def get_named_ground_truth_data(name):
"""Returns ground truth data set based on name.
Args:
name: String with the name of the dataset.
Raises:
ValueError: if an invalid data set name is provided.
"""
if name == "dsprites_full":
return dsprites.DSprites([1, 2, 3, 4, 5])
elif name == "dsprites_noshape":
return dsprites.DSprites([2, 3, 4, 5])
elif name == "color_dsprites":
return dsprites.ColorDSprites([1, 2, 3, 4, 5])
elif name == "noisy_dsprites":
return dsprites.NoisyDSprites([1, 2, 3, 4, 5])
elif name == "scream_dsprites":
return dsprites.ScreamDSprites([1, 2, 3, 4, 5])
elif name == "smallnorb":
return norb.SmallNORB()
elif name == "cars3d":
return cars3d.Cars3D()
elif name == "mpi3d_toy":
return mpi3d.MPI3D(mode="mpi3d_toy")
elif name == "mpi3d_realistic":
return mpi3d.MPI3D(mode="mpi3d_realistic")
elif name == "mpi3d_real":
return mpi3d.MPI3D(mode="mpi3d_real")
elif name == "shapes3d":
return shapes3d.Shapes3D()
elif name == "dummy_data":
return dummy_data.DummyData()
elif name == "modelnet":
return DisLibGroundTruthData(**MODELNET_PARAMS)
elif name == "arrow":
return DisLibGroundTruthData(**ARROW_PARAMS)
elif name == "pixel4":
return DisLibGroundTruthData(**WRAPPED_PIXEL4_PARAMS)
elif name == "pixel8":
return DisLibGroundTruthData(**WRAPPED_PIXEL8_PARAMS)
else:
raise ValueError("Invalid data set name.")
def get_named_ground_truth_data(name):
"""Returns ground truth data set based on name.
Args:
name: String with the name of the dataset.
Raises:
ValueError: if an invalid data set name is provided.
"""
if name == "threeDotsCache":
# a large random sample from ThreeDots
return threeDots.ThreeDotsTrainingCache()
elif name == "threeDots":
return threeDots.ThreeDots()
elif name == "dsprites_full":
return dsprites.DSprites([1, 2, 3, 4, 5])
elif name == "dsprites_noshape":
return dsprites.DSprites([2, 3, 4, 5])
elif name == "color_dsprites":
return dsprites.ColorDSprites([1, 2, 3, 4, 5])
elif name == "noisy_dsprites":
return dsprites.NoisyDSprites([1, 2, 3, 4, 5])
elif name == "scream_dsprites":
return dsprites.ScreamDSprites([1, 2, 3, 4, 5])
elif name == "smallnorb":
return norb.SmallNORB()
elif name == "cars3d":
return cars3d.Cars3D()
elif name == "mpi3d_toy":
return mpi3d.MPI3D(mode="mpi3d_toy")
elif name == "mpi3d_realistic":
return mpi3d.MPI3D(mode="mpi3d_realistic")
elif name == "mpi3d_real":
return mpi3d.MPI3D(mode="mpi3d_real")
elif name == "mpi3d_multi_real":
return mpi3d_multi.MPI3DMulti(mode="mpi3d_real")
elif name == "shapes3d":
return shapes3d.Shapes3D()
elif name == "dummy_data":
return dummy_data.DummyData()
else:
raise ValueError("Invalid data set name.")
def get_named_ground_truth_data(name):
"""Returns ground truth data set based on name.
Args:
name: String with the name of the dataset.
Raises:
ValueError: if an invalid data set name is provided.
"""
if name == "dsprites_full":
return dsprites.DSprites([1, 2, 3, 4, 5])
elif name == "dsprites_noshape":
return dsprites.DSprites([2, 3, 4, 5])
elif name == "color_dsprites":
return dsprites.ColorDSprites([1, 2, 3, 4, 5])
elif name == "noisy_dsprites":
return dsprites.NoisyDSprites([1, 2, 3, 4, 5])
elif name == "scream_dsprites":
return dsprites.ScreamDSprites([1, 2, 3, 4, 5])
elif name == "smallnorb":
return norb.SmallNORB()
elif name == "cars3d":
return cars3d.Cars3D()
elif name == "mpi3d_toy":
return mpi3d.MPI3D(mode="mpi3d_toy")
elif name == "mpi3d_realistic":
return mpi3d.MPI3D(mode="mpi3d_realistic")
elif name == "mpi3d_real":
return mpi3d.MPI3D(mode="mpi3d_real")
elif name == "3dshapes":
return shapes3d.Shapes3D()
elif name == "3dshapes_holdout" or name == "3dshapes_pca_holdout_s5000":
return shapes3d_partial.Shapes3DPartial(name)
elif name == "3dshapes_model_all":
return shapes3d_partial.Shapes3DPartial(name), None
elif name[:8] == "3dshapes":
return shapes3d_partial.Shapes3DPartial(
name + '_train'), shapes3d_partial.Shapes3DPartial(name + '_valid')
elif name == "dummy_data":
return dummy_data.DummyData()
else:
raise ValueError("Invalid data set name.")
def get_dlib_data(task):
ut.log("Loading {}".format(task))
if task == "dsprites":
# 5 factors
return dsprites.DSprites(list(range(1, 6)))
elif task == "shapes3d":
# 6 factors
return shapes3d.Shapes3D()
elif task == "norb":
# 4 factors + 1 nuisance (which we'll handle via n_dim=2)
return norb.SmallNORB()
elif task == "cars3d":
# 3 factors
return cars3d.Cars3D()
elif task == "mpi3d":
# 7 factors
return mpi3d.MPI3D()
elif task == "scream":
# 5 factors + 2 nuisance (handled as n_dim=2)
return dsprites.ScreamDSprites(list(range(1, 6)))
from disentanglement_lib.data.ground_truth import mpi3d
from disentanglement_lib.evaluation.metrics import dci, factor_vae, sap_score, mig, irs, utils
def representation_fn(x):
main_net.eval()
with torch.no_grad():
images = preprocess(x).cuda()
transformed_codes = main_net(images, 1, 1, encoding_only=True)
comp_loss, z_factors = compress_net(transformed_codes)
repr_code = torch.cat([z for k, z in z_factors.items()],
1).cpu().numpy()
return repr_code
ground_truth_data = mpi3d.MPI3D()
print('representation_fn')
scores = factor_vae.compute_factor_vae(ground_truth_data, representation_fn,
np.random.RandomState(0), 64, 10000,
5000, 10000)
print(' factor_vae: %.6f' % scores['eval_accuracy'])
scores = dci.compute_dci(ground_truth_data, representation_fn,
np.random.RandomState(0), 10000, 5000)
print(' dci: %.6f' % scores['disentanglement'])
scores = sap_score.compute_sap(ground_truth_data,
representation_fn,
np.random.RandomState(0),
10000,