Output GIF is All NONE

[quan-meng]

Output GIF is All NONE

[SimonCK666]

Nope

Acturally in the end of training, the loss of discriminator was almost converged to 0.5, and it looked like normal.

All facts show that the camera pose and generator training are fine.
It's just like when doing volume rendering, the rays all missed the drum object.

Today morning, I retried for the dataset:hotdog/. Also keep the batch_size was 6.
In this time, the model seems work all good.

So the problem seems just happened on the dataset:drums/

[quan-meng]

Changing the batch size has a certain effect on the stability of GAN

If the result is empty,
it should mean that GAN training has failed
Generally the loss of generator and discriminator are around 1.
If the training fails at the beginning, I will rerun it.

Is the loss of discriminator in your training close to 0?

[SimonCK666]

Sorry to bother you again,

I found that I cannot train this model successfully unless using the hotdog data.

Do I need to change these params blow, when I use other data to train? Such as drums or lego data.
I found when I change data, the camera pose estimation always gather.

(I put the camera pose estimation when trained lego below)

azim_range: [ 0., 360. ]  # the range of azimuth
elev_range: [ 0., 90. ]   # the range of elevation
radius: [ 4.0, 4.0 ]  # the range of radius

near: 2.0
far: 6.0

[quan-meng]

You don't need to change the params in a same dataset, e.g., Blender (six scenes with upper hemisphere distribution), DTU

I will check it again in my computer with the default setting.

[quan-meng]

You can attach the screenshot of your training curve in tensorboard

so I can better analyze the problem based on my experience

[quan-meng]

I partially tested the code on chair and drums scenes,
the estimated camera parameters, RGB and Depth maps all look right

A successful training curve should look like this