when prediction value is very small, metric logloss calculate error

[alexdyysp]

Describe the bug(问题描述) /usr/local/anaconda3/lib/python3.7/site-packages/sklearn/metrics/classification.py:2174: RuntimeWarning: divide by zero encountered in log loss = -(transformed_labels * np.log(y_pred)).sum(axis=1)

To Reproduce(复现步骤) Steps to reproduce the behavior:

model = xDeepFM(linear_feature_columns,dnn_feature_columns,task='binary',device=device)
model.compile("adam", "binary_crossentropy",
              metrics=['log_loss'], )

Operating environment(运行环境):

• python version [3.7]
• torch version [1.1.0]
• deepctr-torch version 0.1.2
• sklearn version 0.0

Additional context my model and data will predict very small value like 0.0001 and value close to 1 like 0.9998 even appear 0 and 1. so it will make skearn metrci caculate error.

[wzh9969]

Same problem. It is caused by different default float types between pytorch (float32) and numpy/sklearn (float64). I solved by turning the model output to float64 (.astype("float64")). I modified two places in BaseModel.fit and BaseModel.predict methods and the problem is solved in my case.

fit():
    train_result[name].append(metric_fun(
                                    y.cpu().data.numpy(), y_pred.cpu().data.numpy().astype("float64")))
predict():
    return np.concatenate(pred_ans).astype("float64")