def new_exit(pipeline, prefix):
def lab(msg):
return prefix % msg
return {
'pipe': pipeline,
'remove_std0':
hp_TF(lab('remove_std0')),
'varthresh':
hp_lognormal(lab('varthresh'),
np.log(1e-4), np.log(1000)),
'l2_reg': hp_lognormal(lab('l2_reg'),
np.log(1e-5), np.log(1e3)),
'divrowl2': hp_TF(lab('divrowl2')),
}
def many_dists():
a = hp_choice('a', [0, 1, 2])
b = hp_randint('b', 10)
c = hp_uniform('c', 4, 7)
d = hp_loguniform('d', -2, 0)
e = hp_quniform('e', 0, 10, 3)
f = hp_qloguniform('f', 0, 3, 2)
g = hp_normal('g', 4, 7)
h = hp_lognormal('h', -2, 2)
i = hp_qnormal('i', 0, 10, 2)
j = hp_qlognormal('j', 0, 2, 1)
z = a + b + c + d + e + f + g + h + i + j
return {'loss': scope.float(scope.log(1e-12 + z**2))}
def many_dists():
a=hp_choice('a', [0, 1, 2])
b=hp_randint('b', 10)
c=hp_uniform('c', 4, 7)
d=hp_loguniform('d', -2, 0)
e=hp_quniform('e', 0, 10, 3)
f=hp_qloguniform('f', 0, 3, 2)
g=hp_normal('g', 4, 7)
h=hp_lognormal('h', -2, 2)
i=hp_qnormal('i', 0, 10, 2)
j=hp_qlognormal('j', 0, 2, 1)
z = a + b + c + d + e + f + g + h + i + j
return {'loss': scope.float(scope.log(1e-12 + z ** 2))}
def pipeline_extension(prefix, X, n_patches, max_filters):
assert max_filters > 16
f_layer = new_fbncc_layer(prefix, X, n_patches,
n_filters=s_int(
hp_qloguniform('%sfb_nfilters' % prefix,
np.log(8.01), np.log(max_filters), q=16)),
size=rfilter_size('%sfb_size' % prefix, 3, 8),
)
p_layer = partial(slm_lpool,
stride=hp_choice('%sp_stride' % prefix, [1, 2]),
order=hp_choice('%sp_order' % prefix,
[1, 2, hp_lognormal('%sp_order_real' % prefix,
mu=np.log(1), sigma=np.log(3))]),
ker_size=rfilter_size('%sp_size' % prefix, 2, 8))
return [f_layer, p_layer]
def exit_lpool(pipeline, layer_num, Xcm, n_patches, max_n_features):
def lab(msg):
return 'l%i_out_lp_%s' % (layer_num, msg)
fsize = rfilter_size(lab('fsize'), 3, 8)
filtering_res = pyll_getattr(Xcm, 'shape')[2] - fsize + 1
# -- N.B. Xrows depends on other params, so we can't use it to set the
# upper bound on lpsize. We can only sample independently, and
# then fail below with non-positive number of features.
psize = rfilter_size(lab('psize'), 1, 5)
stride = hp_choice(lab('stride'), [1, 2, 3])
pooling_res = scope.ceildiv(filtering_res - psize + 1, stride)
nsize = rfilter_size(lab('nsize'), 1, 5)
norm_res = pooling_res - nsize + 1
# -- raises exception at rec_eval if norm_res is 0
nfilters = max_n_features // (scope.max(norm_res, 0) ** 2)
filtering = new_fbncc_layer(
prefix='l%ielp' % layer_num,
Xcm=Xcm,
n_patches=n_patches,
n_filters=nfilters,
size=fsize,
)
pooling = partial(slm_lpool,
ker_size=psize,
stride=stride,
order=hp_choice(lab('order_choice'), [
1.0, 2.0, logu_range(lab('order_real'), .1, 10.)]))
normalization = partial(slm_lnorm,
ker_size=nsize,
remove_mean=hp_TF(lab('norm_rmean')),
threshold=hp_lognormal(lab('norm_thresh'),
np.log(1.0), np.log(3)),
)
seq = hp_choice(lab('use_norm'), [
[filtering, pooling],
[filtering, pooling, normalization]])
return new_exit(pipeline + seq, lab('%s'))
def new_fbncc_layer(prefix, Xcm, n_patches, n_filters, size,
memlimit=5e8, # -- limit patches array to 500MB
):
def lab(msg):
return '%s_fbncc_%s' % (prefix, msg)
def get_rseed(name, N):
fullname = lab(name)
low = stable_hash(fullname) % (2 ** 31)
rval = hp_choice(fullname, range(low, low + N))
return rval
patches = random_patches(
Xcm, n_patches, size, size,
rng=np_RandomState(get_rseed('patch_rseed', 10)),
channel_major=True,
memlimit=memlimit)
remove_mean = hp_TF(lab('remove_mean'))
beta = hp_lognormal(lab('beta'), np.log(100), np.log(100))
hard_beta = hp_TF(lab('hard'))
# TODO: use different nfilters, beta etc. for each algo
# -- random projections filterbank allocation
random_projections = partial(slm_fbncc_chmaj,
m_fb=slm_uniform_M_FB(
nfilters=n_filters,
size=size,
channels=pyll_getattr(Xcm, 'shape')[1],
rseed=get_rseed('r_rseed', 10),
normalize=hp_TF(lab('r_normalize')),
dtype='float32',
ret_cmajor=True,
),
remove_mean=remove_mean,
beta=beta,
hard_beta=hard_beta)
# -- random whitened projections filterbank allocation
random_whitened_projections = partial(slm_fbncc_chmaj,
m_fb=fb_whitened_projections(patches,
patch_whitening_filterbank_X(patches,
gamma=hp_lognormal(lab('wr_gamma'),
np.log(1e-2), np.log(100)),
o_ndim=2,
remove_mean=remove_mean,
beta=beta,
hard_beta=hard_beta,
),
n_filters=n_filters,
rseed=get_rseed('wr_rseed', 10),
dtype='float32',
),
remove_mean=remove_mean,
beta=beta,
hard_beta=hard_beta)
# -- whitened patches filterbank allocation
whitened_patches = partial(slm_fbncc_chmaj,
m_fb=fb_whitened_patches(patches,
patch_whitening_filterbank_X(patches,
gamma=hp_lognormal(lab('wp_gamma'),
np.log(1e-2), np.log(100)),
o_ndim=2,
remove_mean=remove_mean,
beta=beta,
hard_beta=hard_beta,
),
n_filters=n_filters,
rseed=get_rseed('wp_rseed', 10),
dtype='float32',
),
remove_mean=remove_mean,
beta=beta,
hard_beta=hard_beta)
# --> MORE FB LEARNING ALGOS HERE <--
# TODO: V1-like filterbank (incl. with whitening matrix)
# TODO: sparse coding
# TODO: OMP from Coates 2011
# TODO: K-means
# TODO: RBM
# TODO: DAA
# TODO: ssRBM
rchoice = hp_choice(lab('algo'), [
random_projections,
random_whitened_projections,
whitened_patches,
])
return rchoice
