You must complete the NLU Setup before following this document.

Generating input and output data

GPSR

cd into the gpsr folder:

    $ cd $ROS_WORKSPACE/mbot_natural_language_processing/mbot_nlu_training/common/src/mbot_nlu_training/gpsr        

generate data for intent training:

    $ python /intention_NN/gpsr_data_generator.py 

generate data for slots training:

    $ python /intention_NN/gpsr_data_generator.py 

two files are generated in respective folders:

    input  / inputs_slot_filling
    output / outputs_slot_filling

input are the phrases and output is the ground truth intent or label corresponding to the phrases. The data generator takes a bit of time in less powerful PCs. (tip: Try to run it in Harode or Batatinha)

For ERL data generation, do the above steps in ERL training folder:

    $ cd $ROS_WORKSPACE/mbot_natural_language_processing/mbot_nlu_training/common/src/mbot_nlu_training/erl

Setting up the PC for training

This is usually done in Harode or Batatinha.

1. tensorflow set up using virtual environment === Create a Virtualenv environment by issuing the following command:

   $ virtualenv --system-site-packages -p python3 tensorflow

This needs virtual-env pre installed for it to work. Ask some one with admin privilages to set it up in case not. For reference: tensorFlow installation reference

Activate the Virtualenv environment by issuing one of the following commands:

  $ source ~/tensorflow/bin/activate

The preceding source command should change your prompt to the following:

  (tensorflow)$  

Ensure pip ≥8.1 is installed:

  (tensorflow)$ easy_install -U pip

Issue the following command to install TensorFlow in the active Virtualenv environment:

  (tensorflow)$ pip3 install --upgrade tensorflow-gpu==1.4.0  

When you are done using TensorFlow, you may deactivate the environment by invoking the deactivate function as follows:

  (tensorflow)$ deactivate  

2. Uninstalling tensorFlow ===
   To uninstall TensorFlow, simply remove the tree you created. For example:

   $ rm -r ~/tensorflow

3. Install progressbar ===

Progress bar is used in the training script for feedback.

Install progressbar by invoking the following command:

  (tensorflow)$ easy_install -U progressbar2  

4. Run the training ===

Copy the mbot_nlu_training folder to the training PC using scp.

The training for intent or slots is started by running the train_nn_model.py, which can be found in the gpsr/intention_NN or gpsr/slots_NN folders (mentioned above) respectively.

Before running the code make sure you have the GPU index specified at the last of the training file in the environment variable [CUDA_VISIBLE_DEVICES]:

  (tensorflow)$ python3 train_nn_model.py

5. Results ===

The classifiers from the training are stored after each epoch in latest_intent_classifier and latest_slots_classifiers respectively.

6. Tips ===

• The intent and slots training can be executed in parallel. For this create two virtual environments with separate names. Run intent from virtual environment 1 and slots from virtual environment 2
• Use glances to see the resources consumed by the process
• Use tmux to create sessions for training. This way the training is securely running even though your PC is off/lost connection.
• The training script currently has functionality to select the GPU for training. This is done using the env variable [CUDA_AVAILABLE_DEVICES] at the bottom of the training script.