def predict_one(self, model, datum, vocab):
"""
:param model:
:param datum: a (lemma chars)
:param vocab:
:return:
"""
lemma, feats = datum
x_enc_inputs = BaseMorphData.lemma_chars_to_ids(lemma, vocab)
x_feature_ids = BaseMorphData.feat_d_to_ids(feats, vocab)
# TODO: this works only with batch=1; consider batching and transposition instead of viewing
x_enc_inputs_v = cuda_if_gpu(
Variable(torch.LongTensor(x_enc_inputs)).view(-1, 1))
x_feature_ids_v = cuda_if_gpu(
Variable(torch.LongTensor(x_feature_ids)).view(-1, 1))
pred_form_char_ids = model.predict_one_instance(
(x_enc_inputs_v, x_feature_ids_v))
pred_form_chars = BaseMorphData.unvectorize_seq(
pred_form_char_ids, vocab)
pred_form = ''.join([
ch for ch in pred_form_chars
if not ch in MORPH_OUTPUT_CLASSES_START
])
return pred_form
def predict(self, input_var):
enc_inputs_v, feature_ids_v = input_var
src_len = enc_inputs_v.size()[0]
dec_len = self.max_tgt_len
batch_size = enc_inputs_v.size()[1]
batch_indices = cuda_if_gpu(torch.arange(0, batch_size).long())
# Embedding lookup - chars
char_embeddings = self.embedding_lookup(enc_inputs_v) # SL x B x E
# Encode
encoder_outputs, encoder_hidden = self.encoder(char_embeddings,
batch_size=batch_size)
if self.encoder.rnn.bidirectional:
# squeeze the outputs of a BiRNN encoder
# outputs: # SL x B x enc_dim
# hidden: # 1 x B x enc_dim
encoder_outputs = sum_bidirectional_outputs(
encoder_outputs, self.encoder.rnn)
encoder_hidden = sum_bidirectional_state(encoder_hidden)
# Embedding lookup - features
# 1 x self.num_embedding_feats * self.embedding_dim
feat_embeddings = self.embedding_lookup(feature_ids_v).view(
batch_size, -1)
feat_embeddings = F.relu(self.feat_proj_layer(feat_embeddings))
# Initialize decoder
# 1 x B x enc_dim
dec_hidden = encoder_hidden[:1] # 1 x B x enc_dim
# Decode
prev_y_id = self.BOS_ID
prev_y_var = cuda_if_gpu(
Variable(torch.LongTensor([prev_y_id] * batch_size)))
dec_ids = torch.zeros(dec_len, batch_size)
# attn_ids_v = cuda_if_gpu(torch.zeros(batch_size).long())
for di in range(dec_len):
prev_y_embeddings = self.embedding_lookup(prev_y_var)
# prev_y_embeddings_drop = self.dropout(prev_y)
# 1 x B x dec_input_dim
dec_input = torch.cat((prev_y_embeddings, feat_embeddings), 1)
dec_output, dec_hidden, attn_weights = self.decoder(
dec_input, dec_hidden, encoder_outputs)
unnormalized_scores = self.output_layer(dec_output[0])
logits = self.softmax(unnormalized_scores)
topval, topids = logits.data.topk(1)
prev_y_ids = topids.squeeze(1) # B
# bookeeping
dec_ids[di] = prev_y_ids
# attn_w.append(attn_weights.data)
prev_y_var = cuda_if_gpu(Variable(prev_y_ids))
return dec_ids
def predict_one_instance(self, input_var):
enc_inputs_v, feature_ids_v = input_var
src_len = enc_inputs_v.size()[0]
# Embedding lookup - chars
char_embeddings = self.embedding_lookup(enc_inputs_v) # SL x B x E
# Encode
# SL x B x 2*H, num_layers * num_directions x B x H
enc_outputs, enc_hidden, enc_cell = self.encoder(
rnn_input=char_embeddings, batch_size=1)
# Embedding lookup - features
# 1 x self.num_embedding_feats * self.embedding_dim
feat_embeddings = self.embedding_lookup(feature_ids_v).view(1, -1)
feat_embeddings = F.relu(self.feat_proj_layer(feat_embeddings))
# Initialize decoder
# 1 x B x enc_dim
dec_hidden = sum_bidirectional_state(enc_hidden)
dec_cell = sum_bidirectional_state(enc_cell)
# Decode
dec_ids = []
prev_y_id = BOS_ID
prev_y_var = cuda_if_gpu(Variable(torch.LongTensor([prev_y_id])))
total_decoded_len = 0
attn_idx = 0
while (prev_y_id != EOS_ID
and total_decoded_len < self.max_tgt_len * 2):
prev_y_embeddings = self.embedding_mat(prev_y_var)
# prev_y_embeddings_drop = self.dropout(prev_y)
attended_encoder_output = enc_outputs[attn_idx]
# 1 x B x dec_input_dim
dec_input = torch.cat(
(prev_y_embeddings, feat_embeddings, attended_encoder_output),
1).unsqueeze(0)
dec_output, dec_hidden, dec_cell = self.decoder(
dec_input, dec_hidden, dec_cell)
unnormalized_scores = self.output_layer(dec_output[0])
logits = self.softmax(unnormalized_scores)
topval, topidx = logits.data.topk(1)
prev_y_id = topidx.data.cpu().numpy()[0][0]
total_decoded_len += 1
# store result
dec_ids.append(prev_y_id)
prev_y_var = cuda_if_gpu(Variable(torch.LongTensor([prev_y_id])))
# check if step or char output to promote i.
if prev_y_id == STEP_ID and attn_idx < src_len - 1:
attn_idx += 1
return dec_ids
def predict_one_instance(self, input_var):
enc_inputs_v, feature_ids_v = input_var
src_len = enc_inputs_v.size()[0]
# Embedding lookup - chars
char_embeddings = self.embedding_lookup(enc_inputs_v) # SL x B x E
# Encode
# SL x B x 2*H, num_layers * num_directions x B x H
encoder_outputs, encoder_hidden = self.encoder(char_embeddings,
batch_size=1)
if self.encoder.rnn.bidirectional:
# squeeze the outputs of a BiRNN encoder
# outputs: # SL x B x enc_dim
# hidden: # 1 x B x enc_dim
encoder_outputs = sum_bidirectional_outputs(
encoder_outputs, self.encoder.rnn)
encoder_hidden = sum_bidirectional_state(encoder_hidden)
# Embedding lookup - features
# 1 x self.num_embedding_feats * self.embedding_dim
feat_embeddings = self.embedding_lookup(feature_ids_v).view(1, -1)
feat_embeddings = F.relu(self.feat_proj_layer(feat_embeddings))
# Initialize decoder
# 1 x B x enc_dim
dec_hidden = encoder_hidden[:1] # 1 x B x enc_dim
# Decode
dec_ids = []
prev_y_id = self.BOS_ID
prev_y_var = cuda_if_gpu(Variable(torch.LongTensor([prev_y_id])))
total_decoded_len = 0
while (prev_y_id != self.EOS_ID
and total_decoded_len < self.max_tgt_len):
prev_y_embeddings = self.embedding_mat(prev_y_var)
# prev_y_embeddings_drop = self.dropout(prev_y)
# 1 x B x dec_input_dim
dec_input = torch.cat((prev_y_embeddings, feat_embeddings), 1)
dec_output, dec_hidden, attn_weights = self.decoder(
dec_input, dec_hidden, encoder_outputs)
unnormalized_scores = self.output_layer(dec_output[0])
logits = self.softmax(unnormalized_scores)
topval, topidx = logits.data.topk(1)
prev_y_id = topidx[0][0]
total_decoded_len += 1
# store result
dec_ids.append(prev_y_id)
prev_y_var = cuda_if_gpu(Variable(torch.LongTensor([prev_y_id])))
return dec_ids
def train_step(self, model, np_xy_mat_pair):
x_var = cuda_if_gpu(Variable(torch.from_numpy(np_xy_mat_pair[0])))
y_var = cuda_if_gpu(
Variable(
torch.from_numpy(np_xy_mat_pair[1]).type(
torch.FloatTensor))) # have to cast to float
logits = model.forward(x_var) #
loss_var = self.calc_loss(logits, y_var)
return loss_var
def decode_teacher(self, encoder_hidden, encoder_outputs, batch_x_feat_var,
batch_y_var):
"""
Decoding policy 1: feeding the ground truth label as a target
:param dec_input_var: ground truth labels
:param encoder_hidden: the last hidden state of the Encoder RNN; (num_layers * num_directions) x B x enc_dim
:param encoder_outputs: SL x B x enc_dim
:return:
"""
dec_len = batch_y_var.size()[0]
batch_size = batch_y_var.size()[1]
# 1 x self.num_embedding_feats * self.embedding_dim
feat_embeddings = self.embedding_lookup(batch_x_feat_var).view(
batch_size, -1)
feat_embeddings = F.relu(self.feat_proj_layer(feat_embeddings))
# 1 x B x enc_dim
dec_hidden = encoder_hidden[:
1] # (num_layers * num_directions) x B x enc_dim
# 1 x B
prev_y = cuda_if_gpu(
Variable(torch.LongTensor([self.BOS_ID] * batch_size)))
# TL x B x Output_size
predicted_logits = cuda_if_gpu(
Variable(torch.zeros(dec_len, batch_size, self.output_size)))
# Teacher forcing: feed the target as the next input
for di in range(dec_len):
# embedding lookup of a vector of length = B; result: B x E
prev_y_embeddings = self.embedding_mat(prev_y)
# prev_y = self.dropout(prev_y) # apply dropout
# B x dec_input_dim + num_feats * embed_dim
dec_input = torch.cat((prev_y_embeddings, feat_embeddings), 1)
# 1 x B x dec_output_dim, # 1 x B x dec_output_dim, B x SL
dec_output, dec_hidden, attn_weights = self.decoder(
dec_input,
dec_hidden,
encoder_outputs,
)
unnormalized_scores = self.output_layer(dec_output[0])
logits = self.softmax(unnormalized_scores)
predicted_logits[di] = logits # store this step's output logit
prev_y = batch_y_var[di] # next input
return predicted_logits
def init_rnn_hidden(rnn, num_directions, batch_size):
hidden_state = Variable(
torch.zeros(rnn.num_layers * num_directions, batch_size,
rnn.hidden_size))
return cuda_if_gpu(hidden_state)
def init_gru_state(num_directions, enc_or_dec, batch_size):
hidden = Variable(
torch.zeros(enc_or_dec.num_layers * num_directions, batch_size,
enc_or_dec.hidden_size))
return cuda_if_gpu(hidden)
def decode_teacher(self, encoder_hidden, encoder_cell, encoder_outputs,
attn_ids_v, feat_ids, dec_targets):
"""
Decoding policy 1: feeding the ground truth label as a target
"""
dec_len = dec_targets.size()[0]
batch_size = dec_targets.size()[1]
batch_indices = cuda_if_gpu(torch.arange(0, batch_size).long())
# 1 x self.num_embedding_feats * self.embedding_dim
feat_embeddings = self.embedding_lookup(feat_ids).view(batch_size, -1)
feat_embeddings = F.relu(self.feat_proj_layer(feat_embeddings))
# 1 x B x enc_dim
dec_hidden = sum_bidirectional_state(encoder_hidden)
dec_cell = sum_bidirectional_state(encoder_cell)
# 1 x B
prev_y = cuda_if_gpu(Variable(torch.LongTensor([BOS_ID] * batch_size)))
# TL x B x Output_size
predicted_logits = cuda_if_gpu(
Variable(torch.zeros(dec_len, batch_size, self.output_size)))
# Teacher forcing: feed the target as the next input
for di in range(dec_len):
# embedding lookup of a vector of length = B; result: B x E
prev_y_embeddings = self.embedding_mat(prev_y)
# prev_y = self.dropout(prev_y) # apply dropout
attended_encoder_output = encoder_outputs[attn_ids_v[di],
batch_indices]
# 1 x B x dec_input_dim
dec_input = torch.cat(
(prev_y_embeddings, feat_embeddings, attended_encoder_output),
1).unsqueeze(0)
dec_output, dec_hidden, dec_cell = self.decoder(
dec_input, dec_hidden, dec_cell)
unnormalized_scores = self.output_layer(dec_output[0])
logits = self.softmax(unnormalized_scores)
predicted_logits[di] = logits # store this step's output logit
prev_y = dec_targets[di] # next input
return predicted_logits
def make_binary_prediction(self, x, model):
logit = model(cuda_if_gpu(Variable(torch.from_numpy(x))))
logit_val = logit.cpu().data[0].numpy()
if logit_val >= 0.5:
decision = RIGHT
else:
decision = LEFT
return decision
def batchify_training_data(self, xy_ids, batch_size, is_dev_data):
logger.debug('Batchifying data')
x_data_ids, y_data_ids, forms = xy_ids
data_size = len(x_data_ids)
num_batches = data_size // batch_size
data_indices = self.index_data(data_size, mode='no_shuffling')
batch_pairs = []
for bi in range(num_batches + 1):
batch_x = []
batch_y = []
curr_batch_indices = data_indices[bi * batch_size:(bi + 1) *
batch_size]
for idx in curr_batch_indices:
x_ids = x_data_ids[idx]
y_ids = y_data_ids[idx]
x_enc_ids_copy = copy.deepcopy(x_ids)
batch_x.append(x_enc_ids_copy)
y_ids_copy = copy.deepcopy(y_ids)
batch_y.append(y_ids_copy)
batch_enc_x_var = cuda_if_gpu(Variable(torch.LongTensor(batch_x)))
batch_dec_y_var = cuda_if_gpu(Variable(torch.LongTensor(batch_y)))
batch_pairs.append((batch_enc_x_var, batch_dec_y_var))
if is_dev_data:
self.dev_references = forms
logger.info('Saving dev (training) references to --> %s',
self.fnames.dev_ref_fn)
save_txt(self.dev_references, self.fnames.dev_ref_fn)
assert data_size == len(forms)
return batch_pairs
def make_one_batch(self, sorted_data, data_indices, bi, batch_size):
batch_x_enc = []
batch_x_feat = []
batch_y = []
batch_forms = []
curr_batch_indices = data_indices[bi * batch_size: (bi + 1) * batch_size]
for idx in curr_batch_indices:
x_enc_ids, x_feature_ids, y_ids, forms = sorted_data[idx]
x_enc_ids_copy = copy.deepcopy(x_enc_ids)
batch_x_enc.append(x_enc_ids_copy)
x_feature_ids_copy = copy.deepcopy(x_feature_ids)
batch_x_feat.append(x_feature_ids_copy)
y_ids_copy = copy.deepcopy(y_ids)
batch_y.append(y_ids_copy)
batch_forms.append(forms)
# skipping features, since the num does not change
batch_x_enc_lens = [len(s) for s in batch_x_enc]
batch_y_lens = [len(s) for s in batch_y]
max_enc_seq_len = max(batch_x_enc_lens)
max_dec_seq_len = max(batch_y_lens)
batch_enc_x_padded = [pad_seq(x, max_enc_seq_len, pad_id=self.vocab.PAD_ID) for x in batch_x_enc]
batch_dec_y_padded = [pad_seq(y, max_dec_seq_len, pad_id=self.vocab.PAD_ID) for y in batch_y]
batch_enc_x_var = cuda_if_gpu(Variable(torch.LongTensor(batch_enc_x_padded)).transpose(0, 1))
batch_dec_y_var = cuda_if_gpu(Variable(torch.LongTensor(batch_dec_y_padded)).transpose(0, 1))
batch_x_feat_var = cuda_if_gpu(Variable(torch.LongTensor(batch_x_feat)))
return (batch_enc_x_var, batch_x_feat_var, batch_dec_y_var), batch_forms
def predict_one(self, model, datum, vocab):
"""
:param model:
:param datum: a (lemma chars)
:param vocab:
:return:
"""
lemma, feats = datum
x_ids = MorphMLPData.feat_d_to_ids(feats, vocab)
x_ids[0] = vocab.src_vocab.lookup_tok(lemma)
x_ids_v = cuda_if_gpu(Variable(torch.LongTensor(x_ids)).view(1, -1))
pred_form_id = model.predict(x_ids_v)[0]
pred_form = vocab.tgt_vocab.lookup_id(pred_form_id)
return pred_form
def predict(self, input_var):
enc_inputs_v, feature_ids_v = input_var
src_len = enc_inputs_v.size()[0]
attn_clamp_value = src_len - 1
dec_len = self.max_tgt_len * 2
batch_size = enc_inputs_v.size()[1]
batch_indices = cuda_if_gpu(torch.arange(0, batch_size).long())
# Embedding lookup - chars
char_embeddings = self.embedding_lookup(enc_inputs_v) # SL x B x E
# Encode
# SL x B x 2*H, num_layers * num_directions x B x H
enc_outputs, enc_hidden, enc_cell = self.encoder(
rnn_input=char_embeddings, batch_size=batch_size)
# Embedding lookup - features
# B x self.num_embedding_feats * self.embedding_dim
feat_embeddings = self.embedding_lookup(feature_ids_v).view(
batch_size, -1)
feat_embeddings = F.relu(self.feat_proj_layer(feat_embeddings))
# Initialize decoder
# 1 x B x enc_dim
dec_hidden = sum_bidirectional_state(enc_hidden)
dec_cell = sum_bidirectional_state(enc_cell)
# Decode
prev_y_id = BOS_ID
prev_y_var = cuda_if_gpu(
Variable(torch.LongTensor([prev_y_id] * batch_size)))
dec_ids = torch.zeros(dec_len, batch_size)
attn_ids_v = cuda_if_gpu(torch.zeros(batch_size).long())
# run the decoder through the sequence and predict characters,
# twice max prediction as step outputs are added
for di in range(dec_len):
prev_y_embeddings = self.embedding_mat(prev_y_var) # B x E
# prev_y_embeddings_drop = self.dropout(prev_y)
attended_encoder_output = enc_outputs[
attn_ids_v, batch_indices] # B x enc_dim * enc.num_directions
# 1 x B x dec_input_dim
dec_input = torch.cat(
(prev_y_embeddings, feat_embeddings, attended_encoder_output),
1).unsqueeze(0)
dec_output, dec_hidden, dec_cell = self.decoder(
dec_input, dec_hidden, dec_cell)
unnormalized_scores = self.output_layer(dec_output[0])
logits = self.softmax(unnormalized_scores) # B x TV
topval, topids = logits.data.topk(1)
prev_y_ids = topids.squeeze(1) # B
# store result
dec_ids[di] = prev_y_ids
prev_y_var = cuda_if_gpu(Variable(prev_y_ids))
# check if step or char output to promote i.
attn_mask = prev_y_ids.eq(STEP_ID).long()
attn_ids_v += attn_mask
attn_ids_v = torch.clamp(attn_ids_v, min=0, max=attn_clamp_value)
return dec_ids
