+
Skip to content

Allocating gradients in backward op #663

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

Merged
merged 2 commits into from
Apr 4, 2023
Merged

Allocating gradients in backward op #663

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
47e1257
Allocating gradients in backward op
coreylowman Apr 4, 2023
52db21d
Revert small change to stack
coreylowman Apr 4, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 0 additions & 7 deletions src/tensor/gradients.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -233,7 +233,6 @@ pub trait Tape<E: Unit, D: DeviceStorage>: Default + Merge<Self> + Merge<NoneTap
fn add_backward_op<F>(&mut self, operation: F)
where
F: 'static + FnOnce(&mut Gradients<E, D>) -> Result<(), D::Err>;
fn try_alloc_grad<S: Shape>(&mut self, t: &Tensor<S, E, D>) -> Result<(), D::Err>;
}

impl<E: Unit, D: DeviceStorage> Tape<E, D> for OwnedTape<E, D> {
Expand All @@ -244,9 +243,6 @@ impl<E: Unit, D: DeviceStorage> Tape<E, D> for OwnedTape<E, D> {
{
self.operations.push((unique_id(), Box::new(operation)));
}
fn try_alloc_grad<S: Shape>(&mut self, t: &Tensor<S, E, D>) -> Result<(), D::Err> {
self.gradients.try_alloc_for(t)
}
}

impl<E: Unit, D: DeviceStorage> Tape<E, D> for NoneTape {
Expand All @@ -256,9 +252,6 @@ impl<E: Unit, D: DeviceStorage> Tape<E, D> for NoneTape {
F: 'static + FnOnce(&mut Gradients<E, D>) -> Result<(), D::Err>,
{
}
fn try_alloc_grad<S: Shape>(&mut self, _: &Tensor<S, E, D>) -> Result<(), D::Err> {
Ok(())
}
}

/// Combine two things
Expand Down
6 changes: 3 additions & 3 deletions src/tensor_ops/choose/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,10 +76,10 @@ impl<
let phantom_out = out.clone();

let mut tape = tape.merge(rhs_tape);
tape.try_alloc_grad(&lhs)?;
tape.try_alloc_grad(&rhs)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&lhs)?;
grads.try_alloc_for(&rhs)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_lhs, grad_rhs, grad_out) = grads.muts_and_ref(&lhs, &rhs, &phantom_out);
lhs.device
.backward(&self, &lhs, grad_lhs, &rhs, grad_rhs, grad_out)
Expand Down
6 changes: 3 additions & 3 deletions src/tensor_ops/concat/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -63,10 +63,10 @@ where
let device = lhs.device.clone();
let out = device.forward(&lhs, &rhs)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&lhs)?;
tape.try_alloc_grad(&rhs)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&lhs)?;
grads.try_alloc_for(&rhs)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_a, grad_b, grad_out) = grads.muts_and_ref(&lhs, &rhs, &phantom_out);
device.backward(&lhs, grad_a, &rhs, grad_b, &phantom_out, grad_out)
});
Expand Down
12 changes: 6 additions & 6 deletions src/tensor_ops/conv2d/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -163,10 +163,10 @@ impl<
.alloc((Const, h.convolve_dim(), w.convolve_dim()))?;
lhs.device.forward(op, &lhs, &rhs, &mut out)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&lhs)?;
tape.try_alloc_grad(&rhs)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&lhs)?;
grads.try_alloc_for(&rhs)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_lhs, grad_rhs, grad_out) = grads.muts_and_ref(&lhs, &rhs, &phantom_out);
lhs.device
.backward(op, &lhs, grad_lhs, &rhs, grad_rhs, &phantom_out, grad_out)
Expand Down Expand Up @@ -207,10 +207,10 @@ impl<
let mut tape = ltape.merge(rtape);
lhs.device.forward(op, &lhs, &rhs, &mut out)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&lhs)?;
tape.try_alloc_grad(&rhs)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&lhs)?;
grads.try_alloc_for(&rhs)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_lhs, grad_rhs, grad_out) = grads.muts_and_ref(&lhs, &rhs, &phantom_out);
lhs.device
.backward(op, &lhs, grad_lhs, &rhs, grad_rhs, &phantom_out, grad_out)?;
Expand Down
12 changes: 6 additions & 6 deletions src/tensor_ops/convtrans2d/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,10 +162,10 @@ impl<
.try_zeros_like(&(Const, h.convolve_dim(), w.convolve_dim()))?;
lhs.device.forward(op, &lhs, &rhs, &mut out)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&lhs)?;
tape.try_alloc_grad(&rhs)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&rhs)?;
grads.try_alloc_for(&lhs)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_lhs, grad_rhs, grad_out) = grads.muts_and_ref(&lhs, &rhs, &phantom_out);
lhs.device
.backward(op, &lhs, grad_lhs, &rhs, grad_rhs, &phantom_out, grad_out)
Expand Down Expand Up @@ -207,10 +207,10 @@ impl<
let mut tape = ltape.merge(rtape);
lhs.device.forward(op, &lhs, &rhs, &mut out)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&lhs)?;
tape.try_alloc_grad(&rhs)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&rhs)?;
grads.try_alloc_for(&lhs)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_lhs, grad_rhs, grad_out) = grads.muts_and_ref(&lhs, &rhs, &phantom_out);
lhs.device
.backward(op, &lhs, grad_lhs, &rhs, grad_rhs, &phantom_out, grad_out)?;
Expand Down
7 changes: 3 additions & 4 deletions src/tensor_ops/dropout/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,12 +68,11 @@ impl<S: Shape, E: Dtype, D: DropoutKernel<E>, T: Tape<E, D>> Tensor<S, E, D, T>
let (inp, mut tape) = self.split_tape();
let out = inp.device.forward(op, &inp)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device.backward(op, &inp, grad_inp, grad_out)?;
Ok(())
inp.device.backward(op, &inp, grad_inp, grad_out)
});
Ok(out.put_tape(tape))
}
Expand Down
6 changes: 3 additions & 3 deletions src/tensor_ops/matmul/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,10 +96,10 @@ fn try_binary_op<
let mut tape = ltape.merge(rtape);
let out = fwd(&lhs.device, &lhs, &rhs)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&lhs)?;
tape.try_alloc_grad(&rhs)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&lhs)?;
grads.try_alloc_for(&rhs)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_lhs, grad_rhs, grad_out) =grads.muts_and_ref(&lhs, &rhs, &phantom_out);
bwd(&lhs.device, &lhs, grad_lhs, &rhs, grad_rhs, grad_out)
});
Expand Down
4 changes: 2 additions & 2 deletions src/tensor_ops/max_to/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,9 +69,9 @@ impl<S: Shape, E: Dtype, D: MaxReduceKernel<E>, T: Tape<E, D>> MaxTo for Tensor<
let (inp, mut tape) = self.split_tape();
let out = inp.device.forward(dst, &inp)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device.backward(&inp, grad_inp, &phantom_out, grad_out)
});
Expand Down
4 changes: 2 additions & 2 deletions src/tensor_ops/min_to/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,9 +69,9 @@ impl<S: Shape, E: Dtype, D: MinReduceKernel<E>, T: Tape<E, D>> MinTo for Tensor<
let (inp, mut tape) = self.split_tape();
let out = inp.device.forward(dst, &inp)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device.backward(&inp, grad_inp, &phantom_out, grad_out)
});
Expand Down
8 changes: 4 additions & 4 deletions src/tensor_ops/pool2d/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -109,9 +109,9 @@ macro_rules! pool2d {
.try_zeros_like(&(chan, h.convolve_dim(), w.convolve_dim()))?;
inp.device.forward(op, &inp, &mut out)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device
.backward(op, &inp, grad_inp, &phantom_out, grad_out)
Expand Down Expand Up @@ -150,9 +150,9 @@ macro_rules! pool2d {
))?;
inp.device.forward(op, &inp, &mut out)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device
.backward(op, &inp, grad_inp, &phantom_out, grad_out)
Expand Down
4 changes: 2 additions & 2 deletions src/tensor_ops/reshape_to/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,9 +92,9 @@ impl<S: Shape, E: Dtype, D: ReshapeKernel<E>, T: Tape<E, D>> ReshapeTo for Tenso
let (inp, mut tape) = self.split_tape();
let out = inp.device.forward(dst, &inp)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device.backward(&inp, grad_inp, &phantom_out, grad_out)
});
Expand Down
4 changes: 2 additions & 2 deletions src/tensor_ops/roll/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -74,9 +74,9 @@ impl<S: Shape, E: Dtype, D: RollKernel<E>, T: Tape<E, D>> Roll for Tensor<S, E,
let (t, mut tape) = self.split_tape();
let out = t.device.forward(op, &t)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&t)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&t)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&t, &phantom_out);
t.device.backward(op, &t, grad_inp, grad_out)
});
Expand Down
8 changes: 4 additions & 4 deletions src/tensor_ops/select_and_gather/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -105,9 +105,9 @@ impl<Src: Shape, E: Dtype, D: RemoveDimKernel<E>, T: Tape<E, D>> SelectTo<D>
let (inp, mut tape) = self.split_tape();
let out = inp.device.forward(&inp, &idx)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device
.backward(&inp, grad_inp, &idx, &phantom_out, grad_out)
Expand Down Expand Up @@ -175,9 +175,9 @@ impl<Src: Shape, E: Dtype, D: ReplaceDimKernel<E>, T: Tape<E, D>> GatherTo<D>
let (inp, mut tape) = self.split_tape();
let out = inp.device.forward(&inp, &idx)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device
.backward(&inp, grad_inp, &idx, &phantom_out, grad_out)
Expand Down
4 changes: 2 additions & 2 deletions src/tensor_ops/slice/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,9 +62,9 @@ impl<S: Shape, E: Unit, D: SliceKernel<E>, T: Tape<E, D>> Tensor<S, E, D, T> {
let out = inp.device.forward(&inp, &slice)?;
let phantom_out = out.clone();

tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device.backward(&inp, grad_inp, grad_out, &slice)
});
Expand Down
9 changes: 5 additions & 4 deletions src/tensor_ops/stack/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -150,18 +150,19 @@ where
let shape = *tensors[0].shape();
for t in tensors.iter() {
assert_eq!(t.shape(), &shape);
tape.try_alloc_grad(t)?;
}

// we map to storage refs so kernels don't have to know about tensors
let out = device.forward(new_dim, &tensors)?;

let phantom_out = out.clone();
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
for t in tensors.iter() {
grads.try_alloc_for(t)?;
}
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.many_and_ref(&tensors, &phantom_out);
device.backward(grad_inp, grad_out)?;
Ok(())
device.backward(grad_inp, grad_out)
});
Ok(out.put_tape(tape))
}
Expand Down
4 changes: 2 additions & 2 deletions src/tensor_ops/sum_to/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,9 +66,9 @@ impl<S: Shape, E: Dtype, D: SumKernel<E>, T: Tape<E, D>> SumTo for Tensor<S, E,
let (inp, mut tape) = self.split_tape();
let out = inp.device.forward(dst, &inp)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device.backward(&inp, grad_inp, &phantom_out, grad_out)
});
Expand Down
8 changes: 4 additions & 4 deletions src/tensor_ops/upscale2d/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -185,9 +185,9 @@ impl<
let mut out = inp.device.try_zeros_like(&(chan, out_height, out_width))?;
inp.device.forward(op, &inp, &mut out)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device
.backward(op, &inp, grad_inp, &phantom_out, grad_out)
Expand Down Expand Up @@ -229,9 +229,9 @@ impl<
.try_zeros_like(&(batch, chan, out_height, out_width))?;
inp.device.forward(op, &inp, &mut out)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device
.backward(op, &inp, grad_inp, &phantom_out, grad_out)
Expand Down
5 changes: 4 additions & 1 deletion src/tensor_ops/utilities/backward.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,10 @@ pub trait Backward<E: Dtype, D: DeviceStorage>: HasErr {
impl<E: Dtype, D: OneFillStorage<E>> Backward<E, D> for Tensor<Rank0, E, D, OwnedTape<E, D>> {
fn try_backward(self) -> Result<Gradients<E, D>, Self::Err> {
let (t, mut tape) = self.split_tape();
tape.add_backward_op(move |grads| t.device.try_fill_with_ones(grads.get_mut(&t)));
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&t)?;
t.device.try_fill_with_ones(grads.get_mut(&t))
});
let mut grads = tape.execute()?;
grads.drop_non_leafs();
Ok(grads)
Expand Down
10 changes: 5 additions & 5 deletions src/tensor_ops/utilities/ops.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,9 +51,9 @@ pub(crate) fn try_unary_op<
let (inp, mut tape) = inp.split_tape();
let out = inp.device.forward(op.clone(), &inp)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&inp)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&inp)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_inp, grad_out) = grads.mut_and_ref(&inp, &phantom_out);
inp.device
.backward(op, &inp, grad_inp, &phantom_out, grad_out)
Expand All @@ -79,10 +79,10 @@ pub(crate) fn try_binary_op<
let mut tape = ltape.merge(rtape);
let out = lhs.device.forward(op, &lhs, &rhs)?;
let phantom_out = out.clone();
tape.try_alloc_grad(&lhs)?;
tape.try_alloc_grad(&rhs)?;
tape.try_alloc_grad(&out)?;
tape.add_backward_op(move |grads| {
grads.try_alloc_for(&lhs)?;
grads.try_alloc_for(&rhs)?;
grads.try_alloc_for(&phantom_out)?;
let (grad_lhs, grad_rhs, grad_out) = grads.muts_and_ref(&lhs, &rhs, &phantom_out);
lhs.device
.backward(op, &lhs, grad_lhs, &rhs, grad_rhs, grad_out)
Expand Down
点击 这是indexloc提供的php浏览器服务，不要输入任何密码和下载