def build_ParalCoAtt(task_name, dataset, params):
# w_emb = WordEmbedding(dataset.dictionary.ntoken, 300, 0.0)
# q_emb = QuestionEmbedding(300, num_hid, 1, False, 0.0)
num_hid = params['num_hid']
q_proj = FCNet([768, num_hid])
bi_num_hid = num_hid * 2
co_atts = nn.ModuleList([
ParalCoAttention(dataset.v_dim,
num_hid,
num_hid,
inter_dims=params['scale'],
R=len(params['scale']))
for _ in range(params['reasonSteps'])
])
v_fusion_att = paraAttention(fuse_dim=dataset.v_dim,
glimpses=params['sub_nums'],
inputs_dim=dataset.v_dim,
att_dim=num_hid)
q_fusion_att = paraAttention(fuse_dim=num_hid,
glimpses=params['sub_nums'],
inputs_dim=num_hid,
att_dim=num_hid)
context_gate = FCNet([bi_num_hid, bi_num_hid])
classifier = SimpleClassifier(bi_num_hid, num_hid * 2, 1, 0.5)
return ActionModel(task_name, q_proj, co_atts, q_fusion_att, v_fusion_att,
context_gate, classifier)
def build_temporalAtt(task_name, dataset, params):
# w_emb = WordEmbedding(dataset.dictionary.ntoken, 300, 0.0)
# q_emb = QuestionEmbedding(300, num_hid, 1, False, 0.0)
num_hid = params['num_hid']
q_proj = FCNet([768, num_hid])
bi_num_hid = num_hid*2
co_att = CoAttention(dataset.v_dim, num_hid, bi_num_hid)
v_fusion_att = paraAttention(fuse_dim=dataset.v_dim, glimpses=params['sub_nums'], inputs_dim=dataset.v_dim, att_dim=num_hid)
q_fusion_att = paraAttention(fuse_dim=num_hid, glimpses=params['sub_nums'], inputs_dim=num_hid, att_dim=num_hid)
classifier = SimpleClassifier(
bi_num_hid, num_hid * 2, dataset.num_ans_candidates, 0.5)
return FrameQAModel(task_name, q_proj, co_att, q_fusion_att, v_fusion_att, classifier)
def build_temporalAtt(task_name, dataset, params):
num_hid = params['num_hid']
q_proj = FCNet([768, num_hid])
bi_num_hid = num_hid * 2
co_att = CoAttention(dataset.v_dim, num_hid, bi_num_hid)
v_fusion_att = paraAttention(fuse_dim=dataset.v_dim,
glimpses=params['sub_nums'],
inputs_dim=dataset.v_dim,
att_dim=num_hid)
q_fusion_att = paraAttention(fuse_dim=num_hid,
glimpses=params['sub_nums'],
inputs_dim=num_hid,
att_dim=num_hid)
classifier = SimpleClassifier(2 * num_hid, num_hid * 2, 1, 0.5)
return ActionModel(task_name, q_proj, co_att, q_fusion_att, v_fusion_att,
classifier)