Tf-pose-estimation: parameters used when training from scratch with mobilenet models

[chuan92]

Thanks for your work.
Can you share the parameters used when training from scratch with mobilenet_v1_0.75 models.
It seems a local optimum when I set lr=0.001, batch_size=32

[ouceduxzk]

@chuan92 I have the similar loss, tried even much smaller lr and other optimizers, no luck in getting better loss

[chuan92]

@ildoonet sorry to interrupt you, I would appreciate it if you could share the parameters.

[andyqian2015]

Hi,@chuan92 when you train the mobileNet from scratch,do you encount the problem that ：

    for images_test, heatmaps, vectmaps in df_valid.get_data():
        validation_cache.append((images_test, heatmaps, vectmaps))

take too much time and it result out that the values of images_test,heatmaps and vectmaps all comes to zeros. is it right? if you can reply me . I would be very appreciate.

[ouceduxzk]

is it possible to share the loss plot of your mobilenet based model pose estimation model, such that we can compare at the early stage without running multiple long experiment that end up being a failure

[andyqian2015]

@chuan92 have you worked out this problem,when I train the mobileNet to 10w,the loss is still still to 1000 more or less.

[chuan92]

@andyqian2015 I didn't encounter your problem. But I train the mobileNet to 27000 steps, the loss_valid is 2600 and seems converge.

[blueardour]

I also have the problem of local optimum. Is there any practical experience how to get rid of it and reach more better optimum?

[lingshaokun]

Hi,@chuan92 when you train the mobileNet from scratch,do you encount the problem that ：

    for images_test, heatmaps, vectmaps in df_valid.get_data():
        validation_cache.append((images_test, heatmaps, vectmaps))

take too much time and it result out that the values of images_test,heatmaps and vectmaps all comes to zeros. is it right? if you can reply me . I would be very appreciate.

So, Have you solved this problem?