Running in test mode was not producing any output, so I added an initial implementation of Trainer.test(). This simply autoencodes a batch of inputs, and saves the inputs as well as the autoencoded outputs.

This can be combined with '--split valid' to test autoencoding on validation data. To make these tests more repeatable, shuffling was disabled in test mode and the images are sort ordered.

Sample output attached.

In the BEGAN paper ,it says:"Implementation note: while the updates are made simultaneously, they are still adversarial. As such, it is important to optimize θD and θG independently with respect to their corresponding losses". Maybe this will get better results?

The current training routine is using all of the CelebA images, including the canonical CelebA validation and test data. This change adds support for CelebA splits and by default only trains on only the CelebA training images.

This was done by moving the downloaded data into an "images" subdirectory, and then creating subdirectories of splits with relative symlinks to the original files. This format required only light changes to the config and data_loader.

Note that honoring the standard splits in generative modeling is important for a few reasons. Perhaps most importantly, it allows for fair comparisons across papers when visually inspecting image samples; it is unfair to compare models trained on 200k images versus those only exposed to a 160k subset. But also relevant for models with an encoder like BEGAN - it allows one to evaluate a trained model by performing reconstructions on out of sample validation data.

Excuse me，I try to run the code of BEGAN-pytorch, but the generate images are a mess ,like this:

Could u tell me the reason？ Must I use the BEGAN-tensorflow to get the good result? thanks！

Hi, @dribnet @carpedm20 @scott-vsi

I met this error:

ctilab@ctilab:~/BEGAN-pytorch$ python main.py --dataset=CelebA --num_gpu=1 --use_tensorboard=True Found 162770 images in subfolders of: data/CelebA/splits/train [*] MODEL dir: logs/CelebA_1008_173315 [*] PARAM path: logs/CelebA_1008_173315/params.json 0%| | 0/500000 [00:00<?, ?it/s]/usr/local/lib/python2.7/dist-packages/torch/nn/modules/upsampling.py:135: UserWarning: nn.UpsamplingNearest2d is deprecated. Use nn.Upsample instead. warnings.warn("nn.UpsamplingNearest2d is deprecated. Use nn.Upsample instead.")

Traceback (most recent call last): File "main.py", line 42, in main(config) File "main.py", line 34, in main trainer.train() File "/home/ctilab/BEGAN-pytorch/trainer.py", line 160, in train d_loss_real = l1(AE_x, x) File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/module.py", line 224, in call result = self.forward(*input, **kwargs) File "/home/ctilab/BEGAN-pytorch/models.py", line 117, in forward return backend_fn(self.size_average)(input, target) TypeError: forward() takes at least 3 arguments (2 given) ctilab@ctilab:~/BEGAN-pytorch$

What's wrong with me?

Hi, @dribnet @carpedm20 @scott-vsi

I met this error:

ctilab@ctilab:~/BEGAN-pytorch$ python main.py --dataset=images --num_gpu=0 RuntimeError: module compiled against API version 0xb but this version of numpy is 0xa Traceback (most recent call last): File "main.py", line 1, in import torch File "/usr/local/lib/python2.7/dist-packages/torch/init.py", line 53, in from torch._C import * ImportError: numpy.core.multiarray failed to import

What's wrong with me?

Hi! Refering your code, I make below changes in my code.

I change the way of the initializing at first. Then, I change the way to get the loss. In my code, the both part of the g_loss would be used for learning G's parameters, but in your code it maybe only the AE_G be used for learning G's parameters. I'm not sure about this, but maybe you can try it. And it's not big changes. I'm sorry for my pool English~

Not sure exactly what occurred, but while training with your code with default parameters and tensorboard enabled, the training lost it at 14000/500000.

The network then began outputting completely black images for all three _D, _D, and _D_fake.

Not much more insight from my end I am afraid. Trained on a Titan XP in ubuntu with pytorch 0.1.10 - py27_1cu80 [cuda80] - soumith, torchvision 0.1.6, and python 2.7.13 on ubuntu 16.04.

Loss_D went from 0.0436 to 1.45, and L_x went from 0.0436 to 1.522