Enable eager model test (#40154)

* enable eager model; test=develop

* set bs = 5; test=develop

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

@@ -22,6 +22,7 @@ from paddle.fluid import core
 from paddle.fluid.dygraph.nn import Conv2D, Pool2D, Linear, BatchNorm, Embedding, GRUUnit
 from paddle.fluid.dygraph.base import to_variable
 from test_imperative_base import new_program_scope
+from paddle.fluid.framework import _test_eager_guard
 
 
 class Config(object):
@@ -371,7 +372,7 @@ class OCRAttention(fluid.dygraph.Layer):
 
 
 class TestDygraphOCRAttention(unittest.TestCase):
-    def test_while_op(self):
+    def test_ocr_test(self):
         seed = 90
         epoch_num = 1
         if core.is_compiled_with_cuda():
@@ -400,7 +401,7 @@ class TestDygraphOCRAttention(unittest.TestCase):
                 i * Config.max_length,
                 dtype='int64').reshape([1, Config.max_length])))
 
-        with fluid.dygraph.guard():
+        def run_dygraph():
             fluid.set_flags({'FLAGS_sort_sum_gradient': True})
             paddle.seed(seed)
             paddle.framework.random._manual_program_seed(seed)
@@ -452,6 +453,16 @@ class TestDygraphOCRAttention(unittest.TestCase):
                     for param in ocr_attention.parameters():
                         dy_param_value[param.name] = param.numpy()
 
+            return dy_out, dy_param_init_value, dy_param_value
+
+        with fluid.dygraph.guard():
+            dy_out, dy_param_init_value, dy_param_value = run_dygraph()
+
+        with fluid.dygraph.guard():
+            with _test_eager_guard():
+                eager_out, eager_param_init_value, eager_param_value = run_dygraph(
+                )
+
         with new_program_scope():
             paddle.seed(seed)
             paddle.framework.random._manual_program_seed(seed)
@@ -537,6 +548,17 @@ class TestDygraphOCRAttention(unittest.TestCase):
         for key, value in six.iteritems(static_param_value):
             self.assertTrue(np.allclose(value, dy_param_value[key], rtol=1e-05))
 
+        # check eager here
+        self.assertTrue(np.allclose(static_out, eager_out))
+
+        for key, value in six.iteritems(static_param_init_value):
+            self.assertTrue(np.array_equal(value, eager_param_init_value[key]))
+
+        for key, value in six.iteritems(static_param_value):
+            self.assertTrue(
+                np.allclose(
+                    value, eager_param_value[key], rtol=1e-05))
+
 
 if __name__ == '__main__':
     paddle.enable_static()

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

@@ -27,6 +27,7 @@ from paddle.fluid.dygraph.nn import Conv2D, Pool2D, Linear
 import paddle.fluid.dygraph.nn as nn
 from paddle.fluid.dygraph.base import to_variable
 from test_imperative_base import new_program_scope
+from paddle.fluid.framework import _test_eager_guard
 
 
 class Policy(fluid.dygraph.Layer):
@@ -63,7 +64,7 @@ class TestImperativeMnist(unittest.TestCase):
         mask_list = [[0, 1]]
         mask = np.array(mask_list).astype("float32")
 
-        with fluid.dygraph.guard():
+        def run_dygraph():
             paddle.seed(seed)
             paddle.framework.random._manual_program_seed(seed)
 
@@ -104,6 +105,16 @@ class TestImperativeMnist(unittest.TestCase):
             for param in policy.parameters():
                 dy_param_value[param.name] = param.numpy()
 
+            return dy_out, dy_param_init_value, dy_param_value
+
+        with fluid.dygraph.guard():
+            dy_out, dy_param_init_value, dy_param_value = run_dygraph()
+
+        with fluid.dygraph.guard():
+            with _test_eager_guard():
+                eager_out, eager_param_init_value, eager_param_value = run_dygraph(
+                )
+
         with new_program_scope():
             paddle.seed(seed)
             paddle.framework.random._manual_program_seed(seed)
@@ -171,6 +182,16 @@ class TestImperativeMnist(unittest.TestCase):
         for key, value in six.iteritems(static_param_value):
             self.assertTrue(np.equal(value, dy_param_value[key]).all())
 
+        # check eager
+        for key, value in six.iteritems(static_param_init_value):
+            self.assertTrue(np.equal(value, eager_param_init_value[key]).all())
+
+        self.assertTrue(np.equal(static_out, eager_out).all())
+
+        for key, value in six.iteritems(static_param_value):
+            self.assertTrue(np.equal(value, eager_param_value[key]).all())
+
 
 if __name__ == '__main__':
+    paddle.enable_static()
     unittest.main()

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

@@ -24,6 +24,7 @@ from paddle.fluid.layer_helper import LayerHelper
 from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear
 from paddle.fluid.dygraph.base import to_variable
 from test_imperative_base import new_program_scope
+from paddle.fluid.framework import _test_eager_guard
 
 if fluid.is_compiled_with_cuda():
     fluid.set_flags({'FLAGS_cudnn_deterministic': True})
@@ -310,7 +311,8 @@ class TestImperativeResneXt(unittest.TestCase):
         batch_size = train_parameters["batch_size"]
         batch_num = 1
         epoch_num = 1
-        with fluid.dygraph.guard():
+
+        def run_dygraph():
             paddle.seed(seed)
             paddle.framework.random._manual_program_seed(seed)
 
@@ -371,6 +373,17 @@ class TestImperativeResneXt(unittest.TestCase):
                     for param in se_resnext.parameters():
                         dy_param_value[param.name] = param.numpy()
 
+                    return dy_out, dy_param_init_value, dy_param_value, dy_grad_value
+
+        with fluid.dygraph.guard():
+            dy_out, dy_param_init_value, dy_param_value, dy_grad_value = run_dygraph(
+            )
+
+        with fluid.dygraph.guard():
+            with _test_eager_guard():
+                eager_out, eager_param_init_value, eager_param_value, eager_grad_value = run_dygraph(
+                )
+
         with new_program_scope():
             paddle.seed(seed)
             paddle.framework.random._manual_program_seed(seed)
@@ -479,6 +492,32 @@ class TestImperativeResneXt(unittest.TestCase):
             self.assertTrue(np.isfinite(value.all()))
             self.assertFalse(np.isnan(value.any()))
 
+        # check eager
+        self.assertTrue(
+            np.allclose(static_out, eager_out),
+            "\nstatic_out: {}\neager_out: {}".format(static_out, eager_out))
+
+        self.assertEqual(
+            len(eager_param_init_value), len(static_param_init_value))
+
+        for key, value in six.iteritems(static_param_init_value):
+            self.assertTrue(np.allclose(value, eager_param_init_value[key]))
+
+        self.assertEqual(len(eager_grad_value), len(static_grad_value))
+
+        for key, value in six.iteritems(static_grad_value):
+            self.assertTrue(
+                np.allclose(value, eager_grad_value[key]),
+                "\nstatic_grad_value: {}\neager_grad_value: {}".format(
+                    value, eager_grad_value[key]))
+
+        self.assertEqual(len(eager_param_value), len(static_param_value))
+        for key, value in six.iteritems(static_param_value):
+            self.assertTrue(
+                np.allclose(value, eager_param_value[key]),
+                "\nstatic_param_value: {}\neagear_param_value: {}".format(
+                    value, eager_param_value[key]))
+
 
 if __name__ == '__main__':
     paddle.enable_static()

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

@@ -21,6 +21,7 @@ from paddle.fluid import Embedding, LayerNorm, Linear, Layer
 from paddle.fluid.dygraph import to_variable, guard
 from paddle.fluid.dygraph import TracedLayer
 from test_imperative_base import new_program_scope
+from paddle.fluid.framework import _test_eager_guard
 from paddle.fluid import core
 import numpy as np
 import six
@@ -949,8 +950,7 @@ class TestDygraphTransformerSortGradient(unittest.TestCase):
     def transformer_sort_gradient_float32(self, is_sparse):
         seed = 90
 
-        with guard():
-            fluid.set_flags({'FLAGS_sort_sum_gradient': True})
+        def run_dygraph():
             # NOTE(xiongkun03): In new executor, the inplace strategy is on by default, which will cause result of sumop have some differences. So we disable inplace.
             fluid.set_flags({'FLAGS_new_executor_use_inplace': False})
             paddle.seed(seed)
@@ -998,7 +998,7 @@ class TestDygraphTransformerSortGradient(unittest.TestCase):
 
             for i in range(batch_num):
                 enc_inputs, dec_inputs, label, weights = create_data()
-                if i % 2 == 0:
+                if False:
                     outs, traced_layer = TracedLayer.trace(
                         transformer, [enc_inputs, dec_inputs, label, weights])
 
@@ -1036,6 +1036,14 @@ class TestDygraphTransformerSortGradient(unittest.TestCase):
             dy_predict_value = dy_predict.numpy()
             dy_token_num_value = dy_token_num.numpy()
 
+            return dy_avg_cost_value, dy_sum_cost_value, dy_predict_value, dy_token_num_value, \
+                dy_param_init, dy_param_updated
+
+        with guard():
+            fluid.set_flags({'FLAGS_sort_sum_gradient': True})
+            dy_avg_cost_value, dy_sum_cost_value, dy_predict_value, dy_token_num_value, \
+                dy_param_init, dy_param_updated = run_dygraph()
+
         with new_program_scope():
             paddle.seed(seed)
             paddle.framework.random._manual_program_seed(seed)
@@ -1122,6 +1130,28 @@ class TestDygraphTransformerSortGradient(unittest.TestCase):
         for key, value in six.iteritems(static_param_updated):
             self.assertTrue(np.array_equal(value, dy_param_updated[key]))
 
+        # check eager result
+        with guard():
+            fluid.set_flags({'FLAGS_sort_sum_gradient': False})
+            dy_avg_cost_value, dy_sum_cost_value, dy_predict_value, dy_token_num_value, \
+                dy_param_init, dy_param_updated = run_dygraph()
+
+        with guard():
+            with _test_eager_guard():
+                eager_avg_cost_value, eager_sum_cost_value, eager_predict_value, eager_token_num_value, \
+                    eager_param_init, eager_param_updated = run_dygraph()
+        self.assertTrue(np.allclose(dy_avg_cost_value, eager_avg_cost_value))
+        self.assertTrue(np.allclose(dy_sum_cost_value, eager_sum_cost_value))
+
+        self.assertTrue(np.allclose(dy_predict_value, eager_predict_value))
+        self.assertTrue(np.allclose(dy_token_num_value, eager_token_num_value))
+
+        for key, value in six.iteritems(static_param_init):
+            self.assertTrue(np.array_equal(value, eager_param_init[key]))
+        for key, value in six.iteritems(dy_param_updated):
+            self.assertTrue(np.allclose(value, eager_param_updated[key]))
+
 
 if __name__ == '__main__':
+    paddle.enable_static()
     unittest.main()