Data
From Siqi:
https://www.dropbox.com/s/8wrfbkpaosn23vq/labeledimg1.mat?dl=0
https://www.dropbox.com/s/uw9nx8mp8pqe94e/labeledimg2.mat?dl=0
In the data structure there is vccimg and yagimg which refer to the images on VCC and YAG screens. There is vccbox and yagbox which refer to the box coordinate where it finds the beam, [ystart yend xstart xend]. If it's empty it means it detects no signal. I corrected the mislabeled ones using the fitting method so they should all be labeled correctly now.
There is also a labedimg4.mat:
Some points worth mentioning:
1. the vccImg and yagImg are raw images, i.e. before background subtraction
2. background images are saved in vccbkg and yagbkg.
3. there is a camera gain problem on the VCC camera, so if you need to do background subtraction on VCC you may have to enforce non-negative intensities after subtraction. Background subtraction for YAG images work normally.
4. I also added in vccbeam and yagbeam which give the full beam image on both cameras. When I did the labeling I restrained the search region within the full beam region on these two images, since the vccImg and yagImg are just small portions of the full beam.
From David
I have downloaded the files, they are at (on the psana nodes) /reg/d/ana01/temp/davidsch/mlearn/acc_beam_locate/labeledimg*.mat
Access from Python
scipy can load mat files, here is some code that investigates a little:
| Code Block | ||
|---|---|---|
| ||
In [3]: import scipy.io as sio
In [4]: labeledimg1 = sio.loadmat('labeledimg1.mat')
In [8]: vccImg = labeledimg1['vccImg']
In [18]: vccBox = labeledimg1['vccbox']
# you'll see vccImg and vccBox show up as 1 x 110 arrays of 'object', they are the images and labels for 110 samples
# like Siqi says, a box entry is empty if no beam is present, here we get a count of the non empty boxes, or samples with beam
In [23]: len([bx for bx in vccBox[0,:] if len(bx)>0])
Out[23]: 80
The first entry with a box is 4, so you can plot like
In [24] %pylab
In [26]: imshow(vccImg[0,4])
In [27]: bx = vccBox[0,4]
In [31]: ymin,ymax,xmin,xmax=bx[0,:]
In [32]: plot([xmin,xmin,xmax,xmax,xmin],[ymin,ymax,ymax,ymin,ymin], 'w') |
In which case I see
Description
The page Accelerator Beam finding - Internal Notes talks about the data and code, page has restricted access.
Below is an example of the problem we are trying to solve:
- this is a vcc screen, and the location of the beam has been labeled with the white box.
- We want to use machine learning to predict these box locations.
- The dataset contains vcc screens and YAG screens - two different datasets, currently looking at training two different models - a YAG predictor and a VCC predictor.
Description
There are 3 files, called 1,2 and 4.
- Files 1 and 2 have 142 samples. With file 4, it is the total number of 239 samples is.
- Each sample has a yag, vcc, and box for each - there are also backgrounds to subtract for file 4, and the beam to narrow the searcha oval like image of the entire beam region - one can use that to narrow the search so as to not predict a box in a corner of the yag or vcc screen.
- vcc values are in [0,255], and the boxed beam can get quite brite
- yag values go over 1000, I think, but the boxed value is always dim, like up to 14
First Pass - just files 1 and 2
Given the apparent success of using transfer learning to do Spatial Localization to find the 'fingers' in XTCAV data, we will try the same thing with the accelerator data.
We have to fit the 480 x 640 vcc images, and 1040 x 1392 yag images into the 224 x 224 x 3 RBG sized images that the vgg16 convolutional neural network expects.
I thresholed yag at 255, then made grayscale images for each, using a scipy imresize option.
...