in output:
vector = attend(model, vectors, s)
s = s.add_input(vector)
out_vector = w * s.output() + b
probs = pc.softmax(out_vector)
loss.append(-pc.log(pc.pick(probs, char)))
loss = pc.esum(loss)
return loss
def decode(model, dec_lstm, vectors, output):
output = [EOS] + list(output) + [EOS]
output = [char2int[c] for c in output]
w = pc.parameter(model["decoder_w"])
b = pc.parameter(model["decoder_b"])
s = dec_lstm.initial_state().add_input(pc.vecInput(STATE_SIZE*2))
loss = []
for char in output:
vector = attend(model, vectors, s)
s = s.add_input(vector)
out_vector = w * s.output() + b
probs = pc.softmax(out_vector)
loss.append(-pc.log(pc.pick(probs, char)))
loss = pc.esum(loss)
return loss
def one_word_loss(model, encoder_frnn, encoder_rrnn, decoder_rnn, lemma, feats, word, alphabet_index, aligned_pair,
feat_index, feature_types):
pc.renew_cg()
# read the parameters
char_lookup = model["char_lookup"]
feat_lookup = model["feat_lookup"]
R = pc.parameter(model["R"])
bias = pc.parameter(model["bias"])
padded_lemma = BEGIN_WORD + lemma + END_WORD
# convert characters to matching embeddings
lemma_char_vecs = []
for char in padded_lemma:
try:
lemma_char_vecs.append(char_lookup[alphabet_index[char]])
except KeyError:
# handle UNK
lemma_char_vecs.append(char_lookup[alphabet_index[UNK]])
# convert features to matching embeddings, if UNK handle properly
feat_vecs = []
for feat in sorted(feature_types):
# TODO: is it OK to use same UNK for all feature types? and for unseen feats as well?
# if this feature has a value, take it from the lookup. otherwise use UNK
if feat in feats:
feat_str = feat + ':' + feats[feat]
try:
feat_vecs.append(feat_lookup[feat_index[feat_str]])
except KeyError:
# handle UNK or dropout
feat_vecs.append(feat_lookup[feat_index[UNK_FEAT]])
else:
feat_vecs.append(feat_lookup[feat_index[UNK_FEAT]])
feats_input = pc.concatenate(feat_vecs)
# BiLSTM forward pass
s_0 = encoder_frnn.initial_state()
s = s_0
frnn_outputs = []
for c in lemma_char_vecs:
s = s.add_input(c)
frnn_outputs.append(s.output())
# BiLSTM backward pass
s_0 = encoder_rrnn.initial_state()
s = s_0
rrnn_outputs = []
for c in reversed(lemma_char_vecs):
s = s.add_input(c)
rrnn_outputs.append(s.output())
# BiLTSM outputs
blstm_outputs = []
lemma_char_vecs_len = len(lemma_char_vecs)
for i in xrange(lemma_char_vecs_len):
blstm_outputs.append(pc.concatenate([frnn_outputs[i], rrnn_outputs[lemma_char_vecs_len - i - 1]]))
# initialize the decoder rnn
s_0 = decoder_rnn.initial_state()
s = s_0
# set prev_output_vec for first lstm step as BEGIN_WORD
prev_output_vec = char_lookup[alphabet_index[BEGIN_WORD]]
prev_char_vec = char_lookup[alphabet_index[BEGIN_WORD]]
loss = []
# i is input index, j is output index
i = 0
j = 0
# go through alignments, progress j when new output is introduced, progress i when new char is seen on lemma (no ~)
# TODO: try sutskever flip trick?
# TODO: attention on the lemma chars/feats could help here?
aligned_lemma, aligned_word = aligned_pair
aligned_lemma += END_WORD
aligned_word += END_WORD
# run through the alignments
for index, (input_char, output_char) in enumerate(zip(aligned_lemma, aligned_word)):
possible_outputs = []
# feedback, i, j, blstm[i], feats
decoder_input = pc.concatenate([prev_output_vec,
prev_char_vec,
# char_lookup[alphabet_index[str(i)]],
# char_lookup[alphabet_index[str(j)]],
blstm_outputs[i],
feats_input])
# if reached the end word symbol
if output_char == END_WORD:
s = s.add_input(decoder_input)
decoder_rnn_output = s.output()
probs = pc.softmax(R * decoder_rnn_output + bias)
# compute local loss
loss.append(-pc.log(pc.pick(probs, alphabet_index[END_WORD])))
continue
# if there is no prefix, step
if padded_lemma[i] == BEGIN_WORD and aligned_lemma[index] != ALIGN_SYMBOL:
# perform rnn step
# feedback, i, j, blstm[i], feats
s = s.add_input(decoder_input)
decoder_rnn_output = s.output()
probs = pc.softmax(R * decoder_rnn_output + bias)
# compute local loss
loss.append(-pc.log(pc.pick(probs, alphabet_index[STEP])))
# prepare for the next iteration - "feedback"
prev_output_vec = char_lookup[alphabet_index[STEP]]
prev_char_vec = char_lookup[alphabet_index[EPSILON]]
i += 1
# if there is new output
if aligned_word[index] != ALIGN_SYMBOL:
decoder_input = pc.concatenate([prev_output_vec,
prev_char_vec,
# char_lookup[alphabet_index[str(i)]],
# char_lookup[alphabet_index[str(j)]],
blstm_outputs[i],
feats_input])
# copy i action - maybe model as a single action?
if padded_lemma[i] == aligned_word[j]:
possible_outputs.append(str(i))
possible_outputs.append(padded_lemma[i])
else:
possible_outputs.append(aligned_word[index])
# perform rnn step
s = s.add_input(decoder_input)
decoder_rnn_output = s.output()
probs = pc.softmax(R * decoder_rnn_output + bias)
local_loss = pc.scalarInput(0)
max_output_loss = -pc.log(pc.pick(probs, alphabet_index[possible_outputs[0]]))
max_likelihood_output = possible_outputs[0]
# sum over all correct output possibilities and pick feedback output to be the one with the highest
# probability
for output in possible_outputs:
neg_log_likelihood = -pc.log(pc.pick(probs, alphabet_index[output]))
if neg_log_likelihood < max_output_loss:
max_likelihood_output = output
max_output_loss = neg_log_likelihood
local_loss += neg_log_likelihood
loss.append(local_loss)
# prepare for the next iteration - "feedback"
prev_output_vec = char_lookup[alphabet_index[max_likelihood_output]]
prev_char_vec = char_lookup[alphabet_index[aligned_word[index]]]
j += 1
# now check if it's time to progress on input
if i < len(padded_lemma) - 1 and aligned_lemma[index + 1] != ALIGN_SYMBOL:
# perform rnn step
# feedback, i, j, blstm[i], feats
decoder_input = pc.concatenate([prev_output_vec,
prev_char_vec,
# char_lookup[alphabet_index[str(i)]],
# char_lookup[alphabet_index[str(j)]],
blstm_outputs[i],
feats_input])
s = s.add_input(decoder_input)
decoder_rnn_output = s.output()
probs = pc.softmax(R * decoder_rnn_output + bias)
# compute local loss
loss.append(-pc.log(pc.pick(probs, alphabet_index[STEP])))
# prepare for the next iteration - "feedback"
prev_output_vec = char_lookup[alphabet_index[STEP]]
prev_char_vec = char_lookup[alphabet_index[EPSILON]]
i += 1
# TODO: maybe here a "special" loss function is appropriate?
# loss = esum(loss)
loss = pc.average(loss)
return loss
def train(
feature_mapper,
word_dims,
tag_dims,
lstm_units,
hidden_units,
epochs,
batch_size,
train_data_file,
dev_data_file,
model_save_file,
droprate,
unk_param,
alpha=1.0,
beta=0.0,
):
start_time = time.time()
fm = feature_mapper
word_count = fm.total_words()
tag_count = fm.total_tags()
network = Network(
word_count=word_count,
tag_count=tag_count,
word_dims=word_dims,
tag_dims=tag_dims,
lstm_units=lstm_units,
hidden_units=hidden_units,
struct_out=2,
label_out=fm.total_label_actions(),
droprate=droprate,
)
network.init_params()
print('Hidden units: {}, per-LSTM units: {}'.format(
hidden_units,
lstm_units,
))
print('Embeddings: word={} tag={}'.format(
(word_count, word_dims),
(tag_count, tag_dims),
))
print('Dropout rate: {}'.format(droprate))
print('Parameters initialized in [-0.01, 0.01]')
print('Random UNKing parameter z = {}'.format(unk_param))
print('Exploration: alpha={} beta={}'.format(alpha, beta))
training_data = fm.gold_data_from_file(train_data_file)
num_batches = -(-len(training_data) // batch_size)
print('Loaded {} training sentences ({} batches of size {})!'.format(
len(training_data),
num_batches,
batch_size,
))
parse_every = -(-num_batches // 4)
dev_trees = PhraseTree.load_treefile(dev_data_file)
print('Loaded {} validation trees!'.format(len(dev_trees)))
best_acc = FScore()
for epoch in xrange(1, epochs + 1):
print('........... epoch {} ...........'.format(epoch))
total_cost = 0.0
total_states = 0
training_acc = FScore()
np.random.shuffle(training_data)
for b in xrange(num_batches):
batch = training_data[(b * batch_size):((b + 1) * batch_size)]
explore = [
Parser.exploration(
example,
fm,
network,
alpha=alpha,
beta=beta,
) for example in batch
]
for (_, acc) in explore:
training_acc += acc
batch = [example for (example, _) in explore]
pycnn.renew_cg()
network.prep_params()
errors = []
for example in batch:
## random UNKing ##
for (i, w) in enumerate(example['w']):
if w <= 2:
continue
freq = fm.word_freq_list[w]
drop_prob = unk_param / (unk_param + freq)
r = np.random.random()
if r < drop_prob:
example['w'][i] = 0
fwd, back = network.evaluate_recurrent(
example['w'],
example['t'],
)
for (left,
right), correct in example['struct_data'].items():
scores = network.evaluate_struct(
fwd, back, left, right)
probs = pycnn.softmax(scores)
loss = -pycnn.log(pycnn.pick(probs, correct))
errors.append(loss)
total_states += len(example['struct_data'])
for (left,
right), correct in example['label_data'].items():
scores = network.evaluate_label(fwd, back, left, right)
probs = pycnn.softmax(scores)
loss = -pycnn.log(pycnn.pick(probs, correct))
errors.append(loss)
total_states += len(example['label_data'])
batch_error = pycnn.esum(errors)
total_cost += batch_error.scalar_value()
batch_error.backward()
network.trainer.update()
mean_cost = total_cost / total_states
print(
'\rBatch {} Mean Cost {:.4f} [Train: {}]'.format(
b,
mean_cost,
training_acc,
),
end='',
)
sys.stdout.flush()
if ((b + 1) % parse_every) == 0 or b == (num_batches - 1):
dev_acc = Parser.evaluate_corpus(
dev_trees,
fm,
network,
)
print(' [Val: {}]'.format(dev_acc))
if dev_acc > best_acc:
best_acc = dev_acc
network.save(model_save_file)
print(' [saved model: {}]'.format(model_save_file))
current_time = time.time()
runmins = (current_time - start_time) / 60.
print(' Elapsed time: {:.2f}m'.format(runmins))
def pick_neg_log(self, pred, gold):
return -pycnn.log(pycnn.pick(pred, gold))
def one_word_loss(model, encoder_frnn, encoder_rrnn, decoder_rnn, lemma, feats, word, alphabet_index, aligned_pair,
feat_index, feature_types):
pc.renew_cg()
# read the parameters
char_lookup = model["char_lookup"]
feat_lookup = model["feat_lookup"]
R = pc.parameter(model["R"])
bias = pc.parameter(model["bias"])
padded_lemma = BEGIN_WORD + lemma + END_WORD
# convert characters to matching embeddings
lemma_char_vecs = []
for char in padded_lemma:
try:
lemma_char_vecs.append(char_lookup[alphabet_index[char]])
except KeyError:
# handle UNK
lemma_char_vecs.append(char_lookup[alphabet_index[UNK]])
# convert features to matching embeddings, if UNK handle properly
feat_vecs = []
for feat in sorted(feature_types):
# TODO: is it OK to use same UNK for all feature types? and for unseen feats as well?
# if this feature has a value, take it from the lookup. otherwise use UNK
if feat in feats:
feat_str = feat + ':' + feats[feat]
try:
feat_vecs.append(feat_lookup[feat_index[feat_str]])
except KeyError:
# handle UNK or dropout
feat_vecs.append(feat_lookup[feat_index[UNK_FEAT]])
else:
feat_vecs.append(feat_lookup[feat_index[UNK_FEAT]])
feats_input = pc.concatenate(feat_vecs)
# BiLSTM forward pass
s_0 = encoder_frnn.initial_state()
s = s_0
frnn_outputs = []
for c in lemma_char_vecs:
s = s.add_input(c)
frnn_outputs.append(s.output())
# BiLSTM backward pass
s_0 = encoder_rrnn.initial_state()
s = s_0
rrnn_outputs = []
for c in reversed(lemma_char_vecs):
s = s.add_input(c)
rrnn_outputs.append(s.output())
# BiLTSM outputs
blstm_outputs = []
lemma_char_vecs_len = len(lemma_char_vecs)
for i in xrange(lemma_char_vecs_len):
blstm_outputs.append(pc.concatenate([frnn_outputs[i], rrnn_outputs[lemma_char_vecs_len - i - 1]]))
# initialize the decoder rnn
s_0 = decoder_rnn.initial_state()
s = s_0
# set prev_output_vec for first lstm step as BEGIN_WORD
prev_output_vec = char_lookup[alphabet_index[BEGIN_WORD]]
prev_char_vec = char_lookup[alphabet_index[BEGIN_WORD]]
loss = []
# i is input index, j is output index
i = 0
j = 0
# go through alignments, progress j when new output is introduced, progress i when new char is seen on lemma (no ~)
# TODO: try sutskever flip trick?
# TODO: attention on the lemma chars/feats could help here?
aligned_lemma, aligned_word = aligned_pair
aligned_lemma += END_WORD
aligned_word += END_WORD
# run through the alignments
for index, (input_char, output_char) in enumerate(zip(aligned_lemma, aligned_word)):
possible_outputs = []
# feedback, i, j, blstm[i], feats
decoder_input = pc.concatenate([prev_output_vec,
prev_char_vec,
# char_lookup[alphabet_index[str(i)]],
# char_lookup[alphabet_index[str(j)]],
blstm_outputs[i],
feats_input])
# if reached the end word symbol
if output_char == END_WORD:
s = s.add_input(decoder_input)
decoder_rnn_output = s.output()
probs = pc.softmax(R * decoder_rnn_output + bias)
# compute local loss
loss.append(-pc.log(pc.pick(probs, alphabet_index[END_WORD])))
continue
# if there is no prefix, step
if padded_lemma[i] == BEGIN_WORD and aligned_lemma[index] != ALIGN_SYMBOL:
# perform rnn step
# feedback, i, j, blstm[i], feats
s = s.add_input(decoder_input)
decoder_rnn_output = s.output()
probs = pc.softmax(R * decoder_rnn_output + bias)
# compute local loss
loss.append(-pc.log(pc.pick(probs, alphabet_index[STEP])))
# prepare for the next iteration - "feedback"
prev_output_vec = char_lookup[alphabet_index[STEP]]
prev_char_vec = char_lookup[alphabet_index[EPSILON]]
i += 1
# if there is new output
if aligned_word[index] != ALIGN_SYMBOL:
decoder_input = pc.concatenate([prev_output_vec,
prev_char_vec,
# char_lookup[alphabet_index[str(i)]],
# char_lookup[alphabet_index[str(j)]],
blstm_outputs[i],
feats_input])
# copy i action - maybe model as a single action?
if padded_lemma[i] == aligned_word[j]:
possible_outputs.append(str(i))
possible_outputs.append(padded_lemma[i])
else:
possible_outputs.append(aligned_word[index])
# perform rnn step
s = s.add_input(decoder_input)
decoder_rnn_output = s.output()
probs = pc.softmax(R * decoder_rnn_output + bias)
local_loss = pc.scalarInput(0)
max_output_loss = -pc.log(pc.pick(probs, alphabet_index[possible_outputs[0]]))
max_likelihood_output = possible_outputs[0]
# sum over all correct output possibilities and pick feedback output to be the one with the highest
# probability
for output in possible_outputs:
neg_log_likelihood = -pc.log(pc.pick(probs, alphabet_index[output]))
if neg_log_likelihood < max_output_loss:
max_likelihood_output = output
max_output_loss = neg_log_likelihood
local_loss += neg_log_likelihood
loss.append(local_loss)
# prepare for the next iteration - "feedback"
prev_output_vec = char_lookup[alphabet_index[max_likelihood_output]]
prev_char_vec = char_lookup[alphabet_index[aligned_word[index]]]
j += 1
# now check if it's time to progress on input
if i < len(padded_lemma) - 1 and aligned_lemma[index + 1] != ALIGN_SYMBOL:
# perform rnn step
# feedback, i, j, blstm[i], feats
decoder_input = pc.concatenate([prev_output_vec,
prev_char_vec,
# char_lookup[alphabet_index[str(i)]],
# char_lookup[alphabet_index[str(j)]],
blstm_outputs[i],
feats_input])
s = s.add_input(decoder_input)
decoder_rnn_output = s.output()
probs = pc.softmax(R * decoder_rnn_output + bias)
# compute local loss
loss.append(-pc.log(pc.pick(probs, alphabet_index[STEP])))
# prepare for the next iteration - "feedback"
prev_output_vec = char_lookup[alphabet_index[STEP]]
prev_char_vec = char_lookup[alphabet_index[EPSILON]]
i += 1
# TODO: maybe here a "special" loss function is appropriate?
# loss = esum(loss)
loss = pc.average(loss)
return loss