stitchfix / fauxtograph Goto Github PK

I don't know where this error is coming from. I have a tar.gz file here of a few images that I was hoping to train with your program: https://www.dropbox.com/s/zfekwp1cspfj9r2/images.tar.gz?dl=0

$ fauxtograph train ./images ./models/model_name
Loading Image Files...
Image Files Loaded!

 Training for 200 epochs.

epoch: 1
  0%|                                          | 0/1 [00:00<?, ?it/s]Traceback (most recent call last):
  File "/Users/jonathan/anaconda/bin/fauxtograph", line 11, in <module>
    sys.exit(fauxtograph())
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/click/core.py", line 700, in __call__
    return self.main(*args, **kwargs)
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/click/core.py", line 680, in main
    rv = self.invoke(ctx)
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/click/core.py", line 1027, in invoke
    return _process_result(sub_ctx.command.invoke(sub_ctx))
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/click/core.py", line 873, in invoke
    return ctx.invoke(self.callback, **ctx.params)
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/click/core.py", line 508, in invoke
    return callback(*args, **kwargs)
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/fauxtograph/fauxto.py", line 116, in train
    iae.fit(batch_size=batch, n_epochs=epoch)
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/fauxtograph/fauxtograph.py", line 269, in fit
    r_loss, kl_div, out = self._forward(x_batch)
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/fauxtograph/fauxtograph.py", line 176, in _forward
    encoded = self._encode(batch)
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/fauxtograph/fauxtograph.py", line 162, in _encode
    batch = F.relu(getattr(self.model, 'encode_%i' % i)(batch))
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/chainer/function.py", line 164, in __call__
    self._check_data_type_forward(in_data)
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/chainer/function.py", line 191, in _check_data_type_forward
    self.check_type_forward(in_type)
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/chainer/functions/linear.py", line 105, in check_type_forward
    type_check.Variable(self.W.shape[1], 'W.shape[1]')),
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/chainer/utils/type_check.py", line 457, in expect
    expr.expect()
  File "/Users/jonathan/anaconda/lib/python2.7/site-packages/chainer/utils/type_check.py", line 428, in expect
    '{0} {1} {2}'.format(left, self.inv, right))
chainer.utils.type_check.InvalidType: Expect: prod(in_types[0].shape[1:]) == W.shape[1]
Actual: 30000 != 22500```

While training on my own images, I'm getting this error,

chainer.utils.type_check.InvalidType:
Invalid operation is performed in: LinearFunction (Forward)

Expect: prod(in_types[0].shape[1:]) == in_types[1].shape[1]
Actual: 259656 != 27648

VAE and VAEGAN code is currently using mean squared error as the reconstruction loss function. In most papers / implementations, I'm more used to seeing binary cross entropy with numbers reported in nats.

Curious what we think would be best here. I did do a quick look for this in the chainer docs but didn't see binary cross entropy listed as one of the built in loss functions.

Followed steps outline in the blog. Got an error during training:

fauxtograph train images/ models/model_out
Loading Image Files...
Image Files Loaded!

Training for 200 epochs.

epoch: 1
0%| | 0/12 [00:00<?, ?it/s]Traceback (most recent call last):
File "/usr/local/bin/fauxtograph", line 8, in
load_entry_point('fauxtograph==0.1.8', 'console_scripts', 'fauxtograph')()
File "/Library/Python/2.7/site-packages/click/core.py", line 716, in call
return self.main(_args, *_kwargs)
File "/Library/Python/2.7/site-packages/click/core.py", line 696, in main
rv = self.invoke(ctx)
File "/Library/Python/2.7/site-packages/click/core.py", line 1060, in invoke
return _process_result(sub_ctx.command.invoke(sub_ctx))
File "/Library/Python/2.7/site-packages/click/core.py", line 889, in invoke
return ctx.invoke(self.callback, *_ctx.params)
File "/Library/Python/2.7/site-packages/click/core.py", line 534, in invoke
return callback(_args, **kwargs)
File "/Library/Python/2.7/site-packages/fauxtograph/fauxto.py", line 116, in train
iae.fit(batch_size=batch, n_epochs=epoch)
File "/Library/Python/2.7/site-packages/fauxtograph/fauxtograph.py", line 271, in fit
loss.backward()
File "/Library/Python/2.7/site-packages/chainer/variable.py", line 171, in backward
self.grad = numpy.ones_like(self.data)
File "/Library/Python/2.7/site-packages/chainer/variable.py", line 110, in grad
% (type(self.data), type(g), error_msg))
TypeError: Type of data and grad mismatch: <type 'numpy.ndarray'> != <type 'numpy.float32'>
This error is occured in two cases. The first case is when the user manually
sets the Variable.grad incorrectly. The second case is when some Function
implementation has a bug.

Looks like the BeautifulSoup dependency does not work on py3. Is there a reason for this specific BS version as opposed to bs4?

pip install fauxtograph
Collecting fauxtograph
  Downloading fauxtograph-1.0.3.tar.gz
Collecting chainer==1.6.0 (from fauxtograph)
  Downloading chainer-1.6.0.tar.gz (904kB)
    100% |████████████████████████████████| 905kB 273kB/s
Requirement already satisfied (use --upgrade to upgrade): pillow in ./anaconda/envs/py34/lib/python3.4/site-packages (from fauxtograph)
Collecting joblib (from fauxtograph)
  Downloading joblib-0.9.4-py2.py3-none-any.whl (112kB)
    100% |████████████████████████████████| 114kB 1.7MB/s
Collecting tqdm (from fauxtograph)
  Downloading tqdm-3.8.0-py2.py3-none-any.whl
Collecting BeautifulSoup (from fauxtograph)
  Downloading BeautifulSoup-3.2.1.tar.gz
    Complete output from command python setup.py egg_info:
    Traceback (most recent call last):
      File "<string>", line 20, in <module>
      File "/private/var/folders/b1/jbv4n2f56mz8m3hxkdtc51cr0000gn/T/pip-build-9v2hhigl/BeautifulSoup/setup.py", line 22
        print "Unit tests have failed!"
                                      ^
    SyntaxError: Missing parentheses in call to 'print'

I'm on ubuntu
pip install fauxtograph appears to work fine, however
fauxtograph download images/
says fauxtograph is undefined?
I tried git cloning the code, then cding in and executing this
python fauxtograph.py download ./images
it pauses, but does nothing?

Saw your Deep Style presentation and the generated images based on trained model on Hubble dataset. I was wondering what parameter set you used to train over Hubble dataset. I used the default parameter set and only got black images generated at the end! Also played with image_multiplier argument and got a bunch of similar black images with a gray (lighter) circle in the center.
Any tips here?

I´m testing out the --shape param, attempting to train/generate higher resolution images.
Seems like it does not like square formats, unfortunately resulting in black images.
Getting a bunch of these errors to: "RuntimeWarning: overflow encountered in exp self.y = 1 / (1 + numpy.exp(-x[0]))"

Any tips here?

Hi

Can you possibly add the script to train/generate face images with VAE/GAN?
As you suggested, the hyper-parameters seem to be critical. Thanks!

Hi, @jwhitmire @billeisenhauer @bjallen

When I tried the .ipynb on my dataset. I got the following error:

`
nvcc warning : The 'compute_20', 'sm_20', and 'sm_21' architectures are deprecated, and may be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).
nvcc fatal : Could not open output file '/tmp/tmpxft_00004d81_00000000'

Traceback (most recent call last):
File "see.py", line 17, in
vg.fit(x_all, save_freq=2, pic_freq=30, n_epochs=4, model_path = m_path, img_path=im_path) #, mirroring=True)#, flag_gpu=False)
File "/usr/local/lib/python2.7/dist-packages/fauxtograph/fauxtograph.py", line 1342, in fit
kl_loss, dif_l, disc_rec, disc_batch, disc_samp = self._forward(x_batch)
File "/usr/local/lib/python2.7/dist-packages/fauxtograph/fauxtograph.py", line 1099, in _forward
encoded, means, ln_vars = self._encode(batch, test=test)
File "/usr/local/lib/python2.7/dist-packages/fauxtograph/fauxtograph.py", line 1082, in _encode
x = self.enc(data, test=test)
File "/usr/local/lib/python2.7/dist-packages/fauxtograph/vaegan.py", line 127, in call
batch = self.bn1(batch, test=test)
File "/usr/local/lib/python2.7/dist-packages/chainer/links/normalization/batch_normalization.py", line 93, in call
ret = func(x, self.gamma, self.beta)
File "/usr/local/lib/python2.7/dist-packages/chainer/function.py", line 105, in call
outputs = self.forward(in_data)
File "/usr/local/lib/python2.7/dist-packages/chainer/functions/normalization/batch_normalization.py", line 53, in forward
mean = x.mean(axis=axis)
File "cupy/core/core.pyx", line 679, in cupy.core.core.ndarray.mean (cupy/core/core.cpp:13034)
File "cupy/core/core.pyx", line 687, in cupy.core.core.ndarray.mean (cupy/core/core.cpp:12899)
File "cupy/core/reduction.pxi", line 252, in cupy.core.core.simple_reduction_function.call (cupy/core/core.cpp:44135)
File "cupy/util.pyx", line 36, in cupy.util.memoize.decorator.ret (cupy/util.cpp:1194)
File "cupy/core/reduction.pxi", line 181, in cupy.core.core._get_simple_reduction_function (cupy/core/core.cpp:42698)
File "cupy/core/reduction.pxi", line 108, in cupy.core.core._get_simple_reduction_kernel (cupy/core/core.cpp:40479)
File "cupy/core/carray.pxi", line 87, in cupy.core.core.compile_with_cache (cupy/core/core.cpp:26123)
File "/usr/local/lib/python2.7/dist-packages/cupy/cuda/compiler.py", line 112, in compile_with_cache
base = _empty_file_preprocess_cache[env] = preprocess('', options)
File "/usr/local/lib/python2.7/dist-packages/cupy/cuda/compiler.py", line 75, in preprocess
pp_src = _run_nvcc(cmd, root_dir)
File "/usr/local/lib/python2.7/dist-packages/cupy/cuda/compiler.py", line 37, in _run_nvcc
return subprocess.check_output(cmd, cwd=cwd)
File "/usr/lib/python2.7/subprocess.py", line 573, in check_output
raise CalledProcessError(retcode, cmd, output=output)
subprocess.CalledProcessError: Command '['nvcc', '--preprocess', '/var/tmp/tmphEoIKW/kern.cu']' returned non-zero exit status 1

`

Please help..

Hi,
I have a question confused about the latent variable z

According to the code in class VAEGAN,

    def _forward(self, batch, test=False):

        encoded, means, ln_vars = self._encode(batch, test=test)
        rec = self._decode(encoded, test=test)
        normer = reduce(lambda x, y: x*y, means.data.shape)
        kl_loss = F.gaussian_kl_divergence(means, ln_vars)/normer

        samp_p = np.random.standard_normal(means.data.shape).astype('float32')
        z_p = chainer.Variable(samp_p)

It means that this z is first sampled according to VAE from:

encoded, means, ln_vars = self._encode(batch, test=test)

    def encode(self, data, test=False):
        x = self.enc(data, test=test)
        mean, ln_var = F.split_axis(x, 2, 1)
        samp = np.random.standard_normal(mean.data.shape).astype('float32')
        samp = Variable(samp)
        if self.flag_gpu:
            samp.to_gpu()
        z = samp * F.exp(0.5*ln_var) + mean

which means that the sampler z is s.t:

then, the sampled z is input to the decoder network to generate images, according to

rec = self._decode(encoded, test=test)

then, another z is sampled by:

        samp_p = np.random.standard_normal(means.data.shape).astype('float32')
        z_p = chainer.Variable(samp_p)

which means it is s.t.

and the generated images is created by:

rec_p = self._decode(z_p)

My question is since the two types of z are sampled from different distributions,
why they are input into the same decoder network?
And which output are the generated images from decoder, since they both created by sampling z according to a distribution ?

This is just my opinion.
Please feel free if I make mistakes.
Thank you.

Running on Mac for Hubble dataset:
$ fauxtograph train ./images ./models/model_name
Traceback (most recent call last):
File "/usr/local/bin/fauxtograph", line 9, in
load_entry_point('fauxtograph==1.0.2', 'console_scripts', 'fauxtograph')()
File "/Library/Python/2.7/site-packages/click/core.py", line 716, in call
return self.main(_args, *_kwargs)
File "/Library/Python/2.7/site-packages/click/core.py", line 696, in main
rv = self.invoke(ctx)
File "/Library/Python/2.7/site-packages/click/core.py", line 1060, in invoke
return _process_result(sub_ctx.command.invoke(sub_ctx))
File "/Library/Python/2.7/site-packages/click/core.py", line 889, in invoke
return ctx.invoke(self.callback, *_ctx.params)
File "/Library/Python/2.7/site-packages/click/core.py", line 534, in invoke
return callback(_args, **kwargs)
File "/Library/Python/2.7/site-packages/fauxtograph/fauxto.py", line 122, in train
x_all = vae.load_images(file_paths)
File "/Library/Python/2.7/site-packages/fauxtograph/fauxtograph.py", line 239, in load_images
x_all = np.array([read(fname) for fname in tqdm.tqdm(filepaths)])
ValueError: could not broadcast input array from shape (300,300,3) into shape (300)

The call to self.save in line #347 of fauxtograph.py always sends save_meta as False.

It appears that you should increase save_counter only when epoch % save_freq == 0.

The same issue is present in GAN and VAEGAN.