Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fluid_benchmark_support_recordioreader

Author: yi.wu

benchmark/.gitignore

@@ -7,3 +7,6 @@ paddle/rnn/imdb.pkl
 caffe/image/logs
 tensorflow/image/logs
 tensorflow/rnn/logs
+fluid/models/*.pyc
+fluid/logs
+fluid/nohup.out

benchmark/fluid/fluid_benchmark.py

@@ -40,18 +40,18 @@ def parse_args():
     parser.add_argument(
         '--batch_size', type=int, default=32, help='The minibatch size.')
     parser.add_argument(
-        '--learning_rate',
-        type=float,
-        default=0.001,
-        help='The minibatch size.')
+        '--learning_rate', type=float, default=0.001, help='The learning rate.')
     parser.add_argument(
         '--skip_batch_num',
         type=int,
         default=5,
         help='The first num of minibatch num to skip, for better performance test'
     )
     parser.add_argument(
-        '--iterations', type=int, default=80, help='The number of minibatches.')
+        '--iterations',
+        type=int,
+        default=80,
+        help='The number of minibatches, set to -1 to run all batches.')
     parser.add_argument(
         '--pass_num', type=int, default=100, help='The number of passes.')
     parser.add_argument(
@@ -71,11 +71,12 @@ def parse_args():
         type=int,
         default=1,
         help='If gpus > 1, will use ParallelExecutor to run, else use Executor.')
+    # this option is available only for vgg and resnet.
     parser.add_argument(
         '--data_set',
         type=str,
         default='flowers',
-        choices=['cifar10', 'flowers'],
+        choices=['cifar10', 'flowers', 'imagenet'],
         help='Optional dataset for benchmark.')
     parser.add_argument(
         '--infer_only', action='store_true', help='If set, run forward only.')
@@ -228,27 +229,32 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
     iters, num_samples, start_time = 0, 0, time.time()
     for pass_id in range(args.pass_num):
         train_losses = []
-        for batch_id, data in enumerate(train_reader()):
+        reader_generator = train_reader()
+        batch_id = 0
+        data = None
+        while True:
+            if not args.use_reader_op:
+                data = next(reader_generator, None)
+            if iters == args.iterations or data == None:
+                break
             if iters == args.skip_batch_num:
                 start_time = time.time()
                 num_samples = 0
-            if iters == args.iterations:
-                break
+
             if args.use_reader_op:
                 loss = exe.run(train_prog, fetch_list=[avg_loss])
             else:
                 loss = exe.run(train_prog,
                                feed=feeder.feed(data),
                                fetch_list=[avg_loss])
             iters += 1
-            num_samples += len(data)
+            batch_id += 1
+            # FIXME(wuyi): last batch size maybe different
+            num_samples += len(args.batch_size)
             train_losses.append(loss)
             print("Pass: %d, Iter: %d, Loss: %f\n" %
                   (pass_id, iters, np.mean(train_losses)))
-        train_elapsed = time.time() - start_time
-        examples_per_sec = num_samples / train_elapsed
-        print('\nTotal examples: %d, total time: %.5f, %.5f examples/sec\n' %
-              (num_samples, train_elapsed, examples_per_sec))
+        print_train_time(start_time, time.time(), num_samples)
         print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses)))
         # evaluation
         if not args.no_test and batch_acc != None:
@@ -309,7 +315,14 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
         num_samples = 0
         iters = 0
         start_time = time.time()
-        for batch_id, data in enumerate(train_reader()):
+        reader_generator = train_reader()
+        batch_id = 0
+        data = None
+        while True:
+            if not args.use_reader_op:
+                data = next(reader_generator, None)
+            if iters == args.iterations or data == None:
+                break
             if args.profile and pass_id == 0 and batch_id == 5:
                 profiler.start_profiler("All")
             elif args.profile and pass_id == 0 and batch_id == 10:
@@ -318,8 +331,6 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
             if iters == args.skip_batch_num:
                 start_time = time.time()
                 num_samples = 0
-            if iters == args.iterations:
-                break
             # NOTE: if use reader ops, the input data is not splited to multiple cards
             if args.use_reader_op and iters >= args.iterations / args.gpus:
                 break
@@ -334,12 +345,10 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
             if batch_id % 1 == 0:
                 print("Pass %d, batch %d, loss %s" %
                       (pass_id, batch_id, np.array(loss)))
-        train_elapsed = time.time() - start_time
+            batch_id += 1
         if args.use_reader_op:
             num_samples = num_samples * args.gpus
-        examples_per_sec = num_samples / train_elapsed
-        print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
-              (num_samples, train_elapsed, examples_per_sec))
+        print_train_time(start_time, time.time(), num_samples)
         if not args.no_test and batch_acc != None:
             test_acc = test(startup_exe, infer_prog, test_reader, feeder,
                             batch_acc)
@@ -350,12 +359,19 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
 def print_arguments(args):
     vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
                                 vars(args)['device'] == 'GPU')
-    print('----------- resnet Configuration Arguments -----------')
+    print('----------- Configuration Arguments -----------')
     for arg, value in sorted(vars(args).iteritems()):
         print('%s: %s' % (arg, value))
     print('------------------------------------------------')
 
 
+def print_train_time(start_time, end_time, num_samples):
+    train_elapsed = end_time - start_time
+    examples_per_sec = num_samples / train_elapsed
+    print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
+          (num_samples, train_elapsed, examples_per_sec))
+
+
 def main():
     args = parse_args()
     print_arguments(args)

benchmark/fluid/models/resnet.py

@@ -27,6 +27,7 @@
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 import paddle.fluid.profiler as profiler
+from recordio_converter import imagenet_train, imagenet_test
 
 
 def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'):
@@ -123,13 +124,30 @@ def get_model(args):
         else:
             dshape = [32, 32, 3]
         model = resnet_cifar10
-    else:
+        train_reader = paddle.dataset.cifar.train10()
+        test_reader = paddle.dataset.cifar.test10()
+    elif args.data_set == "flowers":
         class_dim = 102
         if args.data_format == 'NCHW':
             dshape = [3, 224, 224]
         else:
             dshape = [224, 224, 3]
         model = resnet_imagenet
+        train_reader = paddle.dataset.flowers.train()
+        test_reader = paddle.dataset.flowers.test()
+    elif args.data_set == "imagenet":
+        class_dim = 1000
+        if args.data_format == 'NCHW':
+            dshape = [3, 224, 224]
+        else:
+            dshape = [224, 224, 3]
+        model = resnet_imagenet
+        if not args.data_dir:
+            raise Exception(
+                "Must specify --data_dir when training with imagenet")
+        train_reader = imagenet_train(args.data_dir)
+        test_reader = imagenet_test(args.data_dir)
+
     if args.use_reader_op:
         filelist = [
             os.path.join(args.data_path, f) for f in os.listdir(args.data_path)
@@ -163,15 +181,10 @@ def get_model(args):
 
     optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
 
-    train_reader = paddle.batch(
+    batched_train_reader = paddle.batch(
         paddle.reader.shuffle(
-            paddle.dataset.cifar.train10()
-            if args.data_set == 'cifar10' else paddle.dataset.flowers.train(),
-            buf_size=5120),
-        batch_size=args.batch_size)
-    test_reader = paddle.batch(
-        paddle.dataset.cifar.test10()
-        if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
+            train_reader, buf_size=5120),
         batch_size=args.batch_size)
+    batched_test_reader = paddle.batch(train_reader, batch_size=args.batch_size)
 
-    return avg_cost, inference_program, optimizer, train_reader, test_reader, batch_acc
+    return avg_cost, inference_program, optimizer, batched_train_reader, batched_test_reader, batch_acc

benchmark/fluid/recordio_converter.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import os
+import random
 import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
@@ -53,6 +54,13 @@ def prepare_flowers(outpath, batch_size):
                        [1])
 
 
+def default_mapper(sample):
+    img, label = sample
+    img = image.simple_transform(
+        img, 256, 224, True, mean=[103.94, 116.78, 123.68])
+    return img.flatten().astype('float32'), label
+
+
 def imagenet_train(data_dir):
     contents = os.listdir(data_dir)
     if set(contents) != set(
@@ -68,6 +76,8 @@ def imagenet_train(data_dir):
             img, lbl = l[:-1].split(" ")
             img2label[img] = int(lbl)
             imgfilelist.append(img)
+    # shuffle all, this is slow
+    random.shuffle(imgfilelist)
 
     def train_reader():
         for idx, imgfile in enumerate(imgfilelist):
@@ -76,15 +86,36 @@ def train_reader():
             label = [img2label[imgfile], ]
             yield [data, label]
 
-    def default_mapper(sample):
-        img, label = sample
-        img = image.simple_transform(
-            img, 256, 224, True, mean=[103.94, 116.78, 123.68])
-        return img.flatten().astype('float32'), label
-
     return paddle.reader.map_readers(default_mapper, train_reader)
 
 
+def imagenet_test(data_dir):
+    contents = os.listdir(data_dir)
+    if set(contents) != set(
+        ["train", "train.txt", "val", "val_set", "val.txt", "unzip.sh"]):
+        raise Exception("Imagenet data contents error!")
+    img2label = dict()
+    imgfilelist = []
+    with open(os.path.join(data_dir, "val.txt")) as fn:
+        while 1:
+            l = fn.readline()
+            if not l:
+                break
+            img, lbl = l[:-1].split(" ")
+            img2label[img] = int(lbl)
+            imgfilelist.append(img)
+
+    def test_reader():
+        for idx, imgfile in enumerate(imgfilelist):
+            base_path = os.path.join(data_dir, "val", imgfile.split(".")[0])
+            image_path = ".".join([base_path, "jpeg"])
+            data = image.load_image(image_path)
+            label = [img2label[imgfile], ]
+            yield [data, label]
+
+    return paddle.reader.map_readers(default_mapper, test_reader)
+
+
 # FIXME(wuyi): delete this when https://github.com/PaddlePaddle/Paddle/pull/11066 is merged
 def convert_reader_to_recordio_files(
         filename,

benchmark/fluid/run.sh

@@ -2,6 +2,7 @@
 # This script benchmarking the PaddlePaddle Fluid on
 # single thread single GPU.
 
+mkdir -p logs
 #export FLAGS_fraction_of_gpu_memory_to_use=0.0
 export CUDNN_PATH=/paddle/cudnn_v5
 
@@ -35,6 +36,7 @@ nohup stdbuf -oL nvidia-smi \
       --format=csv \
       --filename=mem.log  \
       -l 1 &
+
 # mnist
 # mnist gpu mnist 128
 FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \
@@ -43,7 +45,7 @@ FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \
                --batch_size=128 \
                --skip_batch_num=5 \
                --iterations=500 \
-               2>&1 | tee -a mnist_gpu_128.log
+               2>&1 | tee -a logs/mnist_gpu_128.log
 
 # vgg16
 # gpu cifar10 128
@@ -53,7 +55,7 @@ FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \
                --batch_size=128 \
                --skip_batch_num=5 \
                --iterations=30 \
-               2>&1 | tee -a vgg16_gpu_128.log
+               2>&1 | tee -a logs/vgg16_gpu_128.log
 
 # flowers gpu  128
 FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \
@@ -63,28 +65,28 @@ FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \
                --data_set=flowers \
                --skip_batch_num=5 \
                --iterations=30 \
-               2>&1 | tee -a vgg16_gpu_flowers_32.log
+               2>&1 | tee -a logs/vgg16_gpu_flowers_32.log
 
 # resnet50
 # resnet50 gpu cifar10 128
 FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \
-               --model=resnet50 \
+               --model=resnet \
                --device=GPU \
                --batch_size=128 \
                --data_set=cifar10 \
                --skip_batch_num=5 \
                --iterations=30 \
-               2>&1 | tee -a resnet50_gpu_128.log
+               2>&1 | tee -a logs/resnet50_gpu_128.log
 
 # resnet50 gpu flowers 64
 FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \
-               --model=resnet50 \
+               --model=resnet \
                --device=GPU \
                --batch_size=64 \
                --data_set=flowers \
                --skip_batch_num=5 \
                --iterations=30 \
-               2>&1 | tee -a resnet50_gpu_flowers_64.log
+               2>&1 | tee -a logs/resnet50_gpu_flowers_64.log
 
 # lstm
 # lstm gpu imdb 32 # tensorflow only support batch=32
@@ -94,7 +96,7 @@ FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \
                --batch_size=32 \
                --skip_batch_num=5 \
                --iterations=30 \
-               2>&1 | tee -a lstm_gpu_32.log
+               2>&1 | tee -a logs/lstm_gpu_32.log
 
 # seq2seq
 # seq2seq gpu wmb 128
@@ -104,4 +106,4 @@ FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \
                --batch_size=128 \
                --skip_batch_num=5 \
                --iterations=30 \
-               2>&1 | tee -a lstm_gpu_128.log
+               2>&1 | tee -a logs/lstm_gpu_128.log

cmake/external/grpc.cmake

@@ -33,10 +33,18 @@ ELSE()
   SET(BUILD_CMD make HAS_SYSTEM_PROTOBUF=false -s -j ${NUM_OF_PROCESSOR} static grpc_cpp_plugin)
 ENDIF()
 
+# FIXME(wuyi): do not build zlib cares protobuf twice, find a way to build grpc with them
 ExternalProject_Add(
     extern_grpc
     DEPENDS protobuf zlib
-    URL "http://paddlepaddledeps.bj.bcebos.com/grpc.tar.xz"
+    # NOTE(wuyi):
+    # this package is generated by following steps:
+    # 1. git clone -b v1.8.x https://github.com/grpc/grpc.git
+    # 2. submodule update --init
+    # 3. keep only zlib, cares, protobuf, boringssl under "third_party",
+    #    checkout and clean other dirs under third_party
+    # 4. remove .git, and package the directory.
+    URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.8.x.tar.gz"
     PREFIX          ${GRPC_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CONFIGURE_COMMAND ""
@@ -49,7 +57,6 @@ ExternalProject_Add(
     INSTALL_COMMAND make prefix=${GRPC_INSTALL_DIR} install
 )
 
-# FIXME(typhoonzero): hack to get static lib path, try a better way like merge them.
 ADD_LIBRARY(grpc++_unsecure STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET grpc++_unsecure PROPERTY IMPORTED_LOCATION
              "${GRPC_INSTALL_DIR}/lib/libgrpc++_unsecure.a")