fix bug of celoss when using ignore_index and reduction (#30180)

* fix bug of using ignore_index and reduction,test=develop

* fix bug of celoss when using ignore_index and reduction, test=develop

* improve performance when ignore_index=-100, test=develop

* add test in test_cross_entropy_loss.py for coverage rate, test=develop

* rm comment in test_cross_entropy_loss.py, test=develop

* del  hard code of "float64" in python/paddle/nn/functional/loss.py, test=develop

* change mask to a more simplified implementation, test=develop

* del comment in python/paddle/nn/functional/loss.py, test=develop

* del hard code and change mask to a more simplified implementation, test=develop

* change mask to a more simplified implementation, test=develop

* change mask to a more simplified implementation, test=develop

python/paddle/fluid/tests/unittests/test_cross_entropy_loss.py

@@ -93,6 +93,90 @@ def cross_entropy_loss_2d(input,
 
 
 class CrossEntropyLoss(unittest.TestCase):
+    def test_cross_entropy_loss_1d_with_mean_ignore(self):
+        input_np = np.random.random([2, 4]).astype(np.float64)
+        label_np = np.random.randint(0, 4, size=(2)).astype(np.int64)
+        paddle.enable_static()
+        prog = fluid.Program()
+        startup_prog = fluid.Program()
+        place = fluid.CUDAPlace(0) if fluid.core.is_compiled_with_cuda(
+        ) else fluid.CPUPlace()
+        with fluid.program_guard(prog, startup_prog):
+            input = fluid.data(name='input', shape=[2, 4], dtype='float64')
+            label = fluid.data(name='label', shape=[2], dtype='int64')
+            cross_entropy_loss = paddle.nn.loss.CrossEntropyLoss(ignore_index=0)
+            ret = cross_entropy_loss(input, label)
+
+            exe = fluid.Executor(place)
+            static_ret = exe.run(prog,
+                                 feed={
+                                     'input': input_np,
+                                     'label': label_np,
+                                 },
+                                 fetch_list=[ret])
+            self.assertIsNotNone(static_ret)
+        expected = cross_entropy_loss_1d(input_np, label_np)[0]
+
+        with fluid.dygraph.guard():
+            cross_entropy_loss = paddle.nn.loss.CrossEntropyLoss(
+                axis=1, ignore_index=0)
+            dy_ret = cross_entropy_loss(
+                fluid.dygraph.to_variable(input_np),
+                fluid.dygraph.to_variable(label_np))
+            dy_ret_value = dy_ret.numpy()
+            self.assertIsNotNone(dy_ret_value)
+        expected = cross_entropy_loss_1d(input_np, label_np, ignore_index=0)[0]
+        self.assertTrue(np.allclose(static_ret, dy_ret_value))
+        self.assertTrue(np.allclose(static_ret, expected))
+        self.assertTrue(np.allclose(dy_ret_value, expected))
+
+    def test_cross_entropy_loss_1d_with_weight_mean_ignore(self):
+        input_np = np.random.random([2, 4]).astype(np.float64)
+        label_np = np.random.randint(0, 4, size=(2)).astype(np.int64)
+        weight_np = np.random.random([4]).astype(np.float64)  #shape:C
+        paddle.enable_static()
+        prog = fluid.Program()
+        startup_prog = fluid.Program()
+        place = fluid.CUDAPlace(0) if fluid.core.is_compiled_with_cuda(
+        ) else fluid.CPUPlace()
+        with fluid.program_guard(prog, startup_prog):
+            input = fluid.data(name='input', shape=[2, 4], dtype='float64')
+            label = fluid.data(name='label', shape=[2], dtype='int64')
+            weight = fluid.data(
+                name='weight', shape=[4],
+                dtype='float64')  #weight for each class
+            cross_entropy_loss = paddle.nn.loss.CrossEntropyLoss(
+                weight=weight, ignore_index=0)
+            ret = cross_entropy_loss(input, label)
+
+            exe = fluid.Executor(place)
+            static_ret = exe.run(prog,
+                                 feed={
+                                     'input': input_np,
+                                     'label': label_np,
+                                     "weight": weight_np
+                                 },
+                                 fetch_list=[ret])
+            self.assertIsNotNone(static_ret)
+        expected = cross_entropy_loss_1d(
+            input_np, label_np, weight=weight_np)[0]
+
+        with fluid.dygraph.guard():
+            cross_entropy_loss = paddle.nn.loss.CrossEntropyLoss(
+                weight=fluid.dygraph.to_variable(weight_np),
+                axis=1,
+                ignore_index=0)
+            dy_ret = cross_entropy_loss(
+                fluid.dygraph.to_variable(input_np),
+                fluid.dygraph.to_variable(label_np))
+            dy_ret_value = dy_ret.numpy()
+            self.assertIsNotNone(dy_ret_value)
+        expected = cross_entropy_loss_1d(
+            input_np, label_np, weight=weight_np, ignore_index=0)[0]
+        self.assertTrue(np.allclose(static_ret, dy_ret_value))
+        self.assertTrue(np.allclose(static_ret, expected))
+        self.assertTrue(np.allclose(dy_ret_value, expected))
+
     def test_cross_entropy_loss_1d_with_weight_mean(self):
         input_np = np.random.random([2, 4]).astype(np.float64)
         label_np = np.random.randint(0, 4, size=(2)).astype(np.int64)

python/paddle/nn/functional/loss.py

@@ -1223,22 +1223,50 @@ def cross_entropy(input,
             ignore_index=ignore_index,
             axis=axis)
         if weight is not None:
-            weight_gather = core.ops.gather_nd(weight, label)  #trans to sample
+            weight_gather = core.ops.gather_nd(
+                weight, label)  #trans weight from class to sample, shape:N
             input_shape = list(label.shape)
-            weight_gather_reshape, _ = core.ops.reshape2(weight_gather, None,
-                                                         'shape', input_shape)
+            weight_gather_reshape = reshape(weight_gather, shape=input_shape)
             out = core.ops.elementwise_mul(out, weight_gather_reshape)
 
         if reduction == "sum":
+            #   because of softmax_with_cross_entropy op's inner logic, 
+            #   in the out tensor of this op, the loss of sample with class_index==ignore_index is 0
+            #   so, reduce_sum all directly is ok
             return core.ops.reduce_sum(out, 'reduce_all', True)
         elif reduction == "mean":
-            if weight is not None:
+            #1. if weight==none, 
+            #    numerator: reduce_sum all loss directly is ok causeof softmax_with_cross_entropy's inner logic
+            #    denominator: count sample num with class_index!=ignore_index
+            #2. else
+            #    numerator: loss's weighted sum 
+            #    denominator: cal the sum of weight where the sample's class_index!=ignore_index
+            if ignore_index != -100:
+                out_sum = core.ops.reduce_sum(out, 'reduce_all', True)
+                #for each label[i],set 1 or 0, according to ignore_index
+                #mask[i]=0, if label[i]==ignore_index
+                #mask[i]=1, otherwise 
+                mask = (label != ignore_index)
+                if (weight is None):
+                    mask = paddle.cast(mask, dtype=out_sum.dtype)
+                    count = core.ops.reduce_sum(mask, 'reduce_all', True)
+                    ret = out_sum / count
+                else:
+                    mask = paddle.cast(mask, weight_gather_reshape.dtype)
+                    weight_ignored = core.ops.elementwise_mul(
+                        mask, weight_gather_reshape)
+                    weight_sum = core.ops.reduce_sum(weight_ignored,
+                                                     'reduce_all', True)
+                    ret = out_sum / weight_sum
+                return ret
+            elif weight is not None:
                 out_sum = core.ops.reduce_sum(out, 'reduce_all', True)
                 total_weight = core.ops.reduce_sum(weight_gather_reshape,
                                                    'reduce_all', True)
                 return out_sum / total_weight
             else:
                 return core.ops.mean(out)
+
         else:
             if input_dims - 1 == label_dims:
                 out = paddle.squeeze(out, axis=axis)
@@ -1258,7 +1286,8 @@ def cross_entropy(input,
         fluid.data_feeder.check_variable_and_dtype(
             weight, 'weight', ['float32', 'float64'], 'softmax_cross_entropy')
         weight_name = name if reduction == 'none' else None
-        weight_gather = paddle.gather_nd(weight, label)  #trans to sample
+        weight_gather = paddle.gather_nd(
+            weight, label)  #trans weight from class to sample, shape:N
         input_shape = list(label.shape)
         weight_gather_reshape = reshape(weight_gather, shape=input_shape)
         out = paddle.multiply(out, weight_gather_reshape, name=weight_name)
@@ -1266,12 +1295,29 @@ def cross_entropy(input,
     if reduction == "sum":
         return paddle.sum(out, name=name)
     elif reduction == "mean":
-        if weight is not None:
+        if ignore_index != -100:
+            out_sum = paddle.sum(out, name=name)
+            #for each label[i],set 1 or 0, according to ignore_index
+            #mask[i]=0, if label[i]==ignore_index
+            #mask[i]=1, otherwise 
+            mask = (label != ignore_index)
+            if (weight is None):
+                mask = paddle.cast(mask, dtype=out_sum.dtype)
+                count = paddle.sum(mask, name=name)
+                ret = out_sum / count
+            else:
+                mask = paddle.cast(mask, weight_gather_reshape.dtype)
+                weight_ignored = paddle.multiply(mask, weight_gather_reshape)
+                weight_sum = paddle.sum(weight_ignored, name=name)
+                ret = out_sum / weight_sum
+            return ret
+        elif weight is not None:
             out_sum = paddle.sum(out, name=name)
             total_weight = paddle.sum(weight_gather_reshape)
             return out_sum / total_weight
         else:
             return paddle.mean(out, name=name)
+
     else:
         if input_dims - 1 == label_dims:
             out = paddle.squeeze(out, axis=axis)