def get_image_features(self, task, batched_lmap_speed_thresh=None):
if batched_lmap_speed_thresh is None:
batched_lmap_speed_thresh = self.batched_lmap_speed_thresh
images = task.images
try:
rval, _images, cdict = self.image_features[images]
# -- currently it is true that all tasks should be indexing into
# -- the same set of images. Later when this is not the case,
# -- delete this over-strict check.
assert _images is images
except KeyError:
feature_lmap = self.get_image_features_lmap(task.images,
batched_lmap_speed_thresh)
rval = cache_memmap(
feature_lmap,
self.memmap_name + '_image_features_' + task.name,
del_atexit=self.memmap_del_atexit)
foobar.append_ndarray_signature(rval[0],
'get_image_features features 0', task.name)
foobar.append_ndarray_signature(rval[100],
'get_image_features features 100', task.name)
cdict = {}
self.image_features[images] = rval, images, cdict
return rval, cdict
def alloc_random_uniform_filterbank(n_filters, height, width,
channels, dtype, rseed, normalize=True):
"""
Generate the same weights as are generated by pythor3
"""
if height != width:
raise ValueError('filters must be square')
if channels is None:
filter_shape = [n_filters, height, width]
else:
filter_shape = [n_filters, height, width, channels]
rng = np.random.RandomState(rseed)
foobar.append_randomstate('alloc_random_uniform_filterbank', rng)
fb_data = rng.uniform(size=filter_shape)
# normalize each filter in the bank if needed
if normalize:
# TODO: vectorize these computations, do all at once.
for fidx, filt in enumerate(fb_data):
# normalization here means zero-mean, unit-L2norm
filt -= filt.mean()
filt_norm = np.sqrt((filt * filt).sum())
assert filt_norm != 0
filt /= filt_norm
fb_data[fidx] = filt
foobar.append_ndarray_signature(fb_data, 'alloc_random_uniform_filterbank')
return fb_data.astype(dtype)
def fb_whitened_projections(patches, pwfX, n_filters, rseed, dtype):
"""
pwfX is the output of patch_whitening_filterbank_X with reshape=False
M, and fb will be reshaped to match elements of patches
"""
M, P, patches_cn = pwfX
if patches_cn.ndim != 2:
raise TypeError('wrong shape for pwfX args, should be flattened',
patches_cn.shape)
rng = np.random.RandomState(rseed)
foobar.append_randomstate('fb_whitened_projections', rng)
D = rng.randn(n_filters, patches_cn.shape[1])
D = D / (np.sqrt((D**2).sum(axis=1))[:, None] + 1e-20)
fb = dot_f32(D, P)
fb.shape = (n_filters, ) + patches.shape[1:]
M.shape = patches.shape[1:]
M = M.astype(dtype)
fb = fb.astype(dtype)
if fb.size == 0:
raise ValueError('filterbank had size 0')
foobar.append_ndarray_signature(M, 'fb_whitened_projections M')
foobar.append_ndarray_signature(fb, 'fb_whitened_projections fb')
return M, fb
def alloc_random_uniform_filterbank(n_filters,
height,
width,
channels,
dtype,
rseed,
normalize=True):
"""
Generate the same weights as are generated by pythor3
"""
if height != width:
raise ValueError('filters must be square')
if channels is None:
filter_shape = [n_filters, height, width]
else:
filter_shape = [n_filters, height, width, channels]
rng = np.random.RandomState(rseed)
foobar.append_randomstate('alloc_random_uniform_filterbank', rng)
fb_data = rng.uniform(size=filter_shape)
# normalize each filter in the bank if needed
if normalize:
# TODO: vectorize these computations, do all at once.
for fidx, filt in enumerate(fb_data):
# normalization here means zero-mean, unit-L2norm
filt -= filt.mean()
filt_norm = np.sqrt((filt * filt).sum())
assert filt_norm != 0
filt /= filt_norm
fb_data[fidx] = filt
foobar.append_ndarray_signature(fb_data, 'alloc_random_uniform_filterbank')
return fb_data.astype(dtype)
def get_image_features(self, task, batched_lmap_speed_thresh=None):
if batched_lmap_speed_thresh is None:
batched_lmap_speed_thresh = self.batched_lmap_speed_thresh
images = task.images
try:
rval, _images, cdict = self.image_features[images]
# -- currently it is true that all tasks should be indexing into
# -- the same set of images. Later when this is not the case,
# -- delete this over-strict check.
assert _images is images
except KeyError:
feature_lmap = self.get_image_features_lmap(
task.images, batched_lmap_speed_thresh)
rval = cache_memmap(feature_lmap,
self.memmap_name + '_image_features_' +
task.name,
del_atexit=self.memmap_del_atexit)
foobar.append_ndarray_signature(rval[0],
'get_image_features features 0',
task.name)
foobar.append_ndarray_signature(rval[100],
'get_image_features features 100',
task.name)
cdict = {}
self.image_features[images] = rval, images, cdict
return rval, cdict
def fb_whitened_projections(patches, pwfX, n_filters, rseed, dtype):
"""
pwfX is the output of patch_whitening_filterbank_X with reshape=False
M, and fb will be reshaped to match elements of patches
"""
M, P, patches_cn = pwfX
if patches_cn.ndim != 2:
raise TypeError('wrong shape for pwfX args, should be flattened',
patches_cn.shape)
rng = np.random.RandomState(rseed)
foobar.append_randomstate('fb_whitened_projections', rng)
D = rng.randn(n_filters, patches_cn.shape[1])
D = D / (np.sqrt((D ** 2).sum(axis=1))[:, None] + 1e-20)
fb = dot_f32(D, P)
fb.shape = (n_filters,) + patches.shape[1:]
M.shape = patches.shape[1:]
M = M.astype(dtype)
fb = fb.astype(dtype)
if fb.size == 0:
raise ValueError('filterbank had size 0')
foobar.append_ndarray_signature(M, 'fb_whitened_projections M')
foobar.append_ndarray_signature(fb, 'fb_whitened_projections fb')
return M, fb
def load_ensemble_grams(self, norm_sample, ens, sample1, sample2):
trial_attachments = self.ctrl.trials.trial_attachments
# -- load the gram matrices saved by each ensemble member
for trial in self.history:
trial_norm_key = self.norm_key(norm_sample, tid=trial['tid'])
info('Loading grams from document %i' % trial['tid'])
debug(' .. saved_grams: %s' %
str(trial['result']['grams'][trial_norm_key]))
for (s1, s2) in trial['result']['grams'][trial_norm_key]:
if set([sample1, sample2]) == set([s1, s2]):
if not ens.has_gram(trial_norm_key, s1, s2):
att_key = 'gram_%s_%s_%s.pkl' % (trial_norm_key, s1, s2)
info('retrieving gram_data %i:%s'
% (trial['tid'], att_key))
try:
gram_data = cached_gram_load(trial['tid'], att_key)
except IOError:
gram_data = trial_attachments(trial)[att_key]
cached_gram_save(trial['tid'], att_key, gram_data)
info('retrieved %i bytes' % len(gram_data))
gram = loads_gram(gram_data)
if s1 == sample1:
ens.add_gram(trial_norm_key, sample1, sample2, gram)
else:
ens.add_gram(trial_norm_key, sample1, sample2,
gram.T)
foobar.append_ndarray_signature(
gram,
'load gram', trial_norm_key, sample1, sample2)
info('Loading grams done')
def unsup_images(data_view, trn, N):
"""
Return a block of
"""
if trn == 'DevTrain':
# -- extract training images, and put them into channel-major format
imgs = larray.reindex(data_view.image_pixels,
data_view.dev_train['lpathidx'][0, :N])[:]
imgs = np.asarray(imgs)
assert 'int' in str(imgs.dtype)
foobar.append_ndarray_signature(imgs, 'unsup_images')
foobar.append_trace('unsup_images N', N)
return imgs.transpose(0, 3, 1, 2).copy()
else:
raise NotImplementedError()
def patch_whitening_filterbank_X(patches, o_ndim, gamma,
remove_mean, beta, hard_beta,
):
"""
patches - Image patches (can be uint8 pixels or floats)
o_ndim - 2 to get matrix outputs, 4 to get image-stack outputs
gamma - non-negative real to boost low-principle components
remove_mean - see contrast_normalize
beta - see contrast_normalize
hard_beta - see contrast_normalize
Returns: M, P, X
M - mean of contrast-normalized patches
P - whitening matrix / filterbank for contrast-normalized patches
X - contrast-normalized patches
"""
# Algorithm from Coates' sc_vq_demo.m
# -- patches -> column vectors
X = patches.reshape(len(patches), -1).astype('float64')
X = contrast_normalize(X,
remove_mean=remove_mean,
beta=beta,
hard_beta=hard_beta)
# -- ZCA whitening (with low-pass)
logger.debug('patch_whitening_filterbank_X starting ZCA')
M, _std = mean_and_std(X)
Xm = X - M
assert Xm.shape == X.shape
logger.info('patch_whitening_filterbank_X starting ZCA: dot %s' %
str(Xm.shape))
C = dot_f64(Xm.T, Xm) / (Xm.shape[0] - 1)
logger.debug('patch_whitening_filterbank_X starting ZCA: eigh')
D, V = np.linalg.eigh(C)
logger.debug(
'patch_whitening_filterbank_X starting ZCA: dot %s' % str(V.shape))
P = dot_f32(np.sqrt(1.0 / (D + gamma)) * V, V.T)
# -- return to image space
if o_ndim == 4:
M = M.reshape(patches.shape[1:])
P = P.reshape((P.shape[0],) + patches.shape[1:])
X = X.reshape((len(X),) + patches.shape[1:])
elif o_ndim == 2:
pass
else:
raise ValueError('o_ndim not in (2, 4)', o_ndim)
logger.debug('patch_whitening_filterbank_X -> done')
foobar.append_ndarray_signature(M, 'patch_whitening_filterbank_X M')
foobar.append_ndarray_signature(P, 'patch_whitening_filterbank_X P')
foobar.append_ndarray_signature(X, 'patch_whitening_filterbank_X X')
dtype = patches.dtype
return M.astype(dtype), P.astype(dtype), X.astype(dtype)
def contrast_normalize(patches, remove_mean, beta, hard_beta):
X = patches
if X.ndim != 2:
raise TypeError('contrast_normalize requires flat patches')
if remove_mean:
xm = X.mean(1)
else:
xm = X[:, 0] * 0
Xc = X - xm[:, None]
if 0:
# -- for some reason the following sometimes uses gigs of RAM
l2 = (Xc * Xc).sum(axis=1)
else:
l2 = np.zeros_like(Xc[:, 0])
for i in xrange(Xc.shape[1]):
l2 += Xc[:, i]**2
if hard_beta:
div2 = np.maximum(l2, beta)
else:
div2 = l2 + beta
Xc /= np.sqrt(div2[:, None])
foobar.append_ndarray_signature(Xc, 'contrast_normalize')
return Xc
def fb_whitened_patches(patches, pwfX, n_filters, rseed, dtype):
"""
pwfX is the output of patch_whitening_filterbank_X with reshape=False
M, and fb will be reshaped to match elements of patches
"""
M, P, patches_cn = pwfX
rng = np.random.RandomState(rseed)
foobar.append_randomstate('fb_whitened_patches', rng)
d_elems = rng.randint(len(patches_cn), size=n_filters)
D = dot_f64(patches_cn[d_elems] - M, P)
D = D / (np.sqrt((D**2).sum(axis=1))[:, None] + 1e-20)
fb = dot_f32(D, P)
fb.shape = (n_filters, ) + patches.shape[1:]
M.shape = patches.shape[1:]
M = M.astype(dtype)
fb = fb.astype(dtype)
if fb.size == 0:
raise ValueError('filterbank had size 0')
foobar.append_ndarray_signature(M, 'fb_whitened_patches M')
foobar.append_ndarray_signature(fb, 'fb_whitened_patches fb')
return M, fb
def fb_whitened_patches(patches, pwfX, n_filters, rseed, dtype):
"""
pwfX is the output of patch_whitening_filterbank_X with reshape=False
M, and fb will be reshaped to match elements of patches
"""
M, P, patches_cn = pwfX
rng = np.random.RandomState(rseed)
foobar.append_randomstate('fb_whitened_patches', rng)
d_elems = rng.randint(len(patches_cn), size=n_filters)
D = dot_f64(patches_cn[d_elems] - M, P)
D = D / (np.sqrt((D ** 2).sum(axis=1))[:, None] + 1e-20)
fb = dot_f32(D, P)
fb.shape = (n_filters,) + patches.shape[1:]
M.shape = patches.shape[1:]
M = M.astype(dtype)
fb = fb.astype(dtype)
if fb.size == 0:
raise ValueError('filterbank had size 0')
foobar.append_ndarray_signature(M, 'fb_whitened_patches M')
foobar.append_ndarray_signature(fb, 'fb_whitened_patches fb')
return M, fb
def contrast_normalize(patches, remove_mean, beta, hard_beta):
X = patches
if X.ndim != 2:
raise TypeError('contrast_normalize requires flat patches')
if remove_mean:
xm = X.mean(1)
else:
xm = X[:,0] * 0
Xc = X - xm[:, None]
if 0:
# -- for some reason the following sometimes uses gigs of RAM
l2 = (Xc * Xc).sum(axis=1)
else:
l2 = np.zeros_like(Xc[:, 0])
for i in xrange(Xc.shape[1]):
l2 += Xc[:, i] ** 2
if hard_beta:
div2 = np.maximum(l2, beta)
else:
div2 = l2 + beta
Xc /= np.sqrt(div2[:, None])
foobar.append_ndarray_signature(Xc, 'contrast_normalize')
return Xc
def random_patches(images, N, R, C, rng, channel_major=False, memlimit=None):
"""Return a stack of N image patches (channel major version)"""
def N_with_memlimit():
if memlimit is not None:
# -- memlimit in bytes
sizelimit = memlimit / images.dtype.itemsize
return min(N, sizelimit // (R * C * iF))
else:
return N
if channel_major:
n_imgs, iF, iR, iC = images.shape
N = N_with_memlimit()
rval = np.empty((N, iF, R, C), dtype=images.dtype)
else:
n_imgs, iR, iC, iF = images.shape
N = N_with_memlimit()
rval = np.empty((N, R, C, iF), dtype=images.dtype)
foobar.append_trace('random_patches dims', *rval.shape)
foobar.append_randomstate('random_patches rng', rng)
srcs = rng.randint(n_imgs, size=N)
if R > iR or C > iC:
raise InvalidDescription('cannot extract patches', (R, C))
roffsets = rng.randint(iR - R + 1, size=N)
coffsets = rng.randint(iC - C + 1, size=N)
# TODO: this can be done with one advanced index right?
for rv_i, src_i, ro, co in zip(rval, srcs, roffsets, coffsets):
if channel_major:
rv_i[:] = images[src_i, :, ro: ro + R, co : co + C]
else:
rv_i[:] = images[src_i, ro: ro + R, co : co + C]
foobar.append_ndarray_signature(rval, 'random_patches rval')
return rval
def random_patches(images, N, R, C, rng, channel_major=False, memlimit=None):
"""Return a stack of N image patches (channel major version)"""
def N_with_memlimit():
if memlimit is not None:
# -- memlimit in bytes
sizelimit = memlimit / images.dtype.itemsize
return min(N, sizelimit // (R * C * iF))
else:
return N
if channel_major:
n_imgs, iF, iR, iC = images.shape
N = N_with_memlimit()
rval = np.empty((N, iF, R, C), dtype=images.dtype)
else:
n_imgs, iR, iC, iF = images.shape
N = N_with_memlimit()
rval = np.empty((N, R, C, iF), dtype=images.dtype)
foobar.append_trace('random_patches dims', *rval.shape)
foobar.append_randomstate('random_patches rng', rng)
srcs = rng.randint(n_imgs, size=N)
if R > iR or C > iC:
raise InvalidDescription('cannot extract patches', (R, C))
roffsets = rng.randint(iR - R + 1, size=N)
coffsets = rng.randint(iC - C + 1, size=N)
# TODO: this can be done with one advanced index right?
for rv_i, src_i, ro, co in zip(rval, srcs, roffsets, coffsets):
if channel_major:
rv_i[:] = images[src_i, :, ro:ro + R, co:co + C]
else:
rv_i[:] = images[src_i, ro:ro + R, co:co + C]
foobar.append_ndarray_signature(rval, 'random_patches rval')
return rval
def normalized_image_match_features(self, task, svm_dct, role,
batched_lmap_speed_thresh=None):
assert role in ('train', 'test')
if batched_lmap_speed_thresh is None:
batched_lmap_speed_thresh = self.batched_lmap_speed_thresh
image_features, cdict = self.get_image_features(task,
batched_lmap_speed_thresh=batched_lmap_speed_thresh)
del cdict # -- no longer used (waste of memory)
pipeline = self.pipeline
info('Indexing into image_features of shape %s' %
str(image_features.shape))
comps = [getattr(comparisons, cc)
for cc in self.comparison_names]
n_features = np.prod(image_features.shape[1:])
n_trn = len(task.lidx)
x_trn_shp = (n_trn, len(comps), n_features)
info('Allocating training ndarray of shape %s' % str(x_trn_shp))
x_trn = np.empty(x_trn_shp, dtype='float32')
# -- pre-compute all of the image_features we will need
all_l_features = reindex(image_features, task.lidx)[:]
all_r_features = reindex(image_features, task.ridx)[:]
all_l_features = all_l_features.reshape(len(all_l_features), -1)
all_r_features = all_r_features.reshape(len(all_r_features), -1)
foobar.append_ndarray_signature(all_l_features,
'normalized_image_match l_features', task.name)
foobar.append_ndarray_signature(all_r_features,
'normalized_image_match r_features', task.name)
if role == 'train':
if np.allclose(all_l_features.var(axis=0), 0.0):
raise ValueError(
'Homogeneous features (non-finite features)')
xmean_l, xstd_l = mean_and_std(all_l_features,
remove_std0=pipeline['remove_std0'])
xmean_r, xstd_r = mean_and_std(all_r_features,
remove_std0=pipeline['remove_std0'])
xmean = (xmean_l + xmean_r) / 2.0
# -- this is an ad-hoc way of blending the variances.
xstd = np.sqrt(np.maximum(xstd_l, xstd_r) ** 2
+ pipeline['varthresh'])
foobar.append_ndarray_signature(
xmean, 'normalized_image_match xmean', task.name)
foobar.append_ndarray_signature(
xstd, 'normalized_image_match xstd', task.name)
svm_dct['xmean'] = xmean
svm_dct['xstd'] = xstd
else:
xmean = svm_dct['xmean']
xstd = svm_dct['xstd']
info('Computing comparison features')
# -- now compute the "comparison functions" into x_trn
for jj, (lfeat, rfeat) in enumerate(
zip(all_l_features, all_r_features)):
lfeat_z = (lfeat - xmean) / xstd
rfeat_z = (rfeat - xmean) / xstd
for ci, comp in enumerate(comps):
x_trn[jj, ci, :] = comp(lfeat_z, rfeat_z)
if pipeline['divrowl2']:
info('Dividing by feature norms')
# -- now normalize by average feature norm because some
# comparison functions come out smaller than others
if role == 'train':
svm_dct['divrowl2_avg_nrm'] = {}
for ci, cname in enumerate(self.comparison_names):
avg_nrm = average_row_l2norm(x_trn[:, ci, :]) + 1e-7
svm_dct['divrowl2_avg_nrm'][cname] = avg_nrm
avg_nrm_vec = [svm_dct['divrowl2_avg_nrm'][cname]
for cname in self.comparison_names]
x_trn /= np.asarray(avg_nrm_vec)[None, :, None]
foobar.append_trace('get_normlized_features avg_nrm', avg_nrm_vec)
# -- collapse comparison and feature dimensions
x_trn.shape = (x_trn.shape[0], x_trn.shape[1] * x_trn.shape[2])
foobar.append_ndarray_signature(
x_trn, 'normalized_image_match x_trn', task.name)
info('normalized_image_match_features complete')
return x_trn
def train_image_match_indexed(self, task, valid=None):
pipeline = self.pipeline
info('training svm on %s' % task.name)
ens = EnsembleSVC(task.name)
norm_task = task.name
norm_key = self.norm_key(norm_task)
svm_dct = {
'ens': ens,
'norm_key': norm_key,
'norm_task': task.name,
'task_name': task.name,
}
ens.add_member(norm_key)
ens.add_sample(task.name, task.y)
x_trn = self.normalized_image_match_features(task, svm_dct,
role='train')
ens.add_features(norm_key, task.name, x_trn)
foobar.append_ndarray_signature(x_trn,
'train_image x_trn', norm_key, task.name)
info('computing gram: %s / %s / %s' % (
norm_key, task.name, task.name))
ens.compute_gram(norm_key, task.name, task.name, dtype='float32')
foobar.append_ndarray_signature(
ens._grams[(norm_key, task.name, task.name)],
'train_image train_gram', norm_key, task.name)
if valid is not None:
info('cross-validating svm on %s' % valid.name)
x_val = self.normalized_image_match_features(valid, svm_dct,
role='test',
# -- assume that slow features were caught earlier
batched_lmap_speed_thresh={'seconds': 30, 'elements': 1},
)
foobar.append_ndarray_signature(
x_val,
'train_image x_val', norm_key, valid.name, task.name)
ens.add_sample(valid.name, valid.y)
ens.add_features(norm_key, valid.name, x_val)
info('computing gram: %s / %s / %s' % (
norm_key, valid.name, task.name))
ens.compute_gram(norm_key, valid.name, task.name, dtype='float32')
foobar.append_ndarray_signature(
ens._grams[(norm_key, valid.name, task.name)],
'train_image valid_gram', norm_key, valid.name, task.name)
# -- re-fit the model using best weights on train + valid sets
info('computing gram: %s / %s / %s' % (
norm_key, valid.name, valid.name))
ens.compute_gram(norm_key, valid.name, valid.name, dtype='float32')
train_valid = '%s_%s' % (task.name, valid.name)
ens.add_compound_sample(train_valid, [task.name, valid.name])
def load_history():
info('loading history')
self.load_ensemble_history(
fields=['result.weights','result.grams'])
self.load_ensemble_weights(norm_task, task.name, ens)
self.load_ensemble_grams(norm_task, ens, task.name, task.name)
if valid is not None:
self.load_ensemble_grams(norm_task, ens, valid.name, task.name)
self.load_ensemble_grams(norm_task, ens, valid.name, valid.name)
def train_main():
ens.train_sample = task.name
t0 = time.time()
if valid is None:
svm_dct['l2_reg'] = pipeline['l2_reg']
ens.fit_svm(svm_dct['l2_reg'])
svm_dct['train_error'] = ens.error_rate(task.name)
svm_dct['loss'] = svm_dct['train_error']
else:
#scales = {m: 3.0 for m in ens._weights}
scales = dict([(m, 3.0) for m in ens._weights])
scales[norm_key] = 100.0
info('fit_weights_crossvalid(%s, %i)' % (
valid.name, self.svm_crossvalid_max_evals))
ens.fit_weights_crossvalid(valid.name,
max_evals=self.svm_crossvalid_max_evals,
scales=scales)
foobar.append_trace('xvalid weights', sorted(ens._weights.items()))
svm_dct['task_error'] = ens.error_rate(task.name)
foobar.append_trace('task_error', svm_dct['task_error'])
svm_dct['valid_name'] = valid.name
svm_dct['valid_error'] = ens.error_rate(valid.name)
info('valid_error %f' % svm_dct['valid_error'])
foobar.append_trace('valid_error', svm_dct['valid_error'])
svm_dct['l2_reg'] = None # -- use default when retraining
# -- re-fit the model using best weights on train + valid sets
ens.train_sample = train_valid
ens.fit_svm()
fit_time = time.time() - t0
svm_dct['fit_time'] = fit_time
info('training with just the current features...')
train_main()
svm_dct['task_error_no_ensemble'] = svm_dct['task_error']
svm_dct['valid_error_no_ensemble'] = svm_dct['valid_error']
load_history()
if self.history:
info('training the full ensemble...')
train_main()
try:
print_summary = ens.print_summary
except AttributeError:
print_summary = lambda : None
print_summary()
dct = self._results['train_image_match_indexed']
dct.setdefault(norm_key, {})
if task.name in dct[norm_key]:
warn('Overwriting train_image_match_indexed result: %s'
% task.name)
dct[norm_key][task.name] = svm_dct
return svm_dct
def normalized_image_match_features(self,
task,
svm_dct,
role,
batched_lmap_speed_thresh=None):
assert role in ('train', 'test')
if batched_lmap_speed_thresh is None:
batched_lmap_speed_thresh = self.batched_lmap_speed_thresh
image_features, cdict = self.get_image_features(
task, batched_lmap_speed_thresh=batched_lmap_speed_thresh)
del cdict # -- no longer used (waste of memory)
pipeline = self.pipeline
info('Indexing into image_features of shape %s' %
str(image_features.shape))
comps = [getattr(comparisons, cc) for cc in self.comparison_names]
n_features = np.prod(image_features.shape[1:])
n_trn = len(task.lidx)
x_trn_shp = (n_trn, len(comps), n_features)
info('Allocating training ndarray of shape %s' % str(x_trn_shp))
x_trn = np.empty(x_trn_shp, dtype='float32')
# -- pre-compute all of the image_features we will need
all_l_features = reindex(image_features, task.lidx)[:]
all_r_features = reindex(image_features, task.ridx)[:]
all_l_features = all_l_features.reshape(len(all_l_features), -1)
all_r_features = all_r_features.reshape(len(all_r_features), -1)
foobar.append_ndarray_signature(all_l_features,
'normalized_image_match l_features',
task.name)
foobar.append_ndarray_signature(all_r_features,
'normalized_image_match r_features',
task.name)
if role == 'train':
if np.allclose(all_l_features.var(axis=0), 0.0):
raise ValueError('Homogeneous features (non-finite features)')
xmean_l, xstd_l = mean_and_std(all_l_features,
remove_std0=pipeline['remove_std0'])
xmean_r, xstd_r = mean_and_std(all_r_features,
remove_std0=pipeline['remove_std0'])
xmean = (xmean_l + xmean_r) / 2.0
# -- this is an ad-hoc way of blending the variances.
xstd = np.sqrt(
np.maximum(xstd_l, xstd_r)**2 + pipeline['varthresh'])
foobar.append_ndarray_signature(xmean,
'normalized_image_match xmean',
task.name)
foobar.append_ndarray_signature(xstd,
'normalized_image_match xstd',
task.name)
svm_dct['xmean'] = xmean
svm_dct['xstd'] = xstd
else:
xmean = svm_dct['xmean']
xstd = svm_dct['xstd']
info('Computing comparison features')
# -- now compute the "comparison functions" into x_trn
for jj, (lfeat, rfeat) in enumerate(zip(all_l_features,
all_r_features)):
lfeat_z = (lfeat - xmean) / xstd
rfeat_z = (rfeat - xmean) / xstd
for ci, comp in enumerate(comps):
x_trn[jj, ci, :] = comp(lfeat_z, rfeat_z)
if pipeline['divrowl2']:
info('Dividing by feature norms')
# -- now normalize by average feature norm because some
# comparison functions come out smaller than others
if role == 'train':
svm_dct['divrowl2_avg_nrm'] = {}
for ci, cname in enumerate(self.comparison_names):
avg_nrm = average_row_l2norm(x_trn[:, ci, :]) + 1e-7
svm_dct['divrowl2_avg_nrm'][cname] = avg_nrm
avg_nrm_vec = [
svm_dct['divrowl2_avg_nrm'][cname]
for cname in self.comparison_names
]
x_trn /= np.asarray(avg_nrm_vec)[None, :, None]
foobar.append_trace('get_normlized_features avg_nrm', avg_nrm_vec)
# -- collapse comparison and feature dimensions
x_trn.shape = (x_trn.shape[0], x_trn.shape[1] * x_trn.shape[2])
foobar.append_ndarray_signature(x_trn, 'normalized_image_match x_trn',
task.name)
info('normalized_image_match_features complete')
return x_trn
def patch_whitening_filterbank_X(
patches,
o_ndim,
gamma,
remove_mean,
beta,
hard_beta,
):
"""
patches - Image patches (can be uint8 pixels or floats)
o_ndim - 2 to get matrix outputs, 4 to get image-stack outputs
gamma - non-negative real to boost low-principle components
remove_mean - see contrast_normalize
beta - see contrast_normalize
hard_beta - see contrast_normalize
Returns: M, P, X
M - mean of contrast-normalized patches
P - whitening matrix / filterbank for contrast-normalized patches
X - contrast-normalized patches
"""
# Algorithm from Coates' sc_vq_demo.m
# -- patches -> column vectors
X = patches.reshape(len(patches), -1).astype('float64')
X = contrast_normalize(X,
remove_mean=remove_mean,
beta=beta,
hard_beta=hard_beta)
# -- ZCA whitening (with low-pass)
logger.debug('patch_whitening_filterbank_X starting ZCA')
M, _std = mean_and_std(X)
Xm = X - M
assert Xm.shape == X.shape
logger.info('patch_whitening_filterbank_X starting ZCA: dot %s' %
str(Xm.shape))
C = dot_f64(Xm.T, Xm) / (Xm.shape[0] - 1)
logger.debug('patch_whitening_filterbank_X starting ZCA: eigh')
D, V = np.linalg.eigh(C)
logger.debug('patch_whitening_filterbank_X starting ZCA: dot %s' %
str(V.shape))
P = dot_f32(np.sqrt(1.0 / (D + gamma)) * V, V.T)
# -- return to image space
if o_ndim == 4:
M = M.reshape(patches.shape[1:])
P = P.reshape((P.shape[0], ) + patches.shape[1:])
X = X.reshape((len(X), ) + patches.shape[1:])
elif o_ndim == 2:
pass
else:
raise ValueError('o_ndim not in (2, 4)', o_ndim)
logger.debug('patch_whitening_filterbank_X -> done')
foobar.append_ndarray_signature(M, 'patch_whitening_filterbank_X M')
foobar.append_ndarray_signature(P, 'patch_whitening_filterbank_X P')
foobar.append_ndarray_signature(X, 'patch_whitening_filterbank_X X')
dtype = patches.dtype
return M.astype(dtype), P.astype(dtype), X.astype(dtype)
