def callback(data):
if (data.iteration + 1) % 100 != 0:
return
clf = data.model
trn_data = clf.train_set
val_data = clf.valid_sets[0]
preds = [
ele[2] for ele in clf.eval_train(f1_metric_xgb2) if ele[1] == 'f12'
][0]
labels = trn_data.get_label().astype(np.int)
res = f1_group_idx(labels, preds, list_idx).astype(np.bool)
weight = trn_data.get_weight()
if weight is None:
weight = np.ones(preds.shape[0])
weight[res] = 0.5
weight[~res] = 2
trn_data.set_weight(weight)
preds = [
ele[2] for ele in clf.eval_valid(f1_metric_xgb2) if ele[1] == 'f12'
][0]
labels = val_data.get_label().astype(np.int)
t = time.time()
res = f1_group(labels, preds, list_idx)
sc = np.mean(res)
logger.info('cal [{}] {} {}'.format(data.iteration + 1, sc,
time.time() - t))
def callback(data):
if (data.iteration + 1) % 500 != 0:
return
clf = data.model
trn_data = clf.train_set
val_data = clf.valid_sets[0]
'''
preds = [ele[2] for ele in clf.eval_train(f1_metric_xgb2) if ele[1] == 'f12'][0]
labels = trn_data.get_label().astype(np.int)
res, sc = f1_group_sc(labels, preds, list_idx) # .astype(np.bool)
weight = np.where(res > sc, 1, 1 + (sc - res) * 100)
trn_data.set_weight(weight)
'''
preds = [
ele[2] for ele in clf.eval_valid(f1_metric_xgb2) if ele[1] == 'f12'
][0]
labels = val_data.get_label().astype(np.int)
t = time.time()
res = f1_group(labels, preds, list_idx)
sc = np.mean(res)
logger.info('cal [{}] {} {}'.format(data.iteration + 1, sc,
time.time() - t))
def f1_metric_xgb(pred, dtrain):
label = dtrain.get_label().astype(np.int)
pred = pred.astype(np.float64)
#res = [f1(label.take(i), pred.take(i)) for i in list_idx]
res = f1_group(label, pred, list_idx)
sc = np.mean(res)
logger.debug('f1: %s' % (sc))
return 'f1', -sc
def f1_metric(label, pred):
res = f1_group(label, pred, list_idx)
sc = np.mean(res)
logger.debug('f1: %s' % (sc))
return 'f1', sc, True
