def ensemble(results, ans_unk_idx):
final_result = masked_unk_softmax(results[0], dim=1, mask_idx=ans_unk_idx)
if len(results) == 1:
return final_result
for result in results[1:]:
final_result += masked_unk_softmax(result, dim=1, mask_idx=ans_unk_idx)
return final_result
def ensemble(results, ans_unk_idx):
final_result = masked_unk_softmax(results[0], dim=1, mask_idx=ans_unk_idx)
if len(results) == 1:
return final_result
for result in results[1:]:
final_result += masked_unk_softmax(result, dim=1, mask_idx=ans_unk_idx)
return final_result
def run_model(current_model, data_reader, UNK_idx=0):
softmax_tot = []
q_id_tot = []
start = timeit.default_timer()
for i, batch in enumerate(data_reader):
if (i + 1) % 100 == 0:
end = timeit.default_timer()
time = end - start
start = timeit.default_timer()
print(" process batch %d for test for %.1f s" % (i + 1, time))
sys.stdout.flush()
verbose_info = batch['verbose_info']
q_ids = verbose_info['question_id'].cpu().numpy().tolist()
logit_res = one_stage_run_model(batch, current_model, eval_mode=True)
softmax_res = masked_unk_softmax(logit_res, dim=1, mask_idx=UNK_idx)
softmax_res = softmax_res.data.cpu().numpy().astype(np.float16)
q_id_tot += q_ids
softmax_tot.append(softmax_res)
softmax_result = np.vstack(softmax_tot)
return q_id_tot, softmax_result
def run_model(current_model, data_reader, UNK_idx=0):
softmax_tot = []
q_id_tot = []
start = timeit.default_timer()
for i, batch in enumerate(data_reader):
if (i+1) % 100 == 0:
end = timeit.default_timer()
time = end - start
start = timeit.default_timer()
print(" process batch %d for test for %.1f s" % (i+1, time))
sys.stdout.flush()
verbose_info = batch['verbose_info']
q_ids = verbose_info['question_id'].cpu().numpy().tolist()
logit_res = one_stage_run_model(batch, current_model, eval_mode=True)
softmax_res = masked_unk_softmax(logit_res, dim=1, mask_idx=UNK_idx)
softmax_res = softmax_res.data.cpu().numpy().astype(np.float16)
q_id_tot += q_ids
softmax_tot.append(softmax_res)
softmax_result = np.vstack(softmax_tot)
return q_id_tot, softmax_result
os.makedirs(out_dir, exist_ok=True)
if is_one_stageModel(model_type):
for i, batch in enumerate(data_reader_val):
if i % 100 == 0:
print("process batch %d" % i)
verbose_info = batch["verbose_info"]
answer_scores = batch["ans_scores"]
answer_scores_np = answer_scores.numpy()
for imd, current_model in enumerate(current_models):
# print("process model %d"%imd)
logit_res = one_stage_run_model(batch, current_model)
softmax_res = masked_unk_softmax(logit_res,
dim=1,
mask_idx=ans_dic.UNK_idx)
softmax_res_data = softmax_res.data.cpu().numpy()
with open(os.path.join(out_dir, tmp_model_file_name % imd),
"wb") as w:
pickle.dump(softmax_res_data, w)
ensembled_soft_max_result = ensemble_model(out_dir, num_of_model)
nsample, _ = answer_scores_np.shape
total_sample += nsample
scores = compute_score_with_prob(ensembled_soft_max_result,
answer_scores_np)
total_score += scores
os.makedirs(out_dir, exist_ok=True)
if is_one_stageModel(model_type):
for i, batch in enumerate(data_reader_val):
if i % 100 == 0:
print("process batch %d" % i)
verbose_info = batch['verbose_info']
answer_scores = batch['ans_scores']
answer_scores_np = answer_scores.numpy()
for imd, current_model in enumerate(current_models):
# print("process model %d"%imd)
logit_res = one_stage_run_model(batch, current_model)
softmax_res = masked_unk_softmax(logit_res,
dim=1,
mask_idx=ans_dic.UNK_idx)
softmax_res_data = softmax_res.data.cpu().numpy()
with open(os.path.join(out_dir,
tmp_model_file_name % imd), 'wb') as w:
pickle.dump(softmax_res_data, w)
ensembled_soft_max_result = ensemble_model(out_dir, num_of_model)
nsample, _ = answer_scores_np.shape
total_sample += nsample
scores = compute_score_with_prob(ensembled_soft_max_result,
answer_scores_np)
total_score += scores
