def random_uniform(shape, minval=0.0, maxval=1.0, dtype=None, seed=None):
rng_state = _make_rng_state(seed)
R = edsl.prng(rng_state.tensor, shape)
dtype = dtype or floatx()
if dtype != 'float32':
R = edsl.cast(R, plaidml.DType.from_numpy(dtype))
O = (maxval - minval) * R + minval
return _KerasNode('random_uniform', tensor=O)
def dropout(x, level, noise_shape=None, seed=None):
I = x.tensor
if noise_shape is not None and len(noise_shape) != I.shape.ndims:
raise ValueError('noise_shape ndims doesn\'t match input ndims')
if noise_shape is None:
shape = I.shape.dims
else:
shape = noise_shape
rng_state = _make_rng_state(seed)
R = 1.0 - level
M = 1.0 / R
T = edsl.prng(rng_state.tensor, shape)
O = edsl.select(T < R, I * M, 0.0)
return _KerasNode('dropout', tensor=O)
def dropout(x, level, noise_shape=None, seed=None):
logger.debug('dropout(x: {}, level: {}, noise_shape: {}, seed: {})'.format(
x, level, noise_shape, seed))
I = x.tensor
if noise_shape is not None and len(noise_shape) != I.shape.ndims:
raise ValueError("noise_shape ndims doesn't match input ndims")
if noise_shape is None:
shape = I.shape.dims
else:
shape = noise_shape
rng_state = _make_rng_state(seed)
R = 1.0 - level
M = 1.0 / R
T = edsl.prng(rng_state.tensor, shape)
O = edsl.select(T < R, I * M, 0.0)
return _KerasNode('dropout', tensor=O)
