def main():
""" Process NMEC with precompiled denoiser and Fielnet
"""
def denoising_func(x):
orig_shape = x.shape
x = 1. - x/255.
x = denoiser.predict(x, verbose=0)
x = x.reshape(orig_shape)
return x
try:
featext = load_verbatimnet('fc7', paramsfile='/fileserver/iam/iam-processed/models/fiel_657.hdf5')
featext.compile(loss='mse', optimizer='sgd')
featext_func = lambda x: featext.predict(x, verbose=0.0)
print "Making the denoiser"
denoiser = conv4p_model()
denoiser.load_weights('/fileserver/iam/models/conv4p_linet56-iambin-tifs.hdf5')
hdf5file='/fileserver/nmec-handwriting/flat_nmec_bin_uint8.hdf5'
with h5py.File(hdf5file, "r") as data_file:
features = extract_features_for_corpus(data_file,featext_func,shingle_dim=(56,56),transform=denoising_func)
with h5py.File("output_features.hdf5", "w") as feature_file:
for document_id, document_features in features.iteritems():
feature_file.create_dataset(document_id, data=document_features)
except Exception as e:
print e
pdb.post_mortem()
for i in xrange(1, ccis[1]):
energy = np.array(ccis[0] == i).sum()
if energy > minthresh and energy < maxthresh:
ii = np.where(ccis[0] == i)
xb = ii[0].min()
yb = ii[1].min()
xe = ii[0].max()
ye = ii[1].max()
ccs += [shingler(ccis[0][xb:xe, yb:ye], shingle_dim=(56, 56))]
print "Finished finding " + str(len(ccs)) + " features from image."
return np.expand_dims(np.array(ccs), 1)
# Neural network
from globalclassify.fielutil import load_verbatimnet
featext = load_verbatimnet(
'fc7', paramsfile='/fileserver/iam/iam-processed/models/fiel_657.hdf5')
featext.compile(loss='mse', optimizer='sgd')
# Denoising neural network
from denoiser.noisenet import conv4p_model
denoiser = conv4p_model()
denoiser.load_weights('/work/models/conv4p_linet56-iambin-tifs.hdf5')
hdf5file = '/fileserver/nmec-handwriting/flat_nmec_cropped_bin_uint8.hdf5'
flatnmec = h5py.File(hdf5file, 'r')
outputdir = '/fileserver/nmec-handwriting/localfeatures/second-pass/'
# Extract connected components, and then shingles with minimum threshold 500
for imname in flatnmec.keys():
ccis = connectedcomponents(flatnmec[imname])
for i in xrange(1,ccis[1]):
energy = np.array(ccis[0]==i).sum()
if energy > minthresh and energy < maxthresh:
ii = np.where( ccis[0] == i )
xb = ii[0].min()
yb = ii[1].min()
xe = ii[0].max()
ye = ii[1].max()
ccs += [ shingler( ccis[0][xb:xe, yb:ye], shingle_dim=(56,56) ) ]
print "Finished finding "+str(len(ccs))+" features from image."
return np.expand_dims( np.array( ccs ), 1 )
# Neural network
from globalclassify.fielutil import load_verbatimnet
featext = load_verbatimnet('fc7', paramsfile='/fileserver/iam/iam-processed/models/fiel_657.hdf5')
featext.compile(loss='mse', optimizer='sgd')
# Denoising neural network
from denoiser.noisenet import conv4p_model
denoiser = conv4p_model()
denoiser.load_weights('/work/models/conv4p_linet56-iambin-tifs.hdf5')
hdf5file='/fileserver/nmec-handwriting/flat_nmec_cropped_bin_uint8.hdf5'
flatnmec=h5py.File(hdf5file,'r')
outputdir = '/fileserver/nmec-handwriting/localfeatures/second-pass/'
# Extract connected components, and then shingles with minimum threshold 500
for imname in flatnmec.keys():
ccis = connectedcomponents( flatnmec[imname] )