Fboemer/grappler (#285)

* Reduced usage of command-line flags:
  - NGRAPH_COMPLEX_PACK is now specified as part of the encryption parameters
  - NGRAPH_ENCRYPT_DATA has been replaced by `encrypt` in a `server_config.parameter_map`
  -  NGRAPH_ENCRYPT_MODEL has been removed
  - NGRAPH_TF_BACKEND has been replaced by `ngraph_backend` in a `server_config.parameter_map`
  - NGRAPH_HE_SEAL_CONFIG is still used for debugging, but should be specified via `encryption_parameters` in a `server_config.parameter_map`
  - NGRAPH_ENABLE_CLIENT has been replaced with `enable_client` in a `server_config.parameter_map`

* Use of `HEOpAnnotations` and `PropagateHEAnnotations` pass to label each node as `(encrypted vs. plaintext)`, `(packed vs. unpacked)`, `(from_client vs. from_server)`
  - Not every Node/Op has an annotation (until NervanaSystems/ngraph#3752, NervanaSystems/ngraph#3738), so some logic is redundant / incomplete

* Add plaintexts to protobuf messages

* Fixed clang client build
* Preliminary support for `max` / `divide` / `exp` / `softmax` ops

* MobileNetV2 example is broken due to Power() not being constant-folded (fixed in NervanaSystems/ngraph#3725, but not yet in ngraph-brige)

* Expanded unit-tests for better coverage of plaintext packing / complex packing

* `tf.Keras` remains  unsupported, because the function has different ngraph-tf clusters, each of whose result op will be a plain tensor (created via create_tensor). Then, the result will be plaintext for each cluster, but we want ciphertext for each; one solution would be to just have a single HETensor class

CMakeLists.txt

@@ -20,25 +20,28 @@ cmake_minimum_required(VERSION 3.10)
 # Global project name
 project(he_transformer LANGUAGES CXX)
 
-# CXX flags
-if(DEFINED CMAKE_CXX_CLAGS)
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++1z -std=gnu++1z -Wall")
-else()
-  set(CMAKE_CXX_FLAGS "-std=c++1z -std=gnu++1z -Wall")
-endif()
+# Cmake flags SEAL requires C++17
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CXX_EXTENSIONS OFF)
+set(CMAKE_INSTALL_MESSAGE LAZY)
 set(CMAKE_CXX_FLAGS
-    "${CMAKE_CXX_FLAGS} -march=native -Wno-deprecated-declarations")
-
+    "${CMAKE_CXX_FLAGS} -Wall -march=native -Wno-deprecated-declarations")
 set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS} -g")
 set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS} -O0 -g")
 
+# This allows libhe_seal_backend.so to find libraries in the same directory
+set(CMAKE_INSTALL_RPATH "\$ORIGIN")
+
+# These variables are undocumented but useful.
+set(CMAKE_DISABLE_SOURCE_CHANGES ON)
+set(CMAKE_DISABLE_IN_SOURCE_BUILD ON)
+
 if("${CMAKE_CXX_COMPILER_ID}" MATCHES "^(Apple)?Clang$")
   message(STATUS "Setting clang flags...")
   include(cmake/clang_flags.cmake)
 endif()
 
-set(CMAKE_INSTALL_MESSAGE LAZY)
-
 if(CMAKE_BUILD_TYPE)
   set(RELEASE_TYPES
       Debug
@@ -56,17 +59,10 @@ endif()
 
 include(cmake/sdl.cmake)
 
-# These variables are undocumented but useful.
-set(CMAKE_DISABLE_SOURCE_CHANGES ON)
-set(CMAKE_DISABLE_IN_SOURCE_BUILD ON)
-
 # For ngraph/log
 set(PROJECT_ROOT_DIR "${CMAKE_CURRENT_SOURCE_DIR}")
 add_definitions(-DPROJECT_ROOT_DIR="${PROJECT_ROOT_DIR}")
 
-# This allows libhe_seal_backend.so to find libraries in the same directory
-set(CMAKE_INSTALL_RPATH "\$ORIGIN")
-
 # he-transformer headers
 set(HE_TRANSFORMER_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/src)
 

README.md

@@ -28,8 +28,8 @@ The [examples](https://github.com/NervanaSystems/he-transformer/tree/master/exam
 - virtualenv v16.1.0
 - bazel v0.25.2
 #### The following dependencies are built automatically
-- [nGraph](https://github.com/NervanaSystems/ngraph) - v0.25.0
-- [nGraph-tf](https://github.com/tensorflow/ngraph-bridge) - v0.18.1
+- [nGraph](https://github.com/NervanaSystems/ngraph) - v0.25.1-rc.7
+- [nGraph-tf](https://github.com/tensorflow/ngraph-bridge) - v0.19.0-rc4
 - [SEAL](https://github.com/Microsoft/SEAL) - v3.3.1
 - [TensorFlow](https://github.com/tensorflow/tensorflow) - v1.14.0
 - [Boost](https://github.com/boostorg) v1.69
@@ -77,8 +77,7 @@ make doc
 to create doxygen documentation in `$HE_TRANSFORMER/build/doc/doxygen`.
 
 #### 1b. Python bindings for client
-To build an experimental client-server model with python bindings, see the [python](https://github.com/NervanaSystems/he-transformer/tree/master/python) folder.
-***Note***: This feature is experimental. For best experience, you should skip this step.
+To build an client-server model with python bindings, see the [python](https://github.com/NervanaSystems/he-transformer/tree/master/python) folder.
 
 ### 2. Run C++ unit-tests
 Ensure the virtual environment is active, i.e. run `source $HE_TRANSFORMER/build/external/venv-tf-py3/bin/activate`

cmake/ngraph-tf.cmake

@@ -20,7 +20,7 @@ set(EXTERNAL_NGRAPH_INSTALL_DIR ${EXTERNAL_INSTALL_DIR})
 set(NGRAPH_TF_CMAKE_PREFIX ext_ngraph_tf)
 
 set(NGRAPH_TF_REPO_URL https://github.com/tensorflow/ngraph-bridge.git)
-set(NGRAPH_TF_GIT_LABEL v0.18.1)
+set(NGRAPH_TF_GIT_LABEL cad093d84cc3a1ce212d8a96c67217321b44309b)
 
 set(NGRAPH_TF_SRC_DIR
     ${CMAKE_BINARY_DIR}/${NGRAPH_TF_CMAKE_PREFIX}/src/${NGRAPH_TF_CMAKE_PREFIX})
@@ -56,6 +56,9 @@ if(${USE_PREBUILT_TF})
     )
 endif()
 
+# TODO: enable other options
+set(ng_tf_build_flags "--use_grappler_optimizer")
+
 ExternalProject_Add(ext_ngraph_tf
                     GIT_REPOSITORY ${NGRAPH_TF_REPO_URL}
                     GIT_TAG ${NGRAPH_TF_GIT_LABEL}

configs/he_seal_ckks_config_N11_L1.json

@@ -4,5 +4,6 @@
   "security_level": 128,
   "coeff_modulus": [
     54
-  ]
+  ],
+  "complex_packing": true
 }

examples/MobileNetV2/README.md → examples/ImageNet/MobileNetV2/README.md

@@ -1,18 +1,23 @@
 # MobileNet V2 example
 
 This folder demonstrates an example of inference on MobileNetV2.
-Note: this is a work in progress, and requires ~150GB memory.
+Note: this is a work in progress, and requires ~50GB memory.
 Runtime will be very slow without many cores.
 
-See here: https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet
+See https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet
 for a description.
 
 # Setup
 1. Make sure python env is active, i.e. run
 ```bash
 source $HE_TRANSFORMER/build/external/venv-tf-py3/bin/activate
 ```
-Also be sure the `pyhe_client` wheel has been installed
+Also ensure the `pyhe_client` wheel has been installed (see `python` folder for instructions).
+
+The examples rely on numpy and pillow, so run
+```bash
+pip install numpy pillow
+```
 
 2. Build Tensorflow graph transforms and add them to your path:
 
@@ -32,11 +37,6 @@ export PATH=$HE_TRANSFORMER/build/ext_ngraph_tf/src/ext_ngraph_tf/build_cmake/te
 python get_models.py
 ```
 
-4. To enable image processing, run
-```bash
-pip install pillow
-```
-
 # Image-Net evaluation
 1. First, sign up for an account at image-net.org
 2. Download the 2012 test_images (all tasks)) 13GB MD5: `e64ceb247e473635708aed23ab6d839` file on image-net.org
@@ -61,21 +61,23 @@ For the remaining instructions, run```bash
 export DATA_DIR=path_to_your_data_dir
 ```
 
-4. To run inference using TensorFlow on unencrypted data, call
+## CPU backend
+To run inference using the CPU backend on unencrypted data, call
 ```bash
 python test.py \
   --data_dir=$DATA_DIR \
-  --batch_size=300
+  --batch_size=300 \
+  --backend=CPU
 ```
 
+##
 5. To call inference using HE_SEAL's plaintext operations (for debugging), call
 ```bash
-NGRAPH_TF_BACKEND=HE_SEAL \
 STOP_CONST_FOLD=1 \
 python test.py \
 --data_dir=$DATA_DIR \
---ngraph=true \
---batch_size=300
+--batch_size=300 \
+--backend=HE_SEAL
 ```
 Note, the `STOP_CONST_FOLD` flag will prevent the constant folding graph optimization.
 For large batch sizes, const folding incurs significant overhead during graph compilation, and doesn't result in much runtime speedup.

examples/MobileNetV2/client.py → examples/ImageNet/MobileNetV2/client.py

@@ -36,8 +36,6 @@
 import util
 import numpy as np
 
-FLAGS = None
-
 
 def print_nodes(filename):
     graph_def = read_pb_file(filename)
@@ -68,8 +66,6 @@ def get_imagenet_labels():
 
 
 def main(FLAGS):
-    util.VAL_IMAGE_FLAGS = FLAGS
-
     imagenet_inference_labels = get_imagenet_inference_labels()
     imagenet_training_labels = get_imagenet_training_labels()
     assert (

examples/MobileNetV2/test.py → examples/ImageNet/MobileNetV2/test.py

@@ -19,21 +19,21 @@
 import numpy as np
 import json
 import argparse
-import os
 import time
 import PIL
 from PIL import Image
 import multiprocessing as mp
 import util
+import ngraph_bridge
 
 from util import get_imagenet_inference_labels, \
                  get_imagenet_training_labels, \
                  get_validation_image, \
                  get_validation_images, \
                  get_validation_labels, \
-                 str2bool
-
-FLAGS = None
+                 str2bool, \
+                 server_argument_parser, \
+                 server_config_from_flags
 
 
 def print_nodes(filename):
@@ -56,9 +56,8 @@ def load_model(filename):
 
 
 def main(FLAGS):
-    using_client = (os.environ.get('NGRAPH_ENABLE_CLIENT') is not None)
 
-    if using_client:
+    if FLAGS.enable_client:
         print('Using client')
     else:
         print('Not using client')
@@ -71,25 +70,27 @@ def main(FLAGS):
     assert (
         sorted(imagenet_training_labels) == sorted(imagenet_inference_labels))
 
-    if not using_client:
-        validation_nums = get_validation_labels(FLAGS)
-        x_test = get_validation_images(FLAGS)
-        validation_labels = imagenet_inference_labels[validation_nums]
-    else:
+    validation_nums = get_validation_labels(FLAGS)
+    validation_labels = imagenet_inference_labels[validation_nums]
+
+    if FLAGS.enable_client:
+        # Server input is dummy
         x_test = np.random.rand(FLAGS.batch_size, FLAGS.image_size,
                                 FLAGS.image_size, 3)
+    else:
+        x_test = get_validation_images(FLAGS)
 
-    if FLAGS.ngraph:
-        import ngraph_bridge
-        print(ngraph_bridge.__version__)
+    config = server_config_from_flags(FLAGS, 'input')
 
-    config = tf.compat.v1.ConfigProto()
-    config.intra_op_parallelism_threads = 44
-    config.inter_op_parallelism_threads = 44
-    if FLAGS.ngraph:
-        config = ngraph_bridge.update_config(config)
     sess = tf.compat.v1.Session(config=config)
     graph_def = load_model(FLAGS.model)
+
+    for node in graph_def.node:
+        if 'FusedBatchNorm' in node.name or 'Pow' in node.name:
+            print(node)
+
+    #print('node names', [n.name for n in graph_def.node])
+
     tf.import_graph_def(graph_def, name='')
 
     input_tensor = sess.graph.get_tensor_by_name('input:0')
@@ -111,7 +112,7 @@ def main(FLAGS):
     else:
         top5 = np.flip(y_pred.argsort()[:, -5:], axis=1)
 
-    if not using_client:
+    if not FLAGS.enable_client:
         preds = imagenet_training_labels[top5]
 
         if FLAGS.batch_size < 10:
@@ -124,7 +125,7 @@ def main(FLAGS):
 
 
 if __name__ == '__main__':
-    parser = argparse.ArgumentParser()
+    parser = server_argument_parser()
     parser.add_argument(
         '--data_dir',
         type=str,
@@ -161,7 +162,6 @@ def main(FLAGS):
         help='crop to this size before resizing to image_size')
     parser.add_argument(
         '--ngraph', type=str2bool, default=False, help='use ngraph backend')
-    parser.add_argument('--batch_size', type=int, default=1, help='Batch size')
     parser.add_argument(
         '--start_batch', type=int, default=0, help='Test data start index')