def full_extract(fhs,config_fname):
fh = fhs[0][0][1]
params, featsel = get_params(config_fname)
arr = image2array(params,fh)
orig_imga,orig_imga_conv = image_preprocessing(arr,params)
fvector_l = []
filt_l = get_gabor_filters(params['filter'])
for cidx in xrange(orig_imga_conv.shape[2]):
imga0 = map_preprocessing(orig_imga_conv[:,:,cidx],params)
imga1 = v1f.v1like_norm(imga0[:,:,sp.newaxis], conv_mode, **params['normin'])
imga2 = v1f.v1like_filter(imga1[:,:,0], conv_mode, filt_l)
minout = params['activ']['minout'] # sustain activity
maxout = params['activ']['maxout'] # saturation
imga3 = imga2.clip(minout, maxout)
imga4 = v1f.v1like_norm(imga3, conv_mode, **params['normout'])
if "sparsify" in params and params["sparsify"]:
imga4 = (imga4.max(2)[:,:,None] == imga4)
imga5 = v1f.v1like_pool(imga4, conv_mode, **params['pool'])
output = imga5
fvector = include_map_level_features(imga1,imga2,imga3,imga4,imga5,output,featsel)
fvector_l += [fvector]
# -- include image-level bells & whistles
fvector_l = include_image_level_features(orig_imga,fvector_l, featsel)
# -- reshape
fvector_l = [fvector.ravel() for fvector in fvector_l]
out = sp.concatenate(fvector_l).ravel()
return cPickle.dumps(out)
def normalize(fhs,config_fname,pset):
fh = fhs[0][0][1]
assert pset in ['in','out'], 'pset not recognized'
params, featsel = get_params(config_fname)
conv_mode = params['conv_mode']
input = cPickle.loads(fh.read())
output = {}
for cidx in input.keys():
if len(input[cidx].shape) == 3:
inobj = input[cidx]
else:
inobj = input[cidx][:,:,sp.newaxis]
output[cidx] = v1f.v1like_norm(inobj, conv_mode, **params['norm' + pset])
return cPickle.dumps(output)