Merge branch 'master' into feature/20180401-file-format-converter

sony · Jun 7, 2018 · 491332b · 491332b
2 parents 563925b + 97669f9
commit 491332b
Show file tree

Hide file tree

Showing 35 changed files with 1,515 additions and 40 deletions.
diff --git a/build-tools/code_generator/function_types.yaml b/build-tools/code_generator/function_types.yaml
@@ -368,6 +368,12 @@ DepthwiseDeconvolution:
 Round:
   float: [float]
   half: [Half]
+Ceil:
+  float: [float]
+  half: [Half]
+Floor:
+  float: [float]
+  half: [Half]
 Sin:
   float: [float]
   half: [Half]

diff --git a/build-tools/code_generator/functions.yaml b/build-tools/code_generator/functions.yaml
@@ -700,14 +700,10 @@ Normalization:
       .. math::
           \begin{eqnarray}
             \mu &=& \frac{1}{M} \sum x_i \\
-             rm &=& ({\rm decay\_rate}) rm + (1 - {\rm decay\_rate}) \mu \\
-            y_i &=& x_i - rm
+            y_i &=& x_i - \mu
           \end{eqnarray}
 
-      At validation time, it is defined as
-
-      .. math::
-          y_i = x_i - rm
+      At testing time, the mean values used are those that were computed during training by moving average.
 
       Note:
           The backward performs an approximated differentiation that takes into account only the latest mini-batch.
@@ -1779,6 +1775,52 @@ Math:
       .. math::
           \frac{\partial y_i}{\partial x_i} = 1.
 
+    inputs:
+      x:
+        doc: Input variable
+    arguments: {}
+    outputs:
+      y:
+        doc: N-D array with the same shape as x
+  Ceil:
+    snake_name: ceil
+    doc: |2+
+
+      Element-wise ceil function.
+
+      In the forward pass, this function simply returns the smallest integer which is not less than the input.
+
+      .. math::
+          y_i = ceil(x_i).
+
+      In the backward pass, the simple Straight-Through Estimator (STE) is applied,
+
+      .. math::
+          \frac{\partial y_i}{\partial x_i} = 1.
+
+    inputs:
+      x:
+        doc: Input variable
+    arguments: {}
+    outputs:
+      y:
+        doc: N-D array with the same shape as x
+  Floor:
+    snake_name: floor
+    doc: |2+
+
+      Element-wise floor function.
+
+      In the forward pass, this function simply returns the largest integer which is not greater than the input.
+
+      .. math::
+          y_i = floor(x_i).
+
+      In the backward pass, the simple Straight-Through Estimator (STE) is applied,
+
+      .. math::
+          \frac{\partial y_i}{\partial x_i} = 1.
+
     inputs:
       x:
         doc: Input variable

diff --git a/build-tools/make/build-with-docker.mk b/build-tools/make/build-with-docker.mk
@@ -31,6 +31,7 @@ DOCKER_IMAGE_NAME_BASE ?= nnabla-build
 DOCKER_IMAGE_AUTO_FORMAT ?= $(DOCKER_IMAGE_NAME_BASE)-auto-format
 DOCKER_IMAGE_DOC ?= $(DOCKER_IMAGE_NAME_BASE)-doc
 DOCKER_IMAGE_BUILD ?= $(DOCKER_IMAGE_NAME_BASE)-build
+DOCKER_IMAGE_NNABLA ?= $(DOCKER_IMAGE_NAME_BASE)-nnabla
 
 DOCKER_RUN_OPTS +=--rm
 DOCKER_RUN_OPTS += -v $$(pwd):$$(pwd)
@@ -96,3 +97,20 @@ bwd-nnabla-wheel: docker_image_build
 bwd-nnabla-test: docker_image_build
 	cd $(NNABLA_DIRECTORY) \
 	&& docker run $(DOCKER_RUN_OPTS) $(DOCKER_IMAGE_BUILD) make -f build-tools/make/build.mk nnabla-test-local
+
+.PHONY: bwd-nnabla-shell
+bwd-nnabla-shell: docker_image_build
+	cd $(NNABLA_DIRECTORY) \
+	&& docker run $(DOCKER_RUN_OPTS) -it --rm ${DOCKER_IMAGE_BUILD} make nnabla-shell
+
+########################################################################################################################
+# Docker image with current nnabla
+.PHONY: docker_image_nnabla
+docker_image_nnabla: bwd-nnabla-cpplib bwd-nnabla-wheel
+	docker pull ubuntu:16.04
+	cd $(NNABLA_DIRECTORY) \
+	&& cp docker/development/Dockerfile.build.py$(PYTHON_VERSION_MAJOR)$(PYTHON_VERSION_MINOR) Dockerfile \
+	&& echo ADD $(shell echo build_wheel_py$(PYTHON_VERSION_MAJOR)$(PYTHON_VERSION_MINOR)/dist/*.whl) /tmp/ >>Dockerfile \
+	&& echo RUN pip install /tmp/$(shell basename build_wheel_py$(PYTHON_VERSION_MAJOR)$(PYTHON_VERSION_MINOR)/dist/*.whl) >>Dockerfile \
+	&& docker build $(DOCKER_BUILD_ARGS) -t $(DOCKER_IMAGE_NNABLA) . \
+	&& rm -f Dockerfile
diff --git a/build-tools/make/build.mk b/build-tools/make/build.mk
@@ -92,6 +92,14 @@ nnabla-install:
 	-pip uninstall -y nnabla
 	pip install $(BUILD_DIRECTORY_WHEEL)/dist/*.whl
 
+########################################################################################################################
+# Shell (for rapid development)
+.PHONY: nnabla-shell
+nnabla-shell:
+	PS1="nnabla-build: " bash --norc -i
+
+########################################################################################################################
+# test
 .PHONY: nnabla-test-cpplib
 nnabla-test-cpplib: nnabla-cpplib
 	@$(MAKE) -C $(BUILD_DIRECTORY_CPPLIB) cpplibtest

diff --git a/doc/python/api.rst b/doc/python/api.rst
@@ -16,4 +16,5 @@ Python API Reference
     api/communicator
     api/monitor
     api/data_iterator
+    api/utils
     api/ext
diff --git a/doc/python/api/function.rst b/doc/python/api/function.rst
@@ -36,7 +36,9 @@ Neural Network Layers
 
 .. autofunction:: affine
 .. autofunction:: convolution
+.. autofunction:: depthwise_convolution
 .. autofunction:: deconvolution
+.. autofunction:: depthwise_deconvolution
 .. autofunction:: max_pooling
 .. autofunction:: average_pooling
 .. autofunction:: sum_pooling
@@ -135,6 +137,8 @@ Math
 .. autofunction:: exp
 .. autofunction:: log
 .. autofunction:: round
+.. autofunction:: ceil
+.. autofunction:: floor
 .. autofunction:: identity
 .. autofunction:: matrix_diag
 .. autofunction:: matrix_diag_part

diff --git a/doc/python/api/parametric_function.rst b/doc/python/api/parametric_function.rst
@@ -52,12 +52,18 @@ Here is the list of parametric functions.
 
 .. autofunction:: affine
 .. autofunction:: convolution
+.. autofunction:: depthwise_convolution
 .. autofunction:: deconvolution
+.. autofunction:: depthwise_deconvolution
 .. autofunction:: batch_normalization
+.. autofunction:: mean_subtraction
 
 .. autofunction:: embed
 .. autofunction:: prelu
 
+.. autofunction:: svd_affine
+.. autofunction:: svd_convolution
+.. autofunction:: cpd3_convolution
 .. autofunction:: binary_connect_affine
 .. autofunction:: binary_connect_convolution
 .. autofunction:: binary_weight_affine
@@ -69,6 +75,7 @@ Here is the list of parametric functions.
 .. autofunction:: fixed_point_quantized_convolution
 .. autofunction:: pow2_quantized_affine
 .. autofunction:: pow2_quantized_convolution
+
 .. autofunction:: lstm
 
 .. autoclass:: LSTMCell

diff --git a/doc/python/api/utils.rst b/doc/python/api/utils.rst
@@ -0,0 +1,9 @@
+Utils
+==============
+
+.. toctree::
+    :maxdepth: 1
+
+    utils/data_iterator.rst
+    utils/profiling.rst
+
diff --git a/doc/python/api/data_iterator.rst → doc/python/api/utils/data_iterator.rst b/doc/python/api/data_iterator.rst → doc/python/api/utils/data_iterator.rst
diff --git a/doc/python/api/utils/profiling.rst b/doc/python/api/utils/profiling.rst
@@ -0,0 +1,14 @@
+Profiling
+==============
+
+.. automodule:: nnabla.utils.profiler
+
+Profiler
+----------
+
+.. autoclass:: GraphProfiler
+    :members:
+
+.. autoclass:: GraphProfilerCsvWriter
+    :members:
+
diff --git a/doc/python/install_on_windows.rst b/doc/python/install_on_windows.rst
@@ -104,7 +104,7 @@ Check for running (CUDA/cuDNN).
 
     > ipython
     
-    In [1]: import nnabla_ext.cuda.cudnn
+    In [1]: import nnabla_ext.cudnn
     2017-06-16 18:42:18,881 [nnabla][Level 99]: Initializing CPU extension...
     2017-06-16 18:42:19,923 [nnabla][Level 99]: Initializing CUDA extension...
     2017-06-16 18:42:20,243 [nnabla][Level 99]: Initializing cuDNN extension...

diff --git a/doc/python/tutorial/by_examples.rst b/doc/python/tutorial/by_examples.rst
@@ -576,8 +576,8 @@ CUDA by specifying a context before building a graph. We strongly
 recommend using a CUDNN context that is fast. Although the context class
 can be instantiated by ``nn.Context()``, specifying a context descriptor
 might be a bit complicated for users. There for we recommend create a
-context by using a helper function ``extension_context()`` found in the
-``nnabla.contrib.context`` module. NNabla officially supports ``cpu``
+context by using a helper function ``get_extension_context()`` found in the
+``nnabla.ext_utils`` module. NNabla officially supports ``cpu``
 and ``cudnn`` as a context specifier passed to the first argument
 (extension name). NOTE: By setting the cudnn context as a global default
 context, Functions and solves created are instantiated with CUDNN
@@ -589,9 +589,9 @@ for details.
 .. code-block:: python2
 
     # Run on CUDA
-    from nnabla.contrib.context import extension_context
+    from nnabla.ext_utils import get_extension_context
     cuda_device_id = 0
-    ctx = extension_context('cudnn', device_id=cuda_device_id)
+    ctx = get_extension_context('cudnn', device_id=cuda_device_id)
     print "Context:", ctx
     nn.set_default_context(ctx)  # Set CUDA as a default context.
     y, hs = cnn(x)

diff --git a/doc/python/tutorial/dynamic_and_static_nn.rst b/doc/python/tutorial/dynamic_and_static_nn.rst
@@ -106,10 +106,10 @@ neural network does not change during training.
 
 .. code-block:: python2
 
-    from nnabla.contrib.context import extension_context
+    from nnabla.ext_utils import get_extension_context
     
     # setup cuda extension
-    ctx_cuda = extension_context('cudnn', device_id=GPU)  # replace 'cuda.cudnn' by 'cpu' if you want to run the example on the CPU
+    ctx_cuda = get_extension_context('cudnn', device_id=GPU)  # replace 'cudnn' by 'cpu' if you want to run the example on the CPU
     nn.set_default_context(ctx_cuda)
     
     # create variables for network input and label

diff --git a/doc/python/tutorial/multi_device_training.rst b/doc/python/tutorial/multi_device_training.rst
@@ -88,7 +88,7 @@ Prepare the dependencies
     
     import nnabla as nn
     import nnabla.communicators as C
-    from nnabla.contrib.context import extension_context
+    from nnabla.ext_utils import get_extension_context
     import nnabla.functions as F
     from nnabla.initializer import (
         calc_uniform_lim_glorot,
@@ -104,13 +104,13 @@ Define the communicator for gradients exchange.
 
     %%px
     extension_module = "cudnn"
-    ctx = extension_context(extension_module)
+    ctx = get_extension_context(extension_module)
     comm = C.MultiProcessDataParalellCommunicator(ctx)
     comm.init()
     n_devices = comm.size
     mpi_rank = comm.rank
     device_id = mpi_rank
-    ctx = extension_context(extension_module, device_id=device_id)
+    ctx = get_extension_context(extension_module, device_id=device_id)
 
 Check different ranks are assigned to different devices
 

diff --git a/docker/Dockerfile.py2 → docker/py2/Dockerfile b/docker/Dockerfile.py2 → docker/py2/Dockerfile
diff --git a/docker/Dockerfile.py3 → docker/py3/Dockerfile b/docker/Dockerfile.py3 → docker/py3/Dockerfile
diff --git a/include/nbla/function/ceil.hpp b/include/nbla/function/ceil.hpp
@@ -0,0 +1,45 @@
+// Copyright (c) 2017 Sony Corporation. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+/** Ceil
+ */
+
+#ifndef NBLA_FUNCTION_CEIL_HPP
+#define NBLA_FUNCTION_CEIL_HPP
+
+#include <nbla/function/utils/base_transform_unary.hpp>
+
+#include <cmath>
+
+namespace nbla {
+
+/** @class Ceil
+@brief Ceiling value, defined as
+@f[
+y_i = ceil(x_i),
+@f]
+
+Inputs:
+- N-D array.
+
+Outputs:
+- N-D array.
+
+@tparam T Data type for computation.
+\ingroup FunctionImpGrp
+ */
+
+NBLA_DEFINE_TRANSFORM_UNARY(Ceil, std::ceil(x), dy, false);
+}
+#endif
diff --git a/include/nbla/function/floor.hpp b/include/nbla/function/floor.hpp
@@ -0,0 +1,43 @@
+// Copyright (c) 2017 Sony Corporation. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef NBLA_FUNCTION_FLOOR_HPP
+#define NBLA_FUNCTION_FLOOR_HPP
+
+#include <nbla/function/utils/base_transform_unary.hpp>
+
+#include <cmath>
+
+namespace nbla {
+
+/** @class Floor
+@brief Flooring value, defined as
+@f[
+y_i = floor(x_i),
+@f]
+
+
+Inputs:
+- N-D array.
+
+
+Outputs:
+- N-D array.
+@tparam T Data type for computation.
+\ingroup FunctionImplGrp
+ */
+
+NBLA_DEFINE_TRANSFORM_UNARY(Floor, std::floor(x), dy, false);
+}
+#endif
diff --git a/python/setup.py b/python/setup.py
@@ -228,6 +228,7 @@ def extopts(library_name, library_dir):
                 'nnabla.utils.converter.nnabla',
                 'nnabla.utils.converter.nnablart',
                 'nnabla.utils.converter.onnx',
+                'nnabla.utils.factorization',
                 'nnabla_ext',
                 'nnabla_ext.cpu', ]