fix prune input bug (#30384)

python/paddle/fluid/dygraph/dygraph_to_static/program_translator.py

@@ -470,19 +470,22 @@ class StaticFunction(object):
         cached_program_len = len(self._program_cache)
         # If specific `input_spec`, apply convertion from dygraph layers into static Program.
         if cached_program_len == 0:
-            if input_spec is None:
-                input_spec = self._function_spec.input_spec
-            elif self._function_spec.input_spec is not None:
-                if not input_specs_compatible(
+            desired_input_spec = input_spec
+            if self._function_spec.input_spec is not None:
+                if input_spec is not None and not input_specs_compatible(
                         flatten(input_spec),
                         flatten(self._function_spec.input_spec)):
                     raise ValueError(
                         "The `input_spec`: {} used to construct concrete_program is conflict with the `input_spec`: {} in `@paddle.jit.to_static`".
                         format(input_spec, self._function_spec.input_spec))
+                # NOTE(chenweihang): we should always translated program based on the `input_spec`
+                # decorated on forward if it is valid
+                desired_input_spec = self._function_spec.input_spec
 
-            has_input_spec = (input_spec is not None)
+            has_input_spec = (desired_input_spec is not None)
             if has_input_spec:
-                concrete_program, _ = self.get_concrete_program(*input_spec)
+                concrete_program, _ = self.get_concrete_program(
+                    *desired_input_spec)
                 return concrete_program
             else:
                 raise ValueError(

python/paddle/fluid/dygraph/dygraph_to_static/utils.py

@@ -1222,37 +1222,41 @@ def unwrap(func):
     return unwrapped_f
 
 
-def input_specs_compatible(src_input_specs, other_input_specs):
+def input_specs_compatible(src_input_specs, desired_input_specs):
     """
     Returns True if the two input specs are compatible, otherwise False.
 
     args:
         src_input_spec (list[InputSpec]|tuple(InputSpec)): list/tuple of
             paddle.static.InputSpec
-        other_input_spec (list[InputSpec]|tuple(InputSpec)): list/tuple of
+        desired_input_specs (list[InputSpec]|tuple(InputSpec)): list/tuple of
             paddle.static.InputSpec
     """
     len_specs = len(src_input_specs)
-    if len_specs != len(other_input_specs):
-        return False
-
-    for i in range(len_specs):
-        src_shape = src_input_specs[i].shape
-        other_shape = other_input_specs[i].shape
-        len_shape = len(src_shape)
-        if len_shape != len(other_shape):
-            return False
-        for j in range(len_shape):
-            if src_shape[j] is None or src_shape[j] < 0:
-                continue
-            if other_shape[j] is None or other_shape[j] < 0:
-                continue
-            if src_shape[j] != other_shape[j]:
+    if len_specs != len(desired_input_specs):
+        # NOTE(chenweihang): if the input_spec of jit.save is a subset of
+        # input_spec of to_static, also compatible 
+        for spec in src_input_specs:
+            if spec not in desired_input_specs:
+                return False
+    else:
+        for i in range(len_specs):
+            src_shape = src_input_specs[i].shape
+            other_shape = desired_input_specs[i].shape
+            len_shape = len(src_shape)
+            if len_shape != len(other_shape):
+                return False
+            for j in range(len_shape):
+                if src_shape[j] is None or src_shape[j] < 0:
+                    continue
+                if other_shape[j] is None or other_shape[j] < 0:
+                    continue
+                if src_shape[j] != other_shape[j]:
+                    return False
+
+            src_dtype = convert_dtype(src_input_specs[i].dtype)
+            other_dtype = convert_dtype(desired_input_specs[i].dtype)
+            if src_dtype != other_dtype:
                 return False
-
-        src_dtype = convert_dtype(src_input_specs[i].dtype)
-        other_dtype = convert_dtype(other_input_specs[i].dtype)
-        if src_dtype != other_dtype:
-            return False
 
     return True

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

@@ -95,6 +95,38 @@ class LinerNetWithLabel(paddle.nn.Layer):
         return out, avg_loss
 
 
+class LinerNetWithPruneInput(paddle.nn.Layer):
+    def __init__(self, in_size, out_size):
+        super(LinerNetWithPruneInput, self).__init__()
+        self._linear = Linear(in_size, out_size)
+
+    @declarative(input_spec=[
+        InputSpec(
+            shape=[None, 784], dtype='float32', name="image"), InputSpec(
+                shape=[None, 1], dtype='int64', name="label")
+    ])
+    def forward(self, x, label):
+        out = self._linear(x)
+        loss = fluid.layers.cross_entropy(out, label)
+        avg_loss = fluid.layers.mean(loss)
+        return out
+
+
+class LinerNetWithUselessInput(paddle.nn.Layer):
+    def __init__(self, in_size, out_size):
+        super(LinerNetWithUselessInput, self).__init__()
+        self._linear = Linear(in_size, out_size)
+
+    @declarative(input_spec=[
+        InputSpec(
+            shape=[None, 784], dtype='float32', name="image"), InputSpec(
+                shape=[None, 1], dtype='int64', name="label")
+    ])
+    def forward(self, x, label):
+        out = self._linear(x)
+        return out
+
+
 class LinearNetReturnLoss(fluid.dygraph.Layer):
     def __init__(self, in_size, out_size):
         super(LinearNetReturnLoss, self).__init__()
@@ -627,16 +659,24 @@ class TestJitSaveMultiCases(unittest.TestCase):
         paddle.seed(SEED)
         paddle.framework.random._manual_program_seed(SEED)
 
-    def verify_inference_correctness(self, layer, model_path, with_label=False):
+    def verify_inference_correctness(self,
+                                     layer,
+                                     model_path,
+                                     with_label_and_loss=False,
+                                     with_label=False):
         layer.eval()
         loaded_layer = paddle.jit.load(model_path)
         loaded_layer.eval()
         # inference & compare
         x = paddle.to_tensor(np.random.random((1, 784)).astype('float32'))
-        if with_label:
+        if with_label_and_loss:
             y = paddle.to_tensor(np.random.random((1, 1)).astype('int64'))
             pred, _ = layer(x, y)
             pred = pred.numpy()
+        elif with_label:
+            y = paddle.to_tensor(np.random.random((1, 1)).astype('int64'))
+            pred = layer(x, y)
+            pred = pred.numpy()
         else:
             pred = layer(x).numpy()
         loaded_pred = loaded_layer(x).numpy()
@@ -714,7 +754,8 @@ class TestJitSaveMultiCases(unittest.TestCase):
             ],
             output_spec=[out])
 
-        self.verify_inference_correctness(layer, model_path, True)
+        self.verify_inference_correctness(
+            layer, model_path, with_label_and_loss=True)
 
     def test_prune_to_static_no_train(self):
         layer = LinerNetWithLabel(784, 1)
@@ -732,7 +773,36 @@ class TestJitSaveMultiCases(unittest.TestCase):
             ],
             output_spec=output_spec)
 
-        self.verify_inference_correctness(layer, model_path, True)
+        self.verify_inference_correctness(
+            layer, model_path, with_label_and_loss=True)
+
+    def test_prune_input_to_static_no_train(self):
+        layer = LinerNetWithPruneInput(784, 1)
+
+        model_path = "test_prune_input_to_static_no_train/model"
+        paddle.jit.save(
+            layer,
+            model_path,
+            input_spec=[
+                InputSpec(
+                    shape=[None, 784], dtype='float32', name="image")
+            ])
+
+        self.verify_inference_correctness(layer, model_path, with_label=True)
+
+    def test_prune_useless_input_to_static_no_train(self):
+        layer = LinerNetWithUselessInput(784, 1)
+
+        model_path = "test_prune_useless_input_to_static_no_train/model"
+        paddle.jit.save(
+            layer,
+            model_path,
+            input_spec=[
+                InputSpec(
+                    shape=[None, 784], dtype='float32', name="image")
+            ])
+
+        self.verify_inference_correctness(layer, model_path, with_label=True)
 
     def test_no_prune_input_spec_name_warning(self):
         layer = LinearNetWithInputSpec(784, 1)