Fix some problems in StaticModelRunner (#24082) (#24163)

* adapt old version pretrain model load, test=develop

* fix infer error & multiple input error, test=develop

python/paddle/fluid/dygraph/static_runner.py

@@ -168,8 +168,13 @@ class StaticModelRunner(layers.Layer):
         super(StaticModelRunner, self).__init__()
 
         # Step 0. key variable definitions
-        self._load_program_desc = None
-        self._program_desc = None
+        # loaded inference program desc
+        self._infer_program_desc = None
+        # recovered train program desc
+        self._train_program_desc = None
+        # StaticModelRunner executed program desc,
+        # switch infer or train by train() and eval()
+        self._trace_program_desc = None
         self._inner_scope = core.Scope()
         # the layer outputs var desc
         self._output_descs = []
@@ -182,46 +187,47 @@ class StaticModelRunner(layers.Layer):
 
         # Step 1. load program desc from disk
         # the saved model hold feed, fetch & scale op, no need, can be remove
-        self._load_program_desc = self._load_static_model(model_dir,
-                                                          model_filename)
+        self._infer_program_desc = self._load_static_model(model_dir,
+                                                           model_filename)
 
-        # Step 2. set all `is_test` attributes to False
-        self._change_is_test_status(False)
-
-        # Step 3. load all parameters
+        # Step 2. load all parameters
         self._load_persisitable_dict(model_dir, params_filename)
 
-        # Step 4. generate backwar program desc
-        self._program_desc = self._append_backward_desc()
+        # Step 3. generate backwar program desc
+        self._train_program_desc = self._append_backward_desc()
 
-        # Step 5. recheck parameters stop gradients
+        # Step 4. recheck parameters stop gradients
         self._recheck_stop_gradients()
 
+        # Step 5. set default mode to train
+        self.train()
+
     def train(self):
         self._is_test = False
-        self._change_is_test_status(False)
+        self._trace_program_desc = self._train_program_desc
 
     def eval(self):
         self._is_test = True
-        self._change_is_test_status(True)
+        self._trace_program_desc = self._infer_program_desc
 
-    def forward(self, inputs):
+    def forward(self, *args):
         """
         Executed forward part of StaticModelRunner Layer.
         Generally execute directly using the Layer object.
 
         Args:
-            inputs(np.ndarray|Variable|list[np.ndarray|Variable]): the inputs of StaticModelRunner
+            args(tuple(np.ndarray|Variable)): the inputs of StaticModelRunner.
+                The order of input variables needs to be the same as the order 
+                of feed variables when using `save_inference_model` to save model.
         
         Returns:
             Variable|list[Variable]: The forward outputs of StaticModelRunner Layer.
+                If there is only one output, return Variable;
+                if there are multiple outputs, return list[Variable].
         """
         # Step 1. prepare inputs, outputs, attrs
-        if not isinstance(inputs, (list, tuple)):
-            inputs = [inputs]
-
         input_vars = []
-        for i, value in enumerate(inputs):
+        for i, value in enumerate(args):
             if not isinstance(value, (np.ndarray, core.VarBase)):
                 raise TypeError(
                     "The type of inputs.value in StaticModelRunner.forward must be numpy array or Variable(VarBase), but received %s."
@@ -265,9 +271,9 @@ class StaticModelRunner(layers.Layer):
             outputs={'Out': output_vars,
                      'OutScope': tmp_scope_vec},
             attrs={
-                'global_block': self._program_desc.block(0),
+                'global_block': self._trace_program_desc.block(0),
                 'start_op_index': 0,
-                'end_op_index': self._load_program_desc.block(0).op_size(),
+                'end_op_index': self._infer_program_desc.block(0).op_size(),
                 'is_test': self._is_test
             })
 
@@ -280,7 +286,7 @@ class StaticModelRunner(layers.Layer):
         # be user wanted result.
         for param in params:
             grad_name = param.name + core.grad_var_suffix()
-            grad_var = self._program_desc.block(0).find_var(
+            grad_var = self._trace_program_desc.block(0).find_var(
                 cpt.to_bytes(grad_name))
             # NOTE: cannot find var desc maybe no problem, such as in batch_norm
             if grad_var is None:
@@ -334,11 +340,15 @@ class StaticModelRunner(layers.Layer):
                 self._output_names.append(cpt.to_bytes(op.input('X')[0]))
                 self._output_descs.append(
                     root_block.find_var(cpt.to_bytes(op.input('X')[0])))
-            elif op.type() == 'fetch' and op.input('X')[0].startswith(
-                    'save_infer_model/scale_'):
+            elif op.type() == 'fetch':
                 ops_to_remove.append(i)
                 fetch_var_name = cpt.to_bytes(op.output('Out')[0])
                 root_block._remove_var(fetch_var_name)
+                # NOTE: some old pre-train models have no extra scale_op
+                if not op.input('X')[0].startswith('save_infer_model/scale_'):
+                    self._output_names.append(cpt.to_bytes(op.input('X')[0]))
+                    self._output_descs.append(
+                        root_block.find_var(cpt.to_bytes(op.input('X')[0])))
             else:
                 if op.has_attr("op_callstack"):
                     op.remove_attr("op_callstack")
@@ -346,14 +356,20 @@ class StaticModelRunner(layers.Layer):
         for op_idx in reversed(ops_to_remove):
             root_block._remove_op(op_idx, op_idx + 1)
 
+        # NOTE: reverse feed vars
+        self._input_names.reverse()
+
         return program_desc
 
     @switch_to_static_graph
     def _append_backward_desc(self):
-        assert self._load_program_desc is not None, "The StaticModelRunner not initialized properly."
-        program_desc_copy = core.ProgramDesc(self._load_program_desc)
+        assert self._infer_program_desc is not None, "The StaticModelRunner not initialized properly."
+        program_desc_copy = core.ProgramDesc(self._infer_program_desc)
+
+        # Step 1. set all `is_test` attributes to False
+        self._change_is_test_status(program_desc_copy, False)
 
-        # Step 1. prepare program and related var
+        # Step 2. prepare program and related var
         # NOTE: To reuse backward interfaces, build Program firstly.
         # Originally, there is no need to build a program, but need to almost
         # rewrite a series of methods for append_backward for program_desc. 
@@ -366,15 +382,15 @@ class StaticModelRunner(layers.Layer):
         for out in self._output_descs:
             targets.append(program.global_block().var(out.name()))
 
-        # Step 2. append backward
+        # Step 3. append backward
         backward.gradients(targets=targets, inputs=[])
         return program.desc
 
     def _load_persisitable_dict(self, model_dir, params_filename=None):
         load_dirname = os.path.normpath(model_dir)
-        assert self._load_program_desc is not None, "The StaticModelRunner not initialized properly."
+        assert self._infer_program_desc is not None, "The StaticModelRunner not initialized properly."
 
-        persis_vars = self._get_persis_vars(self._load_program_desc)
+        persis_vars = self._get_persis_vars(self._infer_program_desc)
         load_var_map = {}
         for each_var in persis_vars:
             orig_each_name = each_var.name()
@@ -425,12 +441,12 @@ class StaticModelRunner(layers.Layer):
                 self._param_names.append(param.name)
 
     def _recheck_stop_gradients(self):
-        assert self._program_desc is not None, "The StaticModelRunner not initialized properly."
+        assert self._train_program_desc is not None, "The StaticModelRunner not initialized properly."
         # NOTE: After loading the model, the stop_gradient information 
         # of the original variable is lost, but if a parameter does not
         # have a corresponding @GRAD variable in the backward program,
         # it can be said that it is also stop_gradient
-        all_var_names = self._get_all_var_names(self._program_desc)
+        all_var_names = self._get_all_var_names(self._train_program_desc)
         for param_name in self._parameters:
             param_grad_name = param_name + core.grad_var_suffix()
             if param_grad_name not in all_var_names:
@@ -472,19 +488,19 @@ class StaticModelRunner(layers.Layer):
         return var_desc.persistable()
 
     def _is_parameter(self, persis_var_desc):
-        assert self._load_program_desc is not None, "The StaticModelRunner not initialized properly."
+        assert self._infer_program_desc is not None, "The StaticModelRunner not initialized properly."
         # 1. firstly, param should be input of op
         input_ops = []  # op can be repeated
-        for block_idx in six.moves.range(self._load_program_desc.num_blocks()):
-            block = self._load_program_desc.block(block_idx)
+        for block_idx in six.moves.range(self._infer_program_desc.num_blocks()):
+            block = self._infer_program_desc.block(block_idx)
             for op_idx in six.moves.range(block.op_size()):
                 op = block.op(op_idx)
                 # NOTE: parameter is the input of a certain op
                 if persis_var_desc.name() in op.input_arg_names():
                     input_ops.append(op)
         # 2. secondly, param should not be output of op or be same op's output
-        for block_idx in six.moves.range(self._load_program_desc.num_blocks()):
-            block = self._load_program_desc.block(block_idx)
+        for block_idx in six.moves.range(self._infer_program_desc.num_blocks()):
+            block = self._infer_program_desc.block(block_idx)
             for op_idx in six.moves.range(block.op_size()):
                 op = block.op(op_idx)
                 if persis_var_desc.name() in op.output_arg_names():
@@ -495,11 +511,10 @@ class StaticModelRunner(layers.Layer):
                         return False
         return True
 
-    def _change_is_test_status(self, is_test):
+    def _change_is_test_status(self, program_desc, is_test):
         # change all `is_test` attributes
-        assert self._load_program_desc is not None, "The StaticModelRunner not initialized properly."
-        for i in six.moves.range(self._load_program_desc.num_blocks()):
-            block = self._load_program_desc.block(i)
+        for i in six.moves.range(program_desc.num_blocks()):
+            block = program_desc.block(i)
             for j in six.moves.range(block.op_size()):
                 op = block.op(j)
                 if op.has_attr('is_test'):
@@ -520,8 +535,8 @@ class StaticModelRunner(layers.Layer):
         old_name = param_desc.name()
         new_name = self._append_loaded_suffix(param_desc.name())
         param_desc.set_name(new_name)
-        for block_idx in six.moves.range(self._load_program_desc.num_blocks()):
-            block = self._load_program_desc.block(block_idx)
+        for block_idx in six.moves.range(self._infer_program_desc.num_blocks()):
+            block = self._infer_program_desc.block(block_idx)
             for op_idx in six.moves.range(block.op_size()):
                 op = block.op(op_idx)
                 op._rename_input(old_name, new_name)

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

@@ -149,7 +149,7 @@ class TestImperativeStaticModelRunnerMnist(unittest.TestCase):
                     label = data[1]
                     label.stop_gradient = True
 
-                    cost = mnist(inputs=img)
+                    cost = mnist(img)
 
                     loss = fluid.layers.cross_entropy(cost, label)
                     avg_loss = fluid.layers.mean(loss)
@@ -229,7 +229,67 @@ class TestImperativeStaticModelRunnerMnist(unittest.TestCase):
 
         return static_x_data, static_out, static_param_init_value, static_param_value
 
-    def test_mnist_no_params_filename(self):
+    def load_and_infer_dygraph(self):
+        place = fluid.CUDAPlace(0) if core.is_compiled_with_cuda(
+        ) else fluid.CPUPlace()
+        with fluid.dygraph.guard(place):
+            fluid.default_main_program().random_seed = self.seed
+
+            mnist = fluid.dygraph.static_runner.StaticModelRunner(
+                model_dir=self.save_dirname, model_filename=self.model_filename)
+
+            train_reader = paddle.batch(
+                self.reader_decorator(paddle.dataset.mnist.test()),
+                batch_size=self.batch_size,
+                drop_last=True)
+            train_loader = fluid.io.DataLoader.from_generator(capacity=10)
+            train_loader.set_sample_list_generator(train_reader, places=place)
+
+            mnist.eval()
+
+            for batch_id, data in enumerate(train_loader()):
+                img = data[0]
+                cost = mnist(img)
+
+                if batch_id >= 1:
+                    break
+
+            dy_x_data = img.numpy()
+            dy_out = cost.numpy()
+
+        return dy_x_data, dy_out
+
+    def load_and_infer_static(self):
+        with new_program_scope():
+            place = fluid.CUDAPlace(0) if core.is_compiled_with_cuda(
+            ) else fluid.CPUPlace()
+
+            exe = fluid.Executor(place)
+            [infer_program, feed_target_names,
+             fetch_targets] = fluid.io.load_inference_model(self.save_dirname,
+                                                            exe)
+            infer_program.random_seed = self.seed
+
+            train_reader = paddle.batch(
+                self.reader_decorator(paddle.dataset.mnist.test()),
+                batch_size=self.batch_size,
+                drop_last=True)
+
+            for batch_id, data in enumerate(train_reader()):
+                static_x_data = np.array([x[0] for x in data])
+                out = exe.run(infer_program,
+                              feed={feed_target_names[0]: static_x_data},
+                              fetch_list=fetch_targets)
+
+                if batch_id >= 1:
+                    break
+
+            static_param_value = {}
+            static_out = out[0]
+
+        return static_x_data, static_out
+
+    def test_mnist_train_no_params_filename(self):
         self.save_dirname = "mnist.inference.model.noname"
         self.model_filename = None
         self.params_filename = None
@@ -257,7 +317,7 @@ class TestImperativeStaticModelRunnerMnist(unittest.TestCase):
             key += core.loaded_var_suffix()
             self.assertTrue(np.allclose(value, dy_param_value[key], atol=1e-4))
 
-    def test_mnist_with_params_filename(self):
+    def test_mnist_train_with_params_filename(self):
         self.save_dirname = "mnist.inference.model"
         self.model_filename = "mnist.model"
         self.params_filename = "mnist.params"
@@ -285,6 +345,26 @@ class TestImperativeStaticModelRunnerMnist(unittest.TestCase):
             key += core.loaded_var_suffix()
             self.assertTrue(np.allclose(value, dy_param_value[key], atol=1e-4))
 
+    def test_mnist_infer_no_params_filename(self):
+        self.save_dirname = "mnist.inference.model.noname"
+        self.model_filename = None
+        self.params_filename = None
+        # Phase 1. run and save static model
+        self.train_and_save_model()
+
+        # Phase 2. load model & train dygraph
+        dy_x_data, dy_out = \
+            self.load_and_infer_dygraph()
+
+        static_x_data, static_out = \
+            self.load_and_infer_static()
+
+        # Phase 3. compare
+        self.assertTrue(np.array_equal(static_x_data, dy_x_data))
+
+        np.testing.assert_array_almost_equal(static_out, dy_out)
+        self.assertTrue(np.allclose(static_out, dy_out, atol=1e-04))
+
 
 if __name__ == '__main__':
     unittest.main()

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

@@ -136,7 +136,7 @@ class TestImperativeStaticModelRunnerWhile(unittest.TestCase):
                 label = data[1]
                 label.stop_gradient = True
 
-                cost = while_net(inputs=img)
+                cost = while_net(img)
 
                 loss = fluid.layers.cross_entropy(cost, label)
                 avg_loss = fluid.layers.mean(loss)