Update cntk backend with CNTK 2.2 release (#7907)

* integrate with cntk native batch axis convert api

* support mask in recurrent layer

* update interface name

* add support for case static axis is in front of batch axis

* add reverse support

* using native padding api

* update ci test with cntk 2.1

* fix style issue

* use one hot workaround

* api change

* fix style issue

* add comments

* fix bugs in merge

* add output rank to one hot approach

* use cntk padding

* use native batch reshape

* fix the reversed rnn bug

* update padding interface

* update reshape to support free dimension

* workaround free dimension in squeeze

* enable tests with non-specified shape

* also check free dimension

* remove useless padding code

* fix a free dimension bug

* add backward compatible check

* fix pep8 issue

* update travis with cntk 2.2

.travis.yml

@@ -56,9 +56,9 @@ install:
 
   # install cntk
   - if [[ "$TRAVIS_PYTHON_VERSION" == "2.7" ]]; then
-      pip install https://cntk.ai/PythonWheel/CPU-Only/cntk-2.1-cp27-cp27mu-linux_x86_64.whl;
+      pip install https://cntk.ai/PythonWheel/CPU-Only/cntk-2.2-cp27-cp27mu-linux_x86_64.whl;
     elif [[ "$TRAVIS_PYTHON_VERSION" == "3.5" ]]; then
-      pip install https://cntk.ai/PythonWheel/CPU-Only/cntk-2.1-cp35-cp35m-linux_x86_64.whl;
+      pip install https://cntk.ai/PythonWheel/CPU-Only/cntk-2.2-cp35-cp35m-linux_x86_64.whl;
     fi
 
   # install pydot for visualization tests

keras/backend/cntk_backend.py

@@ -9,7 +9,6 @@
 
 C.set_global_option('align_axis', 1)
 
-
 b_any = any
 
 
@@ -243,7 +242,8 @@ def placeholder(
         if ndim:
             shape = tuple([None for _ in range(ndim)])
 
-    cntk_shape = [C.InferredDimension if s is None else s for s in shape]
+    dynamic_dimension = C.FreeDimension if _get_cntk_version() >= 2.2 else C.InferredDimension
+    cntk_shape = [dynamic_dimension if s is None else s for s in shape]
     cntk_shape = tuple(cntk_shape)
 
     if dynamic_axis_num > len(cntk_shape):
@@ -563,9 +563,12 @@ def gather(reference, indices):
     # There is a bug in cntk gather op which may cause crash.
     # We have made a fix but not catched in CNTK 2.1 release.
     # Will udpate with gather op in next release
-    num_class = reference.shape[0]
-    one_hot_matrix = C.ops.one_hot(indices, num_class)
-    return C.times(one_hot_matrix, reference, output_rank=len(reference.shape) - 1)
+    if _get_cntk_version() >= 2.2:
+        return C.ops.gather(reference, indices)
+    else:
+        num_class = reference.shape[0]
+        one_hot_matrix = C.ops.one_hot(indices, num_class)
+        return C.times(one_hot_matrix, reference, output_rank=len(reference.shape) - 1)
 
 
 def _remove_dims(x, axis, keepdims=False):
@@ -665,7 +668,8 @@ def squeeze(x, axis):
     for _ in sorted(_axis, reverse=True):
         del shape[_]
 
-    new_shape = tuple(shape[nones:])
+    new_shape = shape[nones:]
+    new_shape = tuple([C.InferredDimension if _ == C.FreeDimension else _ for _ in new_shape])
     return C.reshape(x, new_shape)
 
 
@@ -753,7 +757,7 @@ def _reshape_dummy_dim(x, axis):
 
     _axis = [_ + len(shape) if _ < 0 else _ for _ in axis]
 
-    if shape.count(C.InferredDimension) > 1:
+    if shape.count(C.InferredDimension) > 1 or shape.count(C.FreeDimension) > 1:
         result = x
         for index in sorted(_axis, reverse=True):
             result = C.reshape(result,
@@ -765,7 +769,7 @@ def _reshape_dummy_dim(x, axis):
         for index in sorted(_axis, reverse=True):
             del shape[index]
 
-        shape = tuple(shape)
+        shape = [C.InferredDimension if _ == C.FreeDimension else _ for _ in shape]
         return C.reshape(x, shape)
 
 
@@ -1063,6 +1067,7 @@ def flatten(x):
 
 
 def reshape(x, shape):
+    shape = tuple([C.InferredDimension if _ == C.FreeDimension else _ for _ in shape])
     if isinstance(x, C.variables.Parameter):
         return C.reshape(x, shape)
     else:
@@ -1077,7 +1082,7 @@ def reshape(x, shape):
                     'collapse of batch axis with inferred dimension. '
                     'The reshape did not take place.')
                 return x
-            return C.user_function(ReshapeBatch(x, shape[1:]))
+            return _reshape_batch(x, shape)
         else:
             # no collapse, then first need to padding the shape
             if num_dynamic_axis >= len(shape):
@@ -1161,7 +1166,7 @@ def repeat(x, n):
     # return the same x to take cntk broadcast feature
     # to make the recurrent layer work.
     # need to be fixed in GA.
-    if n is C.InferredDimension:
+    if n is C.InferredDimension or n is C.FreeDimension:
         return x
     index = 1 - _get_dynamic_axis_num(x)
     if index < 0 or index > 1:
@@ -1748,7 +1753,7 @@ def _is_input_shape_compatible(input, placeholder):
             input_shape = input.shape[num_dynamic:]
             placeholder_shape = placeholder.shape
             for i, p in zip(input_shape, placeholder_shape):
-                if i != p and p != C.InferredDimension:
+                if i != p and p != C.InferredDimension and p != C.FreeDimension:
                     return False
         return True
 
@@ -1852,10 +1857,16 @@ def temporal_padding(x, padding=(1, 1)):
     base_shape = x.shape
     if num_dynamic_axis > 0:
         assert len(base_shape) == 2
-        x = _padding(x, padding, 0)
+        if hasattr(C, 'pad'):
+            x = C.pad(x, pattern=[padding, (0, 0)])
+        else:
+            x = _padding(x, padding, 0)
     else:
         assert len(base_shape) == 3
-        x = _padding(x, padding, 1)
+        if hasattr(C, 'pad'):
+            x = C.pad(x, pattern=[(0, 0), padding, (0, 0)])
+        else:
+            x = _padding(x, padding, 1)
     return x
 
 
@@ -1872,13 +1883,11 @@ def _padding(x, pattern, axis):
         prefix_shape = tuple(prefix_shape)
         x = C.splice(C.constant(value=0, shape=prefix_shape), x, axis=axis)
         base_shape = x.shape
-
     if pattern[1] > 0:
         postfix_shape = list(base_shape)
         postfix_shape[axis] = pattern[1]
         postfix_shape = tuple(postfix_shape)
         x = C.splice(x, C.constant(value=0, shape=postfix_shape), axis=axis)
-
     return x
 
 
@@ -1896,21 +1905,33 @@ def spatial_2d_padding(x, padding=((1, 1), (1, 1)), data_format=None):
     if data_format == 'channels_first':
         if num_dynamic_axis > 0:
             assert len(base_shape) == 3
-            x = _padding(x, padding[0], 1)
-            x = _padding(x, padding[1], 2)
+            if hasattr(C, 'pad'):
+                x = C.pad(x, pattern=[[0, 0], list(padding[0]), list(padding[1])])
+            else:
+                x = _padding(x, padding[0], 1)
+                x = _padding(x, padding[1], 2)
         else:
             assert len(base_shape) == 4
-            x = _padding(x, padding[0], 2)
-            x = _padding(x, padding[1], 3)
+            if hasattr(C, 'pad'):
+                x = C.pad(x, pattern=[[0, 0], [0, 0], list(padding[0]), list(padding[1])])
+            else:
+                x = _padding(x, padding[0], 2)
+                x = _padding(x, padding[1], 3)
     else:
         if num_dynamic_axis > 0:
             assert len(base_shape) == 3
-            x = _padding(x, padding[0], 0)
-            x = _padding(x, padding[1], 1)
+            if hasattr(C, 'pad'):
+                x = C.pad(x, pattern=[list(padding[0]), list(padding[1]), [0, 0]])
+            else:
+                x = _padding(x, padding[0], 0)
+                x = _padding(x, padding[1], 1)
         else:
             assert len(base_shape) == 4
-            x = _padding(x, padding[0], 1)
-            x = _padding(x, padding[1], 2)
+            if hasattr(C, 'pad'):
+                x = C.pad(x, pattern=[[0, 0], list(padding[0]), list(padding[1]), [0, 0]])
+            else:
+                x = _padding(x, padding[0], 1)
+                x = _padding(x, padding[1], 2)
     return x
 
 
@@ -1929,25 +1950,37 @@ def spatial_3d_padding(x, padding=((1, 1), (1, 1), (1, 1)), data_format=None):
     if data_format == 'channels_first':
         if num_dynamic_axis > 0:
             assert len(base_shape) == 4
-            x = _padding(x, padding[0], 1)
-            x = _padding(x, padding[1], 2)
-            x = _padding(x, padding[2], 3)
+            if hasattr(C, 'pad'):
+                x = C.pad(x, pattern=[[0, 0], list(padding[0]), list(padding[1]), list(padding[2])])
+            else:
+                x = _padding(x, padding[0], 1)
+                x = _padding(x, padding[1], 2)
+                x = _padding(x, padding[2], 3)
         else:
             assert len(base_shape) == 5
-            x = _padding(x, padding[0], 2)
-            x = _padding(x, padding[1], 3)
-            x = _padding(x, padding[2], 4)
+            if hasattr(C, 'pad'):
+                x = C.pad(x, pattern=[[0, 0], [0, 0], list(padding[0]), list(padding[1]), list(padding[2])])
+            else:
+                x = _padding(x, padding[0], 2)
+                x = _padding(x, padding[1], 3)
+                x = _padding(x, padding[2], 4)
     else:
         if num_dynamic_axis > 0:
             assert len(base_shape) == 4
-            x = _padding(x, padding[0], 0)
-            x = _padding(x, padding[1], 1)
-            x = _padding(x, padding[2], 2)
+            if hasattr(C, 'pad'):
+                x = C.pad(x, pattern=[list(padding[0]), list(padding[1]), list(padding[2]), [0, 0]])
+            else:
+                x = _padding(x, padding[0], 0)
+                x = _padding(x, padding[1], 1)
+                x = _padding(x, padding[2], 2)
         else:
             assert len(base_shape) == 5
-            x = _padding(x, padding[0], 1)
-            x = _padding(x, padding[1], 2)
-            x = _padding(x, padding[2], 3)
+            if hasattr(C, 'pad'):
+                x = C.pad(x, pattern=[[0, 0], list(padding[0]), list(padding[1]), list(padding[2]), [0, 0]])
+            else:
+                x = _padding(x, padding[0], 1)
+                x = _padding(x, padding[1], 2)
+                x = _padding(x, padding[2], 3)
     return x
 
 
@@ -2249,6 +2282,29 @@ def reverse(x, axes):
     return C.slice(x, cntk_axes, begin_index, end_index, strides)
 
 
+def _reshape_batch(x, shape):
+    # there is a bug in cntk 2.1's unpack_batch implementation
+    if hasattr(C, 'unpack_batch') and _get_cntk_version() >= 2.2:
+        const_a = C.unpack_batch(x)
+        const_a = C.reshape(const_a, shape)
+        return C.to_batch(const_a)
+    else:
+        return C.user_function(ReshapeBatch(x, shape[1:]))
+
+
+def _get_cntk_version():
+    version = C.__version__
+    if version.endswith('+'):
+        version = version[:-1]
+    try:
+        return float(version)
+    except:
+        warnings.warn(
+            'CNTK backend warning: CNTK version not detected. '
+            'Will using CNTK 2.0 GA as default.')
+        return float(2.0)
+
+
 class ReshapeBatch(C.ops.functions.UserFunction):
     def __init__(self, input, shape, name='reshape_with_batch'):
         super(ReshapeBatch, self).__init__([input], as_numpy=False, name=name)

tests/keras/activations_test.py

@@ -173,11 +173,6 @@ def test_elu():
     assert_allclose(result, test_values, rtol=1e-05)
 
     negative_values = np.array([[-1, -2]], dtype=K.floatx())
-    # cntk can't rebind the input shape, so create the model again to
-    # test different batch size
-    if (K.backend() == 'cntk'):
-        x2 = K.placeholder(ndim=2)
-        f = K.function([x2], [activations.elu(x2, 0.5)])
     result = f([negative_values])[0]
     true_result = (np.exp(negative_values) - 1) / 2
 
@@ -196,12 +191,6 @@ def test_selu():
 
     negative_values = np.array([[-1, -2]], dtype=K.floatx())
 
-    # cntk can't rebind the input shape, so create the model again to
-    # test different batch size
-    if (K.backend() == 'cntk'):
-        x2 = K.placeholder(ndim=2)
-        f = K.function([x2], [activations.selu(x2)])
-
     result = f([negative_values])[0]
     true_result = (np.exp(negative_values) - 1) * scale * alpha