search

ApolloAuto / apollo

An open autonomous driving platform
Apache License 2.0
25.22k stars 9.72k forks source link

Prediction: train cruise mlp model #13357

Open zhangyanide opened 3 years ago

zhangyanide commented 3 years ago

When training cruise mlp model, how to deal with the data imbalance? The down-sampling code has been commented. And my valid loss dropped for a while and then remained unchanged. How to solve it?

Blanksan3 commented 3 years ago

Hi, could you share how to train the cruise mlp model? Did you write your own script? Since I can't run it based on current open source code. Thanks so much!

zhangyanide commented 3 years ago

Hi, could you share how to train the cruise mlp model? Did you write your own script? Since I can't run it based on current open source code. Thanks so much!

generate_cruise_labels.py; merge_label_dicts.py; combine_features_and_labels.py(change some code to fit cruise, according to junction); cruise_h5_preprocessing.py; cruiseMLP_train.py

Blanksan3 commented 3 years ago

Thank you for your response. But the issues are not solved. The first one is that it needs the length of dict_labels bigger than 30 when combining the features and labels(if len(dict_labels[key]) < future_status_label_dim:), but the length of dict_labels for a particular key(obs id + timestamp) represents the number of lane sequence, so it is usually 2 or 3 at a moment. The second one is it needs .h5 file in cruise_h5_preprocessing.py file. But it only creates .np file after running combine_features_and_labels.py file, so I feel merge_h5.py file is needed between two processes. But I failed on the prior step, so I don't have a chance to run them. Can you help me fix these? I will appreciate it very much! Or you can email me in Chinese. (wei.cs@aliyun.com)