def __init__(
self,
glove_path: str,
model_params: ModelParams,
hidden_dimension: int,
):
super().__init__()
self.model_params = model_params
self.word_embedding_net = WordEmbedding(
vocabulary_size=model_params.vocabulary_size,
pretrained_vectors_file=glove_path,
embedding_dimension=model_params.word_embedding_dimension,
dropout=0.25,
)
self.question_embedding_net = QuestionEmbedding(
input_dimension=model_params.word_embedding_dimension,
number_hidden_units=hidden_dimension,
number_of_layers=1,
)
self.question_projection_net = MultiLayerNet(
dimensions=[hidden_dimension, hidden_dimension], dropout=0.5)
self.image_projection_net = MultiLayerNet(
dimensions=[
model_params.object_embedding_dimension,
hidden_dimension,
],
dropout=0.5,
)
if self.model_params.add_self_attention:
self.question_self_attention_net = SelfAttention(hidden_dimension,
dropout=0.3)
self.visual_self_attention_net = SelfAttention(hidden_dimension,
dropout=0.3)
self.question_attention_net = Attention(model_params.number_of_objects,
dropout=0.3)
self.visual_attention_net = Attention(
model_params.question_sequence_length, dropout=0.3)
if model_params.fusion_method == FusionMethod.CONCAT:
factor = 3 if self.model_params.add_self_attention else 2
self.classifier = Classifier(
input_dimension=hidden_dimension * 3,
hidden_dimension=hidden_dimension * 4,
output_dimension=model_params.num_ans_candidates,
dropout=0.5,
)
elif model_params.fusion_method == FusionMethod.HADAMARD:
self.classifier = Classifier(
input_dimension=hidden_dimension,
hidden_dimension=hidden_dimension * 2,
output_dimension=model_params.num_ans_candidates,
dropout=0.5,
)
def __init__(self, args, device, collect_da_predictions=True):
super(MTL_Model3, self).__init__()
self.input_size = args.embedding_dim
self.hidden_size_u = args.lstm_sent_size
self.hidden_size_d = args.lstm_utt_size
self.num_layers = args.lstm_layers
self.num_dialogacts = args.num_classes
self.device = device
self.emb = GloveEmbedding(args)
self.only_da = True if args.loss == 'da' else False
self.bilstm_u = nn.LSTM(self.input_size,
self.hidden_size_u,
self.num_layers,
bidirectional=True,
batch_first=True)
for param in self.bilstm_u.parameters():
if len(param.shape) >= 2:
init.orthogonal_(param.data)
else:
init.normal_(param.data)
self.bilstm_d = nn.LSTM(2 * self.hidden_size_u,
self.hidden_size_d,
self.num_layers,
bidirectional=True,
batch_first=True)
for param in self.bilstm_d.parameters():
if len(param.shape) >= 2:
init.orthogonal_(param.data)
else:
init.normal_(param.data)
self.attn_u = Attention(2 * self.hidden_size_u)
self.attn_d = Attention(2 * self.hidden_size_d)
self.ff_u = nn.Linear(2 * self.hidden_size_u, self.num_dialogacts)
self.ff_d = nn.Linear(2 * self.hidden_size_d, 1)
nn.init.normal_(self.ff_d.weight, mean=0, std=1)
nn.init.normal_(self.ff_u.weight, mean=0, std=1)
self.dropout_u = nn.Dropout(args.dropout_prob)
self.collect_da_predictions = collect_da_predictions
self.da_predictions = []
#add weights to the loss function to account for the distribution of dialog acts in daily dialog
#nll_class_weights = torch.tensor([0.0, 2.1861911569232313, 3.4904300472491396, 6.120629125122877, 10.787031308006435]).to(device)
if args.num_classes == 5:
nll_class_weights = torch.tensor([0.0, 1.0, 1.0, 1.0,
1.0]).to(device)
# self.nll = nn.NLLLoss(weight=nll_class_weights, reduction='none')
self.nll = nn.CrossEntropyLoss(weight=nll_class_weights,
reduction='mean')
else:
self.nll = nn.CrossEntropyLoss(reduction='mean')
def __init__(self,
query_hidden_size,
memory_bank_size,
num_classes,
attn_mode,
dropout=0.0,
ordinal=False,
hr_enc=False):
super(WordMultiHopAttnClassifier, self).__init__()
self.memory_bank_size = memory_bank_size
self.query_hidden_size = query_hidden_size
self.num_classes = num_classes
self.hr_enc = hr_enc
self._query_vector = nn.Parameter(torch.zeros(1, query_hidden_size))
init.uniform_(self._query_vector, -0.1, 0.1)
self.attention_layer = Attention(memory_bank_size,
memory_bank_size,
coverage_attn=False,
attn_mode=attn_mode)
self.glimpse_layer = Attention(query_hidden_size,
memory_bank_size,
coverage_attn=False,
attn_mode=attn_mode)
if self.hr_enc:
self._sent_query_vector = nn.Parameter(
torch.zeros(1, query_hidden_size))
init.uniform_(self._sent_query_vector, -0.1, 0.1)
self.sent_attention_layer = Attention(memory_bank_size,
memory_bank_size,
coverage_attn=False,
attn_mode=attn_mode)
self.sent_glimpse_layer = Attention(query_hidden_size,
memory_bank_size,
coverage_attn=False,
attn_mode=attn_mode)
self.ordinal = ordinal
self.expanded_memory_size = memory_bank_size if not hr_enc else 2 * memory_bank_size
if ordinal:
self.classifier = nn.Sequential(
nn.Linear(self.expanded_memory_size,
self.expanded_memory_size), nn.Dropout(p=dropout),
nn.ReLU(), nn.Linear(self.expanded_memory_size, num_classes),
nn.Sigmoid())
else:
self.classifier = nn.Sequential(
nn.Linear(self.expanded_memory_size,
self.expanded_memory_size), nn.Dropout(p=dropout),
nn.ReLU(), nn.Linear(self.expanded_memory_size, num_classes),
nn.LogSoftmax(dim=1))
def create_model(vocab_size,
tag_size,
max_len,
rnn_units,
embedding_dims,
emb_matrix=None):
inputs = Input(shape=(max_len, ))
embedding_layer = Embedding(
input_dim=vocab_size, output_dim=embedding_dims,
input_length=max_len) if emb_matrix is None else Embedding(
input_dim=vocab_size,
output_dim=embedding_dims,
weights=[emb_matrix],
input_length=max_len,
trainable=False)
bi_rnn = Bidirectional(LSTM(units=rnn_units, return_sequences=True))
attention = Attention(max_len)
classifier = Dense(1, activation="sigmoid")
embedding = embedding_layer(inputs)
x = bi_rnn(embedding)
x = attention(x)
output = classifier(x)
model = Model(inputs, output)
model.compile(optimizer="adam",
loss="binary_crossentropy",
metrics=["accuracy"])
return model
def build_baseline(embeddings, num_ans_candidates):
vision_features = config.output_features
visual_glimpses = config.visual_glimpses
question_features = hidden_features = config.hid_dim
w_emb = WordEmbedding(embeddings, dropout=0.0)
q_emb = QuestionEmbedding(w_dim=300,
hid_dim=question_features,
nlayers=1,
bidirect=False,
dropout=0.0)
v_att = Attention(
v_dim=vision_features,
q_dim=question_features * q_emb.ndirections,
hid_dim=hidden_features,
glimpses=visual_glimpses,
)
classifier = SimpleClassifier(
in_dim=(question_features * q_emb.ndirections, vision_features),
hid_dim=(hidden_features, hidden_features * 2),
out_dim=num_ans_candidates,
dropout=0.5)
return BaseModel(w_emb, q_emb, v_att, classifier)
def __init__(
self, h_enc: int, h_dec: int, label_to_id: Dict, dropout: float = 0.,
teacher_force: float = 0., attention: Dict = None
):
"""Convert label to consequence of sublabels and use lstm cell to predict next
:param h_enc: encoder hidden state, correspond to hidden state of LSTM cell
:param h_dec: size of LSTM cell input/output
:param label_to_id: dict for converting labels to ids
:param dropout: probability of dropout
:param teacher_force: probability of teacher forcing, 0 corresponds to always use previous predicted value
:param attention: if passed, init attention with given args
"""
super().__init__(h_enc, h_dec, label_to_id)
self.teacher_force = teacher_force
self.embedding = nn.Embedding(self.out_size, self.h_dec, padding_idx=self.pad_index)
self.linear = nn.Linear(self.h_enc, self.out_size)
self.dropout = nn.Dropout(dropout)
self.use_attention = attention is not None
if self.use_attention:
self.attention = Attention(self.h_enc, self.h_dec, **attention)
lstm_input_size = self.h_enc + self.h_dec if self.use_attention else self.h_dec
self.lstm = nn.LSTM(input_size=lstm_input_size, hidden_size=self.h_enc)
def __init__(self, args):
super(Decoder, self).__init__()
self.vocab_size = args.vocab_size
self.embed_size = args.embed_size
self.hidden_size = args.hidden_size
self.layer_size = args.layer_size
self.batch_size = 0
self.device = args.device
self.beam_width = args.beam_width
self.teacher_rate = args.teacher_rate
#self.word_embed=nn.Embedding(self.vocab_size, self.embed_size,padding_idx=constants.PAD)
self.word_embed = nn.Embedding(self.vocab_size,
self.embed_size,
padding_idx=constants.PAD)
#self.hidden_exchange=nn.Linear(self.hidden_size*2,self.hidden_size)
self.gru = nn.GRU(self.embed_size,
self.hidden_size,
num_layers=args.layer_size,
bidirectional=False,
dropout=args.dropout,
batch_first=True) #decoderは双方向にできない
self.attention = Attention(args)
self.attention_wight = nn.Linear(self.hidden_size * 3,
self.hidden_size * 3)
self.out = nn.Linear(self.hidden_size * 3, self.vocab_size)
self.dropout = nn.Dropout(args.dropout)
self.activate = nn.Tanh()
def baseline(args, dataset, pretrained=False):
# initialise model
w_emb = WordEmbedding(dataset.dictionary.ntoken, 300, 0.0)
q_emb = QuestionEmbedding(300, args.num_hid, 1, False, 0.0)
v_att = Attention(dataset.v_dim, q_emb.num_hid, args.num_hid)
q_net = FCNet([args.num_hid, args.num_hid])
v_net = FCNet([dataset.v_dim, args.num_hid])
classifier = SimpleClassifier(args.num_hid, 2 * args.num_hid,
dataset.num_ans_candidates, 0.5)
model = BaseModel(args, w_emb, q_emb, v_att, q_net, v_net, classifier)
# load model on device if available
map_location = None
if not model.cuda_available:
map_location = torch.device('cpu')
# download and load pretrained model
if pretrained:
key = 'baseline-vqa'
url = pretrained_urls[key]
model.load_state_dict(download_model(
key, url, map_location=map_location)['model'],
strict=False)
else:
key = 'untrained'
# set model name
model.name = key
return model
def model(self):
# (Bi-GRU) layers
rnn_outputs, _ = bi_rnn(GRUCell(self.hidden_size),
GRUCell(self.hidden_size),
inputs=self.batch_embedded,
dtype=tf.float32)
tf.summary.histogram('RNN_outputs', rnn_outputs)
# Attention layers
with tf.name_scope('Attention_layer'):
attention_ = Attention(rnn_outputs,
self.attention_size,
time_major=False,
return_alphas=True)
self.attention_output, alphas = attention_.attentionModel()
tf.summary.histogram('alphas', alphas)
print('attention_output.shape:', self.attention_output.shape)
def __init__(self, train_dset, word_embedding_size=300, id_embedding_size=32,
hidden_size=100, num_layers=1, bidirectional=False, da=100, r=10, attention_size=32, num_latent=32,
dropout=0.5, soa_size=50):
super(HANCI, self).__init__()
review_hidden_size = hidden_size * (1 + int(bidirectional))
self.user_word_emb = WordEmbedding(train_dset.dictionary.ntoken, word_embedding_size)
self.item_word_emb = WordEmbedding(train_dset.dictionary.ntoken, word_embedding_size)
self.user_review_emb = ReviewLSTMSA(embedding_size=word_embedding_size, dropout=dropout,
hidden_size=hidden_size, num_layers=num_layers, bidirectional=bidirectional,
da=da, r=r)
self.item_review_emb = ReviewLSTMSA(embedding_size=word_embedding_size, dropout=dropout,
hidden_size=hidden_size, num_layers=num_layers, bidirectional=bidirectional,
da=da, r=r)
self.user_id_emb = WordEmbedding(train_dset.user_num, id_embedding_size, dropout=0.)
self.item_id_emb = WordEmbedding(train_dset.item_num, id_embedding_size, dropout=0.)
self.user_rid_emb = WordEmbedding(train_dset.item_num, id_embedding_size, dropout=0.)
self.item_rid_emb = WordEmbedding(train_dset.user_num, id_embedding_size, dropout=0.)
self.user_reviews_att = Attention(review_hidden_size, id_embedding_size, attention_size)
self.item_reviews_att = Attention(review_hidden_size, id_embedding_size, attention_size)
self.user_soft_attention = SoftAttention(review_hidden_size, id_embedding_size, soa_size)
self.item_soft_attention = SoftAttention(review_hidden_size, id_embedding_size, soa_size)
self.relu = nn.ReLU()
self.user_latent = nn.Linear(review_hidden_size, num_latent)
self.item_latent = nn.Linear(review_hidden_size, num_latent)
self.classify = nn.Linear(num_latent, 1, bias=False)
self.dropout = nn.Dropout(dropout)
# 分别用作user和item的偏置
self.user_bias = WordEmbedding(ntoken=train_dset.user_num, emb_dim=1, dropout=0.)
self.item_bias = WordEmbedding(ntoken=train_dset.item_num, emb_dim=1, dropout=0.)
self.global_bias = nn.Parameter(torch.Tensor([0.1]))
# 初始化
self.user_word_emb.init_embedding()
self.item_word_emb.init_embedding()
self.user_bias.init_embedding_with_one(0.1)
self.item_bias.init_embedding_with_one(0.1)
def build_baseline0(dataset, num_hid):
w_emb = WordEmbedding(dataset.dictionary.ntoken, 300, 0.0)
q_emb = QuestionEmbedding(300, num_hid, 1, False, 0.0)
v_att = Attention(dataset.v_dim, q_emb.num_hid, num_hid)
q_net = FCNet([num_hid, num_hid])
v_net = FCNet([dataset.v_dim, num_hid])
classifier = SimpleClassifier(num_hid, 2 * num_hid,
dataset.num_ans_candidates, 0.5)
return BaseModel(w_emb, q_emb, v_att, q_net, v_net, classifier)
def get_model(embedding_matrix, name):
context_embedding_layer = Embedding(input_dim=embedding_matrix.shape[0],
output_dim=embedding_matrix.shape[1],
weights=[embedding_matrix],
input_length=max_context_seq_length,
trainable=False,
name='c_emb' + name)
question_embedding_layer = Embedding(input_dim=embedding_matrix.shape[0],
output_dim=embedding_matrix.shape[1],
weights=[embedding_matrix],
input_length=max_question_seq_length,
trainable=False,
name='q_emb' + name)
# The 2 inputs to our model
c_seq_input = Input(shape=(max_context_seq_length, ),
dtype='int32') # [batch_size, n]
q_seq_input = Input(shape=(max_question_seq_length, ),
dtype='int32') # [batch_size, m]
# Embed both question and context
c_embedded = context_embedding_layer(c_seq_input) # [batch_size, n, d]
q_embedded = question_embedding_layer(q_seq_input) # [batch_size, m, d]
# Bidirectional LSTMs/GRUs used as encoders
c_encoder_out = Bidirectional(LSTM(units, return_sequences=True))(
c_embedded) # [batch_size, n, 2l]
q_encoder_out = Bidirectional(LSTM(units, return_sequences=True))(
q_embedded) # [batch_size, n, 2l]
# Interaction/attention layer, output shape
G = Attention()([c_encoder_out, q_encoder_out]) # [batch_size, n, 4l]
# Modeling layer
m_1 = Bidirectional(LSTM(units,
return_sequences=True))(G) # [batch_size, n, 2l]
# m_2 = Bidirectional(LSTM(units, return_sequences=True, dropout=dropout))(m_1) # [batch_size, n, 2l]
# m_3 = Bidirectional(LSTM(units, return_sequences=True, dropout=dropout))(m_2)
concat1_out = Concatenate(axis=-1)([G, m_1])
ps_start_ = TimeDistributed(Dense(1))(concat1_out) # [batch_size, n, 1]
ps_start_flatten = Flatten()(ps_start_) # [batch_size, n]
ps_start = Activation('softmax')(ps_start_flatten)
# concat2_out = Concatenate(axis=-1)([G, m_3])
ps_end_ = TimeDistributed(Dense(1))(concat1_out) # [batch_size, n, 1]
ps_end_flatten = Flatten()(ps_end_) # [batch_size, n]
ps_end = Activation('softmax')(ps_end_flatten)
model = Model(inputs=[c_seq_input, q_seq_input],
outputs=[ps_start, ps_end])
return model
def __init__(self, char_embedding, input_dim, hidden_size_decoder, shared_context, word_recurrent_cell=None,
drop_out_word_cell=0, timing=False, drop_out_char_embedding_decoder=0,
char_src_attention=False, unrolling_word=False, init_context_decoder=True,
hidden_size_src_word_encoder=None, generator=None, stable_decoding_state=False,
verbose=0):
super(CharDecoder, self).__init__()
self.generator = generator
self.timing = timing
self.char_embedding_decoder = char_embedding
self.shared_context = shared_context
self.unrolling_word = unrolling_word
self.stable_decoding_state = stable_decoding_state
self.init_context_decoder = init_context_decoder
printing("WARNING : stable_decoding_state is {}", var=[stable_decoding_state], verbose_level=0, verbose=verbose)
printing("WARNING : init_context_decoder is {}", var=[init_context_decoder], verbose_level=0, verbose=verbose)
printing("WARNING : DECODER unrolling_word is {}", var=[unrolling_word], verbose_level=0, verbose=verbose)
printing("WARNING : DECODER char_src_attention is {}", var=[char_src_attention], verbose_level=0, verbose=verbose)
self.drop_out_char_embedding_decoder = nn.Dropout(drop_out_char_embedding_decoder)
if word_recurrent_cell is not None:
assert word_recurrent_cell in SUPPORED_WORD_ENCODER, \
"ERROR : word_recurrent_cell should be in {} ".format(SUPPORED_WORD_ENCODER)
if word_recurrent_cell is None:
word_recurrent_cell = nn.GRU
else:
word_recurrent_cell = eval("nn."+word_recurrent_cell)
if isinstance(word_recurrent_cell, nn.LSTM):
printing("WARNING : in the case of LSTM : inital states defined as "
" h_0, c_0 = (zero tensor, source_conditioning) so far (cf. row 70 decoder.py) ",
verbose=self.verbose, verbose_level=0)
printing("WARNING : LSTM only using h_0 for prediction not the cell", verbose=self.verbose, verbose_level=0)
if char_src_attention:
assert hidden_size_src_word_encoder is not None, "ERROR : need hidden_size_src_word_encoder for attention "
# we need to add dimension because of the context vector that is hidden_size encoder projected
#input_dim += hidden_size_src_word_encoder : # NO NEED anymire as we project the all context : same size as currnt
printing("WARNING : DECODER word_recurrent_cell hidden dim will be {} "
"(we added hidden_size_decoder) because of attention", verbose=verbose, verbose_level=0)
# if stable_decoding_state : we add a projection of the attention context vector + the stable one
# TODO : try to project the cat of those three vectors (char, attention context, stable context)
self.context_proj = nn.Linear(hidden_size_decoder*int(stable_decoding_state)+hidden_size_src_word_encoder*int(char_src_attention), char_embedding.embedding_dim) if stable_decoding_state or char_src_attention else None
input_dim = 2*input_dim if stable_decoding_state or char_src_attention else input_dim # because we concat with projected context
self.seq_decoder = word_recurrent_cell(input_size=input_dim, hidden_size=hidden_size_decoder,
num_layers=1,
dropout=drop_out_word_cell,
bias=True, batch_first=True, bidirectional=False)
printing("MODEL Decoder : word_recurrent_cell has been set to {} ".format(str(word_recurrent_cell)),
verbose=verbose, verbose_level=1)
#self.attn_param = nn.Linear(hidden_size_decoder*1) if char_src_attention else None
self.attn_layer = Attention(hidden_size_word_decoder=hidden_size_decoder,
char_embedding_dim=self.char_embedding_decoder.embedding_dim,
time=self.timing,
hidden_size_src_word_encoder=hidden_size_src_word_encoder) if char_src_attention else None
self.verbose = verbose
def __init__(self):
super(Decoder, self).__init__()
self.n_frames_per_step = hps.n_frames_per_step
self.n_mel_channels = hps.n_mel_channels
self.encoder_embedding_dim = hps.encoder_embedding_dim
self.attention_rnn_dim = hps.attention_rnn_dim
self.decoder_rnn_dim = hps.decoder_rnn_dim
self.prenet = Prenet()
self.attention_rnn = nn.LSTMCell(256 + 512, 1024)
self.attention_layer = Attention()
# decoder rnn input : 256 + 512 = 768
# decoder rnn output : 1024
self.decoder_rnn = nn.LSTMCell(256 + 512, 1024, 1)
self.linear_projection = LinearNorm(1024, 80 * 3)
def __init__(self, config, entity_embedding, context_embedding):
super(DKN, self).__init__()
self.config = config
self.kcnn = KCNN(config, entity_embedding, context_embedding)
if self.config.use_attention:
self.attention = Attention(config)
# TODO parameters
self.dnn = nn.Sequential(
nn.Linear(
len(self.config.window_sizes) * 2 * self.config.num_filters,
16), nn.Linear(16, 1))
def __init__(
self,
embedding_cfg={
"input_dim": -1,
"output_dim": 64,
"embeddings_initializer": "uniform",
"embeddings_regularizer": "l2",
"activity_regularizer": "l2",
"mask_zero": True,
"input_length": 40
},
LSTM_cfg={
"units": 64,
"activation": "tanh",
"recurrent_activation": "sigmoid",
"use_bias": True,
"kernel_initializer": "glorot_uniform",
"bias_initializer": "zeros",
"unit_forget_bias": True,
"kernel_regularizer": "l2",
"recurrent_regularizer": "l2",
"bias_regularizer": "l2",
"activity_regularizer": "l2",
"dropout": 0.1,
"recurrent_dropout": 0.1,
"implementation": 2,
"return_sequences": True,
"return_state": True,
"go_backwards": False,
"stateful": False,
"unroll": False
},
num_lstm_layer=1,
max_length=40,
fully_connected_cfg={
"units": -1,
"activation": None,
"use_bias": True
},
attention_cfg={
"hidden_size": 64,
"num_heads": 1
}):
super(Decoder, self).__init__()
self.max_length = 40
self.embedding = tf.keras.layers.Embedding(**embedding_cfg)
self.lstm_layers = []
for _ in range(num_lstm_layer):
self.lstm_layers.append(tf.keras.layers.LSTM(**LSTM_cfg))
self.attention = Attention(**attention_cfg)
self.fully_connected = tf.keras.layers.Dense(**fully_connected_cfg)
def BidLstm(maxlen, max_features, embed_size, embedding_matrix):
inp = Input(shape=(maxlen, ))
x = Embedding(max_features, embed_size, weights=[embedding_matrix],
trainable=False)(inp)
x = Bidirectional(LSTM(300, return_sequences=True, dropout=0.25,
recurrent_dropout=0.25))(x)
x = Attention(maxlen)(x)
x = Dense(256, activation="relu")(x)
x = Dropout(0.25)(x)
x = Dense(256, activation="relu")(x)
x = Dropout(0.25)(x)
x = Dense(6, activation="sigmoid")(x)
model = Model(inputs=inp, outputs=x)
return model
def init_model_from_ckpt():
_, _, _, train_data, valid_data, test_data = get_dataloaders_and_data()
SRC_PAD_IDX = DOC.vocab.stoi[DOC.pad_token]
TRG_PAD_IDX = DOC.vocab.stoi[DOC.pad_token]
INPUT_DIM = len(DOC.vocab)
OUTPUT_DIM = len(DOC.vocab)
attn = Attention(ENC_HID_DIM, DEC_HID_DIM)
enc = Encoder(INPUT_DIM, ENC_EMB_DIM, ENC_HID_DIM, DEC_HID_DIM,
ENC_DROPOUT)
dec = Decoder(OUTPUT_DIM, DEC_EMB_DIM, ENC_HID_DIM, DEC_HID_DIM,
DEC_DROPOUT, attn)
model = Seq2Seq(enc, dec, SRC_PAD_IDX, device).to(device)
most_recent_ckpt = get_most_recent_ckpt('ckpts')
model.load_state_dict(torch.load(most_recent_ckpt))
return model, train_data, valid_data, test_data
def __init__(self, hidden_size, output_size, n_layers=1, dropout=0.1,\
embedding = None):
super(Decoder, self).__init__()
self.output_size = output_size
self.n_layers = n_layers
self.dropout = dropout
if isinstance(embedding, nn.Embedding):
self.embedding = embedding
else:
self.embedding = nn.Embedding(output_size, hidden_size)
if config.use_attn:
self.attention = Attention(hidden_size)
self.rnn = nn.LSTM(hidden_size,
hidden_size,
num_layers=n_layers,
dropout=dropout)
self.out = nn.Linear(hidden_size, output_size)
def __init__(self,
embed_size,
hidden_size,
output_size,
n_layers=1,
dropout=0.2):
super(Decoder, self).__init__()
self.embed_size = embed_size
self.hidden_size = hidden_size
self.output_size = output_size
self.n_layers = n_layers
self.embed = nn.Embedding(output_size, embed_size)
self.dropout = nn.Dropout(dropout, inplace=True)
self.attention = Attention(hidden_size)
self.gru = nn.GRU(hidden_size + embed_size,
hidden_size,
n_layers,
dropout=dropout)
self.out = nn.Linear(hidden_size * 2, output_size)
def Lstm(maxlen, max_features, embed_size, embedding_matrix):
inp = Input(shape=(maxlen, ))
x = Embedding(max_features,
embed_size,
weights=[embedding_matrix],
trainable=False)(inp)
x = LSTM(300, return_sequences=True, dropout=0.25,
recurrent_dropout=0.25)(x)
x = LSTM(300, return_sequences=True, dropout=0.25,
recurrent_dropout=0.25)(x)
x = Attention(maxlen)(x)
x = Dense(256, activation="relu")(x)
inp_convai = Input(shape=(3, ))
x = concatenate([x, inp_convai])
x = Dropout(0.25)(x)
x = Dense(6, activation="sigmoid")(x)
model = Model(inputs=[inp, inp_convai], outputs=x)
return model
def __init__(self,
drug_length,
protein_length,
n_drug,
n_protein,
embd_dim=20,
filter_width=5,
layer_size=1):
"""
Args:
drug_length: drug compound input size, i.e. number of atoms
protein_length: protein input size, i.e number of amino acid
n_drug: length of the drug embedding dictionary
n_protein: length of the protein embedding dictionary (if we only use amino acid as a "word" this is only 20)
embd_dim: embedding vector length
filter_width: convolutional filter width
layer_size: number of abcnn2 layers
"""
super(Model, self).__init__()
self.layer_size = layer_size
# Embedding Layers
self.embd_drug = nn.Embedding(n_drug, embd_dim)
self.embd_protein = nn.Embedding(n_protein, embd_dim)
# ABCNN layers
self.conv = nn.ModuleList([
Convolution(embd_dim, embd_dim, filter_width, 1)
for _ in range(layer_size)
])
self.attn = nn.ModuleList([
Attention(drug_length, protein_length, filter_width)
for _ in range(layer_size)
])
# all-ap average pooling layers
self.ap = AllAP(embd_dim)
# final classifier
self.fc = nn.Linear(embd_dim * 2, 2)
def __init__(self,
config,
no_words,
no_answers,
resnet_model,
lstm_size,
emb_size,
use_pretrained=True):
super(Net, self).__init__()
self.use_pretrained = use_pretrained # whether to use pretrained ResNet
self.word_cnt = no_words # total count of words
self.ans_cnt = no_answers # total count of valid answers
self.lstm_size = lstm_size # lstm emb size to be passed to CBN layer
self.emb_size = emb_size # hidden layer size of MLP used to predict delta beta and gamma parameters
self.config = config # config file containing the values of parameters
self.cbn = config['model']['image']['cbn']['use_cbn']
self.embedding = nn.Embedding(self.word_cnt, self.emb_size)
self.lstm = VariableLengthLSTM(self.config['model']).cuda()
self.net = create_resnet(resnet_model, self.lstm_size, self.emb_size,
self.use_pretrained, self.cbn)
self.attention = Attention(self.config).cuda()
self.que_mlp = nn.Sequential(
nn.Linear(config['model']['no_hidden_LSTM'],
config['model']['no_question_mlp']),
nn.Tanh(),
)
self.img_mlp = nn.Sequential(
nn.Linear(2048, config['model']['no_image_mlp']),
nn.Tanh(),
)
self.dropout = nn.Dropout(config['model']['dropout_keep_prob'])
self.final_mlp = nn.Linear(config['model']['no_hidden_final_mlp'],
self.ans_cnt)
def __init__(self,
items_total: int,
embed_dim: int,
num_attn_queries,
time_encoding='none',
transformer_num_heads=8,
transformer_num_layers=2,
temperature: float=1.0,
dropout=0.1,
set_embed_dropout=0.1):
"""
Args:
pool(str): sum/mean/max
"""
super(SSM, self).__init__()
self.items_total = items_total
self.embed_dim = embed_dim
self.item_embed = nn.Embedding(num_embeddings=items_total, embedding_dim=embed_dim)
self.num_attn_queries = num_attn_queries
if num_attn_queries == -1:
self.attention_pool = Pool(pool='mean')
else:
self.attention_pool = AttentionPool(embed_dim, num_attn_queries, dropout=set_embed_dropout)
self.time_encoding_method = time_encoding
if time_encoding == 'none':
self.time_encoding = None
elif time_encoding == 'positional':
self.time_encoding = PositionalEncoding(embed_dim=embed_dim)
elif time_encoding == 'timestamp':
self.time_encoding = TimestampEncoding(embed_dim=embed_dim)
else:
raise NotImplementedError()
self.transformer_encoder = nn.TransformerEncoder(encoder_layer=nn.TransformerEncoderLayer(d_model=embed_dim,
nhead=transformer_num_heads,
dropout=dropout),
num_layers=transformer_num_layers)
self.attn = Attention(embed_dim=embed_dim, temperature=temperature)
self.items_bias = nn.Parameter(torch.zeros(items_total))
def __init__(self,
items_total: int,
embed_dim: int = 256,
time_encoding='none',
transformer_num_heads: int = 8,
transformer_num_layers: int = 2,
temperature: float = 1.0,
dropout: float = 0.1):
"""
Args:
items_total:
embed_dim:
time_encoding: 'none'/'positional'/'timestamp'
transformer_num_heads:
transformer_num_layers:
dropout:
"""
super(ISM, self).__init__()
self.items_total = items_total
self.embed_dim = embed_dim
self.item_embed = nn.Embedding(num_embeddings=items_total,
embedding_dim=embed_dim)
self.time_encoding_method = time_encoding
if time_encoding == 'none': self.time_encoding = None
elif time_encoding == 'positional':
self.time_encoding = PositionalEncoding(embed_dim=embed_dim)
elif time_encoding == 'timestamp':
self.time_encoding = TimestampEncoding(embed_dim=embed_dim)
else:
raise NotImplementedError()
self.transformer_encoder = nn.TransformerEncoder(
encoder_layer=nn.TransformerEncoderLayer(
d_model=embed_dim,
nhead=transformer_num_heads,
dropout=dropout),
num_layers=transformer_num_layers)
self.attn = Attention(embed_dim=embed_dim, temperature=temperature)
self.items_bias = nn.Parameter(torch.zeros(items_total))
def __init__(self, embedding, cfg):
super(DecoderRNN_Attn, self).__init__()
self.b = cfg.par.beam_size
self.n = cfg.par.n_best
### embedding layer
self.embedding = embedding # [voc_length x emb_size] contains nn.Embedding()
self.V = self.embedding.num_embeddings #vocabulary size
self.E = self.embedding.embedding_dim #embedding size
self.L = cfg.num_layers
self.D = 2 if cfg.bidirectional else 1 ### num of directions
self.H = cfg.hidden_size
self.cuda = cfg.cuda
self.pointer = cfg.pointer
self.coverage = cfg.coverage
self.tt = torch.cuda if self.cuda else torch
### dropout layer to apply on top of the embedding layer
self.dropout = nn.Dropout(cfg.par.dropout)
### set up the RNN
dropout = cfg.par.dropout if self.L > 1 else 0.0 #dropout option adds dropout after all but last recurrent layer, so non-zero dropout expects num_layers greater than 1
if cfg.cell == "lstm":
self.rnn = nn.LSTM(
self.E + self.H, self.H, self.L, dropout=dropout
) #input is embedding+hidden (to allow feed-input)
elif cfg.cell == "gru":
self.rnn = nn.GRU(self.E + self.H, self.H, self.L, dropout=dropout)
else:
sys.exit("error: bad -cell {} option. Use: lstm OR gru\n".format(
cfg.cell))
### Attention mechanism
self.attn = Attention(self.H, cfg.attention, cfg.coverage, cfg.cuda)
### pgen layer
if self.pointer: self.pgen = nn.Linear(self.H * 2 + self.E, 1)
### concat layer
self.concat = nn.Linear(self.H * 2, self.H)
### output layer
self.output = nn.Linear(self.H, self.V)
def __init__(self,
items_total: int,
embed_dim: int,
time_encoding: str,
set_embed_method: str,
num_set_embeds: int,
num_transformer_heads: int,
num_transformer_layers: int,
itm_temperature: float = 1.0,
stm_temperature: float = 1.0,
dropout: float = 0.1,
set_embed_dropout: float = 0.1,
attn_output: bool = True):
"""
Args:
items_total:
embed_dim:
time_encoding:
set_embed_method:
num_transformer_heads:
num_set_embeds:
dropout:
"""
super(DSNTSP, self).__init__()
self.items_total = items_total
self.embed_dim = embed_dim
self.item_embed = nn.Embedding(num_embeddings=items_total,
embedding_dim=embed_dim)
# time encoding
self.time_encoding_method = time_encoding
if time_encoding == 'none':
self.time_encoding = None
elif time_encoding == 'positional':
self.time_encoding = PositionalEncoding(embed_dim=embed_dim)
elif time_encoding == 'timestamp':
self.time_encoding = TimestampEncoding(embed_dim=embed_dim)
else:
raise NotImplementedError()
# set embedding
self.set_embed_method = set_embed_method
if set_embed_method == 'attn_pool':
self.set_embed = AttentionPool(embed_dim,
num_queries=num_set_embeds,
dropout=set_embed_dropout)
else:
raise NotImplementedError()
# Dual Transformer
# self.dual_transformer = DualTransformer(layers=[
# PDTE(embed_dim, num_transformer_heads, dropout=dropout),
# CDTE(embed_dim, num_transformer_heads, dropout=dropout),
# PDTE(embed_dim, num_transformer_heads, dropout=dropout),
# CDTE(embed_dim, num_transformer_heads, dropout=dropout),
# PDTE(embed_dim, num_transformer_heads, dropout=dropout),
# CDTE(embed_dim, num_transformer_heads, dropout=dropout),
# PDTE(embed_dim, num_transformer_heads, dropout=dropout),
# CDTE(embed_dim, num_transformer_heads, dropout=dropout)
# ])
# co-transformer
self.co_transformer = CoTransformer(layer=CoTransformerLayer(
embed_dim, num_transformer_heads, dropout=dropout),
num_layers=num_transformer_layers)
# attention-based prediction
self.item_attn = Attention(embed_dim=embed_dim,
temperature=itm_temperature)
self.set_attn = Attention(embed_dim=embed_dim,
temperature=stm_temperature)
self.items_bias = nn.Parameter(torch.zeros(items_total))
# gate network
if attn_output:
self.gate_net = nn.Sequential(nn.Linear(embed_dim * 3, embed_dim),
nn.Sigmoid())
else:
self.gate_net = nn.Sequential(nn.Linear(embed_dim * 2, embed_dim),
nn.Sigmoid())
self.attn_output = attn_output
embedding_matrix = None
if args.gpt2 == False:
# 加载word_dict
if not os.path.exists(args.data_path) or not os.path.exists(args.word_dict):
write_data(args)
with open(args.word_dict) as f:
for line in f:
word_dict = json.loads(line)
p_np, m_np, r_np = get_pmr(args, word_dict) # (32, 1)
args.vocab_size = len(word_dict)
# model
attention = Attention(args.hidden_size)
encoder = Encoder(args, device, embedding_matrix, args.vocab_size, word_dict, args.embedding_dim, args.hidden_size,
dropout=args.dropout)
decoder = Decoder(args, attention, encoder, word_dict, args.vocab_size, args.embedding_dim, args.char_num,
args.hidden_size, device,
dropout=args.dropout)
model = PreStory(args, word_dict, device, encoder, decoder, args.embedding_dim, args.pmr_size,
args.hidden_size)
if args.use_cuda and args.gpu_para:
model = nn.DataParallel(model, device_ids=[int(i) for i in args.device.split(',')]) # multi-GPU
torch.backends.cudnn.benchmark = True
model = model.to(device)
# 记录模型参数数量
num_parameters = 0
def __init__(self,
train_dset,
word_embedding_size=300,
id_embedding_size=32,
filter_sizes=[3],
num_filters=100,
attention_size=32,
num_latent=32,
dropout=0.5,
soa_size=50):
super(NARRE_SOA, self).__init__()
self.review_num_filters = num_filters * len(filter_sizes)
self.user_word_emb = WordEmbedding(train_dset.dictionary.ntoken,
word_embedding_size)
self.item_word_emb = WordEmbedding(train_dset.dictionary.ntoken,
word_embedding_size)
self.user_review_emb = ReviewCNN(train_dset.user_review_len,
embedding_size=word_embedding_size,
filter_sizes=filter_sizes,
num_filters=num_filters)
self.item_review_emb = ReviewCNN(train_dset.item_review_len,
embedding_size=word_embedding_size,
filter_sizes=filter_sizes,
num_filters=num_filters)
self.user_id_emb = WordEmbedding(train_dset.user_num,
id_embedding_size,
dropout=0.)
self.item_id_emb = WordEmbedding(train_dset.item_num,
id_embedding_size,
dropout=0.)
self.user_rid_emb = WordEmbedding(train_dset.item_num,
id_embedding_size,
dropout=0.)
self.item_rid_emb = WordEmbedding(train_dset.user_num,
id_embedding_size,
dropout=0.)
self.user_reviews_att = Attention(self.review_num_filters,
id_embedding_size, attention_size)
self.item_reviews_att = Attention(self.review_num_filters,
id_embedding_size, attention_size)
self.user_soft_attention = SoftAttention(self.review_num_filters,
id_embedding_size, soa_size)
self.item_soft_attention = SoftAttention(self.review_num_filters,
id_embedding_size, soa_size)
self.user_latent = nn.Linear(self.review_num_filters, num_latent)
self.item_latent = nn.Linear(self.review_num_filters, num_latent)
self.relu = nn.ReLU()
self.classify = nn.Linear(num_latent, 1, bias=False)
self.dropout = nn.Dropout(dropout)
# 分别用作user和item的偏置
self.user_bias = WordEmbedding(ntoken=train_dset.user_num,
emb_dim=1,
dropout=0.)
self.item_bias = WordEmbedding(ntoken=train_dset.item_num,
emb_dim=1,
dropout=0.)
self.global_bias = nn.Parameter(torch.Tensor([0.1]))
# 初始化
self.user_word_emb.init_embedding()
self.item_word_emb.init_embedding()
self.user_bias.init_embedding_with_one(0.1)
self.item_bias.init_embedding_with_one(0.1)
class Decoder(nn.Module):
def __init__(self):
super(Decoder, self).__init__()
self.n_frames_per_step = hps.n_frames_per_step
self.n_mel_channels = hps.n_mel_channels
self.encoder_embedding_dim = hps.encoder_embedding_dim
self.attention_rnn_dim = hps.attention_rnn_dim
self.decoder_rnn_dim = hps.decoder_rnn_dim
self.prenet = Prenet()
self.attention_rnn = nn.LSTMCell(256 + 512, 1024)
self.attention_layer = Attention()
# decoder rnn input : 256 + 512 = 768
# decoder rnn output : 1024
self.decoder_rnn = nn.LSTMCell(256 + 512, 1024, 1)
self.linear_projection = LinearNorm(1024, 80 * 3)
def get_go_frame(self, memory):
batch_size = memory.size(0)
go_frames = Variable(
memory.data.new(batch_size, self.n_frames_per_step *
self.n_mel_channels).zero_())
return go_frames
def parse_decoder_inputs(self, decoder_inputs):
batch_size = decoder_inputs.size(0)
frame_size = decoder_inputs.size(2)
decoder_inputs = decoder_inputs.transpose(1, 2).contiguous()
# print('decoder input transpose : ', decoder_inputs.size())
decoder_inputs = decoder_inputs.view(
batch_size, int(frame_size / self.n_frames_per_step), -1)
# print('decoder input view : ', decoder_inputs.size())
decoder_inputs = decoder_inputs.transpose(0, 1)
# print('decoder input transpose : ', decoder_inputs.size())
return decoder_inputs
def parse_decoder_outputs(self, mel_outputs):
# List[(B, 240) ....] -> (len(List) : 278, B, 240)
mel_outputs = torch.stack(mel_outputs)
# print(mel_outputs.size())
mel_outputs = mel_outputs.transpose(0, 1).contiguous()
# print(mel_outputs.size())
batch_size = mel_outputs.size(0)
mel_outputs = mel_outputs.view(batch_size, -1, 80)
# print(mel_outputs.size())
mel_outputs = mel_outputs.transpose(1, 2)
# print(mel_outputs.size())
return mel_outputs
def initailze_decoder_states(self, memory, mask):
batch_size = memory.size(0)
max_time = memory.size(1)
self.attention_hidden = Variable(
memory.data.new(batch_size, self.attention_rnn_dim).zero_())
self.attention_cell = Variable(
memory.data.new(batch_size, self.attention_rnn_dim).zero_())
self.decoder_hidden = Variable(
memory.data.new(batch_size, self.decoder_rnn_dim).zero_())
self.decoder_cell = Variable(
memory.data.new(batch_size, self.decoder_rnn_dim).zero_())
self.attention_weights = Variable(
memory.data.new(batch_size, max_time).zero_())
self.attention_weights_cum = Variable(
memory.data.new(batch_size, max_time).zero_())
self.attention_context = Variable(
memory.data.new(batch_size, self.encoder_embedding_dim).zero_())
# (B, Seq_len, 512)
self.memory = memory
# (B, Seq_len, 128)
self.processed_memory = self.attention_layer.memory_layer(memory)
self.mask = mask
def decode(self, decoder_input):
# decoder input : (B, 256)
# attention context : (B, 512)
# attention cell input : (B, 256+512)
attention_cell_input = torch.cat(
(decoder_input, self.attention_context), dim=-1)
print('attention cell input : ', attention_cell_input.size())
# attention_hidden : (B, 1024)
# attention_cell : (B, 1024)
self.attention_hidden, self.attention_cell = self.attention_rnn(
attention_cell_input, (self.attention_hidden, self.attention_cell))
print('attention hidden, cell : ', self.attention_hidden.size(),
self.attention_cell.size())
self.attention_hidden = F.dropout(self.attention_hidden, 0.1,
self.training)
attention_weights_cat = torch.cat(
(self.attention_weights.unsqueeze(1),
self.attention_weights_cum.unsqueeze(1)),
dim=1)
print('attention_weights_cat : ', attention_weights_cat.size())
print('attention_weights : ', self.attention_weights.size())
print('attention_weights_cum : ', self.attention_weights_cum.size())
self.attention_context, self.attention_weights = self.attention_layer(
self.attention_hidden, self.memory, self.processed_memory,
attention_weights_cat, self.mask)
return None, None
def forward(self, memory, decoder_inputs, memory_lengths):
# memory : (B, Seq_len, 512) --> encoder outputs
# decoder_inputs : (B, Mel_Channels : 80, frames)
# memory lengths : (B)
decoder_input = self.get_go_frame(memory).unsqueeze(0)
# print('go frames : ', decoder_input.size())
# print('decoder inputs : ', decoder_inputs.size())
decoder_inputs = self.parse_decoder_inputs(decoder_inputs)
decoder_inputs = torch.cat((decoder_input, decoder_inputs), dim=0)
# print('decoder inputs : ', decoder_inputs.size())
decoder_inputs = self.prenet(decoder_inputs)
self.initailze_decoder_states(
memory, mask=~get_mask_from_lengths(memory_lengths))
mel_outputs, alignments = [], []
while len(mel_outputs) < decoder_inputs.size(0) - 1:
decoder_input = decoder_inputs[len(mel_outputs)]
mel_output, attention_weights = self.decode(decoder_input)
# mel_output : (1, B, 240)
mel_outputs.append(mel_output)
break