def init_cnn_lstm(args, layout_vocab_size, object_vocab_size, text_vocab_size):
args.lstm_out = 16
args.cnn_out_dim = 2*args.lstm_out
state_model = models.LookupModel(layout_vocab_size, args.state_embed)
object_model = models.LookupModel(object_vocab_size, args.obj_embed)
lstm = models.TextModel(text_vocab_size, args.lstm_inp, args.lstm_hid, args.lstm_layers, args.lstm_out)
model = cudit(models.CNN_LSTM(state_model, object_model, lstm, args))
return model
def init_nogradient(args, layout_vocab_size, object_vocab_size, text_vocab_size):
args.global_coeffs = 0
args.attention_in_dim = args.obj_embed
args.lstm_out = args.attention_in_dim * args.attention_out_dim * args.attention_kernel**2
state_model = cudit(models.LookupModel(layout_vocab_size, args.state_embed))
object_model = models.LookupModel(object_vocab_size, args.obj_embed)
text_model = models.TextModel(text_vocab_size, args.lstm_inp, args.lstm_hid, args.lstm_layers, args.lstm_out)
heatmap_model = cudit(models.AttentionHeatmap(text_model, args, map_dim=args.map_dim))
model = cudit(models.MultiNoBases(state_model, object_model, heatmap_model, args, map_dim=args.map_dim))
return model
def init_full(args, layout_vocab_size, object_vocab_size, text_vocab_size):
args.global_coeffs = 3
args.attention_in_dim = args.obj_embed
args.lstm_out = args.attention_in_dim * args.attention_out_dim * args.attention_kernel**2 + args.global_coeffs
state_model = models.LookupModel(layout_vocab_size,
args.state_embed).cuda()
object_model = models.LookupModel(object_vocab_size, args.obj_embed)
text_model = models.TextModel(text_vocab_size, args.lstm_inp,
args.lstm_hid, args.lstm_layers,
args.lstm_out)
heatmap_model = models.AttentionGlobal(text_model,
args,
map_dim=args.map_dim).cuda()
model = models.MultiNoRBF(state_model,
object_model,
heatmap_model,
args,
map_dim=args.map_dim).cuda()
return model