This is the problem that came up in the first place. So I printed some information and positioned it here. Can you tell me why?Thank you!!! I use my own database and tensorflow 0.10.0. from tranier.py: self.z: Tensor("random_uniform:0", shape=(?, ?), dtype=float32) from models.py : self.z: Tensor("random_uniform:0", shape=(?, ?), dtype=float32)

Here is the complete error message:

root: /media/z/lhj/images/testData_aligned/testData_aligned/ [16, 64, 64, 3] 3 64 64 self.data_loader: Tensor("ToFloat:0", shape=(16, 3, 64, 64), dtype=float32) self.z_num: 64 trainer x: Tensor("sub:0", shape=(16, 3, 64, 64), dtype=float32) tf.shape(x)[0]: Tensor("strided_slice:0", shape=(), dtype=int32) self.z: Tensor("random_uniform:0", shape=(?, ?), dtype=float32) self.conv_hidden_num: 128 self.channel: 3 self.repeat_num: 4 self.data_format: NCHW from models begin: num_output: 8192 z: Tensor("random_uniform:0", shape=(?, ?), dtype=float32) Traceback (most recent call last): File "main.py", line 44, in main(config) File "main.py", line 32, in main trainer = Trainer(config, data_loader_) File "/home/lhj/software/BEGAN-tensorflow-master/trainer.py", line 93, in init self.build_model() File "/home/lhj/software/BEGAN-tensorflow-master/trainer.py", line 199, in build_model self.repeat_num, self.data_format, reuse=False) File "/home/lhj/software/BEGAN-tensorflow-master/models.py", line 13, in GeneratorCNN x = slim.fully_connected(z, num_output, activation_fn=None) File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/framework/python/ops/arg_scope.py", line 171, in func_with_args return func(*args, **current_args) File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/layers/python/layers/layers.py", line 792, in fully_connected num_input_units = utils.last_dimension(inputs_shape, min_rank=2) File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/layers/python/layers/utils.py", line 199, in last_dimension raise ValueError('last dimension shape must be known but is None') ValueError: last dimension shape must be known but is None

I have had this weird behaviour where after 380000 epochs the generator suddenly mode collapses. For example see:

epoch 350k

epoch 390k

epoch 396k

epoch 400k

epoch 480k

I am suspecting that portrait guy in the middle of the lower row as the poltergeist that caused the mode collapse but I can't prove it :laughing: . Anyone seen this happen before?

You can also see the g_loss and d_loss_fake increasing around epoch 380k

Hi, I put the datasets in the data file, and execute this command：$python main.py --dataset=CelebA --use_gpu=True ,The following bug: Traceback (most recent call last): File "main.py", line 4, in from trainer import Trainer File "E:\product\faceVerification\faceAttributes\BEGAN-tensorflow\trainer.py", line 4, in import StringIO ImportError: No module named 'StringIO'

Thank you for your help!

Couldn't find a reason why, but the model always shows mode collapsing even after learning rate decay one or two times unlike the paper's description. I guess there could be an error on loss but first of all, I don't know how BEGAN tackle this without push-away factor of EBGAN (not sure learning rate decay is enough?).

Hi, sorry to interrupt you , I got another problem when running the program as follows :

Traceback (most recent call last): File "main.py", line 42, in main(config) File "main.py", line 28, in main config.data_format, config.split) File "/home/chikiuso/Downloads/BEGAN-tensorflow/data_loader.py", line 48, in get_loader queue = tf.image.crop_to_bounding_box(queue, 50, 25, 128, 128) File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/image_ops_impl.py", line 473, in crop_to_bounding_box assert_ops += _Check3DImage(image, require_static=False) File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/image_ops_impl.py", line 131, in _Check3DImage raise ValueError("'image' must be three-dimensional.") ValueError: 'image' must be three-dimensional.

I'm new to machine learning. Can anyone tell me how to solve this problem. Thanks ahead.

File "main.py", line 43, in main(config) File "main.py", line 35, in main trainer.train() File "/Users/developer/Documents/git/opensource/BEGAN-tensorflow/BEGAN-tensorflow/trainer.py", line 158, in train self.autoencode(x_fixed, self.model_dir, idx=step, x_fake=x_fake) File "/Users/developer/Documents/git/opensource/BEGAN-tensorflow/BEGAN-tensorflow/trainer.py", line 263, in autoencode x = self.sess.run(self.AE_x, {self.x: img}) File "/Users/developer/anaconda/envs/tensorflow/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 778, in run run_metadata_ptr) File "/Users/developer/anaconda/envs/tensorflow/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 961, in _run % (np_val.shape, subfeed_t.name, str(subfeed_t.get_shape()))) ValueError: Cannot feed value of shape (16, 3, 64, 64) for Tensor 'ToFloat:0', which has shape '(16, 64, 64, 3)'

The model can not show any convergence clues in MNIST dataset during my training. after 1600 batch of training. It just remain no change with serious mode collapse. something wrong?

Hi @carpedm20 - this codebase is now working well for me, wanted to compare notes. with your latest changes I get pretty good results on 64x64 celeba. 128x128 doesn't do as well - probably needs different hyperparams.

Here are some 64x64 results after 130k cycles of training. first x_fixed, D_real :

That looks like pretty good auto-encoder reconstruction results - the resemblance is pretty clear. OK, now here's the D_fake, and G:

Now here my knowledge of this model gets a little hazy. These look fairly good and they look like each other - but they no longer bear very much resemblance x_fixed - should they? If so, then I think this is would likely be a symptom of a pretty significant mode collapse.

Just for comparison, here's what taking the same x_fixed looks like when passed through a trained ALI model:

Some of these come out better than others, but note the overall variation in hair color, skin tone, glasses, hats, backgrounds that seems to be missing from the current BEGAN generator. Curious if you have any thoughts from your own investigations.

Just a word of caution: As far as I understand, the way the interpolation of images is implemented here does not correspond to the way it was done in the paper:

In the paper, they approximate the G's z_1 and z_2 for two real images using Adam (reversing G(z)), then they interpolate between those z_1, z_2 and show G(z_{interp}).

In this implementation, the discriminator's latent codes h_1, h_2 are calculated from the two images, and the discriminator output for the interpolation is shown.

Is it possible to use BEGAN for predicted "B" image based on input "A" image? I know this is generally possible in GAN but in BEGAN implementations look like there is no implicit A-B training and prediction.

method numpy.prod returns a numpy.int64 value, which will cause an ValueError in python3. Refer to the following code block in layers.py$fully_connected()

if not isinstance(num_outputs, six.integer_types): raise ValueError('num_outputs should be int or long, got %s.', num_outputs)

and int() method converts numpy.int64 to be compatible with six.integer_types

This is compatible with python 2

I wanted to pause the training with my dataset

as mentioned on this issues :

I tried to keep the same params python main.py --dataset=mydata

python main.py --dataset=mydata --load_path=mydata_[08...]

It still begin from scratch and erase the old checkpoint.

Since the training is quite long I wonder how I could pause it

Thanks for the awesome repo. Will it be possible for you to release the pre-trained weights for CelebA on 64X64 and 128X128 resolution ?

Thanks, Avisek Lahiri Indian Institute of Technology Kharagpur

I tried to run the test model which I had trained before, but I failed. The results all seemed noise after 'generate()' function. I had check the input z_r, they were numbers between -1~1. (It's correct.) I also print the tensors and variables from checkpoint files, they had numbers inside. Is there any one can help me? thanks!!

i want to use the Wasserstein loss on caffe,as we know BEGAN use tf.train.AdamOptimizer() to min the loss , however i want to to use the Wasserstein loss that 'd_loss =|Y-D(Y)| - |D(G(z)-G(z))|' on the caffe. how to compute the Gradient