Conv+FP pose cnn tensorflow does not learn

[freek.geor...@gmail.com]

Dear Chelsea and Sergey,

Thanks for making your code available!

I have been trying to use image input with MDGPS in the Box2D environment, but can't get the convolutional layers to shift to the target.

When trying to train the convolutional network separately with the feature point layer and a small fully connected network, it will not reliably learn to predict sine and cosine of joint angles and normalized (-1 to 1) position of a target block.

Target prediction is improved when the target block is brighter, but arm pose prediction degrades at the same time.

I have tried a lot of hyperparameter combinations for the network architecture, including those in your BADMM paper.

Could this be due to the feature point layer weakening the gradients?

Did you ever have a similar problem?

Thanks in advance!

Freek

[Chelsea Finn]

Hi Freek,

Are you using GPS or supervised learning for pose prediction? 

Generally, I would suggest trying different initializations for the conv layers. If the spatial softmax is initialized at a very saturated levels, it will have trouble learning.

I would also encourage trying batch normalization.

Chelsea

--
You received this message because you are subscribed to the Google Groups "gps-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to gps-help+unsubscribe@googlegroups.com.
To post to this group, send email to gps-...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/gps-help/720d686a-76ef-4e6a-bbe6-d62b2203473b%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[freek.geor...@gmail.com]

Hi Chelsea,

Thanks for your response!

I am using regular supervised learning for pose prediction with a very simple Tensorflow script.

As for initialization, I did experiment with different standard deviations, but always using the same standard deviation for the fully connected layers as for the convolutional layers.

Now that you mention it, it does make sense to use a lower standard deviation for the conv layers only.

It hadn't occurred to me yet to incorporate batch normailzation, thanks for the tip!

Regards,

Freek

Op vrijdag 17 november 2017 06:58:25 UTC+1 schreef Chelsea Finn:

Hi Freek,

Are you using GPS or supervised learning for pose prediction? 

Generally, I would suggest trying different initializations for the conv layers. If the spatial softmax is initialized at a very saturated levels, it will have trouble learning.

I would also encourage trying batch normalization.

Chelsea

On Wed, Nov 15, 2017 at 4:44 AM, <freek.geor...@gmail.com> wrote:

Dear Chelsea and Sergey,

Thanks for making your code available!

I have been trying to use image input with MDGPS in the Box2D environment, but can't get the convolutional layers to shift to the target.

When trying to train the convolutional network separately with the feature point layer and a small fully connected network, it will not reliably learn to predict sine and cosine of joint angles and normalized (-1 to 1) position of a target block.

Target prediction is improved when the target block is brighter, but arm pose prediction degrades at the same time.

I have tried a lot of hyperparameter combinations for the network architecture, including those in your BADMM paper.

Could this be due to the feature point layer weakening the gradients?

Did you ever have a similar problem?

Thanks in advance!

Freek

--
You received this message because you are subscribed to the Google Groups "gps-help" group.

To unsubscribe from this group and stop receiving emails from it, send an email to gps-help+u...@googlegroups.com.