Training a new PLC model

[dariadiatlova]

Hello,

I am trying to train a new PLC model with train_plc.py and and several things have caused me difficulties:

1. Can I follow the same data preprocessing set up with ./dump_data -train input.s16 features.f32 data.s16 to get input features.f32 for PLC model training?
2. How should lost_file for training look like? Is it a single .txt file - a concatenation of smaller .txt files with one entry per 20ms packet, where 1 means "packet lost" and 0 means "packet not lost"? How to create a single file if original data was augmented after running ./dump_data? Is there any script for it?
3. To close the above questions with lost_file preprocessing, can I just uncomment the line and train the model with random packets marked as lost? Have you noticed any significant degradation in how this works?
4. Following test_plc.py, the output is: features + (1-lost)*out, but the shapes:

• features: [bs, seq_len, nb_used_features+nb_burg_features]
• lost: [bs, seq_len, 1]
• out: [bs, seq_len, nb_used_features]

Did I think of the wrong shapes? What should be the shape of a correct output for writing to output.f32?

Thank you for sharing your code and supporting this repository!

[Janne-byti]

@dariadiatlova Hello, could you please retrain the PLC model? May I ask how to prepare training data?

[RayShing]

Hello, I am also interested in training a new PLC model with a combination of audio features and visual features. I would like to ask how to do that? Any insights are truly appreciated!

[YYX666660]

About the lost_file, I try to use the lossgen model to generate the txt, but I'm not sure how long and how many loss rate should be generated. Very confused about training the PLC model...