def _build(self, batch_size):
src_time_dim = 4
vocab_size = 7
emb = Embeddings(embedding_dim=self.emb_size,
vocab_size=vocab_size,
padding_idx=self.pad_index)
encoder = RecurrentEncoder(emb_size=self.emb_size,
num_layers=self.num_layers,
hidden_size=self.encoder_hidden_size,
bidirectional=True)
decoder = RecurrentDecoder(hidden_size=self.hidden_size,
encoder=encoder,
attention="bahdanau",
emb_size=self.emb_size,
vocab_size=self.vocab_size,
num_layers=self.num_layers,
init_hidden="bridge",
input_feeding=True)
encoder_output = torch.rand(size=(batch_size, src_time_dim,
encoder.output_size))
for p in decoder.parameters():
torch.nn.init.uniform_(p, -0.5, 0.5)
src_mask = torch.ones(size=(batch_size, 1, src_time_dim)) == 1
encoder_hidden = torch.rand(size=(batch_size, encoder.output_size))
return src_mask, emb, decoder, encoder_output, encoder_hidden
def setUp(self):
self.emb_size = 10
self.num_layers = 3
self.hidden_size = 6
self.encoder_hidden_size = 3
self.vocab_size = 5
seed = 42
torch.manual_seed(seed)
bidi_encoder = RecurrentEncoder(emb_size=self.emb_size,
num_layers=self.num_layers,
hidden_size=self.encoder_hidden_size,
bidirectional=True)
uni_encoder = RecurrentEncoder(emb_size=self.emb_size,
num_layers=self.num_layers,
hidden_size=self.encoder_hidden_size*2,
bidirectional=False)
self.encoders = [uni_encoder, bidi_encoder]
def setUp(self):
self.addTypeEqualityFunc(
torch.Tensor, lambda x, y, msg: self.failureException(msg)
if not torch.equal(x, y) else True)
self.emb_size = 10
self.num_layers = 3
self.hidden_size = 7
self.vocab_size = 5
seed = 42
torch.manual_seed(seed)
bidi_encoder = RecurrentEncoder(emb_size=self.emb_size,
num_layers=self.num_layers,
hidden_size=self.hidden_size,
bidirectional=True)
uni_encoder = RecurrentEncoder(emb_size=self.emb_size,
num_layers=self.num_layers,
hidden_size=self.hidden_size,
bidirectional=False)
self.encoders = [uni_encoder, bidi_encoder]
def test_recurrent_encoder_size(self):
for bidirectional in [True, False]:
directional_factor = 2 if bidirectional else 1
encoder = RecurrentEncoder(hidden_size=self.hidden_size,
emb_size=self.emb_size,
num_layers=self.num_layers,
bidirectional=bidirectional)
self.assertEqual(encoder.rnn.hidden_size, self.hidden_size)
# output size is affected by bidirectionality
self.assertEqual(encoder.output_size,
self.hidden_size * directional_factor)
self.assertEqual(encoder.rnn.bidirectional, bidirectional)
def build_model(cfg: dict = None,
src_vocab: Vocabulary = None,
trg_vocab: Vocabulary = None) -> Model:
"""
Build and initialize the model according to the configuration.
:param cfg: dictionary configuration containing model specifications
:param src_vocab: source vocabulary
:param trg_vocab: target vocabulary
:return: built and initialized model
"""
src_padding_idx = src_vocab.stoi[PAD_TOKEN]
trg_padding_idx = trg_vocab.stoi[PAD_TOKEN]
src_embed = Embeddings(**cfg["encoder"]["embeddings"],
vocab_size=len(src_vocab),
padding_idx=src_padding_idx)
if cfg.get("tied_embeddings", False):
if src_vocab.itos == trg_vocab.itos:
# share embeddings for src and trg
trg_embed = src_embed
else:
raise ConfigurationError(
"Embedding cannot be tied since vocabularies differ.")
else:
trg_embed = Embeddings(**cfg["decoder"]["embeddings"],
vocab_size=len(trg_vocab),
padding_idx=trg_padding_idx)
encoder = RecurrentEncoder(**cfg["encoder"],
emb_size=src_embed.embedding_dim)
decoder = RecurrentDecoder(**cfg["decoder"],
encoder=encoder,
vocab_size=len(trg_vocab),
emb_size=trg_embed.embedding_dim)
model = Model(encoder=encoder,
decoder=decoder,
src_embed=src_embed,
trg_embed=trg_embed,
src_vocab=src_vocab,
trg_vocab=trg_vocab)
# custom initialization of model parameters
initialize_model(model, cfg, src_padding_idx, trg_padding_idx)
return model
def test_recurrent_forward(self):
time_dim = 4
batch_size = 2
bidirectional = True
directions = 2 if bidirectional else 1
encoder = RecurrentEncoder(emb_size=self.emb_size,
num_layers=self.num_layers,
hidden_size=self.hidden_size,
bidirectional=bidirectional)
x = torch.rand(size=(batch_size, time_dim, self.emb_size))
# no padding, no mask
x_length = torch.Tensor([time_dim]*batch_size).int()
mask = torch.ones_like(x)
output, hidden, _ = encoder(
embed_src=x, src_length=x_length, mask=mask)
self.assertEqual(output.shape, torch.Size(
[batch_size, time_dim, directions*self.hidden_size]))
self.assertEqual(hidden.shape, torch.Size(
[batch_size, directions*self.hidden_size]))
hidden_target = torch.Tensor(
[[0.1323, 0.0125, 0.2900, -0.0725, -0.0102, -0.4405,
0.1226, -0.3333, -0.3186, -0.2411, 0.1790, 0.1281,
0.0739, -0.0536],
[0.1431, 0.0085, 0.2828, -0.0933, -0.0139, -0.4525,
0.0946, -0.3279, -0.3001, -0.2223, 0.2023, 0.0708,
0.0131, -0.0124]])
output_target = torch.Tensor(
[[[[ 0.0041, 0.0324, 0.0846, -0.0056, 0.0353, -0.2528, 0.0289,
-0.3333, -0.3186, -0.2411, 0.1790, 0.1281, 0.0739, -0.0536],
[ 0.0159, 0.0248, 0.1496, -0.0176, 0.0457, -0.3839, 0.0780,
-0.3137, -0.2731, -0.2310, 0.1866, 0.0758, 0.0366, -0.0069],
[ 0.0656, 0.0168, 0.2182, -0.0391, 0.0214, -0.4389, 0.1100,
-0.2625, -0.1970, -0.2249, 0.1374, 0.0337, 0.0139, 0.0284],
[ 0.1323, 0.0125, 0.2900, -0.0725, -0.0102, -0.4405, 0.1226,
-0.1649, -0.1023, -0.1823, 0.0712, 0.0039, -0.0228, 0.0444]],
[[ 0.0296, 0.0254, 0.1007, -0.0225, 0.0207, -0.2612, 0.0061,
-0.3279, -0.3001, -0.2223, 0.2023, 0.0708, 0.0131, -0.0124],
[ 0.0306, 0.0096, 0.1566, -0.0386, 0.0387, -0.3958, 0.0556,
-0.3034, -0.2701, -0.2165, 0.2061, 0.0364, -0.0012, 0.0184],
[ 0.0842, 0.0075, 0.2181, -0.0696, 0.0121, -0.4389, 0.0874,
-0.2432, -0.1979, -0.2168, 0.1519, 0.0066, -0.0080, 0.0485],
[ 0.1431, 0.0085, 0.2828, -0.0933, -0.0139, -0.4525, 0.0946,
-0.1608, -0.1140, -0.1646, 0.0796, -0.0202, -0.0207, 0.0379]]]])
self.assertTensorAlmostEqual(hidden_target, hidden)
self.assertTensorAlmostEqual(output_target, output)
def build_model(cfg: dict = None,
src_vocab: Vocabulary = None,
trg_vocab: Vocabulary = None):
src_padding_idx = src_vocab.stoi[PAD_TOKEN]
trg_padding_idx = trg_vocab.stoi[PAD_TOKEN]
src_embed = Embeddings(**cfg["encoder"]["embeddings"],
vocab_size=len(src_vocab),
padding_idx=src_padding_idx)
if cfg.get("tied_embeddings", False) \
and src_vocab.itos == trg_vocab.itos:
# share embeddings for src and trg
trg_embed = src_embed
else:
trg_embed = Embeddings(**cfg["decoder"]["embeddings"],
vocab_size=len(trg_vocab),
padding_idx=trg_padding_idx)
encoder = RecurrentEncoder(**cfg["encoder"],
emb_size=src_embed.embedding_dim)
decoder = RecurrentDecoder(**cfg["decoder"],
encoder=encoder,
vocab_size=len(trg_vocab),
emb_size=trg_embed.embedding_dim)
model = Model(encoder=encoder,
decoder=decoder,
src_embed=src_embed,
trg_embed=trg_embed,
src_vocab=src_vocab,
trg_vocab=trg_vocab)
# custom initialization of model parameters
initialize_model(model, cfg, src_padding_idx, trg_padding_idx)
return model
def test_recurrent_freeze(self):
encoder = RecurrentEncoder(freeze=True)
for n, p in encoder.named_parameters():
self.assertFalse(p.requires_grad)
def test_recurrent_input_dropout(self):
drop_prob = 0.5
encoder = RecurrentEncoder(dropout=drop_prob)
input_tensor = torch.Tensor([2, 3, 1, -1])
encoder.train()
dropped = encoder.rnn_input_dropout(input=input_tensor)
# eval switches off dropout
encoder.eval()
no_drop = encoder.rnn_input_dropout(input=input_tensor)
# when dropout is applied, remaining values are divided by drop_prob
self.assertGreaterEqual((no_drop - (drop_prob * dropped)).abs().sum(),
0)
drop_prob = 1.0
encoder = RecurrentEncoder(dropout=drop_prob)
all_dropped = encoder.rnn_input_dropout(input=input_tensor)
self.assertEqual(all_dropped.sum(), 0)
encoder.eval()
none_dropped = encoder.rnn_input_dropout(input=input_tensor)
self.assertTensorEqual(no_drop, none_dropped)
self.assertTensorEqual((no_drop - all_dropped), no_drop)
def test_recurrent_encoder_type(self):
valid_rnn_types = {"gru": GRU, "lstm": LSTM}
for name, obj in valid_rnn_types.items():
encoder = RecurrentEncoder(rnn_type=name)
self.assertEqual(type(encoder.rnn), obj)
def build_model(cfg: dict = None,
src_vocab: Vocabulary = None,
trg_vocab: Vocabulary = None) -> Model:
"""
Build and initialize the model according to the configuration.
:param cfg: dictionary configuration containing model specifications
:param src_vocab: source vocabulary
:param trg_vocab: target vocabulary
:return: built and initialized model
"""
src_padding_idx = src_vocab.stoi[PAD_TOKEN]
trg_padding_idx = trg_vocab.stoi[PAD_TOKEN]
# TODO if continue-us
src_embed = PretrainedEmbeddings(src_vocab,
trg_vocab,
**cfg["encoder"]["embeddings"],
vocab_size=len(src_vocab),
padding_idx=src_padding_idx)
# this ties source and target embeddings
# for softmax layer tying, see further below
if cfg.get("tied_embeddings", False):
if src_vocab.itos == trg_vocab.itos:
# share embeddings for src and trg
trg_embed = src_embed
else:
raise ConfigurationError(
"Embedding cannot be tied since vocabularies differ.")
else:
src_embed = PretrainedEmbeddings(src_vocab,
trg_vocab,
**cfg["encoder"]["embeddings"],
vocab_size=len(src_vocab),
padding_idx=src_padding_idx)
# build encoder
enc_dropout = cfg["encoder"].get("dropout", 0.)
enc_emb_dropout = cfg["encoder"]["embeddings"].get("dropout", enc_dropout)
if cfg["encoder"].get("type", "recurrent") == "transformer":
assert cfg["encoder"]["embeddings"]["embedding_dim"] == \
cfg["encoder"]["hidden_size"], \
"for transformer, emb_size must be hidden_size"
encoder = TransformerEncoder(**cfg["encoder"],
emb_size=src_embed.embedding_dim,
emb_dropout=enc_emb_dropout)
else:
encoder = RecurrentEncoder(**cfg["encoder"],
emb_size=src_embed.embedding_dim,
emb_dropout=enc_emb_dropout)
# build decoder
dec_dropout = cfg["decoder"].get("dropout", 0.)
dec_emb_dropout = cfg["decoder"]["embeddings"].get("dropout", dec_dropout)
if cfg["decoder"].get("type", "recurrent") == "transformer":
decoder = TransformerDecoder(**cfg["decoder"],
encoder=encoder,
vocab_size=len(trg_vocab),
emb_size=trg_embed.embedding_dim,
emb_dropout=dec_emb_dropout)
else:
decoder = RecurrentDecoder(**cfg["decoder"],
encoder=encoder,
vocab_size=len(trg_vocab),
emb_size=trg_embed.embedding_dim,
emb_dropout=dec_emb_dropout)
model = Model(encoder=encoder,
decoder=decoder,
src_embed=src_embed,
trg_embed=trg_embed,
src_vocab=src_vocab,
trg_vocab=trg_vocab)
# tie softmax layer with trg embeddings
"""
if cfg.get("tied_softmax", False):
if trg_embed.lut.weight.shape == \
model.decoder.output_layer.weight.shape:
# (also) share trg embeddings and softmax layer:
model.decoder.output_layer.weight = trg_embed.lut.weight
else:
raise ConfigurationError(
"For tied_softmax, the decoder embedding_dim and decoder "
"hidden_size must be the same."
"The decoder must be a Transformer."
f"shapes: output_layer.weight: {model.decoder.output_layer.weight.shape}; target_embed.lut.weight:{trg_embed.lut.weight.shape}")
"""
# custom initialization of model parameters
initialize_model(model, cfg, src_padding_idx, trg_padding_idx)
return model
def build_model(cfg: dict = None,
src_vocab: Vocabulary = None,
trg_vocab: Vocabulary = None,
trv_vocab: Vocabulary = None,
canonizer=None) -> Model:
"""
Build and initialize the model according to the configuration.
:param cfg: dictionary configuration containing model specifications
:param src_vocab: source vocabulary
:param trg_vocab: target vocabulary
:param trv_vocab: kb true value lookup vocabulary
:return: built and initialized model
"""
src_padding_idx = src_vocab.stoi[PAD_TOKEN]
trg_padding_idx = trg_vocab.stoi[PAD_TOKEN]
if "embedding_files" in cfg.keys(): #init from pretrained
assert not cfg.get(
"tied_embeddings", False
), "TODO implement tied embeddings along with pretrained initialization"
raise NotImplementedError(
"TODO implement kbsrc embed loading for embedding files")
weight_tensors = []
for weight_file in cfg["embedding_files"]:
with open(weight_file, "r") as f:
weight = []
for line in f.readlines():
line = line.split()
line = [float(x) for x in line]
weight.append(line)
weight = FloatTensor(weight)
weight_tensors.append(weight)
# Set source Embeddings to Pretrained Embeddings
src_embed = Embeddings(
int(weight_tensors[0][0].shape[0]),
False, #TODO transformer: change to True
len(weight_tensors[0]),
)
src_embed.lut.weight.data = weight_tensors[0]
# Set target Embeddings to Pretrained Embeddings
trg_embed = Embeddings(
int(weight_tensors[1][0].shape[0]),
False, #TODO transformer: change to True
len(weight_tensors[1]),
)
trg_embed.lut.weight.data = weight_tensors[1]
else:
src_embed = Embeddings(**cfg["encoder"]["embeddings"],
vocab_size=len(src_vocab),
padding_idx=src_padding_idx)
if cfg.get("kb_embed_separate", False):
kbsrc_embed = Embeddings(**cfg["encoder"]["embeddings"],
vocab_size=len(src_vocab),
padding_idx=src_padding_idx)
else:
kbsrc_embed = src_embed
# this ties source and target embeddings
# for softmax layer tying, see further below
if cfg.get("tied_embeddings", False):
if src_vocab.itos == trg_vocab.itos:
# share embeddings for src and trg
trg_embed = src_embed
else:
raise ConfigurationError(
"Embedding cannot be tied since vocabularies differ.")
else:
# Latest TODO: init embeddings with vocab_size = len(trg_vocab joined with kb_vocab)
trg_embed = Embeddings(**cfg["decoder"]["embeddings"],
vocab_size=len(trg_vocab),
padding_idx=trg_padding_idx)
# build encoder
enc_dropout = cfg["encoder"].get("dropout", 0.)
enc_emb_dropout = cfg["encoder"]["embeddings"].get("dropout", enc_dropout)
if cfg["encoder"].get("type", "recurrent") == "transformer":
assert cfg["encoder"]["embeddings"]["embedding_dim"] == \
cfg["encoder"]["hidden_size"], \
"for transformer, emb_size must be hidden_size"
encoder = TransformerEncoder(**cfg["encoder"],
emb_size=src_embed.embedding_dim,
emb_dropout=enc_emb_dropout)
else:
encoder = RecurrentEncoder(**cfg["encoder"],
emb_size=src_embed.embedding_dim,
emb_dropout=enc_emb_dropout)
# retrieve kb task info
kb_task = bool(cfg.get("kb", False))
k_hops = int(
cfg.get("k_hops", 1)
) # k number of kvr attention layers in decoder (eric et al/default: 1)
same_module_for_all_hops = bool(cfg.get("same_module_for_all_hops", False))
do_postproc = bool(cfg.get("do_postproc", True))
copy_from_source = bool(cfg.get("copy_from_source", True))
canonization_func = None if canonizer is None else canonizer(
copy_from_source=copy_from_source)
kb_input_feeding = bool(cfg.get("kb_input_feeding", True))
kb_feed_rnn = bool(cfg.get("kb_feed_rnn", True))
kb_multihead_feed = bool(cfg.get("kb_multihead_feed", False))
posEncKBkeys = cfg.get("posEncdKBkeys", False)
tfstyletf = cfg.get("tfstyletf", True)
infeedkb = bool(cfg.get("infeedkb", False))
outfeedkb = bool(cfg.get("outfeedkb", False))
add_kb_biases_to_output = bool(cfg.get("add_kb_biases_to_output", True))
kb_max_dims = cfg.get("kb_max_dims", (16, 32)) # should be tuple
double_decoder = cfg.get("double_decoder", False)
tied_side_softmax = cfg.get(
"tied_side_softmax",
False) # actually use separate linear layers, tying only the main one
do_pad_kb_keys = cfg.get(
"pad_kb_keys", True
) # doesnt need to be true for 1 hop (=>BIG PERFORMANCE SAVE), needs to be true for >= 2 hops
if hasattr(kb_max_dims, "__iter__"):
kb_max_dims = tuple(kb_max_dims)
else:
assert type(kb_max_dims) == int, kb_max_dims
kb_max_dims = (kb_max_dims, )
assert cfg["decoder"]["hidden_size"]
dec_dropout = cfg["decoder"].get("dropout", 0.)
dec_emb_dropout = cfg["decoder"]["embeddings"].get("dropout", dec_dropout)
if cfg["decoder"].get("type", "recurrent") == "transformer":
if tfstyletf:
decoder = TransformerDecoder(
**cfg["decoder"],
encoder=encoder,
vocab_size=len(trg_vocab),
emb_size=trg_embed.embedding_dim,
emb_dropout=dec_emb_dropout,
kb_task=kb_task,
kb_key_emb_size=kbsrc_embed.embedding_dim,
feed_kb_hidden=kb_input_feeding,
infeedkb=infeedkb,
outfeedkb=outfeedkb,
double_decoder=double_decoder)
else:
decoder = TransformerKBrnnDecoder(
**cfg["decoder"],
encoder=encoder,
vocab_size=len(trg_vocab),
emb_size=trg_embed.embedding_dim,
emb_dropout=dec_emb_dropout,
kb_task=kb_task,
k_hops=k_hops,
kb_max=kb_max_dims,
same_module_for_all_hops=same_module_for_all_hops,
kb_key_emb_size=kbsrc_embed.embedding_dim,
kb_input_feeding=kb_input_feeding,
kb_feed_rnn=kb_feed_rnn,
kb_multihead_feed=kb_multihead_feed)
else:
if not kb_task:
decoder = RecurrentDecoder(**cfg["decoder"],
encoder=encoder,
vocab_size=len(trg_vocab),
emb_size=trg_embed.embedding_dim,
emb_dropout=dec_emb_dropout)
else:
decoder = KeyValRetRNNDecoder(
**cfg["decoder"],
encoder=encoder,
vocab_size=len(trg_vocab),
emb_size=trg_embed.embedding_dim,
emb_dropout=dec_emb_dropout,
k_hops=k_hops,
kb_max=kb_max_dims,
same_module_for_all_hops=same_module_for_all_hops,
kb_key_emb_size=kbsrc_embed.embedding_dim,
kb_input_feeding=kb_input_feeding,
kb_feed_rnn=kb_feed_rnn,
kb_multihead_feed=kb_multihead_feed,
do_pad_kb_keys=do_pad_kb_keys)
# specify generator which is mostly just the output layer
generator = Generator(dec_hidden_size=cfg["decoder"]["hidden_size"],
vocab_size=len(trg_vocab),
add_kb_biases_to_output=add_kb_biases_to_output,
double_decoder=double_decoder)
model = Model(
encoder=encoder, decoder=decoder, generator=generator,
src_embed=src_embed, trg_embed=trg_embed,
src_vocab=src_vocab, trg_vocab=trg_vocab,\
kb_key_embed=kbsrc_embed,\
trv_vocab=trv_vocab,
k_hops=k_hops,
do_postproc=do_postproc,
canonize=canonization_func,
kb_att_dims=len(kb_max_dims),
posEncKBkeys=posEncKBkeys
)
# tie softmax layer with trg embeddings
if cfg.get("tied_softmax", False):
if trg_embed.lut.weight.shape == \
model.generator.output_layer.weight.shape:
# (also) share trg embeddings and softmax layer:
model.generator.output_layer.weight = trg_embed.lut.weight
if model.generator.double_decoder:
# (also also) share trg embeddings and side softmax layer
assert hasattr(model.generator, "side_output_layer")
if tied_side_softmax:
# because of distributivity this becomes O (x_1+x_2) instead of O_1 x_1 + O_2 x_2
model.generator.side_output_layer.weight = trg_embed.lut.weight
else:
raise ConfigurationError(
"For tied_softmax, the decoder embedding_dim and decoder "
"hidden_size must be the same."
"The decoder must be a Transformer.")
# custom initialization of model parameters
initialize_model(model, cfg, src_padding_idx, trg_padding_idx)
return model
def build_model(cfg: dict = None,
src_vocab: Vocabulary = None,
trg_vocab: Vocabulary = None) -> Model:
"""
Build and initialize the model according to the configuration.
:param cfg: dictionary configuration containing model specifications
:param src_vocab: source vocabulary
:param trg_vocab: target vocabulary
:return: built and initialized model
"""
logger.info("Building an encoder-decoder model...")
src_padding_idx = src_vocab.stoi[PAD_TOKEN]
trg_padding_idx = trg_vocab.stoi[PAD_TOKEN]
src_embed = Embeddings(**cfg["encoder"]["embeddings"],
vocab_size=len(src_vocab),
padding_idx=src_padding_idx)
# this ties source and target embeddings
# for softmax layer tying, see further below
if cfg.get("tied_embeddings", False):
if src_vocab.itos == trg_vocab.itos:
# share embeddings for src and trg
trg_embed = src_embed
else:
raise ConfigurationError(
"Embedding cannot be tied since vocabularies differ.")
else:
trg_embed = Embeddings(**cfg["decoder"]["embeddings"],
vocab_size=len(trg_vocab),
padding_idx=trg_padding_idx)
# build encoder
enc_dropout = cfg["encoder"].get("dropout", 0.)
enc_emb_dropout = cfg["encoder"]["embeddings"].get("dropout", enc_dropout)
if cfg["encoder"].get("type", "recurrent") == "transformer":
assert cfg["encoder"]["embeddings"]["embedding_dim"] == \
cfg["encoder"]["hidden_size"], \
"for transformer, emb_size must be hidden_size"
encoder = TransformerEncoder(**cfg["encoder"],
emb_size=src_embed.embedding_dim,
emb_dropout=enc_emb_dropout)
else:
encoder = RecurrentEncoder(**cfg["encoder"],
emb_size=src_embed.embedding_dim,
emb_dropout=enc_emb_dropout)
# build decoder
dec_dropout = cfg["decoder"].get("dropout", 0.)
dec_emb_dropout = cfg["decoder"]["embeddings"].get("dropout", dec_dropout)
if cfg["decoder"].get("type", "recurrent") == "transformer":
decoder = TransformerDecoder(**cfg["decoder"],
encoder=encoder,
vocab_size=len(trg_vocab),
emb_size=trg_embed.embedding_dim,
emb_dropout=dec_emb_dropout)
else:
decoder = RecurrentDecoder(**cfg["decoder"],
encoder=encoder,
vocab_size=len(trg_vocab),
emb_size=trg_embed.embedding_dim,
emb_dropout=dec_emb_dropout)
model = Model(encoder=encoder,
decoder=decoder,
src_embed=src_embed,
trg_embed=trg_embed,
src_vocab=src_vocab,
trg_vocab=trg_vocab)
# tie softmax layer with trg embeddings
if cfg.get("tied_softmax", False):
if trg_embed.lut.weight.shape == \
model.decoder.output_layer.weight.shape:
# (also) share trg embeddings and softmax layer:
model.decoder.output_layer.weight = trg_embed.lut.weight
else:
raise ConfigurationError(
"For tied_softmax, the decoder embedding_dim and decoder "
"hidden_size must be the same."
"The decoder must be a Transformer.")
# custom initialization of model parameters
initialize_model(model, cfg, src_padding_idx, trg_padding_idx)
# initialize embeddings from file
pretrained_enc_embed_path = cfg["encoder"]["embeddings"].get(
"load_pretrained", None)
pretrained_dec_embed_path = cfg["decoder"]["embeddings"].get(
"load_pretrained", None)
if pretrained_enc_embed_path:
logger.info("Loading pretraind src embeddings...")
model.src_embed.load_from_file(pretrained_enc_embed_path, src_vocab)
if pretrained_dec_embed_path and not cfg.get("tied_embeddings", False):
logger.info("Loading pretraind trg embeddings...")
model.trg_embed.load_from_file(pretrained_dec_embed_path, trg_vocab)
logger.info("Enc-dec model built.")
return model
def build_unsupervised_nmt_model(
cfg: dict = None,
src_vocab: Vocabulary = None,
trg_vocab: Vocabulary = None) -> UnsupervisedNMTModel:
"""
Build an UnsupervisedNMTModel.
:param cfg: model configuration
:param src_vocab: Vocabulary for the src language
:param trg_vocab: Vocabulary for the trg language
:return: Unsupervised NMT model as specified in cfg
"""
src_padding_idx = src_vocab.stoi[PAD_TOKEN]
trg_padding_idx = trg_vocab.stoi[PAD_TOKEN]
# build source and target embedding layers
# embeddings in the encoder are pretrained and stay fixed
loaded_src_embed = PretrainedEmbeddings(**cfg["encoder"]["embeddings"],
vocab_size=len(src_vocab),
padding_idx=src_padding_idx,
vocab=src_vocab,
freeze=True)
loaded_trg_embed = PretrainedEmbeddings(**cfg["decoder"]["embeddings"],
vocab_size=len(trg_vocab),
padding_idx=trg_padding_idx,
vocab=trg_vocab,
freeze=True)
# embeddings in the decoder are randomly initialised and will be learned
src_embed = Embeddings(**cfg["encoder"]["embeddings"],
vocab_size=len(src_vocab),
padding_idx=src_padding_idx,
freeze=False)
trg_embed = Embeddings(**cfg["decoder"]["embeddings"],
vocab_size=len(trg_vocab),
padding_idx=trg_padding_idx,
freeze=False)
# build shared encoder
enc_dropout = cfg["encoder"].get("dropout", 0.)
enc_emb_dropout = cfg["encoder"]["embeddings"].get("dropout", enc_dropout)
if cfg["encoder"].get("type", "recurrent") == "transformer":
assert cfg["encoder"]["embeddings"]["embedding_dim"] == \
cfg["encoder"]["hidden_size"], \
"for transformer, emb_size must be hidden_size"
shared_encoder = TransformerEncoder(**cfg["encoder"],
emb_size=src_embed.embedding_dim,
emb_dropout=enc_emb_dropout)
else:
shared_encoder = RecurrentEncoder(**cfg["encoder"],
emb_size=src_embed.embedding_dim,
emb_dropout=enc_emb_dropout)
# build src and trg language decoder
dec_dropout = cfg["decoder"].get("dropout", 0.)
dec_emb_dropout = cfg["decoder"]["embeddings"].get("dropout", dec_dropout)
if cfg["decoder"].get("type", "recurrent") == "transformer":
src_decoder = TransformerDecoder(**cfg["decoder"],
encoder=shared_encoder,
vocab_size=len(src_vocab),
emb_size=src_embed.embedding_dim,
emb_dropout=dec_emb_dropout)
trg_decoder = TransformerDecoder(**cfg["decoder"],
encoder=shared_encoder,
vocab_size=len(trg_vocab),
emb_size=trg_embed.embedding_dim,
emb_dropout=dec_emb_dropout)
else:
src_decoder = RecurrentDecoder(**cfg["decoder"],
encoder=shared_encoder,
vocab_size=len(src_vocab),
emb_size=src_embed.embedding_dim,
emb_dropout=dec_emb_dropout)
trg_decoder = RecurrentDecoder(**cfg["decoder"],
encoder=shared_encoder,
vocab_size=len(trg_vocab),
emb_size=trg_embed.embedding_dim,
emb_dropout=dec_emb_dropout)
# build unsupervised NMT model
model = UnsupervisedNMTModel(loaded_src_embed, loaded_trg_embed, src_embed,
trg_embed, shared_encoder, src_decoder,
trg_decoder, src_vocab, trg_vocab)
# initialise model
# embed_initializer should be none so loaded encoder embeddings won't be overwritten
initialize_model(model.src2src_translator, cfg, src_padding_idx,
src_padding_idx)
initialize_model(model.src2trg_translator, cfg, src_padding_idx,
trg_padding_idx)
initialize_model(model.trg2src_translator, cfg, trg_padding_idx,
src_padding_idx)
initialize_model(model.trg2src_translator, cfg, trg_padding_idx,
trg_padding_idx)
return model