print("Test model", config.model)
model = Transformer(p.vocab, model_file_path=config.save_path, is_eval=False)
# get persona map
filename = 'data/ConvAI2/test_persona_map'
with open(filename, 'rb') as f:
persona_map = pickle.load(f)
#generate
iterations = 11
weights_original = deepcopy(model.state_dict())
tasks = p.get_personas('test')
for per in tqdm(tasks):
num_of_dialog = p.get_num_of_dialog(persona=per, split='test')
for val_dial_index in range(num_of_dialog):
train_iter, val_iter = p.get_data_loader(persona=per,
batch_size=config.batch_size,
split='test',
fold=val_dial_index)
persona = []
for ppp in persona_map[per]:
persona += ppp
persona = list(set(persona))
do_learning(model,
train_iter,
val_iter,
iterations=iterations,
persona=persona)
model.load_state_dict(
{name: weights_original[name]
for name in weights_original})
elif meta_iteration == 10:
m = "select"
else:
m = "selective_training"
print(m)
shuffle(tasks)
for k in range(steps):
st = k * meta_batch_size
ed = st + meta_batch_size
if ed > len(tasks):
ed = len(tasks)
batch_loss = 0
weights_original = deepcopy(meta_net.state_dict())
for i in range(st, ed):
per = tasks[i]
train_iter, val_iter = p.get_data_loader(
persona=per, batch_size=config.batch_size, split='train')
v_loss, v_ppl = do_evaluation(meta_net, val_iter)
train_loss_before.append(math.exp(v_loss))
# Update fast nets
if m != "select":
val_loss, v_ppl = do_learning_fix_step(
meta_net,
train_iter,
val_iter,
iterations=config.meta_iteration,
mode=m,
task=per)
else:
val_loss, v_ppl = do_learning_fix_step(meta_net,
train_iter,
val_iter,