def __init__(self, num_answers, fn_type="softmax"):
super().__init__()
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(args, max_seq_length=MAX_VQA_LENGTH)
hid_dim = self.lxrt_encoder.dim
print("Size of Hidden Dimension:", hid_dim)
fc_dim = int(hid_dim)
print("Size of Hidden Dimension:", fc_dim)
# Type Predictor
self.type_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, 3))
self.sigmoid = nn.Sigmoid()
self.tanh = nn.Tanh()
self.softmax = nn.Softmax()
if fn_type == "tanh":
self.fn = self.tanh
print("FN: TANH")
elif fn_type == "softmax":
self.fn = self.softmax
print("FN: SOFTMAX")
else:
self.fn = self.sigmoid
print("FN: SIGMOID")
# YESNO feedforward
self.yesno_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(2 * hid_dim, fc_dim), GeLU(),
BertLayerNorm(fc_dim, eps=1e-12))
# NUMBER feedforward
self.number_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(2 * hid_dim, fc_dim), GeLU(),
BertLayerNorm(fc_dim, eps=1e-12))
# OTHER feedforward
self.other_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(2 * hid_dim, fc_dim), GeLU(),
BertLayerNorm(fc_dim, eps=1e-12))
# Answering Heads
self.logit_fc1 = nn.Sequential(nn.Linear(4 * fc_dim, hid_dim * 2),
GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, hid_dim))
# Answering Heads
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
def __init__(self, num_answers, model_type='full'):
super().__init__()
self.model_type = model_type
self.lxrt_encoder = LXRTEncoder(args,
max_seq_length=MAX_CLF_LENGTH,
model_type=args.model_type)
hid_dim = self.lxrt_encoder.dim
if num_answers == 2:
output_dim = 1
else:
output_dim = num_answers
if self.model_type != 'concat':
self.logit_fc = nn.Sequential(
nn.Dropout(args.dropout), nn.Linear(hid_dim, hid_dim * 2),
GeLU(), BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Dropout(args.dropout), nn.Linear(hid_dim * 2, output_dim))
else:
linear = nn.Linear(hid_dim, output_dim)
self.logit_fc = nn.Sequential(
nn.Dropout(args.dropout),
linear,
)
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
def __init__(self, num_answers):
super().__init__()
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(args, max_seq_length=MAX_VQA_LENGTH)
hid_dim = self.lxrt_encoder.dim
#Build Decoder with Attention
self.decoder = DecoderWithAttention(attention_dim=hid_dim,
embed_dim=hid_dim,
decoder_dim=hid_dim,
vocab_size=vocab_size,
features_dim=hid_dim,
dropout=0.5)
# VQA Answer heads
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
self.lstm = nn.LSTM(input_size=hid_dim,
hidden_size=hid_dim,
num_layers=1,
batch_first=True)
self.linear = nn.Linear(hid_dim,
vocab_size) #vocab size of bert is 30000
def __init__(self, num_answers):
super().__init__()
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(
args,
max_seq_length=MAX_VQA_LENGTH
)
hid_dim = self.lxrt_encoder.dim
# VQA Answer heads
self.logit_fc = nn.Sequential(
nn.Linear(hid_dim, hid_dim * 2),
GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers)
)
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
# from https://dl.fbaipublicfiles.com/pythia/detectron_model/detectron_model.pth
# https://dl.fbaipublicfiles.com/pythia/detectron_model/detectron_model.yaml
self.args = SimpleNamespace(
# model_file= 'data/faster-rcnn-r101.pth',
model_file='data/R-50-FPN.pth',
config_file='data/R-50-FPN.yaml',
# config_file='../vqa-faster-rcnn/configs/visual_genome_vqa/e2e_faster_rcnn_X-101-64x4d-FPN_1x_MLP_2048_FPN_512_vqa_test.yaml',
batch_size=args.batch_size,
num_features=36,
feature_name="fc6",
confidence_threshold=0,
background=True,
partition=0)
self.detection_model = self._build_detection_model()
def __init__(self, num_answers):
super().__init__()
self.lxrt_encoder = LXRTEncoder(args, max_seq_length=MAX_GQA_LENGTH)
hid_dim = self.lxrt_encoder.dim
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
def __init__(self, metric):
super(MatchingDecoderLV, self).__init__()
HIDDEN_DECODER_SIZE = 256
hid_dim = 768
self.lang_proj = nn.Sequential(
nn.Linear(hid_dim, hid_dim * 2),
GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, HIDDEN_DECODER_SIZE)
)
self.vis_proj = nn.Sequential(
nn.Linear(hid_dim, hid_dim * 2),
GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, HIDDEN_DECODER_SIZE)
)
self.metric = metric
assert metric in ['sdp', 'cosine']
def __init__(self):
super().__init__()
self.lxrt_encoder = LXRTEncoder(args, max_seq_length=20)
self.hid_dim = hid_dim = self.lxrt_encoder.dim
self.logit_fc = nn.Sequential(nn.Linear(hid_dim * 2, hid_dim * 2),
GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, 2))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
def __init__(self, num_answers):
super().__init__()
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(args, max_seq_length=MAX_VQA_LENGTH)
hid_dim = self.lxrt_encoder.dim
# VQA Answer heads
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, 300))
self.logit_fc_ans = nn.Sequential(
nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers))
self.emb_proj = nn.Sequential(nn.Linear(300, hid_dim), GeLU(),
BertLayerNorm(hid_dim, eps=1e-12),
nn.Linear(hid_dim, 300))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
def __init__(self, num_blocks):
super().__init__()
# Build LXRT encoder
#TODO: Make a new class in entry file
self.policy_lxrt_encoder = PolicyLXRTEncoder(
args, max_seq_length=MAX_VQA_LENGTH)
hid_dim = self.policy_lxrt_encoder.dim
# VQA Answer heads
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_blocks))
self.logit_fc.apply(self.policy_lxrt_encoder.model.init_bert_weights)
def __init__(self, model_type='full'):
super().__init__()
self.model_type = model_type
self.lxrt_encoder = LXRTEncoder(args,
max_seq_length=MAX_RANK_LENGTH,
model_type=args.model_type)
self.hid_dim = hid_dim = self.lxrt_encoder.dim
if self.model_type != 'concat':
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim), GeLU(),
BertLayerNorm(hid_dim, eps=1e-12),
nn.Linear(hid_dim, 1))
else:
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, 1), )
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
def __init__(self, num_answers):
super().__init__()
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(args, max_seq_length=MAX_VQA_LENGTH)
hid_dim = self.lxrt_encoder.dim
# VQA Answer heads
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
device = torch.device('cpu')
detection_model = torch.load(args.detection_model, map_location=device)
self.detection_model = detection_model['model'].float().fuse().eval()
def __init__(self, num_answers):
super().__init__()
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(args, max_seq_length=MAX_VQA_LENGTH)
hid_dim = self.lxrt_encoder.dim
# VQA Answer heads
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
self.tokenizer = BertTokenizer.from_pretrained("bert-base-uncased",
do_lower_case=True)
def __init__(self, num_answers, attention=False):
super().__init__()
print(f"Making {__name__}")
self.flag = True
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(args,
max_seq_length=MAX_VQA_LENGTH,
attention=attention)
hid_dim = self.lxrt_encoder.dim
# VQA Answer heads
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
def __init__(self, args, width=100, height=100):
super().__init__()
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(args, max_seq_length=MAX_SEQ_LENGTH)
hid_dim = self.lxrt_encoder.dim
num_logits = width * height
self.width = width
self.height = height
self.n_actions = num_logits
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_logits))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
self.use_detectron = args.use_detectron
if self.use_detectron:
print('Detectron will be used.')
data_path = DATA_PATH
vg_classes = []
with open(os.path.join(data_path, 'objects_vocab.txt')) as f:
for object in f.readlines():
vg_classes.append(object.split(',')[0].lower().strip())
MetadataCatalog.get("vg").thing_classes = vg_classes
yaml_file = DETECTRON2_YAML
cfg = get_cfg()
cfg.merge_from_file(yaml_file)
cfg.MODEL.RPN.POST_NMS_TOPK_TEST = 300
cfg.MODEL.ROI_HEADS.NMS_THRESH_TEST = 0.6
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.6
# VG Weight
cfg.MODEL.WEIGHTS = "http://nlp.cs.unc.edu/models/faster_rcnn_from_caffe.pkl"
self.predictor = DefaultPredictor(cfg)
else:
print('Resnet will be used.')
self.cnn = nn.Sequential(
*(list(models.resnet18(
pretrained=True).children())[:-3])).cuda().eval()
self.cnn2box = nn.Linear(256, 2048)
self.preprocess = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
def __init__(self, num_classes=2):
super().__init__()
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(args,
max_seq_length=MAX_UTTERANCE_LENGTH)
hid_dim = self.lxrt_encoder.dim
# VCSD image features dimensions adjuster
self.adaptive_pool = nn.AdaptiveAvgPool2d((36, 2048))
# VCSD Classification head
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_classes))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
def __init__(self, num_answers, finetune_strategy='standard'):
super().__init__()
#self.finetune_strategy = finetune_strategy
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(args,
max_seq_length=MAX_VQA_LENGTH,
finetune_strategy=finetune_strategy)
hid_dim = self.lxrt_encoder.dim
# VQA Answer heads
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
def create_head(self, num_answers):
hid_dim = self.lxrt_encoder.dim
if self.logit_fc is None:
if self.encoder_type == 'lxrt':
self.logit_fc = nn.Sequential(
nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers))
else:
self.logit_fc = nn.Linear(1024, num_answers)
init_weights = (
self.lxrt_encoder.model.init_bert_weights
if not isinstance(self.lxrt_encoder.model, nn.DataParallel)
else self.lxrt_encoder.model.module.init_bert_weights)
self.logit_fc.apply(init_weights)
return
self.logit_fc[-1] = nn.Linear(hid_dim * 2, num_answers)
def __init__(self, num_answers):
super().__init__()
self.lxrt_encoder = LXRTEncoder(
args,
max_seq_length=MAX_GQA_LENGTH
)
hid_dim = self.lxrt_encoder.dim
self.logit_fc = nn.Sequential(
nn.Linear(hid_dim, hid_dim * 2),
GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers)
)
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
if args.task_pointer != 'none':
self.matching_decoder = MatchingDecoderLV(metric='sdp')
def __init__(self, num_answers):
super().__init__()
self.lxrt_encoder = LXRTEncoder(
args,
max_seq_length=MAX_GQA_LENGTH,
mode='xl'
)
hid_dim = self.lxrt_encoder.dim
self.logit_fc = nn.Sequential(
nn.Linear(hid_dim, hid_dim * 2),
GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers)
)
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
if args.task_nsp_qfpm or args.task_mlm_qfpm:
self.qfpm = BertPreTrainingHeads(BertConfig(vocab_size_or_config_json_file = 30522),
self.lxrt_encoder.model.bert.embeddings.word_embeddings.weight)
def __init__(self, num_answers):
super().__init__()
# Build LXRT encoder
# lxrt.entry.LXRTEncoder -> LXRTFeatureExtraction -> LXRTModel
self.lxrt_encoder = LXRTEncoder(
args,
max_seq_length=MAX_PVQA_LENGTH
)
hid_dim = self.lxrt_encoder.dim
# VQA Answer heads
self.logit_fc = nn.Sequential(
nn.Linear(hid_dim, hid_dim * 2),
GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers)
)
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
def __init__(self):
super(KDDModel, self).__init__()
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(args, mode='lx')
hid_dim = self.lxrt_encoder.dim
self.config = self.lxrt_encoder.model.config
# Image-text heads
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, 2))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
# AMSoftmax loss heads
self.logit_W = torch.nn.Parameter(torch.randn(hid_dim, 2),
requires_grad=True)
nn.init.xavier_normal_(self.logit_W, gain=1)
# MLM heads
self.cls = BertPreTrainingHeads(
self.config,
self.lxrt_encoder.model.bert.embeddings.word_embeddings.weight)
def __init__(self, num_answers, fn_type="softmax"):
super().__init__()
# Build LXRT encoder
self.lxrt_encoder = LXRTEncoder(args, max_seq_length=MAX_VQA_LENGTH)
hid_dim = self.lxrt_encoder.dim
print("Size of Hidden Dimension:", hid_dim)
fc_dim = int(hid_dim)
print("Size of Hidden Dimension:", fc_dim)
self.sigmoid = nn.Sigmoid()
self.tanh = nn.Tanh()
self.softmax = nn.Softmax()
if fn_type == "tanh":
self.fn = self.tanh
print("FN: TANH")
elif fn_type == "softmax":
self.fn = self.softmax
print("FN: SOFTMAX")
else:
self.fn = self.sigmoid
print("FN: SIGMOID")
# YN:AND/OR/NOT/NONE Type Predictor
self.yn_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, 4))
# AND FF
self.and_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(2 * hid_dim, fc_dim), GeLU(),
BertLayerNorm(fc_dim, eps=1e-12))
# OR FF
self.or_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(2 * hid_dim, fc_dim), GeLU(),
BertLayerNorm(fc_dim, eps=1e-12))
# NOT FF
self.not_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(2 * hid_dim, fc_dim), GeLU(),
BertLayerNorm(fc_dim, eps=1e-12))
# NONE FF
self.none_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(2 * hid_dim, fc_dim), GeLU(),
BertLayerNorm(fc_dim, eps=1e-12))
# Answering Heads
self.logit_fc1 = nn.Sequential(nn.Linear(6 * fc_dim, hid_dim * 2),
GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, hid_dim))
self.logit_fc = nn.Sequential(nn.Linear(hid_dim, hid_dim * 2), GeLU(),
BertLayerNorm(hid_dim * 2, eps=1e-12),
nn.Linear(hid_dim * 2, num_answers))
self.logit_fc.apply(self.lxrt_encoder.model.init_bert_weights)
