loss function about synthtext

[ankush-me]

Hi,
The loss function is exactly the same as the one in eq 2 of the following version of the YOLO paper:
https://arxiv.org/pdf/1506.02640v1.pdf

[sebastiangonsal]

But that does not take into account the orientation of boxes in your paper?

[ankush-me]

Those are additional two L2 loss terms on the sine and cosine of the orientation angle.

[sebastiangonsal]

Thanks. My loss oscillates a lot during iterations. Was your loss small enough in the starting iterations already? I am looking at a loss of 150 per image after 170 iterations of batches of size 16 each.

[MansourTrabelsi]

hello please can you help me
i have this error
Traceback (most recent call last):
File "gen.py", line 19, in
from synthgen import *
File "/home/mansour/SynthText/synthgen.py", line 20, in
import text_utils as tu
File "/home/mansour/SynthText/text_utils.py", line 12, in
from pygame import freetype
ImportError: cannot import name freetype

[sebastiangonsal]

@MansourTrabelsi this is unrelated to this issue

[sebastiangonsal]

@ankush-me I am using batchsize of 16, 512 sized images, grid of 32 for each image. I am using weight of 0.01 increasing to 1.0 for grid boxes that do not have object. I increased the weight in a geometric fashion as you have noted in another issue. However, I still end up with zero probabilities for all boxes in the image. Anything that you can think of that can help me ?

[ankush-me]

@sebastiangonsal I did not notice oscillation. First of all, it helps to also monitor the performance during training, by calculating the precision/recall of the predictions in every batch.
Second, if the probability is going to zero, there could be two reasons:

(1) it is a common error in implementing FCRN, that the axes of the labels and the images do not correspond, i.e., the x and y axes are either transposed or flipped. I would first make sure that this is not the case.
(2) the rate of increase is too high. I do not recall at what rate I increased it, but may be try decreasing the rate -- I do remember, if the rate was too high, the network would collapse to all zeros in the beginning itself and wouldn't recover (i.e., recall would go to zero). Are you experiencing that? If so, try lowering the initial 0.01 further. If recall is non-zero in the beginning, then try increasing the weight at a lower rate.

[ankush-me]

@MansourTrabelsi Please check #24.