def _grads(self, delta, delta_space, X):
grad_W = conv.fftconv1d(np.fliplr(delta), X, n_threads=self.n_threads, mode='valid')
grad_W_space = delta_space.with_extents(w=grad_W.shape[1])
grad_W, grad_W_space = grad_W_space.transform(grad_W, grad_W_space.folded_axes)
grad_W = grad_W.sum(axis=grad_W_space.axes.index('b'))
grad_W_space = grad_W_space.without_axes('b')
grad_W, grad_W_space = grad_W_space.transform(grad_W, [('b', 'd', 'f', 'c'), 'w'])
return [grad_W]
def _fprop(self, X, X_space):
K, _ = self._kernel_space.broadcast(np.fliplr(self.W), b=X_space.get_extent('b'))
X = conv.fftconv1d(X, K, n_threads=self.n_threads)
X_space = X_space.with_extents(w=X.shape[1])
X, X_space = X_space.transform(X, (('b', 'd', 'f'), 'c', 'w'))
X = X.sum(axis=X_space.axes.index('c'))
X_space = X_space.without_axes('c')
return X, X_space
def _bprop(self, delta, delta_space):
K, _ = self._kernel_space.broadcast(self.W, b=delta_space.get_extent('b'))
delta = conv.fftconv1d(delta, K, n_threads=self.n_threads, mode='valid')
delta_space = delta_space.with_extents(w=delta.shape[1])
delta, delta_space = delta_space.transform(delta, delta_space.folded_axes)
delta = delta.sum(axis=delta_space.axes.index('f'))
delta_space = delta_space.without_axes('f')
delta_space = delta_space.rename_axes(c='f')
return delta, delta_space
def _fprop(self, X, X_space):
K, _ = self._kernel_space.broadcast(np.fliplr(self.W),
b=X_space.get_extent('b'))
X = conv.fftconv1d(X, K, n_threads=self.n_threads)
X_space = X_space.with_extents(w=X.shape[1])
X, X_space = X_space.transform(X, (('b', 'd', 'f'), 'c', 'w'))
X = X.sum(axis=X_space.axes.index('c'))
X_space = X_space.without_axes('c')
return X, X_space
def _run_fftconv1d(self, n_x, n_k, mode):
n_rows = 2
X = np.random.standard_normal(size=(n_rows, n_x))
K = np.random.uniform(size=(n_rows, n_k))
actual = conv.fftconv1d(X, K, mode=mode)
expected = self.reference_convolve(X, K, mode=mode)
print "X:", X.shape, "K:", K.shape, "A:", actual.shape, "E:", expected.shape
print actual
print
print expected
assert np.allclose(actual, expected)
def _run_fftconv1d(self, n_x, n_k, mode):
n_rows = 2
X = np.random.standard_normal(size=(n_rows, n_x))
K = np.random.uniform(size=(n_rows, n_k))
actual = conv.fftconv1d(X, K, mode=mode)
expected = self.reference_convolve(X, K, mode=mode)
print "X:", X.shape, "K:", K.shape, "A:", actual.shape, "E:", expected.shape
print actual
print
print expected
assert np.allclose(actual, expected)
def _grads(self, delta, delta_space, X):
grad_W = conv.fftconv1d(np.fliplr(delta),
X,
n_threads=self.n_threads,
mode='valid')
grad_W_space = delta_space.with_extents(w=grad_W.shape[1])
grad_W, grad_W_space = grad_W_space.transform(grad_W,
grad_W_space.folded_axes)
grad_W = grad_W.sum(axis=grad_W_space.axes.index('b'))
grad_W_space = grad_W_space.without_axes('b')
grad_W, grad_W_space = grad_W_space.transform(
grad_W, [('b', 'd', 'f', 'c'), 'w'])
return [grad_W]
def _bprop(self, delta, delta_space):
K, _ = self._kernel_space.broadcast(self.W,
b=delta_space.get_extent('b'))
delta = conv.fftconv1d(delta,
K,
n_threads=self.n_threads,
mode='valid')
delta_space = delta_space.with_extents(w=delta.shape[1])
delta, delta_space = delta_space.transform(delta,
delta_space.folded_axes)
delta = delta.sum(axis=delta_space.axes.index('f'))
delta_space = delta_space.without_axes('f')
delta_space = delta_space.rename_axes(c='f')
return delta, delta_space
