Up to this point, we have used circular blur kernel f in pixel coordinates. In this set of experiments, we have used circular blur kernel in physical coordinates. This blog is continuation of previous blogs. Please see the other blogs to learn the experiment procedure.
These results are for far field region. Patches are coming from far field region as in the previous blog.
Note: Best and Worst cases are chosen by comparing the average bubble error in a patch.
Quantitative Metrics: In order to compare success of different experiments, we decide the use the followings:
Quantitative Metric1 = \[ L1Loss(z**f_0 – x**f_0) \]
Quantitative Metric2 = \[ MSELossLoss(z**f_0 – x**f_0) \]
where f_0 is a gaussian kernel with sigma =1.
Note: Both f and f_0 are in physical coordinates.
Experiment 1: learning rate=1e-5, sigma of kernel f=1, regularizer parameter=0.01
Average quantitative metric1 is 21.066962
Average metric1 per bubble is 0.829031
Average quantitative metric2 is 0.805354
Average metric2 per bubble is 0.031346
Worst Case: Center Location(x,z) = 29.61 mm and 48.80 mm , Average Metric1 per bubble: 1.95, Average Metric2 per bubble: 0.133
Best Case: Center Location(x,z) = 17.27 mm and 48.31 mm , Average Metric1 per bubble: 0.3797, Average Metric2 per bubble:0.00542
Experiment 2: learning rate=1e-5, sigma of kernel f=1, regularizer parameter=0.005
Average quantitative metric1 is 21.959468
Average metric1 per bubble is 0.863606
Average quantitative metric2 is 0.812720
Average metric2 per bubble is 0.031575
Worst Case: Center Location(x,z) = 32.07mm and 48.80mm , Average Metric1 per bubble: 1.7694, Average Metric2 per bubble: 0.12264
Best Case: Center Location(x,z) = 19.75mm and 47.32mm , Average Metric1 per bubble: 0.4744, Average Metric2 per bubble: 0.008265
Experiment 3: learning rate=1e-5, sigma of kernel f=1.5, regularizer parameter=0.01
Average quantitative metric1 is 21.031181
Average metric1 per bubble is 0.827787
Average quantitative metric2 is 0.779206
Average metric2 per bubble is 0.030392
Worst Case: Center Location(x,z) = 33.30mm and 48.80mm , Average Metric1 per bubble: 1.828, Average Metric2 per bubble: 0.106
Best Case: Center Location(x,z) = 13.59mm and 36.98mm , Average Metric1 per bubble: 0.514, Average Metric2 per bubble: 0.0099
Experiment 4: learning rate=1e-5, sigma of kernel f=1.5, regularizer parameter=0.005
Average quantitative metric1 is 20.906547
Average metric1 per bubble is 0.823864
Average quantitative metric2 is 0.718547
Average metric2 per bubble is 0.028081
Worst Case: Center Location(x,z) = 29.61mm and 48.80mm , Average Metric1 per bubble: 1.678, Average Metric2 per bubble: 0.1032
Best Case: Center Location(x,z) = 16.05mm and 43.38mm , Average Metric1 per bubble: 0.518, Average Metric2 per bubble: 0.0088
Experiment 5: learning rate=1e-5, sigma of kernel f=2, regularizer parameter=0.01
Average quantitative metric1 is 22.376564
Average metric1 per bubble is 0.881055
Average quantitative metric2 is 0.872184
Average metric2 per bubble is 0.034107
Worst Case: Center Location(x,z) = 33.30mm and 48.80mm , Average Metric1 per bubble: 1.7376, Average Metric2 per bubble: 0.1031
Best Case: Center Location(x,z) = 25.91mm and 45.85mm , Average Metric1 per bubble: 0.606, Average Metric2 per bubble: 0.015455
Experiment 6: learning rate=1e-5, sigma of kernel f=2, regularizer parameter=0.005
Average quantitative metric1 is 23.040157
Average metric1 per bubble is 0.907826
Average quantitative metric2 is 0.863941
Average metric2 per bubble is 0.033746
Worst Case: Center Location(x,z) = 29.61 mm and 48.80 mm , Average Metric1 per bubble: 1.6995, Average Metric2 per bubble: 0.104
Best Case: Center Location(x,z) = 8.66mm and 39.44mm , Average Metric1 per bubble: 0.590, Average Metric2 per bubble: 0.01337
1) I now realized that the quantitative metrics should be normalized by the number of bubbles in the patch (image) that you are processing, so that what you report is the average metric per bubble.
2) I am not clear as to what f you are using for the quantitative metrics, as you are using the same symbol as the f used in computing the loss in training the NN. Please use f_0 for the the one used for the quantitative metrics, and f for the one used for training, and report both. Also report whether each of them is circular in pixel of physical coordinates.
3) For each such set of experiments, it would be good to summarize the results in a table, giving the parameters used (including number of epochs used and batch size – in a caption if same for all experiments), the values of the quantitative metrics, and the % gap between the loss on the training and on the validation data, at the end of training.
1) Done.
2) Done.
3) Here is the summary: https://ufuksoylu.web.illinois.edu/2019/07/25/summary/