def __init__(self, configs):
BaseModel.__init__(self, configs)
self.encoder = TransformerEncoder(configs)
self.pair_scorer = ScoreModule(self.get_pair_embs_size(),
[configs['ffnn_size']] *
configs['ffnn_depth'],
configs['dropout_rate'])
# GENE embeddings (if use_gene_features enabled)
if configs['use_gene_features']:
self.gene_dim = GENE2DIM.get(self.configs['gene_variant'],
GENE_DIM)
self.event2emb = get_event2geneemb(configs['gene_variant'])
for e in self.event2emb:
self.event2emb[e] = self.event2emb[e].to(self.device)
self.defaultgene = nn.Embedding(1, self.gene_dim)
# Initialize embeddings
for name, param in self.named_parameters():
if (not 'transformer'
in name.lower()) and 'embedding' in name.lower():
print('Re-initialize embedding {}'.format(name))
param.data.uniform_(-0.1, 0.1)
# Move model to device
self.to(self.device)
def __init__(self, name, tokenizer, optimizer):
BaseModel.__init__(self, name, tokenizer, optimizer)
def __init__(self, name, tokenizer, optimizer):
BaseModel.__init__(self, name, tokenizer, optimizer)
# Vectorize the data.
self.input_texts = []
self.target_texts = []
self.input_characters = set()
self.target_characters = set()
for ch in self.CHARS_BASIC:
self.input_characters.add(ch)
self.target_characters.add(ch)
lines = data.load_clean_sentences('both')
for line in lines:
input_text = line[1] # Swedish
target_text = line[0] # English
# We use "tab" as the "start sequence" character
# for the targets, and "\n" as "end sequence" character.
target_text = self.CH_START + target_text + self.CH_END
self.input_texts.append(input_text)
self.target_texts.append(target_text)
for char in input_text:
if char not in self.input_characters:
self.input_characters.add(char)
for char in target_text:
if char not in self.target_characters:
self.target_characters.add(char)
self.input_characters = sorted(list(self.input_characters))
self.target_characters = sorted(list(self.target_characters))
self.num_encoder_tokens = len(self.input_characters)
self.num_decoder_tokens = len(self.target_characters)
self.max_encoder_seq_length = max(
[len(txt) for txt in self.input_texts])
self.max_decoder_seq_length = max(
[len(txt) for txt in self.target_texts])
print('Number of samples:', len(self.input_texts))
print('Number of unique input tokens:', self.num_encoder_tokens)
print('Number of unique output tokens:', self.num_decoder_tokens)
print('Max sequence length for inputs:', self.max_encoder_seq_length)
print('Max sequence length for outputs:', self.max_decoder_seq_length)
self.input_token_index = dict([
(char, i) for i, char in enumerate(self.input_characters)
])
self.target_token_index = dict([
(char, i) for i, char in enumerate(self.target_characters)
])
self.encoder_input_data = np.zeros(
(len(self.input_texts), self.max_encoder_seq_length,
self.num_encoder_tokens),
dtype='float32')
self.decoder_input_data = np.zeros(
(len(self.input_texts), self.max_decoder_seq_length,
self.num_decoder_tokens),
dtype='float32')
self.decoder_target_data = np.zeros(
(len(self.input_texts), self.max_decoder_seq_length,
self.num_decoder_tokens),
dtype='float32')
for i, (input_text, target_text) in enumerate(
zip(self.input_texts, self.target_texts)):
for t, char in enumerate(input_text):
self.encoder_input_data[i, t,
self.input_token_index[char]] = 1.
for t, char in enumerate(target_text):
# decoder_target_data is ahead of decoder_input_data by one timestep
self.decoder_input_data[i, t,
self.target_token_index[char]] = 1.
if t > 0:
# decoder_target_data will be ahead by one timestep
# and will not include the start character.
self.decoder_target_data[
i, t - 1, self.target_token_index[char]] = 1.
# Reverse-lookup token index to decode sequences back to
# something readable.
self.reverse_input_char_index = dict(
(i, char) for char, i in self.input_token_index.items())
self.reverse_target_char_index = dict(
(i, char) for char, i in self.target_token_index.items())
def __init__(self,
name,
tokenizer,
optimizer,
include_dropout=False,
latent_dim=256,
reverse_order=False,
bidi=False):
"""
:param reverse_order: If True, reverse the order of input tokens to ease training
"""
BaseModel.__init__(self, name, tokenizer, optimizer)
# Collection all tokens across all input lines
self.include_dropout = include_dropout
self.latent_dim = latent_dim
self.reverse_order = reverse_order
self.bidi = bidi # If true, use a Bidirectional wrapper around the encoder LSTM
self.other_tokens = set() # input
self.eng_tokens = {self.CH_START, self.CH_END} # target
# Collection all tokens across all input lines
for idx, line in enumerate(self.eng_texts):
self.eng_texts[
idx] = self.CH_START + self.eng_texts[idx] + self.CH_END
self.eng_tokenized[idx] = [
self.CH_START
] + self.eng_tokenized[idx] + [self.CH_END]
for token in self.other_tokenized[idx]:
self.other_tokens.add(token)
for token in self.eng_tokenized[idx]:
self.eng_tokens.add(token)
self.other_tokens = sorted(list(self.other_tokens))
self.eng_tokens = sorted(list(self.eng_tokens))
self.num_encoder_tokens = len(self.other_tokens)
self.num_decoder_tokens = len(self.eng_tokens)
self.max_encoder_seq_length = max(
[len(txt) for txt in self.other_tokenized])
self.max_decoder_seq_length = max(
[len(txt) for txt in self.eng_tokenized])
print('Number of samples:', self.num_samples)
print('Number of unique input tokens:', self.num_encoder_tokens)
print('Number of unique output tokens:', self.num_decoder_tokens)
print('Max sequence length for inputs:', self.max_encoder_seq_length)
print('Max sequence length for outputs:', self.max_decoder_seq_length)
self.input_token_index = dict([
(token, i) for i, token in enumerate(self.other_tokens)
])
self.target_token_index = dict([
(token, i) for i, token in enumerate(self.eng_tokens)
])
self.encoder_input_data = numpy.zeros(
(self.num_samples, self.max_encoder_seq_length,
self.num_encoder_tokens),
dtype='uint8')
self.decoder_input_data = numpy.zeros(
(self.num_samples, self.max_decoder_seq_length,
self.num_decoder_tokens),
dtype='uint8')
self.decoder_target_data = numpy.zeros(
(self.num_samples, self.max_decoder_seq_length,
self.num_decoder_tokens),
dtype='uint8')
# Create one-hot encoded values directly
for i, (input_text, target_text) in enumerate(
zip(self.other_tokenized, self.eng_tokenized)):
for t, token in enumerate(input_text):
self.encoder_input_data[i, t,
self.input_token_index[token]] = 1.
if reverse_order:
self.encoder_input_data = numpy.flip(self.encoder_input_data,
1)
for t, token in enumerate(target_text):
# decoder_target_data is ahead of decoder_input_data by one timestep
self.decoder_input_data[i, t,
self.target_token_index[token]] = 1.
if t > 0:
# decoder_target_data will be ahead by one timestep
# and will not include the start character.
self.decoder_target_data[
i, t - 1, self.target_token_index[token]] = 1.
# Reverse-lookup token index to decode sequences back to something readable.
self.reverse_input_token_index = dict(
(i, token) for token, i in self.input_token_index.items())
self.reverse_target_token_index = dict(
(i, token) for token, i in self.target_token_index.items())
def __init__(self):
BaseModel.__init__(self)
self.name = map_collections["scope"]
def __init__(self):
BaseModel.__init__(self)
self.name = map_collections["rol_relation"]
