config.py

'''set this path to the folder containing YUV, MP4 and MAT of our database'''
database_path = '/media/yuhangsong/YuhangSong_1/env/ff/vr_new'

log_dir = "../results/reproduce_17"
'''if clear the logdir before running'''
clear_run = False
import subprocess
if clear_run:
    '''if clear_run, clear the path before create the path'''
    raw_input('You are clearing {}, is that what you want? If not, press Ctrl+C and set clear_run=False in config.py'.format(log_dir))
    subprocess.call(["rm", "-r", log_dir])
subprocess.call(["mkdir", "-p", log_dir])

'''
Description: select mode
Availible: off_line, on_line, data_processor
'''
mode = 'off_line'

if mode in ['off_line']:
    # Note that for online settings, there is no separation for training and teseting
    procedure = 'train'
    if procedure in 'train':
        # train for 1.113M steps
        number_trained_steps = 1113000
    elif procedure in ['test']:
        # set to -1 to make it unlimited
        number_trained_steps = -1

import dataset_config
if mode in ['off_line']:
    if procedure in ['train']:
        game_dic = dataset_config.train_set
    elif procedure in ['test']:
        game_dic = dataset_config.test_set
elif mode in ['on_line']:
    game_dic = dataset_config.test_set
elif mode in ['data_processor']:
    game_dic = dataset_config.train_set + dataset_config.test_set

num_games = len(game_dic)

'''
number of direction to action
this is because we use discrete control
'''
direction_num = 8

if_log_scan_path_real_time = False
if_log_results = False

if if_log_results is True:

    '''number of workflow to produce salmap,
    you set it to any value, but we set it to num_subjects,
    refer to Section xx in https://arxiv.org/abs/1710.10755'''
    predicted_fixation_num = dataset_config.num_subjects
    log_results_interval = 5

# availible: trustworthy_transfer, cc
reward_estimator = 'trustworthy_transfer'

# set to 1.0 to disable reward smooth
reward_smooth_discount_to = 1.0

if_normalize_v_lable = True

if mode is 'data_processor':
    '''
        Description: what data_processor you are doing
        Availible:
            mp4_to_yuv (yuhang)
            generate_groundtruth_heatmaps (yuhang)
            generate_groundtruth_scanpaths (yuhang)
            minglang_mp4_to_yuv (minglang)
            compute_consi (haochen)
            minglang_mp4_to_jpg (minglang)
            minglang_obdl_cfg (minglang)
    '''
    data_processor_id = 'generate_groundtruth_heatmaps'

elif mode is 'on_line':
    '''terminate condition for online training'''

    # set to 1.0 to disable it
    train_to_reward = 0.2

    # set to 1.0 to disable it
    train_to_mo = 0.8

    # too big would make some train hard to end, for some subjects is too hard to learn
    train_to_episode = 500

    '''following are default settings'''
    check_worker_done_time = 5
    worker_done_signal_dir = 'temp/worker_done_signal_dir/'
    worker_done_signal_file = 'worker_done_signal.npz'

data_tensity = 10.0
view_range_lon = 110
view_range_lat = 113
final_discount_to = 10**(-4)

'''set model structure'''
conv_depth = 4
# consi_depth is from another work of mine:
# https://arxiv.org/abs/1710.10036,
# ignore it (keep it 1), if you are not interested
consi_depth = 1
# if set consi_depth=1, only lstm_size[0] matters
lstm_size = [288,128,32]

'''
Description: how many steps do you want to run before update the model
Note: 20 is empirically good for single task,
larger value may benifit the multi-task performence
'''
update_step = 40

import numpy as np
observation_space = np.zeros((42, 42, 1)) # related to model structure, do not change this