-
Notifications
You must be signed in to change notification settings - Fork 1
Parameter description
This page describe the parameters that you need to specify before running the RAM model. The numbers we used kind of works but we never did any grid search or formal tuning. Namely, you might be able to make the model work better play with these parameters.
load_path:(str) the path to saved models
eval_only: (bool) test an existing model if eval_only == true
draw: (bool) do visualization
animate: (bool) visualize the glimpses, glimpse trajectory, and the digits
translateMnist: (bool) translate the raw training images on the fly
eyeCentered: (bool) use eye centered system (under development, just fixed max-min thresholding, haven't tested)
preTraining: (bool) pre-train the model to reconstruct the raw image (under development, currently not tunning the looks
with the reconstruction loss. Also, good reconstruction might require deconvolution.)
preTraining_epoch: (int) number of pre-training epochs
drawReconsturction:(int) visualization that compares the true image versus the reconstructed image
MNIST_SIZE: (int) the side length of the training image (the image doesn't have to be MNIST, by the way)
translated_img_size: (int) the side length of the image after translation (probably should be bigger than MNIST_SIZE)
img_size: (int) actual image side length used when executing the program. You shouldn't need to set this parameter. The code should automatically figure it out based on the translateMnist parameter
n_classes: (int) number of classes for the training data set (e.g. it is 10 for MNIST)
initLr:(float) the initial learning rate
lrDecayRate: (float, from 0 to 1) the decay factor for the learning rate, exponential decay is used
lrDecayFreq: (int) the frequency (in terms of training epoch) for the learning rate decay
momentumValue: (float, from 0 to 1) the momentum value for the gradient decent step
batch_size:(int) just the batch size...
depth: (int) number of "zooms" or granularities
sensorBandwidth: (int) the side length for the smallest zoom (finest granularity)
minRadius: (int) the radius for the smallest zoom. It will be automatically figured out by the code
channels: (int) 1 = grayscale images, i have no idea if the code works on colored images
totalSensorBandwidth: (int) depth x channels x (sensorBandwidth^2) the dimension of the input space
nGlimpses: (int) number of glimpses allowed
loc_sd: (int) the std for the noise, imposed on the glimpse locations
hg_size: (int) the size for the glimpse network
hl_size: (int) the size of the location network
g_size: (int) the size for the glimpse feature
cell_size: (int) the size of the core network (the recurrent part)
cell_out_size: (int) = cell_size
start_step: (int) index for the starting epoch
max_iters: (int) max number of epochs allowed
SMALL_NUM: (int) just a very small number, used to avoid the underflow issue when take log