Shapes3D dataset is not included in the data download script

[alexrakowski]

Also the .npz file listed in

disentanglement_lib/disentanglement_lib/data/ground_truth/shapes3d.py

Line 28 in a070670

SHAPES3D_PATH = os.path.join(

is not to be found anywhere on the web.

[besterma]

Hi,
You can find the Shapes3D dataset on Github linking to here. Save it in your DISENTANGLEMENT_LIB_DATA folder in folder 3dshapes.

To use it in disentanglement_lib, change

disentanglement_lib/disentanglement_lib/data/ground_truth/shapes3d.py

Lines 28 to 32 in a070670

	SHAPES3D_PATH = os.path.join(
	os.environ.get("DISENTANGLEMENT_LIB_DATA", "."), "3dshapes",
	"look-at-object-room_floor-hueXwall-hueXobj-"
	"hueXobj-sizeXobj-shapeXview-azi.npz"
	)

to

SHAPES3D_PATH = os.path.join(
    os.environ.get("DISENTANGLEMENT_LIB_DATA", "."), "3dshapes", "3dshapes.h5"
)

and

disentanglement_lib/disentanglement_lib/data/ground_truth/shapes3d.py

Lines 51 to 56 in a070670

	with tf.gfile.GFile(SHAPES3D_PATH, "rb") as f:
	# Data was saved originally using python2, so we need to set the encoding.
	data = np.load(f, encoding="latin1")
	images = data["images"]
	labels = data["labels"]
	n_samples = np.prod(images.shape[0:6])

to

with h5py.File(SHAPES3D_PATH, 'r') as dataset:
      images = dataset['images'][()]
      labels = dataset['labels'][()]
n_samples = images.shape[0]

Hope that helps.

[alexrakowski]

It helps, I have found the .h5 file myself, but it would be convenient if the appropriate changes to the code were included in the repository.
I am using this as a package and would like to avoid editing the code (or creating a fork just for this purpose).

[neeland]

Convert h5py to npz using the following in a python file:

import h5py
import numpy as np


hf = h5py.File('./3dshapes.h5', 'r')
print(hf)
values_to_save = {"images": hf['images'], "labels": hf['labels']}
np.savez('look-at-object-room_floor-hueXwall-hueXobj-hueXobj-sizeXobj-shapeXview-azi.nzp', **values_to_save)

[noahcao]

@neonkitchen I think you should resize the saved image array. Because the default implementation of Shapes3D dataset requires to reshape the loaded image when using.

[Adamdad]

`import h5py
import numpy as np

hf = h5py.File('./3dshapes.h5', 'r')
print(hf)
values_to_save = {"images": hf['images'].reshape([10,10,10,8,4,15,64,64,3]), "labels": hf['labels']}
np.savez('look-at-object-room_floor-hueXwall-hueXobj-hueXobj-sizeXobj-shapeXview-azi.nzp', **values_to_save)
`