def train_test_all():
preds_test_on_all_models = []
preds_train_on_all_models = []
model_val_loss_dict = {}
weights = []
for i, [
prefix_name, ModelHelper, Model, special_param, PreprocessHelper,
preprocess_param, weight
] in enumerate(model_config):
print_seperater(prefix_name, Model, special_param)
param['output'] = os.path.join(
param['output_root_dir'],
'model_save/output_{}'.format(prefix_name))
param['model'] = Model
param.update(special_param)
param['out_data_dir'] = os.path.join(
param['output_root_dir'],
"out_data_dir_" + PreprocessHelper.__name__)
weights.append(weight)
total_train_preds, total_test_preds = train_test_once(
ModelHelper, PreprocessHelper, preprocess_param)
# 把train数据的预测结果写到文件中
overall_val_loss = write_val_result(total_train_preds,
root_dir=param['root_dir'],
output_dir=param['output'])
model_val_loss_dict[prefix_name] = [overall_val_loss, weight]
# 将每次预测后的结果加到list中
preds_test_on_all_models.append(total_test_preds)
preds_train_on_all_models.append(total_train_preds)
print("preds after repeating -- shape:",
np.shape(preds_test_on_all_models))
# 求均值
preds_test_on_all_models = np.average(np.array(preds_test_on_all_models),
axis=0,
weights=weights)
preds_train_on_all_models = np.average(np.array(preds_train_on_all_models),
axis=0,
weights=weights)
# 将test的预测结果写到文件
write_result(preds_test_on_all_models, root_dir=param['root_dir'])
print("End -- model_val_loss_dict:", model_val_loss_dict)
# 将train的测试结果写到文件
overall_val_loss = write_val_result(preds_train_on_all_models,
root_dir=param['root_dir'])
model_val_loss_dict["overall"] = overall_val_loss
# 将所有的val loss写到文件
with open("val_loss002.pkl", 'wb') as f:
pickle.dump(model_val_loss_dict, f)
def compuate_val_loss():
weights = []
preds_train_on_all_models = []
for i, [
prefix_name, ModelHelper, Model, special_param, PreprocessHelper,
preprocess_param, weight
] in enumerate(model_config):
print_seperater(prefix_name, Model, special_param)
param['output'] = os.path.join(
param['output_root_dir'],
'model_save/output_{}'.format(prefix_name))
weights.append(weight)
submission_csv = os.path.join(param['output'], 'submission_val.csv')
submission = pd.read_csv(submission_csv)
del submission['sig_id']
preds_train_on_all_models.append(np.array(submission.values))
overall_val_loss = calculate_overall_loss(param['root_dir'],
submission_csv)
print("model :{}; loss:{}".format(prefix_name, overall_val_loss))
preds_train_on_all_models = np.average(np.array(preds_train_on_all_models,
dtype=float),
axis=0,
weights=weights)
overall_val_loss = write_val_result(preds_train_on_all_models,
root_dir=param['root_dir'])
print("overall model ; loss:{}".format(overall_val_loss))