def write_result(ids, predicts, ofile, is_infer=True):
x = ids
impression_id = x['impression_id']
prob = gezi.sigmoid(predicts)
position = x['position']
df = pd.DataFrame({
'impression_id': impression_id,
'position': position,
'score': prob
})
if FLAGS.infer_part is not None:
ofile = ofile.replace('.csv', '.{FLAGS.infer_part}.csv')
df.to_csv(ofile, index=False, header=False, sep=' ')
return
df = df.sort_values(['impression_id', 'position'])
tqdm.pandas()
df = df.groupby('impression_id')['score'].progress_apply(list).reset_index(
name='scores')
df['scores2'] = df.scores.progress_apply(
lambda x: '[' + ','.join(map(str, x)) + ']')
df[['impression_id', 'scores2']].to_csv(ofile,
index=False,
header=False,
sep=' ')
if is_infer:
df.scores2 = df.scores.progress_apply(lambda x: '[' + ','.join(
map(str, (-np.asarray(x)).argsort().argsort() + 1)) + ']')
odir = os.path.dirname(ofile)
df[['impression_id', 'scores2']].to_csv(f'{odir}/prediction.txt',
index=False,
header=False,
sep=' ')
os.system(f'cd {odir};zip prediction.zip prediction.txt')
def evaluate(y_true, y_pred, x):
y_prob = gezi.sigmoid(y_pred)
df = pd.DataFrame({
'y_true': y_true,
'y_prob': y_prob,
'impression_id': x['impression_id'],
'uid': x['uid'],
'did': x['did'],
'history_len': x['hist_len'],
'uid_in_train': x['uid_in_train'],
'did_in_train': x['did_in_train']
})
return evaluate_df(df)
def valid_write(ids, labels, predicts, out):
for id, label, predict in zip(ids, labels, predicts):
print('{},{},{:.3f}'.format(id, label, gezi.sigmoid(predict)),
file=out)
def evaluate(y, y_, model_path=None):
y_ = gezi.sigmoid(y_)
auc = roc_auc_score(y, y_)
loss = log_loss(y, y_)
return [auc, loss], ['auc', 'loss']