Ngraph (#1916)

* Added README for the ngraph engine

* Fix --model_save_dir

* Uncomment naive tests

* Update run.sh script

* update to use ngraph

* update readme

* Numactl -l was removed

* update instruction

* download pretrained model for inference only

* fix typo

* update

fluid/PaddleCV/image_classification/README_ngraph.md

+
+# PaddlePaddle inference and training script
+This directory contains configuration and instructions to run the PaddlePaddle + nGraph for a local training and inference.
+
+# How to build PaddlePaddle framework with NGraph engine
+In order to build the PaddlePaddle + nGraph engine and run proper script,  follow up a few steps:
+1. Install PaddlePaddle project
+2. set env exports for nGraph and OpenMP
+3. run the inference/training script
+
+Currently supported models:
+* ResNet50 (inference and training).
+
+Only support Adam optimizer yet.
+
+Short description of aforementioned steps:
+
+## 1. Install PaddlePaddle
+Follow PaddlePaddle [installation instruction](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification#installation) to install PaddlePaddle. If you build PaddlePaddle yourself, please use the following cmake arguments and ensure to set `-DWITH_NGRAPH=ON`.  
+```
+cmake .. -DCMAKE_BUILD_TYPE=Release -DWITH_GPU=OFF -DWITH_MKL=ON -DWITH_MKLDNN=ON  -DWITH_NGRAPH=ON
+```
+Note: MKLDNN and MKL are required.
+
+## 2. Set env exports for nGraph and OMP
+Set the following exports needed for running nGraph:
+```
+export FLAGS_use_ngraph=true
+export OMP_NUM_THREADS=<num_cpu_cores>
+```
+
+Optional exports for better performance:
+```
+export KMP_AFFINITY=granularity=fine,compact,1,0
+export KMP_BLOCKTIME=1
+```
+
+## 3. How the benchmark script might be run.
+If everything built successfully, you can run command in ResNet50 nGraph session in script [run.sh](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleCV/image_classification/run.sh) to start the benchmark job locally. You will need to uncomment the `#ResNet50 nGraph` part of script.
+
+Above is training job using the nGraph, to run the inference job using the nGraph:
+
+Please download the pre-trained resnet50 model from [supported models](https://github.com/PaddlePaddle/models/tree/72dcc7c1a8d5de9d19fbd65b4143bd0d661eee2c/fluid/PaddleCV/image_classification#supported-models-and-performances) for inference script.
+

fluid/PaddleCV/image_classification/run.sh

@@ -192,3 +192,26 @@ python train.py \
 #	--model_category=models_name \
 #       --model_save_dir=output/ \
 #	--l2_decay=3e-4
+
+#ResNet50 nGraph:
+# Training:
+#OMP_NUM_THREADS=`nproc` FLAGS_use_ngraph=true python train.py \
+#    --model=ResNet50 \
+#    --batch_size=128 \
+#    --total_images=1281167 \
+#    --class_dim=1000 \
+#    --image_shape=3,224,224 \
+#    --lr_strategy=none \
+#    --lr=0.001 \
+#    --num_epochs=120 \
+#    --with_mem_opt=False \
+#    --model_category=models_name \
+#    --model_save_dir=output/ \
+#    --lr_strategy=adam \
+#    --use_gpu=False
+# Inference:
+#OMP_NUM_THREADS=`nproc` FLAGS_use_ngraph=true python infer.py  \
+#    --use_gpu=false \
+#    --model=ResNet50 \
+#    --pretrained_model=ResNet50_pretrained
+

fluid/PaddleCV/image_classification/train.py

@@ -116,6 +116,9 @@ def optimizer_setting(params):
             learning_rate=lr,
             momentum=momentum_rate,
             regularization=fluid.regularizer.L2Decay(l2_decay))
+    elif ls["name"] == "adam":
+        lr = params["lr"]
+        optimizer = fluid.optimizer.Adam(learning_rate=lr)
     else:
         lr = params["lr"]
         l2_decay = params["l2_decay"]
@@ -264,14 +267,17 @@ def train(args):
         fluid.io.load_vars(
             exe, pretrained_model, main_program=train_prog, predicate=if_exist)
 
-    visible_device = os.getenv('CUDA_VISIBLE_DEVICES')
-    if visible_device:
-        device_num = len(visible_device.split(','))
+    if args.use_gpu:
+        visible_device = os.getenv('CUDA_VISIBLE_DEVICES')
+        if visible_device:
+            device_num = len(visible_device.split(','))
+        else:
+            device_num = subprocess.check_output(
+                ['nvidia-smi', '-L']).decode().count('\n')
     else:
-        device_num = subprocess.check_output(
-            ['nvidia-smi', '-L']).decode().count('\n')
-
+        device_num = 1
     train_batch_size = args.batch_size / device_num
+
     test_batch_size = 16
     if not args.enable_ce:
         train_reader = paddle.batch(
@@ -292,10 +298,15 @@ def train(args):
 
     train_py_reader.decorate_paddle_reader(train_reader)
     test_py_reader.decorate_paddle_reader(test_reader)
-    train_exe = fluid.ParallelExecutor(
-        main_program=train_prog,
-        use_cuda=bool(args.use_gpu),
-        loss_name=train_cost.name)
+
+    use_ngraph = os.getenv('FLAGS_use_ngraph')
+    if not use_ngraph:
+        train_exe = fluid.ParallelExecutor(
+            main_program=train_prog,
+            use_cuda=bool(args.use_gpu),
+            loss_name=train_cost.name)
+    else:
+        train_exe = exe
 
     train_fetch_list = [
         train_cost.name, train_acc1.name, train_acc5.name, global_lr.name
@@ -314,9 +325,13 @@ def train(args):
         try:
             while True:
                 t1 = time.time()
-                loss, acc1, acc5, lr = train_exe.run(
-                    fetch_list=train_fetch_list)
 
+                if use_ngraph:
+                    loss, acc1, acc5, lr = train_exe.run(
+                        train_prog, fetch_list=train_fetch_list)
+                else:
+                    loss, acc1, acc5, lr = train_exe.run(
+                        fetch_list=train_fetch_list)
                 t2 = time.time()
                 period = t2 - t1
                 loss = np.mean(np.array(loss))