Generative adversarial network

[wangjiawen2013]

Hi,
I am working on a GAN recently like this:

In each epoch, I must train the discriminator for many times and then the generator for one time. I find it's hard to implement GAN using learnerbuilder similar to the code in burn example:

Does learnerbuilder applies only to simple network ? Do I need to write customized loss function and customized training loop ?

[laggui]

Training GANs requires training a generator and a discriminator, with one optimizer for each.

This is a bit more complex than the typical training loop implemented by the learner so you would have to roll with your own training loop 🙂

[wangjiawen2013]

I am re-implementing pytorch wgan with burn recently days, this is the github repository: https://github.com/wangjiawen2013/wgan

I want to merge it into burn's github examples after finishing it. But some problems are still needed to addressed, so may I ask your help here? I hope that you can have a look at it and find the failure reasons.

I have run pytorch wgan sucessfully, this is the output:

But for burn wgan, it is just meanless and repeated patterns:

There must be something wrong with my code. I have tested a lot of times and find that most things are as expected. I have strong doubts that the loss or the optimizer.step() don't work (src/training.rs line 121-130 and line 135-142). It is expected that the loss will be updated fastly when running optimizer.step() many times. In pytorch wgan, it worked as expected. But in burn wgan, the loss (D loss: discriminator loss, G loss: generator loss) updated slowly when running optimizer.step(), no matter it is let loss_d = -discriminator.forward(batch.images).mean() + discriminator.forward(fake_images).mean(); or let loss_d = discriminator.forward(batch.images).mean() - discriminator.forward(fake_images).mean(); !!! And the D loss and G loss differed a lot from that of pytorch, approximate several hundreds times lower than pytorch. Here is the training process:

[Epoch 1/3] [Batch 0/938] [D loss: -0.09546152] [G loss: -0.010683101]
[Epoch 1/3] [Batch 10/938] [D loss: -0.09064701] [G loss: -0.010889139]
[Epoch 1/3] [Batch 20/938] [D loss: -0.09109107] [G loss: -0.011392368]
[Epoch 1/3] [Batch 30/938] [D loss: -0.0905629] [G loss: -0.012158312]
[Epoch 1/3] [Batch 40/938] [D loss: -0.08930381] [G loss: -0.013122454]
[Epoch 1/3] [Batch 50/938] [D loss: -0.088740215] [G loss: -0.014366182]
[Epoch 1/3] [Batch 60/938] [D loss: -0.0866037] [G loss: -0.016056392]
[Epoch 1/3] [Batch 70/938] [D loss: -0.08377892] [G loss: -0.018405747]
[Epoch 1/3] [Batch 80/938] [D loss: -0.08068731] [G loss: -0.021662386]
[Epoch 1/3] [Batch 90/938] [D loss: -0.07828338] [G loss: -0.026396101]
[Epoch 1/3] [Batch 100/938] [D loss: -0.067133166] [G loss: -0.03337042]
[Epoch 1/3] [Batch 110/938] [D loss: -0.057932906] [G loss: -0.04426057]
[Epoch 1/3] [Batch 120/938] [D loss: -0.04228705] [G loss: -0.059334144]
[Epoch 1/3] [Batch 130/938] [D loss: -0.021414958] [G loss: -0.08037117]
[Epoch 1/3] [Batch 140/938] [D loss: 0.0035974532] [G loss: -0.10367289]
[Epoch 1/3] [Batch 150/938] [D loss: 0.021838397] [G loss: -0.1235547]
[Epoch 1/3] [Batch 160/938] [D loss: 0.034707725] [G loss: -0.13705881]
[Epoch 1/3] [Batch 170/938] [D loss: 0.046433717] [G loss: -0.14530241]
[Epoch 1/3] [Batch 180/938] [D loss: 0.048114464] [G loss: -0.15006739]
[Epoch 1/3] [Batch 190/938] [D loss: 0.05085115] [G loss: -0.1528728]
[Epoch 1/3] [Batch 200/938] [D loss: 0.051142164] [G loss: -0.15466571]

I know it is a time-consuming labour to read through these code, while I hope that you could give some suggestions to help it works in burn.

[wangjiawen2013]

Here is the training process of pytorch wgan, it worked as expected: