def transduce(
self, es: 'expression_seqs.ExpressionSequence'
) -> 'expression_seqs.ExpressionSequence':
mask = es.mask
# first layer
forward_es = self.forward_layers[0].transduce(es)
rev_backward_es = self.backward_layers[0].transduce(
expression_seqs.ReversedExpressionSequence(es))
for layer_i in range(1, len(self.forward_layers)):
new_forward_es = self.forward_layers[layer_i].transduce([
forward_es,
expression_seqs.ReversedExpressionSequence(rev_backward_es)
])
rev_backward_es = expression_seqs.ExpressionSequence(
self.backward_layers[layer_i].transduce([
expression_seqs.ReversedExpressionSequence(forward_es),
rev_backward_es
]).as_list(),
mask=mask)
forward_es = new_forward_es
self._final_states = [
transducers.FinalTransducerState(dy.concatenate([self.forward_layers[layer_i].get_final_states()[0].main_expr(),
self.backward_layers[layer_i].get_final_states()[
0].main_expr()]),
dy.concatenate([self.forward_layers[layer_i].get_final_states()[0].cell_expr(),
self.backward_layers[layer_i].get_final_states()[
0].cell_expr()])) \
for layer_i in range(len(self.forward_layers))]
return expression_seqs.ExpressionSequence(expr_list=[
dy.concatenate([forward_es[i], rev_backward_es[-i - 1]])
for i in range(len(forward_es))
],
mask=mask)
def transduce(self, es: expression_seqs.ExpressionSequence) -> expression_seqs.ExpressionSequence:
"""
returns the list of output Expressions obtained by adding the given inputs
to the current state, one by one, to both the forward and backward RNNs,
and concatenating.
Args:
es: an ExpressionSequence
"""
es_list = [es]
for layer_i, (fb, bb) in enumerate(self.builder_layers):
reduce_factor = self._reduce_factor_for_layer(layer_i)
if es_list[0].mask is None: mask_out = None
else: mask_out = es_list[0].mask.lin_subsampled(reduce_factor)
if self.downsampling_method=="concat" and es_list[0].sent_len() % reduce_factor != 0:
raise ValueError(f"For 'concat' subsampling, sequence lengths must be multiples of the total reduce factor, "
f"but got sequence length={es_list[0].sent_len()} for reduce_factor={reduce_factor}. "
f"Set Batcher's pad_src_to_multiple argument accordingly.")
fs = fb.transduce(es_list)
bs = bb.transduce([expression_seqs.ReversedExpressionSequence(es_item) for es_item in es_list])
if layer_i < len(self.builder_layers) - 1:
if self.downsampling_method=="skip":
es_list = [expression_seqs.ExpressionSequence(expr_list=fs[::reduce_factor], mask=mask_out),
expression_seqs.ExpressionSequence(expr_list=bs[::reduce_factor][::-1], mask=mask_out)]
elif self.downsampling_method=="concat":
es_len = es_list[0].sent_len()
es_list_fwd = []
es_list_bwd = []
for i in range(0, es_len, reduce_factor):
for j in range(reduce_factor):
if i==0:
es_list_fwd.append([])
es_list_bwd.append([])
es_list_fwd[j].append(fs[i+j])
es_list_bwd[j].append(bs[es_list[0].sent_len()-reduce_factor+j-i])
es_list = [expression_seqs.ExpressionSequence(expr_list=es_list_fwd[j], mask=mask_out) for j in range(reduce_factor)] + \
[expression_seqs.ExpressionSequence(expr_list=es_list_bwd[j], mask=mask_out) for j in range(reduce_factor)]
else:
raise RuntimeError(f"unknown downsampling_method {self.downsampling_method}")
else:
# concat final outputs
ret_es = expression_seqs.ExpressionSequence(
expr_list=[tt.concatenate([f, b]) for f, b in zip(fs, expression_seqs.ReversedExpressionSequence(bs))], mask=mask_out)
self._final_states = [transducers.FinalTransducerState(tt.concatenate([fb.get_final_states()[0].main_expr(),
bb.get_final_states()[0].main_expr()]),
tt.concatenate([fb.get_final_states()[0].cell_expr(),
bb.get_final_states()[0].cell_expr()])) \
for (fb, bb) in self.builder_layers]
return ret_es
def transduce(
self, src: expression_seqs.ExpressionSequence
) -> expression_seqs.ExpressionSequence:
src_tensor = src.as_tensor()
out_mask = src.mask
if self.downsample_by > 1:
assert len(src_tensor.dim()[0])==2, \
f"Downsampling only supported for tensors of order two. Found dims {src_tensor.dim()}"
(hidden_dim, seq_len), batch_size = src_tensor.dim()
if seq_len % self.downsample_by != 0:
raise ValueError(
"For downsampling, sequence lengths must be multiples of the total reduce factor. "
"Configure batcher accordingly.")
src_tensor = dy.reshape(src_tensor,
(hidden_dim * self.downsample_by,
seq_len // self.downsample_by),
batch_size=batch_size)
if out_mask:
out_mask = out_mask.lin_subsampled(
reduce_factor=self.downsample_by)
output = self.transform.transform(src_tensor)
if self.downsample_by == 1:
if len(output.dim()) != src_tensor.dim(
): # can happen with seq length 1
output = dy.reshape(output,
src_tensor.dim()[0],
batch_size=src_tensor.dim()[1])
output_seq = expression_seqs.ExpressionSequence(expr_tensor=output,
mask=out_mask)
self._final_states = [FinalTransducerState(output_seq[-1])]
return output_seq
def embed_sent(self, x) -> expression_seqs.ExpressionSequence:
"""Embed a full sentence worth of words. By default, just do a for loop.
Args:
x: This will generally be a list of word IDs, but could also be a list of strings or some other format.
It could also be batched, in which case it will be a (possibly masked) :class:`xnmt.batcher.Batch` object
Returns:
An expression sequence representing vectors of each word in the input.
"""
# single mode
if not batchers.is_batched(x):
embeddings = [self.embed(word) for word in x]
# minibatch mode
else:
embeddings = []
seq_len = x.sent_len()
for single_sent in x:
assert single_sent.sent_len() == seq_len
for word_i in range(seq_len):
batch = batchers.mark_as_batch(
[single_sent[word_i] for single_sent in x])
embeddings.append(self.embed(batch))
return expression_seqs.ExpressionSequence(
expr_list=embeddings,
mask=x.mask if batchers.is_batched(x) else None)
def transduce(
self, expr_seq: expression_seqs.ExpressionSequence
) -> expression_seqs.ExpressionSequence:
if expr_seq.dim()[1] > 1:
raise ValueError(
f"LatticeLSTMTransducer requires batch size 1, got {expr_seq.dim()[1]}"
)
lattice = self.cur_src[0]
Wx_iog = dy.parameter(self.p_Wx_iog)
Wh_iog = dy.parameter(self.p_Wh_iog)
b_iog = dy.parameter(self.p_b_iog)
Wx_f = dy.parameter(self.p_Wx_f)
Wh_f = dy.parameter(self.p_Wh_f)
b_f = dy.parameter(self.p_b_f)
h = []
c = []
batch_size = expr_seq.dim()[1]
if self.dropout_rate > 0.0 and self.train:
self.set_dropout_masks(batch_size=batch_size)
for node_i in range(lattice.sent_len()):
cur_node = lattice.nodes[node_i]
val = expr_seq[node_i]
if self.dropout_rate > 0.0 and self.train:
val = dy.cmult(val, self.dropout_mask_x)
i_ft_list = []
if len(cur_node.nodes_prev) == 0:
tmp_iog = dy.affine_transform([b_iog, Wx_iog, val])
else:
h_tilde = sum(h[pred] for pred in cur_node.nodes_prev)
tmp_iog = dy.affine_transform(
[b_iog, Wx_iog, val, Wh_iog, h_tilde])
for pred in cur_node.nodes_prev:
i_ft_list.append(
dy.logistic(
dy.affine_transform(
[b_f, Wx_f, val, Wh_f, h[pred]])))
i_ait = dy.pick_range(tmp_iog, 0, self.hidden_dim)
i_aot = dy.pick_range(tmp_iog, self.hidden_dim,
self.hidden_dim * 2)
i_agt = dy.pick_range(tmp_iog, self.hidden_dim * 2,
self.hidden_dim * 3)
i_it = dy.logistic(i_ait)
i_ot = dy.logistic(i_aot)
i_gt = dy.tanh(i_agt)
if len(cur_node.nodes_prev) == 0:
c.append(dy.cmult(i_it, i_gt))
else:
fc = dy.cmult(i_ft_list[0], c[cur_node.nodes_prev[0]])
for i in range(1, len(cur_node.nodes_prev)):
fc += dy.cmult(i_ft_list[i], c[cur_node.nodes_prev[i]])
c.append(fc + dy.cmult(i_it, i_gt))
h_t = dy.cmult(i_ot, dy.tanh(c[-1]))
if self.dropout_rate > 0.0 and self.train:
h_t = dy.cmult(h_t, self.dropout_mask_h)
h.append(h_t)
self._final_states = [transducers.FinalTransducerState(h[-1], c[-1])]
return expression_seqs.ExpressionSequence(expr_list=h)
def embed_sent(
self,
sent_len: numbers.Integral) -> expression_seqs.ExpressionSequence:
embeddings = dy.strided_select(dy.parameter(self.embeddings), [1, 1],
[0, 0], [self.emb_dim, sent_len])
return expression_seqs.ExpressionSequence(expr_tensor=embeddings,
mask=None)
def embed_sent(self, x):
speech_x = x.batches[0]
factor_x = x.batches[1]
# if speech_x.sent_len()!=factor_x.sent_len():
# print(speech_x.sent_len())
# print(factor_x.sent_len())
# if speech_x.sent_len()!=factor_x.sent_len()+4:
# print("PROBLEM !!!!!!!!") #ah this is due to concatenated sentences which don't have both phonemes for both parts. shouldn't happen with the kaldi phones i don't think
# print("---")
# if speech_x.sent_len()==factor_x.sent_len():
# speech_xs = self.embed_speech_sent(speech_x)
# factor_xs = self.embed_factor_sent(factor_x, speech_x.sent_len())
# elif speech_x.sent_len()+4==factor_x.sent_len():
# speech_xs = self.embed_speech_sent(speech_x)
# factor_xs = self.embed_factor_sent(factor_x, speech_x.sent_len())
# elif speech_x.sent_len() > factor_x.sent_len():
# speech_xs = self.embed_speech_sent(speech_x)
# factor_xs = self.embed_factor_sent(factor_x, speech_x.sent_len())
# else:
# raise ValueError("!! unforseen sent mismatch in factor embedder")
speech_xs = self.embed_speech_sent(speech_x)
factor_xs = self.embed_factor_sent(factor_x, speech_x.sent_len())
catted = dy.concatenate([speech_xs.as_tensor(), factor_xs.as_tensor()])
output_seq = expression_seqs.ExpressionSequence(expr_tensor=catted,
mask=speech_x.mask)
return output_seq
def embed_sent(self,
x: sent.Sentence) -> expression_seqs.ExpressionSequence:
# TODO refactor: seems a bit too many special cases that need to be distinguished
batched = batchers.is_batched(x)
first_sent = x[0] if batched else x
if hasattr(first_sent, "get_array"):
if not batched:
return expression_seqs.LazyNumpyExpressionSequence(
lazy_data=x.get_array())
else:
return expression_seqs.LazyNumpyExpressionSequence(
lazy_data=batchers.mark_as_batch([s for s in x]),
mask=x.mask)
else:
if not batched:
embeddings = [self.embed(word) for word in x]
else:
embeddings = []
for word_i in range(x.sent_len()):
embeddings.append(
self.embed(
batchers.mark_as_batch(
[single_sent[word_i] for single_sent in x])))
return expression_seqs.ExpressionSequence(expr_list=embeddings,
mask=x.mask)
def transduce(self, expr_seq: 'expression_seqs.ExpressionSequence'):
"""
transduce the sequence
Args:
expr_seq: expression sequence
Returns:
expression sequence
"""
batch_size = expr_seq[0].dim()[1]
seq_len = len(expr_seq)
output_exps = []
for pos_i in range(seq_len):
input_i = expr_seq[pos_i]
affine = self.linear_layer(input_i)
# affine = dy.affine_transform([dy.parameter(self.p_b), dy.parameter(self.p_W), input_i])
if self.train and self.dropout_rate:
affine = dy.dropout(affine, self.dropout_rate)
if self.gumbel:
affine = affine + dy.random_gumbel(dim=affine.dim()[0],
batch_size=batch_size)
softmax_out = dy.softmax(affine)
# embedded = self.emb_layer(softmax_out)
embedded = dy.parameter(self.p_E) * softmax_out
if self.residual:
embedded = embedded + input_i
output_exps.append(embedded)
self._final_states = [
transducers.FinalTransducerState(main_expr=embedded)
]
return expression_seqs.ExpressionSequence(expr_list=output_exps,
mask=expr_seq.mask)
def embed_sent(self, x: Any):
embeddings = [embedder.embed_sent(x) for embedder in self.embedders]
ret = []
for j in range(len(embeddings[0])):
ret.append(
dy.esum([embeddings[i][j] for i in range(len(embeddings))]))
return expression_seqs.ExpressionSequence(expr_list=ret,
mask=embeddings[0].mask)
def transduce(
self, expr_seq: expression_seqs.ExpressionSequence
) -> expression_seqs.ExpressionSequence:
"""
transduce the sequence
Args:
expr_seq: expression sequence or list of expression sequences (where each inner list will be concatenated)
Returns:
expression sequence
"""
# Start with a [(length, model_size) x batch] tensor
# B x T x H -> B x H x T
x = expr_seq.as_tensor()
x_len = x.size()[1]
x_batch = x.size()[0]
# Get the query key and value vectors
q = self.lin_q(x).transpose(1, 2).contiguous()
k = self.lin_k(x).transpose(1, 2).contiguous()
v = self.lin_v(x).transpose(1, 2).contiguous()
# q = bq + x * Wq
# k = bk + x * Wk
# v = bv + x * Wv
# Split to batches [(length, head_dim) x batch * num_heads] tensor
q, k, v = [
temp.view((x_batch * self.num_heads, self.head_dim, x_len))
for temp in (q, k, v)
]
# Do scaled dot product [batch*num_heads, length, length], rows are keys, columns are queries
attn_score = torch.matmul(k.transpose(1, 2), q) / sqrt(self.head_dim)
if expr_seq.mask is not None:
mask = torch.Tensor(
np.repeat(expr_seq.mask.np_arr, self.num_heads, axis=0) *
-1e10).to(xnmt.device)
attn_score = attn_score + mask.unsqueeze(2)
attn_prob = torch.nn.Softmax(dim=1)(attn_score)
# attn_prob = dy.softmax(attn_score, d=1)
if self.train and self.dropout > 0.0:
attn_prob = tt.dropout(attn_prob, self.dropout)
# Reduce using attention and resize to match [(length, model_size) x batch]
o = torch.matmul(v, attn_prob).view(x_batch, self.input_dim,
x_len).transpose(1, 2)
# Final transformation
o = self.lin_o(o)
# o = bo + o * Wo
expr_seq = expression_seqs.ExpressionSequence(expr_tensor=o,
mask=expr_seq.mask)
self._final_states = [
transducers.FinalTransducerState(expr_seq[-1], None)
]
return expr_seq
def calculate_baseline(
self, input_states: expr_seq.ExpressionSequence
) -> expr_seq.ExpressionSequence:
transform_seq = []
for input_state in input_states:
transform_seq.append(
self.transform.transform(dy.nobackprop(input_state)))
return expr_seq.ExpressionSequence(expr_list=transform_seq,
mask=input_states.mask)
def transduce(self, es: expression_seqs.ExpressionSequence) -> expression_seqs.ExpressionSequence:
mask = es.mask
sent_len = len(es)
es_expr = es.as_transposed_tensor()
batch_size = es_expr.dim()[1]
es_chn = dy.reshape(es_expr, (sent_len, self.freq_dim, self.chn_dim), batch_size=batch_size)
h_out = {}
for direction in ["fwd", "bwd"]:
# input convolutions
gates_xt_bias = dy.conv2d_bias(es_chn, dy.parameter(self.params["x2all_" + direction]),
dy.parameter(self.params["b_" + direction]), stride=(1, 1), is_valid=False)
gates_xt_bias_list = [dy.pick_range(gates_xt_bias, i, i + 1) for i in range(sent_len)]
h = []
c = []
for input_pos in range(sent_len):
directional_pos = input_pos if direction == "fwd" else sent_len - input_pos - 1
gates_t = gates_xt_bias_list[directional_pos]
if input_pos > 0:
# recurrent convolutions
gates_h_t = dy.conv2d(h[-1], dy.parameter(self.params["h2all_" + direction]), stride=(1, 1), is_valid=False)
gates_t += gates_h_t
# standard LSTM logic
if len(c) == 0:
c_tm1 = dy.zeros((self.freq_dim * self.num_filters,), batch_size=batch_size)
else:
c_tm1 = c[-1]
gates_t_reshaped = dy.reshape(gates_t, (4 * self.freq_dim * self.num_filters,), batch_size=batch_size)
c_t = dy.reshape(dy.vanilla_lstm_c(c_tm1, gates_t_reshaped), (self.freq_dim * self.num_filters,),
batch_size=batch_size)
h_t = dy.vanilla_lstm_h(c_t, gates_t_reshaped)
h_t = dy.reshape(h_t, (1, self.freq_dim, self.num_filters,), batch_size=batch_size)
if mask is None or np.isclose(np.sum(mask.np_arr[:, input_pos:input_pos + 1]), 0.0):
c.append(c_t)
h.append(h_t)
else:
c.append(
mask.cmult_by_timestep_expr(c_t, input_pos, True) + mask.cmult_by_timestep_expr(c[-1], input_pos, False))
h.append(
mask.cmult_by_timestep_expr(h_t, input_pos, True) + mask.cmult_by_timestep_expr(h[-1], input_pos, False))
h_out[direction] = h
ret_expr = []
for state_i in range(len(h_out["fwd"])):
state_fwd = h_out["fwd"][state_i]
state_bwd = h_out["bwd"][-1 - state_i]
output_dim = (state_fwd.dim()[0][1] * state_fwd.dim()[0][2],)
fwd_reshape = dy.reshape(state_fwd, output_dim, batch_size=batch_size)
bwd_reshape = dy.reshape(state_bwd, output_dim, batch_size=batch_size)
ret_expr.append(dy.concatenate([fwd_reshape, bwd_reshape], d=0 if self.reshape_output else 2))
return expression_seqs.ExpressionSequence(expr_list=ret_expr, mask=mask)
# TODO: implement get_final_states()
def compose(
self, embeds: Union[dy.Expression,
List[dy.Expression]]) -> dy.Expression:
if type(embeds) != list:
embeds = [
dy.pick_batch_elem(embeds, i) for i in range(embeds.dim()[1])
]
self.seq_transducer.transduce(
expr_seq.ExpressionSequence(expr_list=embeds))
return self.seq_transducer.get_final_states()[-1].main_expr()
def embed_sent(self, x: Any) -> expression_seqs.ExpressionSequence:
"""Embed a full sentence worth of words. By default, just do a for loop.
Args:
x: This will generally be a list of word IDs, but could also be a list of strings or some other format.
It could also be batched, in which case it will be a (possibly masked) :class:`xnmt.batcher.Batch` object
Returns:
An expression sequence representing vectors of each word in the input.
"""
# single mode
if not batchers.is_batched(x):
expr = expression_seqs.ExpressionSequence(
expr_list=[self.embed(word) for word in x])
# minibatch mode
elif type(self) == LookupEmbedder:
embeddings = []
for word_i in range(x.sent_len()):
batch = batchers.mark_as_batch(
[single_sent[word_i] for single_sent in x])
embeddings.append(self.embed(batch))
expr = expression_seqs.ExpressionSequence(expr_list=embeddings,
mask=x.mask)
else:
assert type(
x[0]
) == sent.SegmentedSentence, "Need to use CharFromWordTextReader for non standard embeddings."
embeddings = []
all_embeddings = []
for sentence in x:
embedding = []
for i in range(sentence.len_unpadded()):
embed_word = self.embed(sentence.words[i])
embedding.append(embed_word)
all_embeddings.append(embed_word)
embeddings.append(embedding)
# Useful when using dy.autobatch
dy.forward(all_embeddings)
all_embeddings.clear()
# Pad the results
expr = batchers.pad_embedding(embeddings)
return expr
def transduce(self, expr_seq: 'expression_seqs.ExpressionSequence') -> 'expression_seqs.ExpressionSequence':
"""
transduce the sequence, applying masks if given (masked timesteps simply copy previous h / c)
Args:
expr_seq: expression sequence or list of expression sequences (where each inner list will be concatenated)
Returns:
expression sequence
"""
if isinstance(expr_seq, expression_seqs.ExpressionSequence):
expr_seq = [expr_seq]
batch_size = expr_seq[0][0].dim()[1]
seq_len = len(expr_seq[0])
if self.dropout_rate > 0.0 and self.train:
self.set_dropout_masks(batch_size=batch_size)
cur_input = expr_seq
self._final_states = []
for layer_i in range(self.num_layers):
h = [dy.zeroes(dim=(self.hidden_dim,), batch_size=batch_size)]
c = [dy.zeroes(dim=(self.hidden_dim,), batch_size=batch_size)]
for pos_i in range(seq_len):
x_t = [cur_input[j][pos_i] for j in range(len(cur_input))]
if isinstance(x_t, dy.Expression):
x_t = [x_t]
elif type(x_t) != list:
x_t = list(x_t)
if sum([x_t_i.dim()[0][0] for x_t_i in x_t]) != self.total_input_dim:
found_dim = sum([x_t_i.dim()[0][0] for x_t_i in x_t])
raise ValueError(f"VanillaLSTMGates: x_t has inconsistent dimension {found_dim}, expecting {self.total_input_dim}")
if self.dropout_rate > 0.0 and self.train:
# apply dropout according to https://arxiv.org/abs/1512.05287 (tied weights)
gates_t = dy.vanilla_lstm_gates_dropout_concat(x_t,
h[-1],
self.Wx[layer_i],
self.Wh[layer_i],
self.b[layer_i],
self.dropout_mask_x[layer_i],
self.dropout_mask_h[layer_i],
self.weightnoise_std if self.train else 0.0)
else:
gates_t = dy.vanilla_lstm_gates_concat(x_t, h[-1], self.Wx[layer_i], self.Wh[layer_i], self.b[layer_i], self.weightnoise_std if self.train else 0.0)
c_t = dy.vanilla_lstm_c(c[-1], gates_t)
h_t = dy.vanilla_lstm_h(c_t, gates_t)
if expr_seq[0].mask is None or np.isclose(np.sum(expr_seq[0].mask.np_arr[:,pos_i:pos_i+1]), 0.0):
c.append(c_t)
h.append(h_t)
else:
c.append(expr_seq[0].mask.cmult_by_timestep_expr(c_t,pos_i,True) + expr_seq[0].mask.cmult_by_timestep_expr(c[-1],pos_i,False))
h.append(expr_seq[0].mask.cmult_by_timestep_expr(h_t,pos_i,True) + expr_seq[0].mask.cmult_by_timestep_expr(h[-1],pos_i,False))
self._final_states.append(transducers.FinalTransducerState(h[-1], c[-1]))
cur_input = [h[1:]]
return expression_seqs.ExpressionSequence(expr_list=h[1:], mask=expr_seq[0].mask)
def transduce(
self, expr_seq: 'expression_seqs.ExpressionSequence'
) -> 'expression_seqs.ExpressionSequence':
"""
transduce the sequence, applying masks if given (masked timesteps simply copy previous h / c)
Args:
expr_seq: expression sequence or list of expression sequences (where each inner list will be concatenated)
Returns:
expression sequence
"""
if isinstance(expr_seq, expression_seqs.ExpressionSequence):
expr_seq = [expr_seq]
concat_inputs = len(expr_seq) >= 2
batch_size = tt.batch_size(expr_seq[0][0])
seq_len = expr_seq[0].sent_len()
mask = expr_seq[0].mask
if self.dropout_rate > 0.0 and self.train:
self.set_dropout_masks(batch_size=batch_size)
cur_input = expr_seq
self._final_states = []
for layer_i in range(self.num_layers):
h = [tt.zeroes(hidden_dim=self.hidden_dim, batch_size=batch_size)]
c = [tt.zeroes(hidden_dim=self.hidden_dim, batch_size=batch_size)]
for pos_i in range(seq_len):
if concat_inputs and layer_i == 0:
x_t = tt.concatenate(
[cur_input[i][pos_i] for i in range(len(cur_input))])
else:
x_t = cur_input[0][pos_i]
h_tm1 = h[-1]
if self.dropout_rate > 0.0 and self.train:
# apply dropout according to https://arxiv.org/abs/1512.05287 (tied weights)
x_t = torch.mul(x_t, self.dropout_mask_x[layer_i])
h_tm1 = torch.mul(h_tm1, self.dropout_mask_h[layer_i])
h_t, c_t = self.layers[layer_i](x_t, (h_tm1, c[-1]))
if mask is None or np.isclose(
np.sum(mask.np_arr[:, pos_i:pos_i + 1]), 0.0):
c.append(c_t)
h.append(h_t)
else:
c.append(
mask.cmult_by_timestep_expr(c_t, pos_i, True) +
mask.cmult_by_timestep_expr(c[-1], pos_i, False))
h.append(
mask.cmult_by_timestep_expr(h_t, pos_i, True) +
mask.cmult_by_timestep_expr(h[-1], pos_i, False))
self._final_states.append(
transducers.FinalTransducerState(h[-1], c[-1]))
cur_input = [h[1:]]
return expression_seqs.ExpressionSequence(expr_list=h[1:], mask=mask)
def transduce(self, x: expression_seqs.ExpressionSequence) -> expression_seqs.ExpressionSequence:
x_T = x.as_transposed_tensor()
scores = x_T * dy.parameter(self.W)
if x.mask is not None:
scores = x.mask.add_to_tensor_expr(scores, multiplicator=-100.0, time_first=True)
if self.pos_enc_max:
seq_len = x_T.dim()[0][0]
pos_enc = self.pos_enc[:seq_len,:]
scores = dy.cmult(scores, dy.inputTensor(pos_enc))
attention = dy.softmax(scores)
output_expr = x.as_tensor() * attention
return expression_seqs.ExpressionSequence(expr_tensor=output_expr, mask=None)
def transduce(self, es):
mask = es.mask
# first layer
forward_es = self.forward_layers[0].transduce(es)
rev_backward_es = self.backward_layers[0].transduce(
expression_seqs.ReversedExpressionSequence(es))
# TODO: concat input of each layer to its output; or, maybe just add standard residual connections
for layer_i in range(1, len(self.forward_layers)):
new_forward_es = self.forward_layers[layer_i].transduce([
forward_es,
expression_seqs.ReversedExpressionSequence(rev_backward_es)
])
mask_out = mask
if mask_out is not None and new_forward_es.mask.np_arr.shape != mask_out.np_arr.shape:
mask_out = mask_out.lin_subsampled(trg_len=len(new_forward_es))
rev_backward_es = expression_seqs.ExpressionSequence(
self.backward_layers[layer_i].transduce([
expression_seqs.ReversedExpressionSequence(forward_es),
rev_backward_es
]).as_list(),
mask=mask_out)
forward_es = new_forward_es
self._final_states = [
transducers.FinalTransducerState(dy.concatenate([self.forward_layers[layer_i].get_final_states()[0].main_expr(),
self.backward_layers[layer_i].get_final_states()[
0].main_expr()]),
dy.concatenate([self.forward_layers[layer_i].get_final_states()[0].cell_expr(),
self.backward_layers[layer_i].get_final_states()[
0].cell_expr()])) \
for layer_i in range(len(self.forward_layers))]
mask_out = mask
if mask_out is not None and forward_es.mask.np_arr.shape != mask_out.np_arr.shape:
mask_out = mask_out.lin_subsampled(trg_len=len(forward_es))
return expression_seqs.ExpressionSequence(expr_list=[
dy.concatenate([forward_es[i], rev_backward_es[-i - 1]])
for i in range(len(forward_es))
],
mask=mask_out)
def exprseq_pooling(self, exprseq):
# Reduce to vector
exprseq = expression_seqs.ExpressionSequence(
expr_tensor=exprseq.mask.add_to_tensor_expr(
exprseq.as_tensor(), -1e10),
mask=exprseq.mask)
if exprseq.expr_tensor is not None:
if len(exprseq.expr_tensor.dim()[0]) > 1:
return dy.max_dim(exprseq.expr_tensor, d=1)
else:
return exprseq.expr_tensor
else:
return dy.emax(exprseq.expr_list)
def transduce(self, src: expression_seqs.ExpressionSequence) -> expression_seqs.ExpressionSequence:
sent_len = len(src)
embeddings = dy.strided_select(dy.parameter(self.embedder), [1,1], [0,0], [self.input_dim, sent_len])
if self.op == 'sum':
output = embeddings + src.as_tensor()
elif self.op == 'concat':
output = dy.concatenate([embeddings, src.as_tensor()])
else:
raise ValueError(f'Illegal op {op} in PositionalTransducer (options are "sum"/"concat")')
if self.train and self.dropout > 0.0:
output = dy.dropout(output, self.dropout)
output_seq = expression_seqs.ExpressionSequence(expr_tensor=output, mask=src.mask)
self._final_states = [transducers.FinalTransducerState(output_seq[-1])]
return output_seq
def transduce(
self, expr_sequence: expression_seqs.ExpressionSequence
) -> expression_seqs.ExpressionSequence:
# first layer
forward_es = self.forward_layers[0].transduce(expr_sequence)
rev_backward_es = self.backward_layers[0].transduce(
expression_seqs.ReversedExpressionSequence(expr_sequence))
for layer_i in range(1, len(self.forward_layers)):
concat_fwd = expression_seqs.ExpressionSequence(expr_list=[
dy.concatenate([fwd_expr, bwd_expr])
for fwd_expr, bwd_expr in zip(
forward_es.as_list(), reversed(rev_backward_es.as_list()))
])
concat_bwd = expression_seqs.ExpressionSequence(expr_list=[
dy.concatenate([fwd_expr, bwd_expr])
for fwd_expr, bwd_expr in zip(reversed(forward_es.as_list()),
rev_backward_es.as_list())
])
new_forward_es = self.forward_layers[layer_i].transduce(concat_fwd)
rev_backward_es = self.backward_layers[layer_i].transduce(
concat_bwd)
forward_es = new_forward_es
self._final_states = [
transducers.FinalTransducerState(dy.concatenate([self.forward_layers[layer_i].get_final_states()[0].main_expr(),
self.backward_layers[layer_i].get_final_states()[
0].main_expr()]),
dy.concatenate([self.forward_layers[layer_i].get_final_states()[0].cell_expr(),
self.backward_layers[layer_i].get_final_states()[
0].cell_expr()])) \
for layer_i in range(len(self.forward_layers))]
return expression_seqs.ExpressionSequence(expr_list=[
dy.concatenate([forward_es[i], rev_backward_es[-i - 1]])
for i in range(len(forward_es))
])
def transduce(self, expr_seq: expression_seqs.ExpressionSequence) -> expression_seqs.ExpressionSequence:
"""
transduce the sequence
Args:
expr_seq: expression sequence or list of expression sequences (where each inner list will be concatenated)
Returns:
expression sequence
"""
Wq, Wk, Wv, Wo = [dy.parameter(x) for x in (self.pWq, self.pWk, self.pWv, self.pWo)]
bq, bk, bv, bo = [dy.parameter(x) for x in (self.pbq, self.pbk, self.pbv, self.pbo)]
# Start with a [(length, model_size) x batch] tensor
x = expr_seq.as_transposed_tensor()
x_len = x.dim()[0][0]
x_batch = x.dim()[1]
# Get the query key and value vectors
# TODO: do we need bias broadcasting in DyNet?
# q = dy.affine_transform([bq, x, Wq])
# k = dy.affine_transform([bk, x, Wk])
# v = dy.affine_transform([bv, x, Wv])
q = bq + x * Wq
k = bk + x * Wk
v = bv + x * Wv
# Split to batches [(length, head_dim) x batch * num_heads] tensor
q, k, v = [dy.reshape(x, (x_len, self.head_dim), batch_size=x_batch * self.num_heads) for x in (q,k,v)]
# Do scaled dot product [(length, length) x batch * num_heads], rows are queries, columns are keys
attn_score = q * dy.transpose(k) / sqrt(self.head_dim)
if expr_seq.mask is not None:
mask = dy.inputTensor(np.repeat(expr_seq.mask.np_arr, self.num_heads, axis=0).transpose(), batched=True) * -1e10
attn_score = attn_score + mask
attn_prob = dy.softmax(attn_score, d=1)
if self.train and self.dropout > 0.0:
attn_prob = dy.dropout(attn_prob, self.dropout)
# Reduce using attention and resize to match [(length, model_size) x batch]
o = dy.reshape(attn_prob * v, (x_len, self.input_dim), batch_size=x_batch)
# Final transformation
# o = dy.affine_transform([bo, attn_prob * v, Wo])
o = bo + o * Wo
expr_seq = expression_seqs.ExpressionSequence(expr_transposed_tensor=o, mask=expr_seq.mask)
self._final_states = [transducers.FinalTransducerState(expr_seq[-1], None)]
return expr_seq
def pad_embedding(embeddings) -> expression_seqs.ExpressionSequence:
max_col = max(len(xs) for xs in embeddings)
p0 = dy.zeros(embeddings[0][0].dim()[0][0])
masks = np.zeros((len(embeddings), max_col), dtype=int)
modified = False
ret = []
for xs, mask in zip(embeddings, masks):
deficit = max_col - len(xs)
if deficit > 0:
xs = xs + ([p0] * deficit)
mask[-deficit:] = 1
modified = True
ret.append(dy.concatenate_cols(xs))
mask = Mask(masks) if modified else None
return expression_seqs.ExpressionSequence(
expr_tensor=dy.concatenate_to_batch(ret), mask=mask)
def embed_speech_sent(self, x):
# TODO refactor: seems a bit too many special cases that need to be distinguished
# x = x.batches[0]
batched = batchers.is_batched(x)
first_sent = x[0] if batched else x
if hasattr(first_sent, "get_array"):
if not batched:
return expression_seqs.LazyNumpyExpressionSequence(
lazy_data=x.get_array())
else:
return expression_seqs.LazyNumpyExpressionSequence(
lazy_data=batchers.mark_as_batch([s for s in x]),
mask=x.mask)
else:
raise ValueError("!! Expected to use above")
return expression_seqs.ExpressionSequence(expr_list=embeddings,
mask=x.mask)
def transduce(
self, es: 'expression_seqs.ExpressionSequence'
) -> 'expression_seqs.ExpressionSequence':
batch_size = tt.batch_size(es.as_tensor())
if es.mask:
seq_lengths = es.mask.seq_lengths()
else:
seq_lengths = [es.sent_len()] * batch_size
# Sort the input and lengths as the descending order
seq_lengths = torch.LongTensor(seq_lengths).to(xnmt.device)
lengths, perm_index = seq_lengths.sort(0, descending=True)
sorted_input = es.as_tensor()[perm_index]
perm_index_rev = [-1] * len(lengths)
for i in range(len(lengths)):
perm_index_rev[perm_index[i]] = i
perm_index_rev = torch.LongTensor(perm_index_rev).to(xnmt.device)
packed_input = nn.utils.rnn.pack_padded_sequence(sorted_input,
list(lengths.data),
batch_first=True)
state_size = self.num_dir * self.num_layers, batch_size, self.hidden_dim // self.num_dir
h0 = sorted_input.new_zeros(*state_size)
c0 = sorted_input.new_zeros(*state_size)
output, (final_hiddens,
final_cells) = self.lstm(packed_input, (h0, c0))
output = nn.utils.rnn.pad_packed_sequence(
output, batch_first=True, total_length=es.sent_len())[0]
# restore the sorting
decoded = output[perm_index_rev]
self._final_states = []
for layer_i in range(self.num_layers):
final_hidden = final_hiddens.view(
self.num_layers, self.num_dir, batch_size,
-1)[layer_i].transpose(0, 1).contiguous().view(batch_size, -1)
final_hidden = final_hidden[perm_index_rev]
self._final_states.append(
transducers.FinalTransducerState(final_hidden))
ret = expression_seqs.ExpressionSequence(expr_tensor=decoded,
mask=es.mask)
return ret
def transduce(
self, seq: expression_seqs.ExpressionSequence
) -> expression_seqs.ExpressionSequence:
if self.train and self.dropout > 0.0:
seq_tensor = dy.dropout(
self.child.transduce(seq).as_tensor(),
self.dropout) + seq.as_tensor()
else:
seq_tensor = self.child.transduce(
seq).as_tensor() + seq.as_tensor()
if self.layer_norm:
d = seq_tensor.dim()
seq_tensor = dy.reshape(seq_tensor, (d[0][0], ),
batch_size=d[0][1] * d[1])
seq_tensor = dy.layer_norm(seq_tensor, self.ln_g, self.ln_b)
seq_tensor = dy.reshape(seq_tensor, d[0], batch_size=d[1])
return expression_seqs.ExpressionSequence(expr_tensor=seq_tensor)
def transduce(
self, x: expression_seqs.ExpressionSequence
) -> expression_seqs.ExpressionSequence:
expr = x.as_transposed_tensor()
batch_size, hidden_dim, seq_len = expr.size()
expr = expr.view((batch_size, self.in_channels,
hidden_dim // self.in_channels, seq_len))
expr = self.cnn_layer(expr)
if self.use_pooling:
expr = self.pooling_layer(expr)
expr = self.activation_fct(expr)
batch_size, out_chn, out_h, seq_len = expr.size()
expr = expr.view((batch_size, out_chn * out_h, seq_len))
output_seq = expression_seqs.ExpressionSequence(
expr_transposed_tensor=expr,
mask=x.mask.lin_subsampled(trg_len=seq_len) if x.mask else None)
self._final_states = [transducers.FinalTransducerState(output_seq[-1])]
return output_seq
def embed_factor_sent(self, x, speech_len):
# single mode
if not batchers.is_batched(x):
embeddings = [self.embed_factor(word) for word in x]
# minibatch mode
else:
embeddings = []
seq_len = x.sent_len()
for single_sent in x:
assert single_sent.sent_len() == seq_len
# for word_i in range(seq_len):
for word_i in range(speech_len):
batch = batchers.mark_as_batch(
[single_sent[word_i] for single_sent in x])
embeddings.append(self.embed_factor(batch))
return expression_seqs.ExpressionSequence(
expr_list=embeddings,
mask=x.mask if batchers.is_batched(x) else None)
def transduce(
self, seq: expression_seqs.ExpressionSequence
) -> expression_seqs.ExpressionSequence:
if self.train and self.dropout > 0.0:
seq_tensor = tt.dropout(
self.child.transduce(seq).as_tensor(),
self.dropout) + seq.as_tensor()
else:
seq_tensor = self.child.transduce(
seq).as_tensor() + seq.as_tensor()
if self.layer_norm:
batch_size = tt.batch_size(seq_tensor)
merged_seq_tensor = tt.merge_time_batch_dims(seq_tensor)
transformed_seq_tensor = self.layer_norm_component.transform(
merged_seq_tensor)
seq_tensor = tt.unmerge_time_batch_dims(transformed_seq_tensor,
batch_size)
return expression_seqs.ExpressionSequence(expr_tensor=seq_tensor)
