refine benchmark

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

@@ -40,10 +40,7 @@ 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.')
     # TODO(wuyi): add "--use_fake_data" option back.
     parser.add_argument(
         '--skip_batch_num',
@@ -72,6 +69,11 @@ def parse_args():
         type=int,
         default=1,
         help='If gpus > 1, will use ParallelExecutor to run, else use Executor.')
+    parser.add_argument(
+        '--cpus',
+        type=int,
+        default=1,
+        help='If cpus > 1, will use ParallelDo to run, else use Executor.')
     parser.add_argument(
         '--data_set',
         type=str,
@@ -231,10 +233,7 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
             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:
@@ -315,10 +314,7 @@ 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
-        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)
@@ -329,12 +325,19 @@ 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)
@@ -342,7 +345,7 @@ def main():
     # the unique trainer id, starting from 0, needed by trainer
     # only
     nccl_id_var, num_trainers, trainer_id = (
-        None, 1, int(os.getenv("PADDLE_TRAINER_ID", "-1")))
+        None, 1, int(os.getenv("PADDLE_TRAINER_ID", "0")))
 
     if args.use_cprof:
         pr = cProfile.Profile()

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