def _make_init_state(dream):
r"""
CommandLine:
python -m ibeis_cnn.draw_net _make_init_state --show
Example:
>>> # DISABLE_DOCTEST
>>> # Assumes mnist is trained
>>> from ibeis_cnn.draw_net import * # NOQA
>>> from ibeis_cnn.models import mnist
>>> model, dataset = mnist.testdata_mnist(dropout=.5)
>>> model.init_arch()
>>> dream = Dream(model, init='rgauss', niters=200)
>>> ut.quit_if_noshow()
>>> import plottool as pt
>>> import vtool as vt
>>> #pt.qt4ensure()
>>> initial_state = dream._make_init_state().transpose((0, 2, 3, 1))[0]
>>> pt.imshow(initial_state, pnum=(1, 2, 1), fnum=1)
>>> pt.imshow(vt.norm01(initial_state), pnum=(1, 2, 2), fnum=1)
>>> ut.show_if_requested()
"""
init = dream.init
input_shape = dream.model.input_shape
b, c, w, h = input_shape
rng = np.random.RandomState(0)
if init is None or init == 'zeros':
# intializing to zeros seems to do nothing on mnist data
initial_state = np.zeros(input_shape, dtype=np.float32)
if init in ['rand', 'random']:
initial_state = rng.rand(*input_shape)
elif init in ['randn']:
initial_state = np.abs(rng.randn(*input_shape)) / 6
initial_state = np.clip(initial_state, 0, 1)
elif init in ['gauss']:
import vtool as vt
initial_state = np.array([[vt.gaussian_patch((h, w), sigma=None)
for _ in range(c)]] * b)
#initial_state /= initial_state.max()
elif init in ['rgauss']:
import vtool as vt
initial_state = np.array([[vt.gaussian_patch((h, w), sigma=None)
for _ in range(c)]] * b)
#initial_state /= initial_state.max()
raug = (np.abs(rng.randn(*input_shape)) * (initial_state.max() / 12))
initial_state += raug
initial_state = np.clip(initial_state, 0, 1)
elif init in ['perlin']:
import vtool as vt
b, c, w, h = input_shape
initial_state = np.array([[vt.perlin_noise((w, h), rng=rng)
for _ in range(c)]] * b)
initial_state = initial_state.astype(np.float32) / 255
initial_state = initial_state.astype(np.float32)
return initial_state
