candle_core/
backprop.rs

1//! Methods for backpropagation of gradients.
2use crate::op::{BinaryOp, Op, ReduceOp, UnaryOp};
3use crate::{Error, Result, Tensor, TensorId};
4use std::collections::{hash_map::Entry, HashMap};
5
6// arg has been reduced to node via reduce_dims, expand it back to arg.
7// This has to handle keepdims.
8fn broadcast_back(arg: &Tensor, node: &Tensor, reduced_dims: &[usize]) -> Result<Tensor> {
9    if arg.rank() == node.rank() {
10        // keepdim = true
11        node.broadcast_as(arg.shape())
12    } else {
13        // keepdim = false
14        // first expand the reduced dims.
15        node.reshape(reduced_dims)?.broadcast_as(arg.shape())
16    }
17}
18
19thread_local! {
20    static CANDLE_GRAD_DO_NOT_DETACH: bool = {
21        match std::env::var("CANDLE_GRAD_DO_NOT_DETACH") {
22            Ok(s) => {
23                !s.is_empty() && s != "0"
24            },
25            Err(_) => false,
26        }
27    }
28}
29
30impl Tensor {
31    /// Return all the nodes that lead to this value in a topologically sorted vec, the first
32    /// elements having dependencies on the latter ones, e.g. the first element if any is the
33    /// argument.
34    /// This assumes that the op graph is a DAG.
35    pub fn sorted_nodes(&self) -> Vec<&Tensor> {
36        // The vec of sorted nodes is passed as an owned value rather than a mutable reference
37        // to get around some lifetime limitations.
38        fn walk<'a>(
39            node: &'a Tensor,
40            nodes: Vec<&'a Tensor>,
41            already_seen: &mut HashMap<TensorId, bool>,
42        ) -> (bool, Vec<&'a Tensor>) {
43            if let Some(&tg) = already_seen.get(&node.id()) {
44                return (tg, nodes);
45            }
46            let mut track_grad = false;
47            let mut nodes = if node.is_variable() {
48                // Do not call recursively on the "leaf" nodes.
49                track_grad = true;
50                nodes
51            } else if node.dtype().is_int() {
52                nodes
53            } else if let Some(op) = node.op() {
54                match op {
55                    Op::IndexAdd(t1, t2, t3, _)
56                    | Op::Scatter(t1, t2, t3, _)
57                    | Op::ScatterAdd(t1, t2, t3, _)
58                    | Op::CustomOp3(t1, t2, t3, _)
59                    | Op::WhereCond(t1, t2, t3) => {
60                        let (tg, nodes) = walk(t1, nodes, already_seen);
61                        track_grad |= tg;
62                        let (tg, nodes) = walk(t2, nodes, already_seen);
63                        track_grad |= tg;
64                        let (tg, nodes) = walk(t3, nodes, already_seen);
65                        track_grad |= tg;
66                        nodes
67                    }
68                    Op::Conv1D {
69                        arg: lhs,
70                        kernel: rhs,
71                        ..
72                    }
73                    | Op::ConvTranspose1D {
74                        arg: lhs,
75                        kernel: rhs,
76                        ..
77                    }
78                    | Op::Conv2D {
79                        arg: lhs,
80                        kernel: rhs,
81                        ..
82                    }
83                    | Op::ConvTranspose2D {
84                        arg: lhs,
85                        kernel: rhs,
86                        ..
87                    }
88                    | Op::CustomOp2(lhs, rhs, _)
89                    | Op::Binary(lhs, rhs, _)
90                    | Op::Gather(lhs, rhs, _)
91                    | Op::IndexSelect(lhs, rhs, _)
92                    | Op::Matmul(lhs, rhs)
93                    | Op::SliceScatter0(lhs, rhs, _) => {
94                        let (tg, nodes) = walk(lhs, nodes, already_seen);
95                        track_grad |= tg;
96                        let (tg, nodes) = walk(rhs, nodes, already_seen);
97                        track_grad |= tg;
98                        nodes
99                    }
100                    Op::Cat(args, _) => args.iter().fold(nodes, |nodes, arg| {
101                        let (tg, nodes) = walk(arg, nodes, already_seen);
102                        track_grad |= tg;
103                        nodes
104                    }),
105                    Op::Affine { arg, mul, .. } => {
106                        if *mul == 0. {
107                            nodes
108                        } else {
109                            let (tg, nodes) = walk(arg, nodes, already_seen);
110                            track_grad |= tg;
111                            nodes
112                        }
113                    }
114                    Op::Unary(_node, UnaryOp::Ceil)
115                    | Op::Unary(_node, UnaryOp::Floor)
116                    | Op::Unary(_node, UnaryOp::Round)
117                    | Op::Unary(_node, UnaryOp::Sign) => nodes,
118                    Op::Reshape(node)
119                    | Op::UpsampleNearest1D { arg: node, .. }
120                    | Op::UpsampleNearest2D { arg: node, .. }
121                    | Op::UpsampleBilinear2D { arg: node, .. }
122                    | Op::AvgPool2D { arg: node, .. }
123                    | Op::MaxPool2D { arg: node, .. }
124                    | Op::Copy(node)
125                    | Op::Broadcast(node)
126                    | Op::Cmp(node, _)
127                    | Op::Reduce(node, ReduceOp::Min | ReduceOp::Sum | ReduceOp::Max, _)
128                    | Op::ToDevice(node)
129                    | Op::Transpose(node, _, _)
130                    | Op::Permute(node, _)
131                    | Op::Narrow(node, _, _, _)
132                    | Op::Unary(node, _)
133                    | Op::Elu(node, _)
134                    | Op::Powf(node, _)
135                    | Op::CustomOp1(node, _) => {
136                        let (tg, nodes) = walk(node, nodes, already_seen);
137                        track_grad |= tg;
138                        nodes
139                    }
140                    Op::ToDType(node) => {
141                        if node.dtype().is_float() {
142                            let (tg, nodes) = walk(node, nodes, already_seen);
143                            track_grad |= tg;
144                            nodes
145                        } else {
146                            nodes
147                        }
148                    }
149                    Op::Reduce(_, ReduceOp::ArgMin | ReduceOp::ArgMax, _) => nodes,
150                }
151            } else {
152                nodes
153            };
154            already_seen.insert(node.id(), track_grad);
155            if track_grad {
156                nodes.push(node);
157            }
158            (track_grad, nodes)
159        }
160        let (_tg, mut nodes) = walk(self, vec![], &mut HashMap::new());
161        nodes.reverse();
162        nodes
163    }
164
165    pub fn backward(&self) -> Result<GradStore> {
166        let sorted_nodes = self.sorted_nodes();
167        let mut grads = GradStore::new();
168        grads.insert(self, self.ones_like()?.contiguous()?);
169        for node in sorted_nodes.iter() {
170            if node.is_variable() {
171                continue;
172            }
173            let grad = grads
174                .remove(node)
175                .expect("candle internal error - grad not populated");
176            // https://github.com/huggingface/candle/issues/1241
177            // Ideally, we would make these operations in place where possible to ensure that we
178            // do not have to allocate too often. Here we just call `.detach` to avoid computing
179            // the backprop graph of the backprop itself. This would be an issue for second order
180            // derivatives but these are out of scope at the moment.
181            let do_not_detach = CANDLE_GRAD_DO_NOT_DETACH.with(|b| *b);
182            let grad = if do_not_detach { grad } else { grad.detach() };
183            if let Some(op) = node.op() {
184                match op {
185                    Op::Binary(lhs, rhs, BinaryOp::Add) => {
186                        let lhs_sum_grad = grads.or_insert(lhs)?;
187                        *lhs_sum_grad = lhs_sum_grad.add(&grad)?;
188                        let rhs_sum_grad = grads.or_insert(rhs)?;
189                        *rhs_sum_grad = rhs_sum_grad.add(&grad)?;
190                    }
191                    Op::Binary(lhs, rhs, BinaryOp::Sub) => {
192                        let lhs_sum_grad = grads.or_insert(lhs)?;
193                        *lhs_sum_grad = lhs_sum_grad.add(&grad)?;
194                        let rhs_sum_grad = grads.or_insert(rhs)?;
195                        *rhs_sum_grad = rhs_sum_grad.sub(&grad)?;
196                    }
197                    Op::Binary(lhs, rhs, BinaryOp::Mul) => {
198                        let lhs_grad = grad.mul(rhs)?;
199                        let lhs_sum_grad = grads.or_insert(lhs)?;
200                        *lhs_sum_grad = lhs_sum_grad.add(&lhs_grad)?;
201                        let rhs_grad = grad.mul(lhs)?;
202                        let rhs_sum_grad = grads.or_insert(rhs)?;
203                        *rhs_sum_grad = rhs_sum_grad.add(&rhs_grad)?;
204                    }
205                    Op::Binary(lhs, rhs, BinaryOp::Div) => {
206                        let lhs_grad = grad.div(rhs)?;
207                        let lhs_sum_grad = grads.or_insert(lhs)?;
208                        *lhs_sum_grad = lhs_sum_grad.add(&lhs_grad)?;
209                        let rhs_grad = grad.mul(lhs)?.div(&rhs.sqr()?)?;
210                        let rhs_sum_grad = grads.or_insert(rhs)?;
211                        *rhs_sum_grad = rhs_sum_grad.sub(&rhs_grad)?;
212                    }
213                    Op::Binary(lhs, rhs, BinaryOp::Minimum)
214                    | Op::Binary(lhs, rhs, BinaryOp::Maximum) => {
215                        let mask_lhs = node.eq(lhs)?.to_dtype(grad.dtype())?;
216                        let mask_rhs = node.eq(rhs)?.to_dtype(grad.dtype())?;
217
218                        // If both masks are 1 one the same point, we want to scale the
219                        // gradient by 0.5 rather than 1.
220                        let lhs_grad = mask_lhs.mul(&grad)?.div(&(&mask_rhs + 1.)?)?;
221                        let lhs_sum_grad = grads.or_insert(lhs)?;
222                        *lhs_sum_grad = lhs_sum_grad.add(&lhs_grad)?;
223
224                        let rhs_grad = mask_rhs.mul(&grad)?.div(&(&mask_lhs + 1.)?)?;
225                        let rhs_sum_grad = grads.or_insert(rhs)?;
226                        *rhs_sum_grad = rhs_sum_grad.add(&rhs_grad)?;
227                    }
228                    Op::WhereCond(pred, t, f) => {
229                        let zeros = grad.zeros_like()?;
230                        let t_sum_grad = grads.or_insert(t)?;
231                        let t_grad = pred.where_cond(&grad, &zeros)?;
232                        *t_sum_grad = t_sum_grad.add(&t_grad)?;
233                        let f_sum_grad = grads.or_insert(f)?;
234                        let f_grad = pred.where_cond(&zeros, &grad)?;
235                        *f_sum_grad = f_sum_grad.add(&f_grad)?;
236                    }
237                    Op::Conv1D {
238                        arg,
239                        kernel,
240                        padding,
241                        stride,
242                        dilation,
243                    } => {
244                        // The output height for conv_transpose1d is:
245                        // (l_in - 1) * stride - 2 * padding + dilation * (k_size - 1) + out_padding + 1
246                        let grad_l_in = grad.dim(2)?;
247                        let k_size = kernel.dim(2)?;
248                        let out_size =
249                            (grad_l_in - 1) * stride + dilation * (k_size - 1) + 1 - 2 * padding;
250                        let out_padding = arg.dim(2)? - out_size;
251                        let grad_arg = grad.conv_transpose1d(
252                            kernel,
253                            *padding,
254                            out_padding,
255                            *stride,
256                            *dilation,
257                            /* groups */ 1,
258                        )?;
259                        let sum_grad = grads.or_insert(arg)?;
260                        *sum_grad = sum_grad.add(&grad_arg)?;
261
262                        let grad_kernel = arg
263                            .transpose(0, 1)?
264                            .conv1d(&grad.transpose(0, 1)?, *padding, *dilation, *stride, 1)?
265                            .transpose(0, 1)?;
266                        let sum_grad = grads.or_insert(kernel)?;
267                        let (_, _, k0) = kernel.dims3()?;
268                        let (_, _, g_k0) = grad_kernel.dims3()?;
269                        let grad_kernel = if g_k0 != k0 {
270                            grad_kernel.narrow(2, 0, k0)?
271                        } else {
272                            grad_kernel
273                        };
274                        *sum_grad = sum_grad.add(&grad_kernel)?;
275                    }
276                    Op::Conv2D {
277                        arg,
278                        kernel,
279                        padding,
280                        stride,
281                        dilation,
282                    } => {
283                        // The output height for conv_transpose2d is:
284                        // (i_h - 1) * stride - 2 * padding + dilation * (k_h - 1) + out_padding + 1
285                        let grad_h = grad.dim(2)?;
286                        let k_h = kernel.dim(2)?;
287                        let out_size =
288                            (grad_h - 1) * stride + dilation * (k_h - 1) + 1 - 2 * padding;
289                        let out_padding = arg.dim(2)? - out_size;
290                        let grad_arg = grad.conv_transpose2d(
291                            kernel,
292                            *padding,
293                            out_padding,
294                            *stride,
295                            *dilation,
296                        )?;
297                        let sum_grad = grads.or_insert(arg)?;
298                        *sum_grad = sum_grad.add(&grad_arg)?;
299
300                        let grad_kernel = arg
301                            .transpose(0, 1)?
302                            .conv2d(&grad.transpose(0, 1)?, *padding, *dilation, *stride, 1)?
303                            .transpose(0, 1)?;
304                        let sum_grad = grads.or_insert(kernel)?;
305                        let (_, _, k0, k1) = kernel.dims4()?;
306                        let (_, _, g_k0, g_k1) = grad_kernel.dims4()?;
307                        let grad_kernel = if g_k0 != k0 || g_k1 != k1 {
308                            grad_kernel.narrow(2, 0, k0)?.narrow(3, 0, k1)?
309                        } else {
310                            grad_kernel
311                        };
312                        *sum_grad = sum_grad.add(&grad_kernel)?;
313                    }
314                    Op::ConvTranspose1D { .. } => Err(Error::BackwardNotSupported {
315                        op: "conv-transpose1d",
316                    })?,
317                    Op::ConvTranspose2D {
318                        arg,
319                        kernel,
320                        padding,
321                        stride,
322                        dilation,
323                        output_padding: _output_padding,
324                    } => {
325                        let grad_arg = grad.conv2d(kernel, *padding, *stride, *dilation, 1)?;
326                        let sum_grad = grads.or_insert(arg)?;
327                        *sum_grad = sum_grad.add(&grad_arg)?;
328
329                        let grad_kernel = grad
330                            .transpose(0, 1)?
331                            .conv2d(&arg.transpose(0, 1)?, *padding, *dilation, *stride, 1)?
332                            .transpose(0, 1)?;
333                        let sum_grad = grads.or_insert(kernel)?;
334                        let (_, _, k0, k1) = kernel.dims4()?;
335                        let (_, _, g_k0, g_k1) = grad_kernel.dims4()?;
336                        let grad_kernel = if g_k0 != k0 || g_k1 != k1 {
337                            grad_kernel.narrow(2, 0, k0)?.narrow(3, 0, k1)?
338                        } else {
339                            grad_kernel
340                        };
341                        *sum_grad = sum_grad.add(&grad_kernel)?;
342                    }
343                    Op::AvgPool2D {
344                        arg,
345                        kernel_size,
346                        stride,
347                    } => {
348                        if kernel_size != stride {
349                            crate::bail!("backward not supported for avgpool2d if ksize {kernel_size:?} != stride {stride:?}")
350                        }
351                        let (_n, _c, h, w) = arg.dims4()?;
352                        let grad_arg = grad.upsample_nearest2d(h, w)?;
353                        let grad_arg =
354                            (grad_arg * (1f64 / (kernel_size.0 * kernel_size.1) as f64))?;
355                        let sum_grad = grads.or_insert(arg)?;
356                        *sum_grad = sum_grad.add(&grad_arg)?;
357                    }
358                    Op::MaxPool2D {
359                        arg,
360                        kernel_size,
361                        stride,
362                    } => {
363                        if kernel_size != stride {
364                            crate::bail!("backward not supported for maxpool2d if ksize {kernel_size:?} != stride {stride:?}")
365                        }
366                        let (_n, _c, h, w) = arg.dims4()?;
367                        // For computing the max-pool gradient, we compute a mask where a 1 means
368                        // that the element is the maximum, then we apply this mask to the
369                        // upsampled gradient (taking into account that multiple max may exist so
370                        // we scale the gradient for this case).
371                        let node_upsampled = node.upsample_nearest2d(h, w)?;
372                        let mask = arg.eq(&node_upsampled)?.to_dtype(arg.dtype())?;
373                        let avg = mask.avg_pool2d_with_stride(*kernel_size, *stride)?;
374                        let grad_arg = ((grad * avg)?.upsample_nearest2d(h, w)? * mask)?;
375                        let sum_grad = grads.or_insert(arg)?;
376                        *sum_grad = sum_grad.add(&grad_arg)?;
377                    }
378                    Op::UpsampleNearest1D { arg, target_size } => {
379                        let (_n, c, size) = arg.dims3()?;
380                        if target_size % size != 0 {
381                            crate::bail!("backward not supported for non integer upscaling factors")
382                        }
383                        let scale = target_size / size;
384
385                        let kernel = Tensor::ones((c, 1, scale), arg.dtype(), arg.device())?;
386                        let conv_sum = grad.conv1d(&kernel, 0, scale, 1, c)?;
387                        let sum_grad = grads.or_insert(arg)?;
388                        *sum_grad = conv_sum;
389                    }
390                    Op::UpsampleNearest2D {
391                        arg,
392                        target_h,
393                        target_w,
394                    } => {
395                        let (_n, c, h, w) = arg.dims4()?;
396                        if target_h % h != 0 || target_w % w != 0 {
397                            crate::bail!("backward not supported for non integer upscaling factors")
398                        }
399                        let scale_h = target_h / h;
400                        let scale_w = target_w / w;
401
402                        if scale_h != scale_w {
403                            crate::bail!("backward not supported for non uniform upscaling factors")
404                        };
405                        let kernel =
406                            Tensor::ones((c, 1, scale_h, scale_w), arg.dtype(), arg.device())?;
407                        let conv_sum = grad.conv2d(&kernel, 0, scale_h, 1, c)?;
408                        let sum_grad = grads.or_insert(arg)?;
409                        *sum_grad = conv_sum;
410                    }
411                    Op::UpsampleBilinear2D { .. } => {
412                        crate::bail!("backward not supported for upsample_bilinear2d")
413                    }
414                    Op::SliceScatter0(lhs, rhs, start_rhs) => {
415                        let rhs_sum_grad = grads.or_insert(rhs)?;
416                        let rhs_grad = grad.narrow(0, *start_rhs, rhs.dim(0)?)?;
417                        *rhs_sum_grad = rhs_sum_grad.add(&rhs_grad)?;
418
419                        let lhs_sum_grad = grads.or_insert(lhs)?;
420                        let lhs_grad = grad.slice_scatter0(&rhs.zeros_like()?, *start_rhs)?;
421                        *lhs_sum_grad = lhs_sum_grad.add(&lhs_grad)?
422                    }
423                    Op::Gather(arg, indexes, dim) => {
424                        let sum_grad = grads.or_insert(arg)?;
425                        *sum_grad = sum_grad.scatter_add(indexes, &grad, *dim)?;
426                    }
427                    Op::Scatter(init, indexes, src, dim) => {
428                        let init_sum_grad = grads.or_insert(init)?;
429                        *init_sum_grad = init_sum_grad.add(&grad)?;
430
431                        let src_grad = grad.gather(indexes, *dim)?;
432                        let src_sum_grad = grads.or_insert(src)?;
433                        *src_sum_grad = src_sum_grad.add(&src_grad)?;
434                    }
435                    Op::ScatterAdd(init, indexes, src, dim) => {
436                        let init_sum_grad = grads.or_insert(init)?;
437                        let mask = init.ones_like()?;
438                        let mask = mask.scatter(indexes, &mask.zeros_like()?, *dim)?;
439                        *init_sum_grad = init_sum_grad.add(&grad.mul(&mask)?)?;
440
441                        let src_grad = grad.gather(indexes, *dim)?;
442                        let src_sum_grad = grads.or_insert(src)?;
443                        *src_sum_grad = src_sum_grad.add(&src_grad)?;
444                    }
445                    Op::IndexAdd(init, indexes, src, dim) => {
446                        let init_sum_grad = grads.or_insert(init)?;
447                        *init_sum_grad = init_sum_grad.add(&grad)?;
448
449                        let src_grad = grad.index_select(indexes, *dim)?;
450                        let src_sum_grad = grads.or_insert(src)?;
451                        *src_sum_grad = src_sum_grad.add(&src_grad)?;
452                    }
453                    Op::IndexSelect(arg, indexes, dim) => {
454                        let sum_grad = grads.or_insert(arg)?;
455                        // index_select accepts strided ids but index_add requires contiguous ones.
456                        let indexes = indexes.contiguous()?;
457                        *sum_grad = sum_grad.index_add(&indexes, &grad, *dim)?;
458                    }
459                    Op::Matmul(lhs, rhs) => {
460                        // Skipping checks, the op went ok, we can skip
461                        // the matmul size checks for now.
462
463                        let lhs_grad = grad.matmul(&rhs.t()?)?;
464                        let lhs_sum_grad = grads.or_insert(lhs)?;
465                        *lhs_sum_grad = lhs_sum_grad.add(&lhs_grad)?;
466
467                        let rhs_grad = lhs.t()?.matmul(&grad)?;
468                        let rhs_sum_grad = grads.or_insert(rhs)?;
469                        *rhs_sum_grad = rhs_sum_grad.add(&rhs_grad)?;
470                    }
471                    Op::Cat(args, dim) => {
472                        let mut start_idx = 0;
473                        for arg in args {
474                            let len = arg.dims()[*dim];
475                            let arg_grad = grad.narrow(*dim, start_idx, len)?;
476                            let sum_grad = grads.or_insert(arg)?;
477                            *sum_grad = sum_grad.add(&arg_grad)?;
478                            start_idx += len;
479                        }
480                    }
481                    Op::Broadcast(arg) => {
482                        let arg_dims = arg.dims();
483                        let node_dims = node.dims();
484                        // The number of dims that have been inserted on the left.
485                        let left_dims = node_dims.len() - arg_dims.len();
486                        let mut sum_dims: Vec<usize> = (0..left_dims).collect();
487                        for (dim, (node_dim, arg_dim)) in node_dims[left_dims..]
488                            .iter()
489                            .zip(arg_dims.iter())
490                            .enumerate()
491                        {
492                            if node_dim != arg_dim {
493                                sum_dims.push(dim + left_dims)
494                            }
495                        }
496
497                        let mut arg_grad = grad.sum_keepdim(sum_dims.as_slice())?;
498                        for _i in 0..left_dims {
499                            arg_grad = arg_grad.squeeze(0)?
500                        }
501                        let sum_grad = grads.or_insert(arg)?;
502                        *sum_grad = sum_grad.add(&arg_grad.broadcast_as(sum_grad.dims())?)?;
503                    }
504                    Op::Reduce(arg, ReduceOp::Sum, reduced_dims) => {
505                        let grad = broadcast_back(arg, &grad, reduced_dims)?;
506                        let sum_grad = grads.or_insert(arg)?;
507                        *sum_grad = sum_grad.add(&grad)?;
508                    }
509                    Op::Reduce(arg, ReduceOp::Max, reduced_dims) => {
510                        let node = broadcast_back(arg, node, reduced_dims)?;
511                        let grad = broadcast_back(arg, &grad, reduced_dims)?;
512                        let grad = node.eq(arg)?.to_dtype(grad.dtype())?.mul(&grad)?;
513                        let sum_grad = grads.or_insert(arg)?;
514                        *sum_grad = sum_grad.add(&grad.broadcast_as(sum_grad.dims())?)?;
515                    }
516                    Op::Reduce(arg, ReduceOp::Min, reduced_dims) => {
517                        let node = broadcast_back(arg, node, reduced_dims)?;
518                        let grad = broadcast_back(arg, &grad, reduced_dims)?;
519                        let grad = node.eq(arg)?.to_dtype(grad.dtype())?.mul(&grad)?;
520                        let sum_grad = grads.or_insert(arg)?;
521                        *sum_grad = sum_grad.add(&grad.broadcast_as(sum_grad.dims())?)?;
522                    }
523                    Op::ToDType(arg) => {
524                        let sum_grad = grads.or_insert(arg)?;
525                        *sum_grad = sum_grad.add(&grad.to_dtype(arg.dtype())?)?
526                    }
527                    Op::Copy(arg) => {
528                        let sum_grad = grads.or_insert(arg)?;
529                        *sum_grad = sum_grad.add(&grad)?
530                    }
531                    Op::Affine { arg, mul, .. } => {
532                        let arg_grad = grad.affine(*mul, 0.)?;
533                        let sum_grad = grads.or_insert(arg)?;
534                        *sum_grad = sum_grad.add(&arg_grad)?
535                    }
536                    Op::Unary(arg, UnaryOp::Log) => {
537                        let sum_grad = grads.or_insert(arg)?;
538                        *sum_grad = sum_grad.add(&(grad / arg)?)?
539                    }
540                    Op::Unary(arg, UnaryOp::Sin) => {
541                        let sum_grad = grads.or_insert(arg)?;
542                        *sum_grad = sum_grad.add(&(&grad * arg.cos())?)?
543                    }
544                    Op::Unary(arg, UnaryOp::Cos) => {
545                        let sum_grad = grads.or_insert(arg)?;
546                        *sum_grad = sum_grad.sub(&(&grad * arg.sin())?)?
547                    }
548                    Op::Unary(arg, UnaryOp::Tanh) => {
549                        let sum_grad = grads.or_insert(arg)?;
550                        let minus_dtanh = (node.sqr()? - 1.)?;
551                        *sum_grad = sum_grad.sub(&(&grad * &minus_dtanh)?)?
552                    }
553                    Op::Unary(arg, UnaryOp::Abs) => {
554                        let sum_grad = grads.or_insert(arg)?;
555                        let ones = arg.ones_like()?;
556                        let abs_grad = arg.ge(&arg.zeros_like()?)?.where_cond(&ones, &ones.neg()?);
557                        *sum_grad = sum_grad.add(&(&grad * abs_grad)?)?
558                    }
559                    Op::Unary(arg, UnaryOp::Exp) => {
560                        let sum_grad = grads.or_insert(arg)?;
561                        *sum_grad = sum_grad.add(&(&grad * *node)?)?
562                    }
563                    Op::Unary(arg, UnaryOp::Neg) => {
564                        let sum_grad = grads.or_insert(arg)?;
565                        *sum_grad = sum_grad.sub(&grad)?
566                    }
567                    Op::Unary(arg, UnaryOp::Recip) => {
568                        let sum_grad = grads.or_insert(arg)?;
569                        let grad = (grad / arg.sqr()?)?;
570                        *sum_grad = sum_grad.sub(&grad)?
571                    }
572                    &Op::Narrow(ref arg, dim, start_idx, len) => {
573                        let arg_dims = arg.dims();
574                        let left_pad = if start_idx == 0 {
575                            None
576                        } else {
577                            let mut dims = arg_dims.to_vec();
578                            dims[dim] = start_idx;
579                            Some(Tensor::zeros(dims, grad.dtype(), grad.device())?)
580                        };
581                        let right_pad = arg_dims[dim] - start_idx - len;
582                        let right_pad = if right_pad == 0 {
583                            None
584                        } else {
585                            let mut dims = arg_dims.to_vec();
586                            dims[dim] = right_pad;
587                            Some(Tensor::zeros(dims, grad.dtype(), grad.device())?)
588                        };
589                        let arg_grad = match (left_pad, right_pad) {
590                            (None, None) => grad,
591                            (Some(l), None) => Tensor::cat(&[&l, &grad], dim)?,
592                            (None, Some(r)) => Tensor::cat(&[&grad, &r], dim)?,
593                            (Some(l), Some(r)) => Tensor::cat(&[&l, &grad, &r], dim)?,
594                        };
595                        let sum_grad = grads.or_insert(arg)?;
596                        *sum_grad = sum_grad.add(&arg_grad)?
597                    }
598                    Op::Unary(_, UnaryOp::Floor)
599                    | Op::Unary(_, UnaryOp::Round)
600                    | Op::Reduce(_, ReduceOp::ArgMin, _)
601                    | Op::Reduce(_, ReduceOp::ArgMax, _)
602                    | Op::Unary(_, UnaryOp::Sign)
603                    | Op::Cmp(_, _) => {}
604                    Op::Reshape(arg) => {
605                        let arg_grad = grad.reshape(arg.dims())?;
606                        let sum_grad = grads.or_insert(arg)?;
607                        *sum_grad = sum_grad.add(&arg_grad)?
608                    }
609                    Op::Unary(_, UnaryOp::Ceil) => Err(Error::BackwardNotSupported { op: "ceil" })?,
610                    Op::Unary(arg, UnaryOp::Gelu) => {
611                        let sum_grad = grads.or_insert(arg)?;
612                        let cube = arg.powf(3.)?;
613                        let tanh = (0.0356774 * &cube + (0.797885 * arg)?)?.tanh()?;
614                        let gelu_grad = (((0.5 * &tanh)?
615                            + (0.0535161 * cube + (0.398942 * arg)?)? * (1. - tanh.powf(2.)?))?
616                            + 0.5)?;
617                        *sum_grad = sum_grad.add(&(&grad * gelu_grad)?)?
618                    }
619                    Op::Unary(arg, UnaryOp::Erf) => {
620                        let sum_grad = grads.or_insert(arg)?;
621                        // d/dx erf(x) = 2/sqrt(pi) * e^(-x^2)
622                        let erf_grad =
623                            (2. / std::f64::consts::PI.sqrt()) * (arg.sqr()?.neg()?).exp()?;
624                        *sum_grad = sum_grad.add(&(&grad * erf_grad)?)?
625                    }
626                    Op::Unary(arg, UnaryOp::GeluErf) => {
627                        let sum_grad = grads.or_insert(arg)?;
628                        // d/dx gelu_erf(x) = 0.5 + 0.398942 e^(-x^2/2) x + 0.5 erf(x/sqrt(2))
629                        let neg_half_square = (arg.sqr()?.neg()? / 2.)?;
630                        let scaled_exp_arg = (0.398942 * neg_half_square.exp()? * arg)?;
631                        let arg_scaled_sqrt = (arg / 2f64.sqrt())?;
632                        let erf_scaled_sqrt = (0.5 * arg_scaled_sqrt.erf()?)?;
633                        let gelu_erf_grad = (0.5 + scaled_exp_arg + erf_scaled_sqrt)?;
634                        *sum_grad = sum_grad.add(&(&grad * gelu_erf_grad)?)?;
635                    }
636                    Op::Unary(arg, UnaryOp::Relu) => {
637                        let sum_grad = grads.or_insert(arg)?;
638                        let relu_grad = arg.ge(&arg.zeros_like()?)?.to_dtype(arg.dtype())?;
639                        *sum_grad = sum_grad.add(&(&grad * relu_grad)?)?
640                    }
641                    Op::Unary(arg, UnaryOp::Silu) => {
642                        let sum_grad = grads.or_insert(arg)?;
643                        // d/dx silu = sigmoid(x) * (1 + x * (1 - sigmoid(x))) = sigmoid(x) * (1 - node) + node
644                        let sigmoid_arg = (arg.neg()?.exp()? + 1.)?.recip()?;
645                        let silu_grad = &sigmoid_arg * (1. - *node) + *node;
646                        *sum_grad = sum_grad.add(&(&grad * silu_grad)?)?
647                    }
648                    Op::Elu(arg, alpha) => {
649                        // d/dx elu(x) = 1 for x > 0, alpha * e^x for x <= 0
650                        let sum_grad = grads.or_insert(arg)?;
651                        let zeros = arg.zeros_like()?;
652                        let positive_mask = arg.gt(&zeros)?.to_dtype(arg.dtype())?;
653                        let negative_mask = arg.le(&zeros)?.to_dtype(arg.dtype())?;
654                        // node == alpha * (e^x - 1) for x <= 0, reuse it
655                        let negative_exp_mask = (negative_mask * (*node + *alpha))?;
656                        let combined_mask = (positive_mask + negative_exp_mask)?;
657                        *sum_grad = sum_grad.add(&(grad * combined_mask)?)?
658                    }
659                    Op::Powf(arg, e) => {
660                        let arg_grad = (&(grad * arg.powf(e - 1.)?)? * *e)?;
661                        let sum_grad = grads.or_insert(arg)?;
662                        *sum_grad = sum_grad.add(&arg_grad)?
663                    }
664                    Op::CustomOp1(arg, c) => {
665                        if let Some(arg_grad) = c.bwd(arg, node, &grad)? {
666                            let sum_grad = grads.or_insert(arg)?;
667                            *sum_grad = sum_grad.add(&arg_grad)?
668                        }
669                    }
670                    Op::CustomOp2(arg1, arg2, c) => {
671                        let (arg_grad1, arg_grad2) = c.bwd(arg1, arg2, node, &grad)?;
672                        if let Some(arg_grad1) = arg_grad1 {
673                            let sum_grad = grads.or_insert(arg1)?;
674                            *sum_grad = sum_grad.add(&arg_grad1)?
675                        }
676                        if let Some(arg_grad2) = arg_grad2 {
677                            let sum_grad = grads.or_insert(arg2)?;
678                            *sum_grad = sum_grad.add(&arg_grad2)?
679                        }
680                    }
681                    Op::CustomOp3(arg1, arg2, arg3, c) => {
682                        let (arg_grad1, arg_grad2, arg_grad3) =
683                            c.bwd(arg1, arg2, arg3, node, &grad)?;
684                        if let Some(arg_grad1) = arg_grad1 {
685                            let sum_grad = grads.or_insert(arg1)?;
686                            *sum_grad = sum_grad.add(&arg_grad1)?
687                        }
688                        if let Some(arg_grad2) = arg_grad2 {
689                            let sum_grad = grads.or_insert(arg2)?;
690                            *sum_grad = sum_grad.add(&arg_grad2)?
691                        }
692                        if let Some(arg_grad3) = arg_grad3 {
693                            let sum_grad = grads.or_insert(arg3)?;
694                            *sum_grad = sum_grad.add(&arg_grad3)?
695                        }
696                    }
697                    Op::Unary(arg, UnaryOp::Sqr) => {
698                        let arg_grad = arg.mul(&grad)?.affine(2., 0.)?;
699                        let sum_grad = grads.or_insert(arg)?;
700                        *sum_grad = sum_grad.add(&arg_grad)?
701                    }
702                    Op::Unary(arg, UnaryOp::Sqrt) => {
703                        let arg_grad = grad.div(node)?.affine(0.5, 0.)?;
704                        let sum_grad = grads.or_insert(arg)?;
705                        *sum_grad = sum_grad.add(&arg_grad)?
706                    }
707                    Op::ToDevice(arg) => {
708                        let sum_grad = grads.or_insert(arg)?;
709                        let arg_grad = grad.to_device(sum_grad.device())?;
710                        *sum_grad = sum_grad.add(&arg_grad)?
711                    }
712                    Op::Transpose(arg, dim1, dim2) => {
713                        let arg_grad = grad.transpose(*dim1, *dim2)?;
714                        let sum_grad = grads.or_insert(arg)?;
715                        *sum_grad = sum_grad.add(&arg_grad)?
716                    }
717                    Op::Permute(arg, dims) => {
718                        let mut inv_dims = vec![0; dims.len()];
719                        for (i, &dim_idx) in dims.iter().enumerate() {
720                            inv_dims[dim_idx] = i
721                        }
722                        let arg_grad = grad.permute(inv_dims)?;
723                        let sum_grad = grads.or_insert(arg)?;
724                        *sum_grad = sum_grad.add(&arg_grad)?
725                    }
726                };
727            }
728        }
729        Ok(grads)
730    }
731}
732
733/// A store for gradients, associating a tensor id to the corresponding gradient tensor, used for back propagation.
734#[derive(Default, Debug)]
735pub struct GradStore(HashMap<TensorId, Tensor>);
736
737impl GradStore {
738    /// Create a new gradient store
739    fn new() -> Self {
740        Self::default()
741    }
742
743    /// Get the gradient tensor corresponding to the given tensor id
744    pub fn get_id(&self, id: TensorId) -> Option<&Tensor> {
745        self.0.get(&id)
746    }
747
748    /// Get the gradient tensor associated with the given tensor
749    pub fn get(&self, tensor: &Tensor) -> Option<&Tensor> {
750        self.0.get(&tensor.id())
751    }
752
753    /// Remove the gradient tensor associated with the given tensor, returning it if it exists
754    pub fn remove(&mut self, tensor: &Tensor) -> Option<Tensor> {
755        self.0.remove(&tensor.id())
756    }
757
758    /// Insert a gradient tensor associated with the given tensor, returning the previous gradient tensor if it existed
759    pub fn insert(&mut self, tensor: &Tensor, grad: Tensor) -> Option<Tensor> {
760        self.0.insert(tensor.id(), grad)
761    }
762
763    /// Insert a gradient tensor associated with the given tensor id, returning the previous gradient tensor if it existed
764    pub fn insert_id(&mut self, id: TensorId, grad: Tensor) -> Option<Tensor> {
765        self.0.insert(id, grad)
766    }
767
768    /// Get the gradient tensor associated with the given tensor, or, if it does not exist,
769    /// insert a tensor of zeroes, with the same shape and type as the given tensors and return it
770    fn or_insert(&mut self, tensor: &Tensor) -> Result<&mut Tensor> {
771        let grad = match self.0.entry(tensor.id()) {
772            Entry::Occupied(entry) => entry.into_mut(),
773            Entry::Vacant(entry) => {
774                let grad = tensor.zeros_like()?;
775                entry.insert(grad)
776            }
777        };
778        Ok(grad)
779    }
780
781    /// Extend this gradient store with the contents of another.
782    /// If an entry is already occupied, updates and ensures new tensor follows correct detach semantics.
783    /// Otherwise simply inserts.
784    pub fn extend(&mut self, other: Self) -> Result<()> {
785        for (id, grad) in other.0 {
786            match self.0.entry(id) {
787                Entry::Occupied(mut entry) => {
788                    let new_grad = entry.get().add(&grad)?;
789
790                    let do_not_detach = CANDLE_GRAD_DO_NOT_DETACH.with(|b| *b);
791                    let new_grad = if do_not_detach {
792                        new_grad
793                    } else {
794                        new_grad.detach()
795                    };
796
797                    *entry.get_mut() = new_grad;
798                }
799                Entry::Vacant(entry) => {
800                    entry.insert(grad);
801                }
802            }
803        }
804        Ok(())
805    }
806
807    /// Get the tensor ids of the stored gradient tensors
808    pub fn get_ids(&self) -> impl Iterator<Item = &TensorId> {
809        self.0.keys()
810    }
811}
candle_core/backprop.rs

candle_core/
backprop.rs
