Use inputs() and tf.placeholder in ModelDesc (#318)

docs/tutorial/training-interface.md

@@ -17,8 +17,8 @@ expects 4 arguments to setup the graph: `InputDesc`, `InputSource`, get_cost fun
 
 ```python
 class MyModel(ModelDesc):
-  def _get_inputs(self):
-    return [InputDesc(...), InputDesc(...)]
+  def inputs(self):
+    return [tf.placeholder(dtype, shape, name), tf.placeholder(dtype, shape, name), ... ]
 
   def _build_graph(self, inputs):
     tensorA, tensorB = inputs

examples/basics/cifar-convnet.py

@@ -26,10 +26,9 @@ class Model(ModelDesc):
         super(Model, self).__init__()
         self.cifar_classnum = cifar_classnum
 
-    def _get_inputs(self):
-        return [InputDesc(tf.float32, (None, 30, 30, 3), 'input'),
-                InputDesc(tf.int32, (None,), 'label')
-                ]
+    def inputs(self):
+        return [tf.placeholder(tf.float32, (None, 30, 30, 3), 'input'),
+                tf.placeholder(tf.int32, (None,), 'label')]
 
     def _build_graph(self, inputs):
         image, label = inputs

examples/basics/mnist-convnet.py

@@ -20,13 +20,12 @@ IMAGE_SIZE = 28
 
 
 class Model(ModelDesc):
-    def _get_inputs(self):
+    def inputs(self):
         """
-        Define all the inputs (with type, shape, name) that
-        the graph will need.
+        Define all the inputs (with type, shape, name) that the graph will need.
         """
-        return [InputDesc(tf.float32, (None, IMAGE_SIZE, IMAGE_SIZE), 'input'),
-                InputDesc(tf.int32, (None,), 'label')]
+        return [tf.placeholder(tf.float32, (None, IMAGE_SIZE, IMAGE_SIZE), 'input'),
+                tf.placeholder(tf.int32, (None,), 'label')]
 
     def _build_graph(self, inputs):
         """This function should build the model which takes the input variables

tensorpack/graph_builder/model_desc.py

@@ -3,11 +3,10 @@
 # File: model_desc.py
 
 
-from abc import ABCMeta, abstractmethod
 from collections import namedtuple
 import tensorflow as tf
-import six
 
+from ..utils import logger
 from ..utils.argtools import memoized
 from ..utils.develop import log_deprecated
 from ..tfutils.gradproc import FilterNoneGrad
@@ -38,9 +37,10 @@ class InputDesc(
         if any(k in name for k in [':', '/', ' ']):
             raise ValueError("Invalid InputDesc name: '{}'".format(name))
         self = super(InputDesc, cls).__new__(cls, type, shape, name)
-        self._cached_placeholder = None
+        self._cached_placeholder = {}
         return self
 
+    # TODO this method seems unused outside this class
     def build_placeholder(self):
         """
         Build a tf.placeholder from the metadata.
@@ -51,8 +51,7 @@ class InputDesc(
         with tf.name_scope(None):   # clear any name scope it might get called in
             ret = tf.placeholder(
                 self.type, shape=self.shape, name=self.name)
-        if self._cached_placeholder is None:
-            self._cached_placeholder = ret  # cached_placeholder only caches the prefix='' case
+        self._register_cached_placeholder(ret)
         return ret
 
     # cannot memoize here, because InputDesc is hashed by its fields.
@@ -63,28 +62,67 @@ class InputDesc(
         Returns:
             tf.Tensor:
         """
-        if self._cached_placeholder is not None:
-            return self._cached_placeholder
-        return self.build_placeholder()
+        g = tf.get_default_graph()
+        if g in self._cached_placeholder:
+            return self._cached_placeholder[g]
+        else:
+            return self.build_placeholder()
+
+    def _register_cached_placeholder(self, placeholder):
+        graph = placeholder.graph
+        assert graph not in self._cached_placeholder, \
+            "Placeholder for this InputDesc had been created before! This is a bug."
+        self._cached_placeholder[graph] = placeholder
+
+    @staticmethod
+    def from_placeholder(placeholder):
+        name = placeholder.op.name
+        if name.endswith('_1') or name.endswith('_2'):
+            logger.error("Creating InputDesc from a placeholder named {}.".format(name))
+            logger.error("You might have mistakenly created this placeholder multiple times!")
+        ret = InputDesc(
+            placeholder.dtype,
+            tuple(placeholder.shape.as_list()),
+            name)
+        ret._register_cached_placeholder(placeholder)
+        return ret
 
 
-@six.add_metaclass(ABCMeta)
 class ModelDescBase(object):
-    """ Base class for a model description.
     """
+    Base class for a model description.
+    """
+
     @memoized
     def get_inputs_desc(self):
         """
         Returns:
-            list[:class:`InputDesc`]: list of the underlying :class:`InputDesc`.
+            a list of :class:`InputDesc`.
         """
-        return self._get_inputs()
+        try:
+            return self._get_inputs()
+        except NotImplementedError:
+            with tf.Graph().as_default():   # create these placeholder in a temporary graph
+                inputs = self.inputs()
+                return [InputDesc.from_placeholder(p) for p in inputs]
 
-    @abstractmethod
     def _get_inputs(self):
         """
-        :returns: a list of InputDesc
+        Returns:
+            a list of :class:`InputDesc`.
+        """
+        raise NotImplementedError()
+
+    def inputs(self):
+        """
+        __Create__ and returns a list of placeholders.
+        To be implemented by subclass.
+        The placeholders __have to__ be created inside this function.
+
+        Returns:
+            a list of `tf.placeholder`, to be converted to :class:`InputDesc`.
         """
+        raise NotImplementedError()
 
     def build_graph(self, *args):
         """
@@ -93,13 +131,12 @@ class ModelDescBase(object):
         By default it will call :meth:`_build_graph` with a list of input tensors.
 
         Args:
-            args ([tf.Tensor]): tensors that matches the list of
-                :class:`InputDesc` defined by ``_get_inputs``.
+            args ([tf.Tensor]): tensors that matches the list of inputs defined by ``inputs()``.
 
         Returns:
             In general it returns nothing, but a subclass (e.g.
-            :class:`ModelDesc` may require it to return necessary information
-            to build the trainer.
+            :class:`ModelDesc`) may require it to return necessary information
+            (e.g. cost) to build the trainer.
         """
         if len(args) == 1:
             arg = args[0]
@@ -122,7 +159,8 @@ class ModelDescBase(object):
 
     def _build_graph(self, inputs):
         """
-        This is an old interface which takes a list of tensors, instead of positional arguments.
+        This is an alternative interface which takes a list of tensors, instead of positional arguments.
+        By default :meth:`build_graph` will call this method.
         """
         pass
 

tensorpack/input_source/input_source.py

@@ -41,7 +41,7 @@ class PlaceholderInput(InputSource):
     Just produce placeholders as input tensors.
     """
     def _setup(self, inputs):
-        self._all_placehdrs = [v.build_placeholder() for v in inputs]
+        self._all_placehdrs = [v.build_placeholder_reuse() for v in inputs]
 
     def _get_input_tensors(self):
         return self._all_placehdrs