config.py

config = {}

# parameters of controlling the robot
config['make_dataset_mode'] = 'prepare_data_for_blocks' # prepare_data_for_chainer or prepare_data_for_blocks
config['robot'] = 'al5d' # al5d, mico
config['task_to_perform'] = 3005
config['camera1_image_topic'] = '/kinect2/qhd/image_color_rect'
config['batch_size'] = 128  # number of samples taken per each update
config['hidden_size'] = 100
config['learning_rate'] = .005
config['training_optimizer'] = 'RMSProp'  # Adam, RMSProp, Adagrad, Adadelta
config['learning_rate_decay'] = 0.5  # set to 0 to not decay learning rate
config['lr_decay_every_n_epochs'] = 100
config['decay_rate'] = 0.999  # decay rate for rmsprop and AdaDelta
config['step_clipping'] = 1.0  # clip norm of gradients at this value
config['dropout'] = .5
config['max_norm_threshold'] = 4
config['nepochs'] = 50000  # number of full passes through the training data
config['seq_length'] = 50  # number of time-steps in the sequence (for truncated BPTT)
config['seq_redundancy'] = 50
config['linear_before_recurrent_size'] = 0
config['use_layer_norm'] = True

config['waypoint_resolution'] = 8
config['resolution_step'] = 1
config['hierarchy_resolutions'] = [8]
config['plot_hidden_states'] = False
config['layer_models'] = ['lstm','lstm','lstm'] # feedforward, lstm, rnn
config['num_recurrent_layers'] = len(config['layer_models'])

# parameters of vision
config['pixel_depth'] = 255
config['image_shape'] = (128, 128) # height, width
config['num_channels'] = 3

# parameters of cost function
config['cost_mode'] = 'RL-MDN'  # MDN, RL-MDN, MSE

# Reinforcement learning parameters
config['gamma'] = 1.0

# parameters of MDN
config['components_size'] = 50
config['sampling_bias'] = 1.0
config['seed'] = 66478

# outputting one dimension at a time parameters
config['single_dim_out'] = True

# parameters about auxiliary predictions - The idea is to have some layers to predict another related prediction, for instance, predict the pose of object or the pose of gripper in next 4 timesteps
config['future_predictions'] = [1]
config['prediction_cost_weights'] = [1]

# parameters of multi-task learning : The idea is to train a network on data of multiple tasks. The ID if the task to be executed is given as an input in each time-step
config['multi_task_mode'] = 'ID'  # ID, goal
config['task_ID_type'] = 'one-hot' # feedforward, one-hot
config['task_ID_FF_dims'] = [65, 20]
config['game_tasks'] = [3001,3002,3003,3005,3006]
config['task_weights'] = [1 for x in range(len(config['game_tasks']))]
config['trajs_paths'] = 'trajectories/al5d'

# parameters of multi-timescale learning : The idea is to enable different layers of LSTM or RNN to work at different time-scales
config['layer_resolutions'] = [1, 1, 1]
config['layer_execution_time_offset'] = [0, 0, 0]

# We make a stacked RNN with the following skip connections added if the corresponding parameter is True
config['connect_x_to_h'] = True
config['connect_h_to_h'] = 'one-previous'  # all-previous , two-previous, one-previous
config['connect_h_to_o'] = True

config['hdf5_file'] = 'input'  + '-'.join(str(x) for x in config['game_tasks']) + '.hdf5'  # hdf5 file with Fuel format

config['input_columns'] = []
config['output_columns'] = ['joint1', 'joint2', 'joint3', 'joint4','joint5','gripper']

for i in config['game_tasks']:
    config['input_columns'] += ['task' + str(i)]
if config['cost_mode'] == 'RL-MDN':
    config['output_columns'] += ['steps_to_goal']
config['train_size'] = 0.80  # fraction of data that goes into train set
config[
    'save_path'] = 'models/{0}_{1}_{2}_{3}_{4}_{5}_{6}_{7}_{8}_best.pkl'.format(
    '-'.join(str(x) for x in config['prediction_cost_weights']),
    '-'.join(str(x) for x in config['game_tasks']),
    config['single_dim_out'], config['gamma'], config['components_size'],
    config['num_recurrent_layers'], config['hidden_size'], config['batch_size'],
    config['seq_length'])  # path to best model file
config[
    'last_path'] = 'models/{0}_{1}_{2}_{3}_{4}_{5}_{6}_{7}_{8}_last.pkl'.format(
    '-'.join(str(x) for x in config['prediction_cost_weights']),
    '-'.join(str(x) for x in config['game_tasks']),
    config['single_dim_out'], config['gamma'], config['components_size'],
    config['num_recurrent_layers'], config['hidden_size'], config['batch_size'],
    config['seq_length'])  # path to last model file
config['load_path'] = config['save_path']