Update hapi.model._save_inference_model by using new features of dy2stat in 2.0-beta API

python/paddle/hapi/model.py

@@ -799,8 +799,7 @@ class Model(object):
             paddle.nn.Layer.
         inputs (InputSpec|list|dict|None): `inputs`, entry points of network,
             could be a InputSpec instance, or lits of InputSpec instances,
-            or dict ({name: InputSpec}), or None. For static graph,
-            inputs must be set. For dynamic graph, it could be None.
+            or dict ({name: InputSpec}), and it couldn't be None.
         labels (InputSpec|list|None): `labels`, entry points of network,
             could be a InputSpec instnace or lits of InputSpec instances,
             or None. For static graph, if labels is required in loss,
@@ -849,10 +848,9 @@ def __init__(self, network, inputs=None, labels=None):
         self._optimizer = None
         self._test_dataloader = None
 
-        if not in_dygraph_mode():
-            if not isinstance(inputs, (list, dict, Input)):
-                raise TypeError(
-                    "'inputs' must be list or dict in static graph mode")
+        if not isinstance(inputs, (list, dict, Input)):
+            raise TypeError(
+                "'inputs' must be list or dict, and couldn't be None.")
         self._inputs = self._verify_spec(inputs, True)
         self._labels = self._verify_spec(labels)
 
@@ -1649,10 +1647,6 @@ def _save_inference_model(self,
                               model_only=False):
         """
         Save inference model can be in static or dynamic mode.
-        It should be noted that before using `save_inference_model`, you should
-        run the model, and the shape you saved is as same as the input of its
-        running. `@paddle.jit.to_static` must be added on `forward` function of
-        your layer in dynamic mode now and these will be optimized later.
 
         Args:
             save_dir (str): The directory path to save the inference model.
@@ -1678,14 +1672,11 @@ def get_inout_spec(all_vars, return_name=False):
 
             return result_list
 
-        # TODO:
-        # 1. Make it Unnecessary to run model before calling `save_inference_model` for users in dygraph.
-        # 2. Save correct shape of input, now the interface stores the shape that the user sent to 
-        #    the inputs of the model in running.
-        # 3. Make it Unnecessary to add `@paddle.jit.to_static` for users in dynamic mode.
         if fluid.in_dygraph_mode():
             with fluid.framework._dygraph_guard(None):
                 layer = self.network
+                layer.forward = paddle.jit.to_static(
+                    layer.forward, input_spec=self._inputs)
 
                 # 1. input check
                 prog_translator = ProgramTranslator()
@@ -1879,18 +1870,7 @@ def summary(self, input_size=None, batch_size=None, dtype=None):
     def _verify_spec(self, specs, is_input=False):
         out_specs = []
 
-        if specs is None:
-            # Note(Aurelius84): If not specific specs of `Input`, using argument names of `forward` function
-            # to generate `Input`. But how can we know the actual shape of each input tensor?
-            if is_input:
-                out_specs = [
-                    Input(
-                        name=n, shape=[None])
-                    for n in extract_args(self.network.forward) if n != 'self'
-                ]
-            else:
-                out_specs = to_list(specs)
-        elif isinstance(specs, dict):
+        if isinstance(specs, dict):
             assert is_input == False
             out_specs = [specs[n] \
                 for n in extract_args(self.network.forward) if n != 'self']

python/paddle/tests/test_model.py

@@ -67,35 +67,6 @@ def forward(self, inputs):
         return x
 
 
-class LeNetDeclarative(fluid.dygraph.Layer):
-    def __init__(self, num_classes=10, classifier_activation=None):
-        super(LeNetDeclarative, self).__init__()
-        self.num_classes = num_classes
-        self.features = Sequential(
-            Conv2d(
-                1, 6, 3, stride=1, padding=1),
-            ReLU(),
-            Pool2D(2, 'max', 2),
-            Conv2d(
-                6, 16, 5, stride=1, padding=0),
-            ReLU(),
-            Pool2D(2, 'max', 2))
-
-        if num_classes > 0:
-            self.fc = Sequential(
-                Linear(400, 120), Linear(120, 84), Linear(84, 10),
-                Softmax())  #Todo: accept any activation
-
-    @declarative
-    def forward(self, inputs):
-        x = self.features(inputs)
-
-        if self.num_classes > 0:
-            x = fluid.layers.flatten(x, 1)
-            x = self.fc(x)
-        return x
-
-
 class MnistDataset(MNIST):
     def __init__(self, mode, return_label=True, sample_num=None):
         super(MnistDataset, self).__init__(mode=mode)
@@ -444,7 +415,9 @@ def test_dynamic_save_static_load(self):
         # dynamic saving
         device = paddle.set_device('cpu')
         fluid.enable_dygraph(device)
-        model = Model(MyModel(classifier_activation=None))
+        inputs = [InputSpec([None, 20], 'float32', 'x')]
+        labels = [InputSpec([None, 1], 'int64', 'label')]
+        model = Model(MyModel(classifier_activation=None), inputs, labels)
         optim = fluid.optimizer.SGD(learning_rate=0.001,
                                     parameter_list=model.parameters())
         model.prepare(optimizer=optim, loss=CrossEntropyLoss(reduction="sum"))
@@ -543,11 +516,10 @@ def test_summary_error(self):
 
     def test_export_deploy_model(self):
         for dynamic in [True, False]:
-            fluid.enable_dygraph() if dynamic else None
-            # paddle.disable_static() if dynamic else None
+            paddle.disable_static() if dynamic else None
             prog_translator = ProgramTranslator()
             prog_translator.enable(False) if not dynamic else None
-            net = LeNetDeclarative()
+            net = LeNet()
             inputs = [InputSpec([None, 1, 28, 28], 'float32', 'x')]
             model = Model(net, inputs)
             model.prepare()
@@ -556,8 +528,9 @@ def test_export_deploy_model(self):
                 os.makedirs(save_dir)
             tensor_img = np.array(
                 np.random.random((1, 1, 28, 28)), dtype=np.float32)
-            ori_results = model.test_batch(tensor_img)
+
             model.save(save_dir, training=False)
+            ori_results = model.test_batch(tensor_img)
             fluid.disable_dygraph() if dynamic else None
 
             place = fluid.CPUPlace() if not fluid.is_compiled_with_cuda(
@@ -574,6 +547,7 @@ def test_export_deploy_model(self):
                 np.testing.assert_allclose(
                     results, ori_results, rtol=1e-5, atol=1e-7)
                 shutil.rmtree(save_dir)
+            paddle.enable_static()
 
 
 class TestRaiseError(unittest.TestCase):
@@ -585,6 +559,14 @@ def test_input_without_name(self):
         with self.assertRaises(ValueError):
             model = Model(net, inputs, labels)
 
+    def test_input_without_input_spec(self):
+        for dynamic in [True, False]:
+            paddle.disable_static() if dynamic else None
+            net = MyModel(classifier_activation=None)
+            with self.assertRaises(TypeError):
+                model = Model(net)
+            paddle.enable_static()
+
 
 if __name__ == '__main__':
     unittest.main()