Our bots are running on telegram
| Profile | Name | id |
|---|---|---|
 |
Eric Cartman | sp_cartman_bot |
 |
Stan Marsh | sp_stan_bot |
 |
Kyle Broflovski | sp_kyle_bot |
 |
Butters Stotch | sp_butters_bot |
 |
Randy Marsh | sp_randy_bot |
 |
Mysterion | sp_mysterion_bot |
The code requires python3. Install dependencies by
pip install -r requirements.txt
Pretrain the model with anonymous dialogues, which keeps the initial personas stable.
The model can restore the previously trained weights and continue training up on that.
python3 src/chatbot.py --mode=pretrainFinetune the model with personalized dialogues, which aims to train the personas.
python3 src/chatbot.py --mode=finetune --checkpoint=pretrain_1000Start the command line interaction with the chat bot.
python3 src/chatbot.py --mode=chat --checkpoint=finetune_1000 --speaker=cartmanThe model settings can be modified in src/config.py
| Parameters | Description |
|---|---|
| DATA_MODE | choose which data set to train on: general_data/persona_data |
| MAX_LENGTH | maximum sentence length to consider |
| N_ITER | training iterations |
| RNN_TYPE | model of encoder & decode, support LSTM & GRU |
| ATTN_TYPE | type of the attention model: dot/general/concat |
| HIDDEN_SIZE | size of the word embedding & number of hidden units in GRU |
| PERSONA_EMBEDDING_SIZE | size of the persona embedding |
| ENCODER_N_LAYERS | number of layers in bi-GRU encoder |
| DECODER_N_LAYERS | number of layers in GRU decoder |
| ENCODER_DROPOUT_RATE | dropout rate in bi-GRU encoder |
| DECODER_DROPOUT_RATE | dropout rate in GRU decoder |
| TEACHER_FORCING_RATIO | the ratio that decoder learns from ground truth instead of last output |
| LR | encoder learning rate |
| DECODER_LR | decoder learning rate: LR*DECODER_LR |
| TF_RATE_DECAY_FACTOR | k in the inverse sigmoid decay func of the teacher force rate |
If you want to run it on the school server, run the following command to load the some modules first
source server_env