# when initializing gradient descent : epochs = 1e6, momentum = 0.999
f, g, _, _ = bib_data.get_tomo_JET(fname, faulty=False, flatten=True)
g = np.transpose(g) * 1e3
f = np.transpose(f) * 1e3
print 'g:', g.shape, g.dtype
print 'f:', f.shape, f.dtype
# ------------------------------------------------------------------------
# Divide into training and validation set
# (no test set needed since it won't overfit)
# if one wants the validation set can also be disregarded by setting ratio=[1.,0.]
i_train, i_valid, _ = bib_utils.divide_data(g.shape[1],
ratio=[.9, .1],
test_set=False,
random=False)
g_valid = g[:, i_valid]
f_valid = f[:, i_valid]
g = g[:, i_train]
f = f[:, i_train]
print 'g_train:', g.shape, g.dtype
print 'f_train:', f.shape, f.dtype
print 'g_valid:', g_valid.shape, g_valid.dtype
print 'f_valid:', f_valid.shape, f_valid.dtype
np.savez(save_path + 'i_divided', i_train=i_train, i_valid=i_valid)
# ------------------------------------------------------------------------
f, g, _, _ = bib_data.get_tomo_JET(fname,
faulty=True,
flatten=False,
clip_tomo=True)
# need to reshape image to match NN dimensions
g = bib_utils.resize_NN_image(g, training=True)
print 'g:', g.shape, g.dtype
print 'f:', f.shape, f.dtype
# ------------------------------------------------------------------------
# Divide into training, validation and test set
i_train, i_valid, i_test = bib_utils.divide_data(g.shape[0],
ratio=[.8, .1, .1],
test_set=True,
random=False)
f_valid = f[i_valid]
g_valid = g[i_valid]
f_train = f[i_train]
g_train = g[i_train]
print 'f_train:', f_train.shape
print 'g_train:', g_train.shape
print 'f_valid:', f_valid.shape
print 'g_valid:', g_valid.shape
np.savez(save_path + 'i_divided',
i_train=i_train,
i_valid=i_valid,