__best_threshold = best_th
f1_best = get_f1_threshold(weighted_valid, valid_ohs, __best_threshold)
print(__best_threshold)
print('f1_best=', f1_best)
f1_best = get_f1_threshold(weighted_valid, valid_ohs,
__best_threshold - 0.1)
print('f1_best(-0.1)=', f1_best)
f1_best = get_f1_threshold(weighted_valid, valid_ohs,
__best_threshold + 0.1)
print('f1_best(+0.1)=', f1_best)
f1_best = get_f1_threshold(weighted_valid, valid_ohs,
__best_threshold_prev)
print(__best_threshold_prev)
print('f1_best(prev)=', f1_best)
weighted_test = test_merged * best_w
weighted_test = np.sum(weighted_test, axis=1)
weighted_test = np.clip(weighted_test, 0.0, 1.0)
output = 'asset/ensemble_nn2.csv'
save_pred(ids_test, weighted_test, th=__best_threshold, fname=output)
output = 'asset/ensemble_nn2_thval.csv'
save_pred(ids_test, weighted_test, th=__best_threshold_prev, fname=output)
output = 'asset/ensemble_nn2_th-1.csv'
save_pred(ids_test, weighted_test, th=__best_threshold - 0.1, fname=output)
output = 'asset/ensemble_nn2_th+1.csv'
save_pred(ids_test, weighted_test, th=__best_threshold + 0.1, fname=output)
'nasnet_fold0_lr0.0001_relu_bce',
'nasnet_fold2_lr0.0001_relu_bce',
]
ensemble_list = model_selection
ensemble_th = 3
C.get()['cv_fold'] = 0
_, _, _, ids_test = get_dataset()
votes = []
for _ in range(len(ids_test)):
vote = [0] * len(name_label_dict)
votes.append(vote)
for model_name in ensemble_list:
if '.csv' in model_name:
path = 'asset_v3/%s' % model_name
else:
path = 'asset_v3/%s.aug.csv' % model_name
print('load... %s' % path)
labels = pd.read_csv(path).set_index('Id')
for idx, s in enumerate(labels['Predicted']):
if not isinstance(s, str) and math.isnan(s):
continue
for pred_idx in [int(i) for i in str(s).split()]:
votes[idx][pred_idx] += 1
output = 'asset/ensemble_vote_th%d.csv' % (ensemble_th)
save_pred(ids_test, votes, th=ensemble_th, fname=output)
print('done. %s' % output)
print(__best_threshold)
print('f1_best(valid)=', f1_best)
best_th = threshold_search(out, valid_ohs)
f1_best = get_f1_threshold(out, valid_ohs, best_th)
print(best_th)
print('f1_best(valid, th searched)=', f1_best)
out_list = []
for batch, logit in zip(grouper(test_merged, bs), grouper(test_logit, bs)):
batch = [
b if isinstance(b, torch.Tensor) else torch.from_numpy(b)
for b in batch if b is not None
]
logit = [
b if isinstance(b, torch.Tensor) else torch.from_numpy(b)
for b in logit if b is not None
]
out_batch = net(
torch.stack(batch, dim=0).cuda(),
torch.stack(logit, dim=0).cuda())
out_list.append(out_batch.detach().cpu().numpy())
out = np.concatenate(out_list, axis=0)
output = 'asset/ensemble_nn4.csv'
save_pred(ids_test, out, th=__best_threshold, fname=output)
print('done-')
p, r = valid_result['stats']
for class_idx in range(num_class()):
writer.add_scalar(
'precision_valid/class%d' % class_idx,
p[class_idx], curr_epoch)
writer.add_scalar('recall_valid/class%d' % class_idx,
r[class_idx], curr_epoch)
# run on test set
preds_test = run_epoch(model,
d_tests,
title='test',
aug=False)['prediction']
save_pred(ids_test,
preds_test,
th=__best_threshold,
fname='asset/%s.csv' % C.get()['name'],
valid_pred=valid_result)
torch.save(
{
'epoch': curr_epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
}, 'asset/%s.best.pt' % C.get()['name'])
torch.save(
{
'epoch': curr_epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),