Merge pull request #24 from LaurentMazare/more-grads

Support gradients for reshape and where_cond.

2 files changed, 17 insertions, 6 deletions

diff --git a/candle-core/src/backprop.rs b/candle-core/src/backprop.rs
index bc6740cf..ef15e65f 100644
--- a/candle-core/src/backprop.rs
+++ b/candle-core/src/backprop.rs
@@ -106,9 +106,8 @@ impl Tensor {
             }
             let grad = grads.remove(node).unwrap();
             // TODO: We should perform all these operations in place (or at least not track the
-            // whole graph).
-            // The only drawback would be if we wanted to support grad of grad but this is out of
-            // scope.
+            // whole graph). The only drawback would be if we wanted to support grad of grad but
+            // this is out of scope.
             if let Some(op) = node.op() {
                 match op {
                     Op::Add(lhs, rhs) => {
@@ -139,8 +138,14 @@ impl Tensor {
                         let rhs_sum_grad = grads.or_insert(rhs)?;
                         *rhs_sum_grad = rhs_sum_grad.add(&rhs_grad)?;
                     }
-                    Op::WhereCond(_pred, _t, _f) => {
-                        return Err(Error::BackwardNotSupported { op: "where_cond" })
+                    Op::WhereCond(pred, t, f) => {
+                        let zeros = grad.zeros_like()?;
+                        let t_sum_grad = grads.or_insert(t)?;
+                        let t_grad = pred.where_cond(&grad, &zeros)?;
+                        *t_sum_grad = t_sum_grad.add(&t_grad)?;
+                        let f_sum_grad = grads.or_insert(f)?;
+                        let f_grad = pred.where_cond(&zeros, &grad)?;
+                        *f_sum_grad = f_sum_grad.add(&f_grad)?;
                     }
                     Op::Embedding(_lhs, _rhs) => {
                         return Err(Error::BackwardNotSupported { op: "embedding" })
@@ -209,7 +214,11 @@ impl Tensor {
                     Op::Softmax(_arg, _) => {
                         return Err(Error::BackwardNotSupported { op: "softmax" })
                     }
-                    Op::Reshape(_arg) => return Err(Error::BackwardNotSupported { op: "reshape" }),
+                    Op::Reshape(arg) => {
+                        let arg_grad = grad.reshape(arg.dims())?;
+                        let sum_grad = grads.or_insert(arg)?;
+                        *sum_grad = sum_grad.add(&arg_grad)?
+                    }
                     Op::Gelu(_) => return Err(Error::BackwardNotSupported { op: "gelu" }),
                     Op::Relu(_) => return Err(Error::BackwardNotSupported { op: "relu" }),
                     Op::Sqr(arg) => {
diff --git a/candle-core/src/tensor.rs b/candle-core/src/tensor.rs
index 4cff4efc..feb59d3c 100644
--- a/candle-core/src/tensor.rs
+++ b/candle-core/src/tensor.rs
@@ -121,6 +121,7 @@ fn from_storage<S: Into<Shape>>(
 }
 
 impl Tensor {
+    // TODO: Maybe this should be a broadcast rather than actually creating the full tensor.
     fn ones_impl<S: Into<Shape>>(
         shape: S,
         dtype: DType,
@@ -144,6 +145,7 @@ impl Tensor {
         Tensor::ones(self.shape(), self.dtype(), &self.device())
     }
 
+    // TODO: Maybe this should be a broadcast rather than actually creating the full tensor.
     fn zeros_impl<S: Into<Shape>>(
         shape: S,
         dtype: DType,