def transduce(self, sent: ExpressionSequence) -> ExpressionSequence:
if self.pos_encoding_type == "trigonometric":
if self.position_encoding_block is None or self.position_encoding_block.shape[
2] < len(sent):
self.initialize_position_encoding(
int(len(sent) * 1.2),
self.input_dim if self.pos_encoding_combine == "add" else
self.pos_encoding_size)
encoding = dy.inputTensor(
self.position_encoding_block[0, :, :len(sent)])
elif self.pos_encoding_type == "embedding":
encoding = self.positional_embedder.embed_sent(
len(sent)).as_tensor()
if self.pos_encoding_type:
if self.pos_encoding_combine == "add":
sent = ExpressionSequence(expr_tensor=sent.as_tensor() +
encoding,
mask=sent.mask)
else: # concat
sent = ExpressionSequence(expr_tensor=dy.concatenate(
[sent.as_tensor(), encoding]),
mask=sent.mask)
elif self.pos_encoding_type:
raise ValueError(f"unknown encoding type {self.pos_encoding_type}")
for module in self.modules:
enc_sent = module.transduce(sent)
sent = enc_sent
self._final_states = [transducers.FinalTransducerState(sent[-1])]
return sent
def transduce(self, src: ExpressionSequence) -> 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")')
output_seq = ExpressionSequence(expr_tensor=output, mask=src.mask)
self._final_states = [FinalTransducerState(output_seq[-1])]
return output_seq
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 transduce(self, src: ExpressionSequence) -> ExpressionSequence:
src = src.as_tensor()
src_height = src.dim()[0][0]
src_width = src.dim()[0][1]
# src_channels = 1
batch_size = src.dim()[1]
# convolution and pooling layers
# src dim is ((40, 1000), 128)
src = padding(src, self.filter_width[0]+3)
l1 = dy.rectify(dy.conv2d(src, dy.parameter(self.filters1), stride = [self.stride[0], self.stride[0]], is_valid = True)) # ((1, 1000, 64), 128)
pool1 = dy.maxpooling2d(l1, (1, 4), (1,2), is_valid = True) #((1, 499, 64), 128)
pool1 = padding(pool1, self.filter_width[1]+3)
l2 = dy.rectify(dy.conv2d(pool1, dy.parameter(self.filters2), stride = [self.stride[1], self.stride[1]], is_valid = True))# ((1, 499, 512), 128)
pool2 = dy.maxpooling2d(l2, (1, 4), (1,2), is_valid = True)#((1, 248, 512), 128)
pool2 = padding(pool2, self.filter_width[2])
l3 = dy.rectify(dy.conv2d(pool2, dy.parameter(self.filters3), stride = [self.stride[2], self.stride[2]], is_valid = True))# ((1, 248, 1024), 128)
pool3 = dy.max_dim(l3, d = 1)
my_norm = dy.l2_norm(pool3) + 1e-6
output = dy.cdiv(pool3,my_norm)
output = dy.reshape(output, (self.num_filters[2],), batch_size = batch_size)
return ExpressionSequence(expr_tensor=output)
def transduce(self, embed_sent: ExpressionSequence) -> ExpressionSequence:
src = embed_sent.as_tensor()
sent_len = src.dim()[0][1]
batch_size = src.dim()[1]
pad_size = (self.window_receptor -
1) / 2 #TODO adapt it also for even window size
src = dy.concatenate([
dy.zeroes((self.input_dim, pad_size), batch_size=batch_size), src,
dy.zeroes((self.input_dim, pad_size), batch_size=batch_size)
],
d=1)
padded_sent_len = sent_len + 2 * pad_size
conv1 = dy.parameter(self.pConv1)
bias1 = dy.parameter(self.pBias1)
src_chn = dy.reshape(src, (self.input_dim, padded_sent_len, 1),
batch_size=batch_size)
cnn_layer1 = dy.conv2d_bias(src_chn, conv1, bias1, stride=[1, 1])
hidden_layer = dy.reshape(cnn_layer1, (self.internal_dim, sent_len, 1),
batch_size=batch_size)
if self.non_linearity is 'linear':
hidden_layer = hidden_layer
elif self.non_linearity is 'tanh':
hidden_layer = dy.tanh(hidden_layer)
elif self.non_linearity is 'relu':
hidden_layer = dy.rectify(hidden_layer)
elif self.non_linearity is 'sigmoid':
hidden_layer = dy.logistic(hidden_layer)
for conv_hid, bias_hid in self.builder_layers:
hidden_layer = dy.conv2d_bias(hidden_layer,
dy.parameter(conv_hid),
dy.parameter(bias_hid),
stride=[1, 1])
hidden_layer = dy.reshape(hidden_layer,
(self.internal_dim, sent_len, 1),
batch_size=batch_size)
if self.non_linearity is 'linear':
hidden_layer = hidden_layer
elif self.non_linearity is 'tanh':
hidden_layer = dy.tanh(hidden_layer)
elif self.non_linearity is 'relu':
hidden_layer = dy.rectify(hidden_layer)
elif self.non_linearity is 'sigmoid':
hidden_layer = dy.logistic(hidden_layer)
last_conv = dy.parameter(self.last_conv)
last_bias = dy.parameter(self.last_bias)
output = dy.conv2d_bias(hidden_layer,
last_conv,
last_bias,
stride=[1, 1])
output = dy.reshape(output, (sent_len, self.output_dim),
batch_size=batch_size)
output_seq = ExpressionSequence(expr_tensor=output)
self._final_states = [FinalTransducerState(output_seq[-1])]
return output_seq
def transduce(self, seq: ExpressionSequence) -> ExpressionSequence:
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 ExpressionSequence(expr_tensor=seq_tensor)
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 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, 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)
def transduce(self, src: ExpressionSequence) -> ExpressionSequence:
src = src.as_tensor()
# convolutional layer
src = padding(src,
src.dim()[0][0],
src.dim()[0][1], self.filter_width, self.stride,
src.dim()[1])
l1 = dy.rectify(
dy.conv2d(src,
dy.parameter(self.filter_conv),
stride=[self.stride, self.stride],
is_valid=True))
timestep = l1.dim()[0][1]
features = l1.dim()[0][2]
batch_size = l1.dim()[1]
# transpose l1 to be (timesetp, dim), but keep the batch_size.
rhn_in = dy.reshape(l1, (timestep, features), batch_size=batch_size)
rhn_in = [dy.pick(rhn_in, i) for i in range(timestep)]
for l in range(self.rhn_num_hidden_layers):
rhn_out = []
# initialize a random vector for the first state vector, keep the same batch size.
prev_state = dy.parameter(self.init[l])
# begin recurrent high way network
for t in range(timestep):
for m in range(0, self.rhn_microsteps):
H = dy.affine_transform([
dy.parameter(self.recur[l][m][1]),
dy.parameter(self.recur[l][m][0]), prev_state
])
T = dy.affine_transform([
dy.parameter(self.recur[l][m][3]),
dy.parameter(self.recur[l][m][2]), prev_state
])
if m == 0:
H += dy.parameter(self.linear[l][0]) * rhn_in[t]
T += dy.parameter(self.linear[l][1]) * rhn_in[t]
H = dy.tanh(H)
T = dy.logistic(T)
prev_state = dy.cmult(1 - T, prev_state) + dy.cmult(
T, H) # ((1024, ), batch_size)
rhn_out.append(prev_state)
if self.residual and l > 0:
rhn_out = [sum(x) for x in zip(rhn_out, rhn_in)]
rhn_in = rhn_out
# Compute the attention-weighted average of the activations
rhn_in = dy.concatenate_cols(rhn_in)
scores = dy.transpose(dy.parameter(self.attention[0][1])) * dy.tanh(
dy.parameter(self.attention[0][0]) *
rhn_in) # ((1,510), batch_size)
scores = dy.reshape(scores, (scores.dim()[0][1], ),
batch_size=scores.dim()[1])
attn_out = rhn_in * dy.softmax(
scores
) # # rhn_in.as_tensor() is ((1024,510), batch_size) softmax is ((510,), batch_size)
return ExpressionSequence(expr_tensor=attn_out)
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 exprseq_pooling(self, exprseq):
# Reduce to vector
exprseq = 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 = src.sent_len()
batch_size = tt.batch_size(src[0])
embeddings = self.embeddings(
torch.tensor([list(range(sent_len))] * batch_size).to(xnmt.device))
# 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 = tt.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 = tt.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, src: ExpressionSequence) -> ExpressionSequence:
src = src.as_tensor()
src_height = src.dim()[0][0]
src_width = 1
batch_size = src.dim()[1]
W = dy.parameter(self.pW)
b = dy.parameter(self.pb)
src = dy.reshape(src, (src_height, src_width), batch_size=batch_size) # ((276, 80, 3), 1)
# convolution and pooling layers
l1 = (W*src)+b
output = dy.cdiv(l1,dy.sqrt(dy.squared_norm(l1)))
return ExpressionSequence(expr_tensor=output)
def __call__(self, x: ExpressionSequence) -> tt.Tensor:
"""
Move the time-dimension of an input expression into the batch dimension via a reshape.
Args:
x: expression of dimensions ((hidden, timesteps), batch_size)
Returns:
expression of dimensions ((hidden,), timesteps*batch_size)
"""
batch_size = x[0].dim()[1]
model_dim = x[0].dim()[0][0]
seq_len = x.sent_len()
total_words = seq_len * batch_size
input_tensor = x.as_tensor()
return dy.reshape(input_tensor, (model_dim, ), batch_size=total_words)
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 src_tensor.dim()==3, \
f"Downsampling only supported for tensors of order two (+ batch). Found dims {src_tensor.size()}"
batch_size, seq_len, hidden_dim = src_tensor.size()
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 = src_tensor.view(
(batch_size, seq_len // self.downsample_by,
hidden_dim * self.downsample_by))
if out_mask:
out_mask = out_mask.lin_subsampled(
reduce_factor=self.downsample_by)
output = self.transform.transform(src_tensor)
output_seq = expression_seqs.ExpressionSequence(expr_tensor=output,
mask=out_mask)
self._final_states = [FinalTransducerState(output_seq[-1])]
return output_seq
def transduce(
self, es: expression_seqs.ExpressionSequence
) -> expression_seqs.ExpressionSequence:
output = self.transform(es.as_tensor(), es.mask)
return expression_seqs.ExpressionSequence(expr_tensor=output,
mask=es.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 (will be accessed via tensor_expr)
Return:
expression sequence
"""
if isinstance(expr_seq, list):
mask_out = expr_seq[0].mask
seq_len = len(expr_seq[0])
batch_size = expr_seq[0].dim()[1]
tensors = [e.as_tensor() for e in expr_seq]
input_tensor = dy.reshape(dy.concatenate(tensors),
(seq_len, 1, self.input_dim),
batch_size=batch_size)
else:
mask_out = expr_seq.mask
seq_len = len(expr_seq)
batch_size = expr_seq.dim()[1]
input_tensor = dy.reshape(dy.transpose(expr_seq.as_tensor()),
(seq_len, 1, self.input_dim),
batch_size=batch_size)
if self.dropout > 0.0 and self.train:
input_tensor = dy.dropout(input_tensor, self.dropout)
proj_inp = dy.conv2d_bias(input_tensor,
dy.parameter(self.p_f),
dy.parameter(self.p_b),
stride=(self.stride, 1),
is_valid=False)
reduced_seq_len = proj_inp.dim()[0][0]
proj_inp = dy.transpose(
dy.reshape(proj_inp, (reduced_seq_len, self.hidden_dim * 3),
batch_size=batch_size))
# proj_inp dims: (hidden, 1, seq_len), batch_size
if self.stride > 1 and mask_out is not None:
mask_out = mask_out.lin_subsampled(trg_len=reduced_seq_len)
h = [dy.zeroes(dim=(self.hidden_dim, 1), batch_size=batch_size)]
c = [dy.zeroes(dim=(self.hidden_dim, 1), batch_size=batch_size)]
for t in range(reduced_seq_len):
f_t = dy.logistic(
dy.strided_select(proj_inp, [], [0, t],
[self.hidden_dim, t + 1]))
o_t = dy.logistic(
dy.strided_select(proj_inp, [], [self.hidden_dim, t],
[self.hidden_dim * 2, t + 1]))
z_t = dy.tanh(
dy.strided_select(proj_inp, [], [self.hidden_dim * 2, t],
[self.hidden_dim * 3, t + 1]))
if self.dropout > 0.0 and self.train:
retention_rate = 1.0 - self.dropout
dropout_mask = dy.random_bernoulli((self.hidden_dim, 1),
retention_rate,
batch_size=batch_size)
f_t = 1.0 - dy.cmult(
dropout_mask, 1.0 - f_t
) # TODO: would be easy to make a zoneout dynet operation to save memory
i_t = 1.0 - f_t
if t == 0:
c_t = dy.cmult(i_t, z_t)
else:
c_t = dy.cmult(f_t, c[-1]) + dy.cmult(i_t, z_t)
h_t = dy.cmult(
o_t, c_t) # note: LSTM would use dy.tanh(c_t) instead of c_t
if mask_out is None or np.isclose(
np.sum(mask_out.np_arr[:, t:t + 1]), 0.0):
c.append(c_t)
h.append(h_t)
else:
c.append(
mask_out.cmult_by_timestep_expr(c_t, t, True) +
mask_out.cmult_by_timestep_expr(c[-1], t, False))
h.append(
mask_out.cmult_by_timestep_expr(h_t, t, True) +
mask_out.cmult_by_timestep_expr(h[-1], t, False))
self._final_states = [transducers.FinalTransducerState(dy.reshape(h[-1], (self.hidden_dim,), batch_size=batch_size), \
dy.reshape(c[-1], (self.hidden_dim,),
batch_size=batch_size))]
return expression_seqs.ExpressionSequence(expr_list=h[1:],
mask=mask_out)