Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Fused LSTM grad-grad #3256

Merged
merged 3 commits into from
Oct 14, 2017
Merged

Fused LSTM grad-grad #3256

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
250c970
Make cuda.fuse and use it in LSTM
unnonouno Aug 24, 2017
f9d08b3
Use empty_like
unnonouno Aug 27, 2017
fee7d2d
Use cuda.fusion
unnonouno Sep 4, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
16 changes: 16 additions & 0 deletions chainer/cuda.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -490,3 +490,19 @@ def set_max_workspace_size(size):
"""
global _max_workspace_size
_max_workspace_size = size


def fuse(*args, **kwargs):
"""Function fusing decorator.

It calls :func:`cupy.fuse` when CuPy is available to make fused function
and does nothing otherwise.

.. seealso::
:func:`cupy.fuse`

"""
if available:
return cupy.fuse(*args, **kwargs)
else:
return lambda f: f
107 changes: 62 additions & 45 deletions chainer/functions/activation/lstm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,13 +180,18 @@ def backward(self, inputs, grads):
c_prev, x, c, gc, gh = inputs
ggc_prev, ggx = grads

gc_prev = xp.zeros_like(c_prev)
gx = xp.zeros_like(x)
gc_next = xp.zeros_like(c)
ggc = ggc_prev.copy()
ggh = xp.zeros_like(gh)
gc_prev = xp.empty_like(c_prev)
gx = xp.empty_like(x)
gc_next = xp.empty_like(c)
ggc = xp.empty_like(ggc_prev)
ggh = xp.empty_like(gh)

batch = len(x)
gc_prev[batch:] = 0
gc_next[batch:] = 0
ggc[batch:] = ggc_prev[batch:]
ggh[batch:] = 0

c_prev = c_prev[:batch]
c = c[:batch]
gc = gc[:batch]
Expand All @@ -195,51 +200,63 @@ def backward(self, inputs, grads):

a, i, f, o = _extract_gates(x)
gga, ggi, ggf, ggo = _extract_gates(ggx)

ga, gi, gf, go = _extract_gates(gx)

sig_o = _sigmoid(o, xp)
gsig_o = _grad_sigmoid(sig_o)
ggsig_o = _grad_grad_sigmoid(sig_o)
sig_i = _sigmoid(i, xp)
gsig_i = _grad_sigmoid(sig_i)
ggsig_i = _grad_grad_sigmoid(sig_i)
sig_f = _sigmoid(f, xp)
gsig_f = _grad_sigmoid(sig_f)
ggsig_f = _grad_grad_sigmoid(sig_f)
tanh_a = xp.tanh(a)
gtanh_a = _grad_tanh(tanh_a)
ggtanh_a = _grad_grad_tanh(tanh_a, gtanh_a)
tanh_c = xp.tanh(c)
gtanh_c = _grad_tanh(tanh_c)
ggtanh_c = _grad_grad_tanh(tanh_c, gtanh_c)

gc_bar = gh * sig_o * gtanh_c + gc

gc_prev[:batch] = ggf * gc_bar * gsig_f
ga[:] = (gga * sig_i * ggtanh_a +
ggi * gtanh_a * gsig_i) * gc_bar
gi[:] = (gga * gtanh_a * gsig_i +
ggi * tanh_a * ggsig_i) * gc_bar
gf[:] = (ggc_prev * (gh * sig_o * gtanh_c + gc) * gsig_f +
ggf * gc_bar * c_prev[:batch] * ggsig_f)

ggc_this = (
ggc_prev * sig_f +
gga * sig_i * gtanh_a +
ggi * tanh_a * gsig_i +
ggf * c_prev[:batch] * gsig_f)
ggc[:batch] = ggc_this

dgc_do = gh * gsig_o * gtanh_c
go[:] = ggc_this * dgc_do + ggo * gh * tanh_c * ggsig_o
dgc_dc = gh * sig_o * ggtanh_c
gc_next[:batch] = ggc_this * dgc_dc + ggo * gh * gtanh_c * gsig_o
ggh[:batch] = ggc_this * sig_o * gtanh_c + ggo * tanh_c * gsig_o

gc_prev[:batch], ga[:], gi[:], gf[:], go[:], gc_next[:batch], \
ggc[:batch], ggh[:batch] \
= lstm_grad_grad(
c_prev, a, i, f, o, c, gc, gh, ggc_prev, gga, ggi, ggf, ggo)
return gc_prev, gx, gc_next, ggc, ggh


def _cupy_sigmoid(x):
half = x.dtype.type(0.5)
return cuda.fusion.tanh(x * half) * half + half


@cuda.fuse()
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

How about using input_num?
lstm_grad_grad will accept writable arguments.

@cuda.fuse(input_num=13)
def lstm_grad_grad(
        c_prev, a, i, f, o, c, gc, gh, ggc_prev, gga, ggi, ggf, ggo,
        gc_prev, ga, gi, gf, go, gc_next, ggc, ggh):

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

How to fix?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I wrote sample. That kernel will reduce array copy.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sorry, This is my mistake.

def lstm_grad_grad(
c_prev, a, i, f, o, c, gc, gh, ggc_prev, gga, ggi, ggf, ggo):
sig_o = _cupy_sigmoid(o)
gsig_o = _grad_sigmoid(sig_o)
ggsig_o = _grad_grad_sigmoid(sig_o)
sig_i = _cupy_sigmoid(i)
gsig_i = _grad_sigmoid(sig_i)
ggsig_i = _grad_grad_sigmoid(sig_i)
sig_f = _cupy_sigmoid(f)
gsig_f = _grad_sigmoid(sig_f)
ggsig_f = _grad_grad_sigmoid(sig_f)
tanh_a = cuda.fusion.tanh(a)
gtanh_a = _grad_tanh(tanh_a)
ggtanh_a = _grad_grad_tanh(tanh_a, gtanh_a)
tanh_c = cuda.fusion.tanh(c)
gtanh_c = _grad_tanh(tanh_c)
ggtanh_c = _grad_grad_tanh(tanh_c, gtanh_c)

gc_bar = gh * sig_o * gtanh_c + gc

gc_prev = ggf * gc_bar * gsig_f
ga = (gga * sig_i * ggtanh_a +
ggi * gtanh_a * gsig_i) * gc_bar
gi = (gga * gtanh_a * gsig_i +
ggi * tanh_a * ggsig_i) * gc_bar
gf = (ggc_prev * (gh * sig_o * gtanh_c + gc) * gsig_f +
ggf * gc_bar * c_prev * ggsig_f)

ggc = (
ggc_prev * sig_f +
gga * sig_i * gtanh_a +
ggi * tanh_a * gsig_i +
ggf * c_prev * gsig_f)

dgc_do = gh * gsig_o * gtanh_c
go = ggc * dgc_do + ggo * gh * tanh_c * ggsig_o
dgc_dc = gh * sig_o * ggtanh_c
gc_next = ggc * dgc_dc + ggo * gh * gtanh_c * gsig_o
ggh = ggc * sig_o * gtanh_c + ggo * tanh_c * gsig_o
return gc_prev, ga, gi, gf, go, gc_next, ggc, ggh


def lstm(c_prev, x):
"""Long Short-Term Memory units as an activation function.

Expand Down