[Cherry-Pick]Fix bug where embedding can‘t be processed correctly in reducer (#29490)

* fix the bug of reducer in embedding

python/paddle/fluid/dygraph/parallel.py

@@ -24,8 +24,8 @@ from paddle.fluid.dygraph import layers
 from paddle.fluid.dygraph import parallel_helper
 from paddle.fluid.dygraph import to_variable, no_grad
 from paddle.utils import deprecated
-from paddle.fluid.dygraph import nn
 import warnings
+import paddle
 
 __all__ = ["prepare_context", "ParallelEnv", "DataParallel"]
 
@@ -419,9 +419,13 @@ class DataParallel(layers.Layer):
         # NOTE(shenliang03): Here we can only use the attributes to judge whether
         # parameter is sparse(or SelectedRows). The reason is that the sparse message
         # can't be obtained when bp hasn't happened yet. So if layer supports sparse parameter,
-        # we should add the layer here like "nn.Embedding".
+        # we should add the layer here like "paddle.nn.layer.common.Embedding".
         def check_layer_sparse(sublayer):
-            if isinstance(sublayer, nn.Embedding):
+            if isinstance(sublayer, paddle.nn.layer.common.Embedding):
+                return sublayer._sparse
+            # NOTE(shenliang03):This is for compatibility. If paddle.fluid.dygraph.Embedding 
+            # is removed in the future, the judgment will also be removed here.
+            if isinstance(sublayer, paddle.fluid.dygraph.Embedding):
                 return sublayer._is_sparse
             return False
 

python/paddle/fluid/tests/unittests/parallel_dygraph_sparse_embedding_fp64.py

@@ -15,14 +15,60 @@
 from __future__ import print_function
 
 import numpy as np
-
 import paddle
-import paddle.fluid as fluid
-from paddle.fluid.dygraph.nn import Embedding
-from paddle.fluid.dygraph.base import to_variable
 
 from test_dist_base import runtime_main, TestParallelDyGraphRunnerBase
-from parallel_dygraph_sparse_embedding import SimpleNet, fake_sample_reader, TestSparseEmbedding
+from paddle.nn import Layer, Embedding
+paddle.set_default_dtype("float64")
+
+
+class SimpleNet(Layer):
+    def __init__(self,
+                 hidden_size,
+                 vocab_size,
+                 num_steps=20,
+                 init_scale=0.1,
+                 is_sparse=False,
+                 dtype="float64"):
+        super(SimpleNet, self).__init__()
+        self.hidden_size = hidden_size
+        self.vocab_size = vocab_size
+        self.init_scale = init_scale
+        self.num_steps = num_steps
+        self.embedding = Embedding(
+            self.vocab_size,
+            self.hidden_size,
+            sparse=True,
+            weight_attr=paddle.ParamAttr(
+                name='embedding_param',
+                initializer=paddle.nn.initializer.Uniform(
+                    low=-init_scale, high=init_scale)))
+        self.softmax_weight = self.create_parameter(
+            attr=paddle.ParamAttr(),
+            shape=[self.hidden_size, self.vocab_size],
+            dtype=dtype,
+            default_initializer=paddle.nn.initializer.Uniform(
+                low=-self.init_scale, high=self.init_scale))
+        self.softmax_bias = self.create_parameter(
+            attr=paddle.ParamAttr(),
+            shape=[self.vocab_size],
+            dtype=dtype,
+            default_initializer=paddle.nn.initializer.Uniform(
+                low=-self.init_scale, high=self.init_scale))
+
+    def forward(self, input, label):
+        x_emb = self.embedding(input)
+        fc = paddle.matmul(x_emb, self.softmax_weight)
+        fc = paddle.add(fc, self.softmax_bias)
+        projection = paddle.reshape(fc, shape=[-1, self.vocab_size])
+        loss = paddle.nn.functional.softmax_with_cross_entropy(
+            logits=projection, label=label, soft_label=False)
+        loss = paddle.reshape(loss, shape=[-1, self.num_steps])
+        loss = paddle.mean(loss, axis=[0])
+        loss = paddle.sum(loss)
+
+        return loss
+
 
 # global configs
 batch_size = 4
@@ -33,24 +79,46 @@ num_steps = 3
 init_scale = 0.1
 
 
-class TestSparseEmbeddingFP64(TestSparseEmbedding):
+def fake_sample_reader():
+    def __reader__():
+        for i in range(batch_num):
+            x_data = np.arange(num_steps).astype('int64')
+            y_data = np.arange(1, 1 + num_steps).astype('int64')
+            yield x_data, y_data
+
+    return __reader__
+
+
+class TestSparseEmbeddingFP64(TestParallelDyGraphRunnerBase):
     def get_model(self):
         model = SimpleNet(
             hidden_size=hidden_size,
             vocab_size=vocab_size,
             num_steps=num_steps,
             init_scale=init_scale,
-            is_sparse=True,
-            dtype="float64")
+            is_sparse=True)
 
         train_reader = paddle.batch(
             fake_sample_reader(), batch_size=batch_size, drop_last=True)
 
-        optimizer = fluid.optimizer.SGD(learning_rate=0.001,
-                                        parameter_list=model.parameters())
+        optimizer = paddle.optimizer.SGD(learning_rate=0.001,
+                                         parameters=model.parameters())
 
         return model, train_reader, optimizer
 
+    def run_one_loop(self, model, optimizer, batch):
+        x_data = np.array([x[0].reshape(3) for x in batch]).astype('int64')
+        y_data = np.array([x[1].reshape(3) for x in batch]).astype('int64')
+        x_data = x_data.reshape((-1, num_steps, 1))
+        y_data = y_data.reshape((-1, 1))
+
+        x = paddle.to_tensor(x_data)
+        y = paddle.to_tensor(y_data)
+
+        dy_loss = model(x, y)
+
+        return dy_loss
+
 
 if __name__ == "__main__":
     runtime_main(TestSparseEmbeddingFP64)