config.lua

--- All parameters goes here
config = config or {}

function config.parse(arg)
	local cmd = torch.CmdLine()
	cmd:text()
	cmd:text('Fast R-CNN for Torch')
	cmd:text()
	cmd:text('')
	-- Parameters
	cmd:option('-resume_training',false, 'True if you are resuming the training from a FRCNN model or false when starting from the imagenet model')
	cmd:option('-pre_trained_file','./data/torch_imagenet_models/imgnet_alexnet.t7','Path to the pretrained weights (used for training)')
	cmd:option('-model_def','./models/AlexNet/FRCNN.lua','Path to the FRCNN model definition')
	cmd:option('-model_weights','./data/trained_models/frcnn_alexnet_VOC2007_iter_40000.t7','Path to the FRCNN weights (used for testing)')
	cmd:option('-use_difficult_objs', true, 'Whether to load the difficult examples or not')
	cmd:option('-scale', 600, 'Scale used for training and testing, currently only single scale is supported.')
	cmd:option('-max_size', 1000, 'Max pixel size of the longest side of a scaled input image')
	cmd:option('-img_per_batch', 2, 'Images per minibatch')
	cmd:option('-GPU_ID',1,'Main GPU ID to be used')
	cmd:option('-n_threads',1, 'Number of threads used for training (In Multi GPU mode)')
	cmd:option('-nGPU',1,'Number of GPUs to be used for training (not completely tested yet)')
	cmd:option('-roi_per_img', 64, 'Minibatch size')
	cmd:option('-fg_fraction', 0.25, 'Fraction of the minibatch that is labeled as foreground (i.e. class > 0)')
	cmd:option('-fg_threshold', 0.5, 'IoU threshold for a ROI to be considered as foreground (if >= FG_THRESH)')
	cmd:option('-bg_threshold_hi', 0.5, 'High IoU threshold for a ROI to be considered as background')
	cmd:option('-bg_threshold_lo',0.1, 'Low IoU threshold for a ROI to be considered as background')
	cmd:option('-use_flipped', true, 'Use horizontally-flipped images during training if true')
	cmd:option('-bbox_threshold', 0.5, 'IoU required between a ROI and a ground-truth box in order for that ROI to be used as a bounding-box regression training example')
	cmd:option('-nms', 0.3, 'Overlap threshold used for non-maxima suppression (suppress boxes with IoU >= this value)')
	cmd:option('-pixel_means', {102.9801,115.9465,122.7717}, 'Pixel mean values (BGR order)')
	cmd:option('-eps', 1e-14, 'Epsilon')
	cmd:option('-log_path','./cache','Path used for saving log data')
	cmd:option('-dataset','voc_2007','Dataset to be used')
	cmd:option('-dataset_path','./data/datasets','Path to the dataset root folder')
	cmd:option('-test_img_set','test','Image set to be used for testing')
	cmd:option('-train_img_set','trainval','Image set to be used for training')
	cmd:option('-cache','./cache','Directory used for saving cache data')
	cmd:option('-optim_momentum',0.9,'Momentum used for sgd optimizer')
	cmd:option('-optim_lr_decay_policy','fixed', 'Learning rate decay policy, can be \'fixed\' or \'exp\', if you are using exp then the second column in the optim_regime should be a table with two elements: the start and the end lr for that row')
	cmd:option('-optim_regimes',{{30000,1e-3,5e-4},{10000,1e-4,5e-4}},'Optimization regime, each row is the number of iterations, the learning rate, and the weight decay')
	cmd:option('-optim_snapshot_iters', 10000, 'Iterations between snapshots (used for saving the network)')
	cmd:option('-save_path','./data/trained_models','Path to be used for saving the trained models')
	-- Parsing the command line 
	config = cmd:parse(arg or {})
	return config
end



return config