Run ./dnn_trainner (without arguments to see help message for arguments)

• The test_data and train_data are the files created by scripts/cifaren.py.
• The structure of the neural network is passed in a json file, they are in rhoban/deep_vision/configs. The parameters are the followings:
  • networks: contains the list of the networks you want to train.
    • The training of each network will take a core.
    • They are two kinds of networks, one layer or two layer. Note: Since one layer and two layers networks have different learning rates they shouldn't be trained with the same config file.
    • There are some examples at the end of the README.md.
  • nb_minibatch: for example 10
  • nb_train_epochs: for example 10 for fast learning 30 for more long learning
  • In order to find the "best" learning rate, a dichotomy is performed. learning_rate_start and learning_rate_end gives the initial boudaries of the dichotomy. dichotomy_depth specifies the depth at which we will stop searching When you find the best structure and the best learning rate, you should put: learning_rate_start=learning_rate_end
• Acceptable rates around 95% of recognition rate is the minimal acceptable value.
• For each network the results are in the folder results/network_name. It contains:
  • A file csv fileresults_...csv.
    • learning_rate: the learning rate used for the optimisation.
    • valisation_score: the validation score of the last epoch or overfit if the learning overfitted the data.
    • last_improvement: the improvement the validation score between the last and the second to last epochs.
    • last_epoch: the epoch at which the optimisation stopped.
    • learning_time: the time needed for the optimisation.
  • Two files corresponding to the neural network:
    • The architecture (.json)
    • The coefficients of the network (.bin)

When you find the best one, relunch the learning with only one network and the specific learning rate.