def __init__(self):
super(AttBasicModel, self).__init__()
self.ss_prob = 0.0 # Schedule sampling probability
self.vocab_size = cfg.MODEL.VOCAB_SIZE + 1 # include <BOS>/<EOS>
self.att_dim = cfg.MODEL.ATT_FEATS_EMBED_DIM \
if cfg.MODEL.ATT_FEATS_EMBED_DIM > 0 else cfg.MODEL.ATT_FEATS_DIM
# word embed
sequential = [nn.Embedding(self.vocab_size, cfg.MODEL.WORD_EMBED_DIM)]
sequential.append(utils.activation(cfg.MODEL.WORD_EMBED_ACT))
if cfg.MODEL.WORD_EMBED_NORM == True:
sequential.append(nn.LayerNorm(cfg.MODEL.WORD_EMBED_DIM))
if cfg.MODEL.DROPOUT_WORD_EMBED > 0:
sequential.append(nn.Dropout(cfg.MODEL.DROPOUT_WORD_EMBED))
self.word_embed = nn.Sequential(*sequential)
# global visual feat embed
sequential = []
if cfg.MODEL.GVFEAT_EMBED_DIM > 0:
sequential.append(
nn.Linear(cfg.MODEL.GVFEAT_DIM, cfg.MODEL.GVFEAT_EMBED_DIM))
sequential.append(utils.activation(cfg.MODEL.GVFEAT_EMBED_ACT))
if cfg.MODEL.DROPOUT_GV_EMBED > 0:
sequential.append(nn.Dropout(cfg.MODEL.DROPOUT_GV_EMBED))
self.gv_feat_embed = nn.Sequential(
*sequential) if len(sequential) > 0 else None
# attention feats embed
sequential = []
if cfg.MODEL.ATT_FEATS_EMBED_DIM > 0:
sequential.append(
nn.Linear(cfg.MODEL.ATT_FEATS_DIM,
cfg.MODEL.ATT_FEATS_EMBED_DIM))
sequential.append(utils.activation(cfg.MODEL.ATT_FEATS_EMBED_ACT))
if cfg.MODEL.DROPOUT_ATT_EMBED > 0:
sequential.append(nn.Dropout(cfg.MODEL.DROPOUT_ATT_EMBED))
if cfg.MODEL.ATT_FEATS_NORM == True:
sequential.append(torch.nn.LayerNorm(
cfg.MODEL.ATT_FEATS_EMBED_DIM))
self.att_embed = nn.Sequential(
*sequential) if len(sequential) > 0 else None
self.dropout_lm = nn.Dropout(
cfg.MODEL.DROPOUT_LM) if cfg.MODEL.DROPOUT_LM > 0 else None
self.logit = nn.Linear(cfg.MODEL.RNN_SIZE, self.vocab_size)
self.p_att_feats = nn.Linear(self.att_dim, cfg.MODEL.ATT_HIDDEN_SIZE) \
if cfg.MODEL.ATT_HIDDEN_SIZE > 0 else None
# bilinear
if cfg.MODEL.BILINEAR.DIM > 0:
self.p_att_feats = None
self.encoder_layers = blocks.create(
cfg.MODEL.BILINEAR.ENCODE_BLOCK,
embed_dim=cfg.MODEL.BILINEAR.DIM,
att_type=cfg.MODEL.BILINEAR.ATTTYPE,
att_heads=cfg.MODEL.BILINEAR.HEAD,
att_mid_dim=cfg.MODEL.BILINEAR.ENCODE_ATT_MID_DIM,
att_mid_drop=cfg.MODEL.BILINEAR.ENCODE_ATT_MID_DROPOUT,
dropout=cfg.MODEL.BILINEAR.ENCODE_DROPOUT,
layer_num=cfg.MODEL.BILINEAR.ENCODE_LAYERS)
def __init__(
self,
embed_dim,
dropout,
att_type,
att_heads,
att_mid_dim,
att_mid_drop,
bifeat_emb_act,
bifeat_emb_drop,
ff_dropout
):
super(EncoderLayer, self).__init__()
self.encoder_attn = LowRank(
embed_dim = embed_dim,
att_type = att_type,
att_heads = att_heads,
att_mid_dim = att_mid_dim,
att_mid_drop = att_mid_drop)
self.dropout = nn.Dropout(dropout)
self.bifeat_emb = nn.Sequential(
nn.Linear(2 * embed_dim, embed_dim),
utils.activation(bifeat_emb_act),
nn.Dropout(bifeat_emb_drop)
)
self.layer_norm = torch.nn.LayerNorm(embed_dim)
self.ff_layer = blocks.create(
'FeedForward',
embed_dim = embed_dim,
ffn_embed_dim = embed_dim * 4,
relu_dropout = ff_dropout,
dropout = ff_dropout)
def initialize():
# Data folder
if not os.path.isdir('data'):
os.mkdir('data')
# Initialization & settings
server.initialize()
ban.load()
server.set_name(SERV_NAME)
server.set_port(SERV_PORT)
server.set_max_players(SERV_MAX_PLAYERS)
server.set_timeout(SERV_TIMEOUT)
# Callback
server.call_on_message_received(on_message_received)
server.set_interpretor_function(exec_instruction)
server.call_to_validate_player(validate_player)
# Blocks
blocks.create()
def __init__(
self,
embed_dim,
dropout,
att_type,
att_heads,
att_mid_dim,
att_mid_drop,
bifeat_emb_act,
bifeat_emb_drop,
ff_dropout,
last_layer = False
):
super(DecoderLayer, self).__init__()
self.last_layer = last_layer
self.word_attn = LowRank(
embed_dim = embed_dim,
att_type = att_type,
att_heads = att_heads,
att_mid_dim = att_mid_dim,
att_mid_drop = att_mid_drop)
self.word_dropout = nn.Dropout(dropout)
self.cross_att = LowRank(
embed_dim = embed_dim,
att_type = att_type,
att_heads = att_heads,
att_mid_dim = att_mid_dim,
att_mid_drop = att_mid_drop)
self.cross_dropout = nn.Dropout(dropout)
self.layer_norm_cross = torch.nn.LayerNorm(embed_dim)
if self.last_layer == False:
self.bifeat_emb = nn.Sequential(
nn.Linear(2 * embed_dim, embed_dim),
utils.activation(bifeat_emb_act),
nn.Dropout(bifeat_emb_drop)
)
self.layer_norm_x = torch.nn.LayerNorm(embed_dim)
self.ff_layer = blocks.create(
'FeedForward',
embed_dim = embed_dim,
ffn_embed_dim = embed_dim * 4,
relu_dropout = ff_dropout,
dropout = ff_dropout)
self.layer_norm_gx = torch.nn.LayerNorm(embed_dim)
def __init__(self, args):
super(XLAN, self).__init__()
self.num_layers = 2
# First LSTM layer
rnn_input_size = cfg.MODEL.RNN_SIZE + cfg.MODEL.BILINEAR.DIM
self.att_lstm = nn.LSTMCell(rnn_input_size, cfg.MODEL.RNN_SIZE)
self.ctx_drop = nn.Dropout(cfg.MODEL.DROPOUT_LM)
self.attention = blocks.create(
cfg.MODEL.BILINEAR.DECODE_BLOCK,
embed_dim=cfg.MODEL.BILINEAR.DIM,
att_type=cfg.MODEL.BILINEAR.ATTTYPE,
att_heads=cfg.MODEL.BILINEAR.HEAD,
att_mid_dim=cfg.MODEL.BILINEAR.DECODE_ATT_MID_DIM,
att_mid_drop=cfg.MODEL.BILINEAR.DECODE_ATT_MID_DROPOUT,
dropout=cfg.MODEL.BILINEAR.DECODE_DROPOUT,
layer_num=cfg.MODEL.BILINEAR.DECODE_LAYERS)
self.att2ctx = nn.Sequential(
nn.Linear(cfg.MODEL.BILINEAR.DIM + cfg.MODEL.RNN_SIZE,
2 * cfg.MODEL.RNN_SIZE), nn.GLU())