def __init__(self, config, img_feature_dim=2052):
# Features + Positions (2048 + 4)
config.img_feature_dim = img_feature_dim
config.img_feature_type = "faster_r-cnn"
config.code_voc = 512
# Original Repo uses 0.3 dropout
config.hidden_dropout_prob = 0.3
super(BertO, self).__init__(config)
self.embeddings = BertEmbeddings(config)
self.encoder = CaptionBertEncoder(config)
self.pooler = BertPooler(config)
self.img_dim = config.img_feature_dim
logger.info('BertImgModel Image Dimension: {}'.format(self.img_dim))
self.img_feature_type = config.img_feature_type
if hasattr(config, 'use_img_layernorm'):
self.use_img_layernorm = config.use_img_layernorm
else:
self.use_img_layernorm = None
if config.img_feature_type == 'dis_code':
self.code_embeddings = nn.Embedding(config.code_voc,
config.code_dim,
padding_idx=0)
self.img_embedding = nn.Linear(config.code_dim,
self.config.hidden_size,
bias=True)
elif config.img_feature_type == 'dis_code_t': # transpose
self.code_embeddings = nn.Embedding(config.code_voc,
config.code_dim,
padding_idx=0)
self.img_embedding = nn.Linear(config.code_size,
self.config.hidden_size,
bias=True)
elif config.img_feature_type == 'dis_code_scale': # scaled
self.input_embeddings = nn.Linear(config.code_dim,
config.code_size,
bias=True)
self.code_embeddings = nn.Embedding(config.code_voc,
config.code_dim,
padding_idx=0)
self.img_embedding = nn.Linear(config.code_dim,
self.config.hidden_size,
bias=True)
else:
self.img_embedding = nn.Linear(self.img_dim,
self.config.hidden_size,
bias=True)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
if self.use_img_layernorm:
self.LayerNorm = BertLayerNorm(config.hidden_size,
eps=config.img_layer_norm_eps)
self.init_weights()
def __init__(self, config):
super(BertPredictionHeadTransform, self).__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
if isinstance(config.hidden_act,
str) or (sys.version_info[0] == 2
and isinstance(config.hidden_act, unicode)):
self.transform_act_fn = ACT2FN[config.hidden_act]
else:
self.transform_act_fn = config.hidden_act
self.LayerNorm = BertLayerNorm(config.hidden_size, eps=1e-12)
def __init__(self, config):
super(BertOutput, self).__init__()
self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
self.LayerNorm = BertLayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)