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 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 build_regat(dataset, args):
print("Building ReGAT model with %s relation and %s fusion method" %
(args.relation_type, args.fusion))
w_emb = WordEmbedding(dataset.dictionary.ntoken, 300, .0, args.op)
q_emb = QuestionEmbedding(300 if 'c' not in args.op else 600, args.num_hid,
1, False, .0)
q_att = QuestionSelfAttention(args.num_hid, .2)
if args.relation_type == "semantic":
v_relation = ExplicitRelationEncoder(
dataset.v_dim,
args.num_hid,
args.relation_dim,
args.dir_num,
args.sem_label_num,
num_heads=args.num_heads,
num_steps=args.num_steps,
nongt_dim=args.nongt_dim,
residual_connection=args.residual_connection,
label_bias=args.label_bias)
elif args.relation_type == "spatial":
v_relation = ExplicitRelationEncoder(
dataset.v_dim,
args.num_hid,
args.relation_dim,
args.dir_num,
args.spa_label_num,
num_heads=args.num_heads,
num_steps=args.num_steps,
nongt_dim=args.nongt_dim,
residual_connection=args.residual_connection,
label_bias=args.label_bias)
else:
v_relation = ImplicitRelationEncoder(
dataset.v_dim,
args.num_hid,
args.relation_dim,
args.dir_num,
args.imp_pos_emb_dim,
args.nongt_dim,
num_heads=args.num_heads,
num_steps=args.num_steps,
residual_connection=args.residual_connection,
label_bias=args.label_bias)
classifier = SimpleClassifier(args.num_hid, args.num_hid * 2,
dataset.num_ans_candidates, 0.5)
gamma = 0
if args.fusion == "ban":
joint_embedding = BAN(args.relation_dim, args.num_hid, args.ban_gamma)
gamma = args.ban_gamma
elif args.fusion == "butd":
joint_embedding = BUTD(args.relation_dim, args.num_hid, args.num_hid)
else:
joint_embedding = MuTAN(args.relation_dim, args.num_hid,
dataset.num_ans_candidates, args.mutan_gamma)
gamma = args.mutan_gamma
classifier = None
return ReGAT(dataset, w_emb, q_emb, q_att, v_relation, joint_embedding,
classifier, gamma, args.fusion, args.relation_type)
def build_baseline0_newatt(dataset, num_hid):
w_emb = WordEmbedding(dataset.dictionary.ntoken, 300, 0.0)
q_emb = QuestionEmbedding(300, num_hid, 1, False, 0.0)
v_att = NewAttention(dataset.v_dim, q_emb.num_hid, num_hid)
q_net = FCNet([q_emb.num_hid, num_hid])
v_net = FCNet([dataset.v_dim, num_hid])
classifier = SimpleClassifier(num_hid, num_hid * 2,
dataset.num_ans_candidates, 0.5)
return BaseModel(w_emb, q_emb, v_att, q_net, v_net, classifier)
def build_baseline_with_onestep(embeddings,
num_ans_candidates,
debias_mode='LearnedMixin'):
assert debias_mode in [
'BiasProduct', 'ReweightByInvBias', 'LearnedMixin', 'Plain'
]
vision_features = config.output_features
visual_glimpses = config.visual_glimpses
hidden_features = config.hid_dim
question_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.Attention(
v_dim=vision_features,
q_dim=question_features,
hid_dim=hidden_features,
glimpses=visual_glimpses,
)
classifier = SimpleClassifier(in_dim=(question_features,
visual_glimpses * vision_features),
hid_dim=(hidden_features,
hidden_features * 2),
out_dim=num_ans_candidates,
dropout=0.5)
# mask_v_att = attention.Attention(
# v_dim=vision_features,
# q_dim=question_features,
# hid_dim=hidden_features,
# glimpses=visual_glimpses,
# )
#
# mask_classifier = SimpleClassifier(
# in_dim=(question_features, vision_features),
# hid_dim=(hidden_features, hidden_features * 2),
# out_dim=num_ans_candidates,
# dropout=0.5
# )
# Add the loss_fn based our arguments
debias_loss_fn = eval(debias_mode)()
return BaseModel_with_Onestep(w_emb, q_emb, v_att, classifier,
debias_loss_fn)
def build_baseline_with_twostep(embeddings,
num_ans_candidates,
debias_mode='LearnedMixin'):
assert debias_mode in [
'BiasProduct', 'ReweightByInvBias', 'LearnedMixin', 'Plain'
]
vision_features = config.output_features
visual_glimpses = config.visual_glimpses
hidden_features = config.hid_dim
question_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.Attention(
v_dim=vision_features,
q_dim=question_features,
hid_dim=hidden_features,
glimpses=visual_glimpses,
)
classifier = SimpleClassifier(in_dim=(question_features,
visual_glimpses * vision_features),
hid_dim=(hidden_features,
hidden_features * 2),
out_dim=num_ans_candidates,
dropout=0.5)
if config.use_rubi:
c1 = MLP(
input_dim=question_features,
dimensions=[1024, 1024, num_ans_candidates],
)
c2 = nn.Linear(num_ans_candidates, num_ans_candidates)
else:
c1, c2 = None, None
# Add the loss_fn based our arguments
debias_loss_fn = eval(debias_mode)(hidden_features if config.fusion_type ==
'mul' else hidden_features * 2)
return BaseModel_with_Twostep(w_emb, q_emb, v_att, classifier,
debias_loss_fn, c1, c2)
def build_regat(dataset, args):
print("Building ReGAT model with %s relation and %s fusion method" %
(args.relation_type, args.fusion))
# 词嵌入向量模型
# WordEmbedding(
# (emb): Embedding(19902, 300, padding_idx=19901)
# (emb_): Embedding(19902, 300, padding_idx=19901)
# (dropout): Dropout(p=0.0, inplace=False)
# )
w_emb = WordEmbedding(dataset.dictionary.ntoken, 300, .0,
args.op) #调用WordEmbding词嵌入方法
#问题嵌入
# QuestionEmbedding(
# (rnn): GRU(600, 1024, batch_first=True) )
q_emb = QuestionEmbedding(300 if 'c' not in args.op else 600, args.num_hid,
1, False, .0)
#问题自注意力
# QuestionSelfAttention(
# (drop): Dropout(p=0.2, inplace=False)
# (W1_self_att_q): FCNet(
# (main): Sequential(
# (0): Dropout(p=0.2, inplace=False)
# (1): Linear(in_features=1024, out_features=1024, bias=True)
# )
# )
# (W2_self_att_q): FCNet(
# (main): Sequential(
# (0): Linear(in_features=1024, out_features=1, bias=True)
# )
# )
# )
q_att = QuestionSelfAttention(args.num_hid, .2)
if args.relation_type == "semantic": #如果关系类型是语义的
v_relation = ExplicitRelationEncoder(
dataset.v_dim,
args.num_hid,
args.relation_dim,
args.dir_num,
args.sem_label_num,
num_heads=args.num_heads,
num_steps=args.num_steps,
nongt_dim=args.nongt_dim,
residual_connection=args.residual_connection,
label_bias=args.label_bias)
elif args.relation_type == "spatial": #如果关系类型是空间的
v_relation = ExplicitRelationEncoder(
dataset.v_dim,
args.num_hid,
args.relation_dim,
args.dir_num,
args.spa_label_num,
num_heads=args.num_heads,
num_steps=args.num_steps,
nongt_dim=args.nongt_dim,
residual_connection=args.residual_connection,
label_bias=args.label_bias)
else: #否则是隐式关系
v_relation = ImplicitRelationEncoder(
dataset.v_dim,
args.num_hid,
args.relation_dim,
args.dir_num,
args.imp_pos_emb_dim,
args.nongt_dim,
num_heads=args.num_heads,
num_steps=args.num_steps,
residual_connection=args.residual_connection,
label_bias=args.label_bias)
#分类器
classifier = SimpleClassifier(args.num_hid, args.num_hid * 2,
dataset.num_ans_candidates, 0.5)
gamma = 0
#采用融合方法
if args.fusion == "ban":
joint_embedding = BAN(args.relation_dim, args.num_hid, args.ban_gamma)
gamma = args.ban_gamma
elif args.fusion == "butd":
joint_embedding = BUTD(args.relation_dim, args.num_hid, args.num_hid)
else:
joint_embedding = MuTAN(args.relation_dim, args.num_hid,
dataset.num_ans_candidates, args.mutan_gamma)
gamma = args.mutan_gamma
classifier = None
return ReGAT(dataset, w_emb, q_emb, q_att, v_relation, joint_embedding,
classifier, gamma, args.fusion, args.relation_type)