Releases · hellomoto-ai/splatoon2-ml

30 May 05:54

moto-hellomoto-ai

vae-gan-v1.5

48b3a7f

VAE-GAN with Adoptive Beta 2 Latest

Latest

This model is at Epoch 31, where it's GAN part starts to collapse.

The detail can be found here.

Model: VaeGan(
  (vae): VAE(
    (encoder): Encoder(
      (convs): Sequential(
        (0): EncoderBlock(
          (0): ReflectionPad2d((2, 2, 2, 2))
          (1): Conv2d(3, 64, kernel_size=(5, 5), stride=(2, 2), bias=False)
          (2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
          (3): ReLU()
        )
        (1): EncoderBlock(
          (0): ReflectionPad2d((2, 2, 2, 2))
          (1): Conv2d(64, 128, kernel_size=(5, 5), stride=(2, 2), bias=False)
          (2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
          (3): ReLU()
        )
        (2): EncoderBlock(
          (0): ReflectionPad2d((2, 2, 2, 2))
          (1): Conv2d(128, 256, kernel_size=(5, 5), stride=(2, 2), bias=False)
          (2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
          (3): ReLU()
        )
      )
      (map): Linear(in_features=36864, out_features=2048, bias=True)
    )
    (decoder): Decoder(
      (map): Linear(in_features=1024, out_features=36864, bias=True)
      (convs): Sequential(
        (0): DecoderBlock(
          (0): ConvTranspose2d(256, 256, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), bias=False)
          (1): InstanceNorm2d(256, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)
          (2): LeakyReLU(negative_slope=0.01)
        )
        (1): DecoderBlock(
          (0): ConvTranspose2d(256, 128, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), bias=False)
          (1): InstanceNorm2d(128, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)
          (2): LeakyReLU(negative_slope=0.01)
        )
        (2): DecoderBlock(
          (0): ConvTranspose2d(128, 64, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), bias=False)
          (1): InstanceNorm2d(64, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)
          (2): LeakyReLU(negative_slope=0.01)
        )
        (3): ReflectionPad2d((2, 2, 2, 2))
        (4): Conv2d(64, 3, kernel_size=(5, 5), stride=(1, 1))
      )
    )
  )
  (discriminator): Discriminator(
    (convs): Sequential(
      (0): ReflectionPad2d((2, 2, 2, 2))
      (1): Conv2d(3, 32, kernel_size=(5, 5), stride=(1, 1))
      (2): ReflectionPad2d((2, 2, 2, 2))
      (3): Conv2d(32, 128, kernel_size=(5, 5), stride=(2, 2))
      (4): ReflectionPad2d((2, 2, 2, 2))
      (5): Conv2d(128, 256, kernel_size=(5, 5), stride=(2, 2))
      (6): ReflectionPad2d((2, 2, 2, 2))
      (7): Conv2d(256, 256, kernel_size=(5, 5), stride=(2, 2))
    )
    (fc): Sequential(
      (0): Linear(in_features=36864, out_features=512, bias=True)
      (1): ReLU(inplace)
      (2): Linear(in_features=512, out_features=1, bias=True)
      (3): Sigmoid()
    )
  )
)
Optimizers: encoder: Adam (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    eps: 1e-08
    lr: 0.0001
    weight_decay: 0
)
decoder: Adam (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    eps: 1e-08
    lr: 0.0001
    weight_decay: 0
)
discriminator: Adam (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    eps: 1e-08
    lr: 0.0001
    weight_decay: 0
)
Beta: Beta: 10.0
Beta Step: 0.1
Target KLD: 0.1

Assets 3

24 May 04:09

moto-hellomoto-ai

vae-gan-v1.4

4cc2090

VAE-GAN with Adoptive Beta

This model is at Epoch 30, where it's GAN part starts to collapse.

The detail can be found here

Model: VaeGan(
  (vae): VAE(
    (encoder): Encoder(
      (convs): Sequential(
        (0): EncoderBlock(
          (0): ReflectionPad2d((2, 2, 2, 2))
          (1): Conv2d(3, 64, kernel_size=(5, 5), stride=(2, 2), bias=False)
          (2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
          (3): ReLU()
        )
        (1): EncoderBlock(
          (0): ReflectionPad2d((2, 2, 2, 2))
          (1): Conv2d(64, 128, kernel_size=(5, 5), stride=(2, 2), bias=False)
          (2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
          (3): ReLU()
        )
        (2): EncoderBlock(
          (0): ReflectionPad2d((2, 2, 2, 2))
          (1): Conv2d(128, 256, kernel_size=(5, 5), stride=(2, 2), bias=False)
          (2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
          (3): ReLU()
        )
      )
      (map): Linear(in_features=36864, out_features=2048, bias=True)
    )
    (decoder): Decoder(
      (map): Linear(in_features=1024, out_features=36864, bias=True)
      (convs): Sequential(
        (0): DecoderBlock(
          (0): ConvTranspose2d(256, 256, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), bias=False)
          (1): InstanceNorm2d(256, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)
          (2): LeakyReLU(negative_slope=0.01)
        )
        (1): DecoderBlock(
          (0): ConvTranspose2d(256, 128, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), bias=False)
          (1): InstanceNorm2d(128, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)
          (2): LeakyReLU(negative_slope=0.01)
        )
        (2): DecoderBlock(
          (0): ConvTranspose2d(128, 64, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), bias=False)
          (1): InstanceNorm2d(64, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)
          (2): LeakyReLU(negative_slope=0.01)
        )
        (3): ReflectionPad2d((2, 2, 2, 2))
        (4): Conv2d(64, 3, kernel_size=(5, 5), stride=(1, 1))
      )
    )
  )
  (discriminator): Discriminator(
    (convs): Sequential(
      (0): ReflectionPad2d((2, 2, 2, 2))
      (1): Conv2d(3, 32, kernel_size=(5, 5), stride=(1, 1))
      (2): ReflectionPad2d((2, 2, 2, 2))
      (3): Conv2d(32, 128, kernel_size=(5, 5), stride=(2, 2))
      (4): ReflectionPad2d((2, 2, 2, 2))
      (5): Conv2d(128, 256, kernel_size=(5, 5), stride=(2, 2))
      (6): ReflectionPad2d((2, 2, 2, 2))
      (7): Conv2d(256, 256, kernel_size=(5, 5), stride=(2, 2))
    )
    (fc): Sequential(
      (0): Linear(in_features=36864, out_features=512, bias=True)
      (1): ReLU(inplace)
      (2): Linear(in_features=512, out_features=1, bias=True)
      (3): Sigmoid()
    )
  )
)
Initial Beta: 10.0
Beta Step: 0.1
Target KLD: 0.1

Assets 3

16 May 16:12

moto-hellomoto-ai

vae-gan-v1.3

dd01034

AE-GAN with BN Cheating

AE with Batch Normalization at the last layer of the encoder.

The detail can be found here.

This model is Epoch 43, which still produces some fake images.

Model: AeGan(
  (ae): AE(
    (encoder): Encoder(
      (convs): Sequential(
        (0): EncoderBlock(
          (0): ReflectionPad2d((2, 2, 2, 2))
          (1): Conv2d(3, 64, kernel_size=(5, 5), stride=(2, 2), bias=False)
          (2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
          (3): ReLU()
        )
        (1): EncoderBlock(
          (0): ReflectionPad2d((2, 2, 2, 2))
          (1): Conv2d(64, 128, kernel_size=(5, 5), stride=(2, 2), bias=False)
          (2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
          (3): ReLU()
        )
        (2): EncoderBlock(
          (0): ReflectionPad2d((2, 2, 2, 2))
          (1): Conv2d(128, 256, kernel_size=(5, 5), stride=(2, 2), bias=False)
          (2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
          (3): ReLU()
        )
      )
      (map): Sequential(
        (0): Linear(in_features=36864, out_features=1024, bias=False)
        (1): BatchNorm1d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (decoder): Decoder(
      (map): Linear(in_features=1024, out_features=36864, bias=True)
      (convs): Sequential(
        (0): DecoderBlock(
          (0): ConvTranspose2d(256, 256, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), bias=False)
          (1): InstanceNorm2d(256, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)
          (2): LeakyReLU(negative_slope=0.01)
        )
        (1): DecoderBlock(
          (0): ConvTranspose2d(256, 128, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), bias=False)
          (1): InstanceNorm2d(128, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)
          (2): LeakyReLU(negative_slope=0.01)
        )
        (2): DecoderBlock(
          (0): ConvTranspose2d(128, 64, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), bias=False)
          (1): InstanceNorm2d(64, eps=1e-05, momentum=0.1, affine=False, track_running_stats=False)
          (2): LeakyReLU(negative_slope=0.01)
        )
        (3): ReflectionPad2d((2, 2, 2, 2))
        (4): Conv2d(64, 3, kernel_size=(5, 5), stride=(1, 1))
      )
    )
  )
  (discriminator): Discriminator(
    (convs): Sequential(
      (0): ReflectionPad2d((2, 2, 2, 2))
      (1): Conv2d(3, 32, kernel_size=(5, 5), stride=(1, 1))
      (2): ReflectionPad2d((2, 2, 2, 2))
      (3): Conv2d(32, 128, kernel_size=(5, 5), stride=(2, 2))
      (4): ReflectionPad2d((2, 2, 2, 2))
      (5): Conv2d(128, 256, kernel_size=(5, 5), stride=(2, 2))
      (6): ReflectionPad2d((2, 2, 2, 2))
      (7): Conv2d(256, 256, kernel_size=(5, 5), stride=(2, 2))
    )
    (fc): Sequential(
      (0): Linear(in_features=36864, out_features=512, bias=True)
      (1): ReLU(inplace)
      (2): Linear(in_features=512, out_features=1, bias=True)
      (3): Sigmoid()
    )
  )
)

Assets 3

19 Apr 04:01

moto-hellomoto-ai

vae-gan-v1.2

09a2fe7

Simple VAE-GAN v3

The model structure is the same as the previous ones, but the update methods are tweaked and improvement is more stable and yields better quality.

The detail can be found here.

Assets 3

16 Apr 05:00

moto-hellomoto-ai

vae-gan-v1.1

ca2e0dd

Simple VAE-GAN v2

Trained with the same setting as v1.0 except fake images generated from randomly sampled latent variables are removed.

No noticeable difference was observed. The detail can be found here

Assets 3

13 Apr 16:14

moto-hellomoto-ai

vae-gan-v1.0

978e9f0

VAE-GAN - Initial Model

First successfully trained VAE-GAN model.

Results are described here

Trained for 61 epochs, 38735 steps with batch size 64.

Torch version: 1.0.1.post2

Note on implementation difference from VAEGAN-PYTORCH

Batch normalization was replaced with instance normalization.
No non-linear activation in feature extraction part of the discriminator.
No loss-based balancing
No gradient clipping

Model: VaeGan(
  (vae): VAE(
    (encoder): Encoder(
      (convs): Sequential(
        (0): EncoderBlock(
          (0): ReflectionPad2d((2, 2, 2, 2))
          (1): Conv2d(3, 64, kernel_size=(5, 5), stride=(2, 2), bias=False)
          (2): InstanceNorm2d(64, eps=1e-05, momentum=0.9, affine=False, track_running_stats=False)
          (3): ReLU()
        )
        (1): EncoderBlock(
          (0): ReflectionPad2d((2, 2, 2, 2))
          (1): Conv2d(64, 128, kernel_size=(5, 5), stride=(2, 2), bias=False)
          (2): InstanceNorm2d(128, eps=1e-05, momentum=0.9, affine=False, track_running_stats=False)
          (3): ReLU()
        )
        (2): EncoderBlock(
          (0): ReflectionPad2d((2, 2, 2, 2))
          (1): Conv2d(128, 256, kernel_size=(5, 5), stride=(2, 2), bias=False)
          (2): InstanceNorm2d(256, eps=1e-05, momentum=0.9, affine=False, track_running_stats=False)
          (3): ReLU()
        )
      )
      (map): Linear(in_features=36864, out_features=2048, bias=True)
    )
    (decoder): Decoder(
      (map): Linear(in_features=1024, out_features=36864, bias=True)
      (convs): Sequential(
        (0): DecoderBlock(
          (0): ConvTranspose2d(256, 256, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), bias=False)
          (1): InstanceNorm2d(256, eps=1e-05, momentum=0.9, affine=False, track_running_stats=False)
          (2): ReLU()
        )
        (1): DecoderBlock(
          (0): ConvTranspose2d(256, 128, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), bias=False)
          (1): InstanceNorm2d(128, eps=1e-05, momentum=0.9, affine=False, track_running_stats=False)
          (2): ReLU()
        )
        (2): DecoderBlock(
          (0): ConvTranspose2d(128, 64, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), bias=False)
          (1): InstanceNorm2d(64, eps=1e-05, momentum=0.9, affine=False, track_running_stats=False)
          (2): ReLU()
        )
        (3): ReflectionPad2d((2, 2, 2, 2))
        (4): Conv2d(64, 3, kernel_size=(5, 5), stride=(1, 1))
      )
    )
  )
  (discriminator): Discriminator(
    (convs): Sequential(
      (0): ReflectionPad2d((2, 2, 2, 2))
      (1): Conv2d(3, 32, kernel_size=(5, 5), stride=(1, 1))
      (2): ReflectionPad2d((2, 2, 2, 2))
      (3): Conv2d(32, 128, kernel_size=(5, 5), stride=(2, 2))
      (4): ReflectionPad2d((2, 2, 2, 2))
      (5): Conv2d(128, 256, kernel_size=(5, 5), stride=(2, 2))
      (6): ReflectionPad2d((2, 2, 2, 2))
      (7): Conv2d(256, 256, kernel_size=(5, 5), stride=(2, 2))
    )  # -> Output used for feature matching
    # ReLU applied here before fc
    (fc): Sequential(
      (0): Linear(in_features=36864, out_features=512, bias=True)
      (1): ReLU(inplace)
      (2): Linear(in_features=512, out_features=1, bias=True)
      (3): Sigmoid()
    )
  )
)
Optimizers: encoder: Adam (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    eps: 1e-08
    lr: 0.0001
    weight_decay: 0
)
decoder: Adam (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    eps: 1e-08
    lr: 0.0001
    weight_decay: 0
)
discriminator: Adam (
Parameter Group 0
    amsgrad: False
    betas: (0.9, 0.999)
    eps: 1e-08
    lr: 0.0001
    weight_decay: 0
)

Assets 3

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Releases: hellomoto-ai/splatoon2-ml

VAE-GAN with Adoptive Beta 2

VAE-GAN with Adoptive Beta

AE-GAN with BN Cheating

Simple VAE-GAN v3

Simple VAE-GAN v2

VAE-GAN - Initial Model