Merge branch 'modify_dev' into modify_readers_to_fit_parallel_executor

Author: JiayiFeng

benchmark/fluid/machine_translation.py

@@ -48,6 +48,13 @@
     type=int,
     default=16,
     help="The sequence number of a mini-batch data. (default: %(default)d)")
+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.')
 parser.add_argument(
     "--dict_size",
     type=int,
@@ -72,16 +79,21 @@
     default=3,
     help="The width for beam searching. (default: %(default)d)")
 parser.add_argument(
-    "--use_gpu",
-    type=distutils.util.strtobool,
-    default=True,
-    help="Whether to use gpu. (default: %(default)d)")
+    '--device',
+    type=str,
+    default='GPU',
+    choices=['CPU', 'GPU'],
+    help="The device type.")
 parser.add_argument(
     "--max_length",
     type=int,
     default=250,
     help="The maximum length of sequence when doing generation. "
     "(default: %(default)d)")
+parser.add_argument(
+    '--with_test',
+    action='store_true',
+    help='If set, test the testset during training.')
 
 
 def lstm_step(x_t, hidden_t_prev, cell_t_prev, size):
@@ -281,7 +293,7 @@ def train():
             paddle.dataset.wmt14.test(args.dict_size), buf_size=1000),
         batch_size=args.batch_size)
 
-    place = core.CUDAPlace(0) if args.use_gpu else core.CPUPlace()
+    place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
     exe = Executor(place)
     exe.run(framework.default_startup_program())
 
@@ -307,14 +319,20 @@ def do_validation():
 
         return total_loss / count
 
+    iters, num_samples, start_time = 0, 0, time.time()
     for pass_id in xrange(args.pass_num):
-        pass_start_time = time.time()
-        words_seen = 0
+        train_accs = []
+        train_losses = []
         for batch_id, data in enumerate(train_batch_generator()):
+            if iters == args.skip_batch_num:
+                start_time = time.time()
+                num_samples = 0
+            if iters == args.iterations:
+                break
             src_seq, word_num = to_lodtensor(map(lambda x: x[0], data), place)
-            words_seen += word_num
+            num_samples += word_num
             trg_seq, word_num = to_lodtensor(map(lambda x: x[1], data), place)
-            words_seen += word_num
+            num_samples += word_num
             lbl_seq, _ = to_lodtensor(map(lambda x: x[2], data), place)
 
             fetch_outs = exe.run(framework.default_main_program(),
@@ -325,24 +343,36 @@ def do_validation():
                                  },
                                  fetch_list=[avg_cost])
 
-            avg_cost_val = np.array(fetch_outs[0])
-            print('pass_id=%d, batch_id=%d, train_loss: %f' %
-                  (pass_id, batch_id, avg_cost_val))
+            iters += 1
+            loss = np.array(fetch_outs[0])
+            print(
+                "Pass = %d, Iter = %d, Loss = %f" % (pass_id, iters, loss)
+            )  # The accuracy is the accumulation of batches, but not the current batch.
 
-        pass_end_time = time.time()
-        test_loss = do_validation()
-        time_consumed = pass_end_time - pass_start_time
-        words_per_sec = words_seen / time_consumed
-        print("pass_id=%d, test_loss: %f, words/s: %f, sec/pass: %f" %
-              (pass_id, test_loss, words_per_sec, time_consumed))
+        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))
+        # evaluation
+        if args.with_test:
+            test_loss = do_validation()
+        exit(0)
 
 
 def infer():
     pass
 
 
+def print_arguments(args):
+    print('----------- seq2seq Configuration Arguments -----------')
+    for arg, value in sorted(vars(args).iteritems()):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------------')
+
+
 if __name__ == '__main__':
     args = parser.parse_args()
+    print_arguments(args)
     if args.infer_only:
         infer()
     else:

benchmark/fluid/mnist.py

@@ -35,6 +35,12 @@ def parse_args():
     parser = argparse.ArgumentParser("mnist model benchmark.")
     parser.add_argument(
         '--batch_size', type=int, default=128, help='The minibatch size.')
+    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=35, help='The number of minibatches.')
     parser.add_argument(
@@ -53,19 +59,14 @@ def parse_args():
         '--use_nvprof',
         action='store_true',
         help='If set, use nvprof for CUDA.')
+    parser.add_argument(
+        '--with_test',
+        action='store_true',
+        help='If set, test the testset during training.')
     args = parser.parse_args()
     return args
 
 
-def print_arguments(args):
-    vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
-                                vars(args)['device'] == 'GPU')
-    print('-----------  Configuration Arguments -----------')
-    for arg, value in sorted(vars(args).iteritems()):
-        print('%s: %s' % (arg, value))
-    print('------------------------------------------------')
-
-
 def cnn_model(data):
     conv_pool_1 = fluid.nets.simple_img_conv_pool(
         input=data,
@@ -161,38 +162,59 @@ def run_benchmark(model, args):
         paddle.dataset.mnist.train(), batch_size=args.batch_size)
 
     accuracy = fluid.average.WeightedAverage()
+    iters, num_samples, start_time = 0, 0, time.time()
     for pass_id in range(args.pass_num):
         accuracy.reset()
-        pass_start = time.time()
+        train_accs = []
+        train_losses = []
         for batch_id, data in enumerate(train_reader()):
+            if iters == args.skip_batch_num:
+                start_time = time.time()
+                num_samples = 0
+            if iters == args.iterations:
+                break
             img_data = np.array(
                 map(lambda x: x[0].reshape([1, 28, 28]), data)).astype(DTYPE)
             y_data = np.array(map(lambda x: x[1], data)).astype("int64")
             y_data = y_data.reshape([len(y_data), 1])
 
-            start = time.time()
             outs = exe.run(
                 fluid.default_main_program(),
                 feed={"pixel": img_data,
                       "label": y_data},
                 fetch_list=[avg_cost, batch_acc, batch_size_tensor]
             )  # The accuracy is the accumulation of batches, but not the current batch.
             accuracy.add(value=outs[1], weight=outs[2])
-            end = time.time()
+            iters += 1
+            num_samples += len(y_data)
             loss = np.array(outs[0])
             acc = np.array(outs[1])
-            print("pass=%d, batch=%d, loss=%f, error=%f, elapse=%f" %
-                  (pass_id, batch_id, loss, 1 - acc, (end - start) / 1000))
+            train_losses.append(loss)
+            train_accs.append(acc)
+            print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" %
+                  (pass_id, iters, loss, acc))
+
+        print("Pass: %d, Loss: %f, Train Accuray: %f\n" %
+              (pass_id, np.mean(train_losses), np.mean(train_accs)))
+        train_elapsed = time.time() - start_time
+        examples_per_sec = num_samples / train_elapsed
 
-        pass_end = time.time()
+        print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
+              (num_samples, train_elapsed, examples_per_sec))
+        # evaluation
+        if args.with_test:
+            test_avg_acc = eval_test(exe, batch_acc, batch_size_tensor,
+                                     inference_program)
+        exit(0)
 
-        train_avg_acc = accuracy.eval()
-        test_avg_acc = eval_test(exe, batch_acc, batch_size_tensor,
-                                 inference_program)
 
-        print("pass=%d, train_avg_acc=%f, test_avg_acc=%f, elapse=%f" %
-              (pass_id, train_avg_acc, test_avg_acc,
-               (pass_end - pass_start) / 1000))
+def print_arguments(args):
+    vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
+                                vars(args)['device'] == 'GPU')
+    print('----------- mnist Configuration Arguments -----------')
+    for arg, value in sorted(vars(args).iteritems()):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------------')
 
 
 if __name__ == '__main__':

benchmark/fluid/resnet.py

@@ -87,15 +87,6 @@ def parse_args():
     return args
 
 
-def print_arguments(args):
-    vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
-                                vars(args)['device'] == 'GPU')
-    print('-----------  Configuration Arguments -----------')
-    for arg, value in sorted(vars(args).iteritems()):
-        print('%s: %s' % (arg, value))
-    print('------------------------------------------------')
-
-
 def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'):
     conv1 = fluid.layers.conv2d(
         input=input,
@@ -279,32 +270,31 @@ def test(exe):
                       'label': label},
                 fetch_list=[avg_cost, batch_acc, batch_size_tensor])
             iters += 1
-            num_samples += label[0]
+            num_samples += len(label)
             accuracy.add(value=acc, weight=weight)
             train_losses.append(loss)
             train_accs.append(acc)
             print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" %
                   (pass_id, iters, loss, acc))
-        pass_train_acc = accuracy.eval()
-        # evaluation
-        if args.with_test:
-            pass_test_acc = test(exe)
-        train_elapsed = time.time() - start_time
         print("Pass: %d, Loss: %f, Train Accuray: %f\n" %
               (pass_id, np.mean(train_losses), np.mean(train_accs)))
-
+        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))
+        # evaluation
+        if args.with_test:
+            pass_test_acc = test(exe)
+        exit(0)
 
-    if args.use_cprof:
-        pr.disable()
-        s = StringIO.StringIO()
-        sortby = 'cumulative'
-        ps = pstats.Stats(pr, stream=s).sort_stats(sortby)
-        ps.print_stats()
-        print(s.getvalue())
+
+def print_arguments(args):
+    vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
+                                vars(args)['device'] == 'GPU')
+    print('----------- resnet Configuration Arguments -----------')
+    for arg, value in sorted(vars(args).iteritems()):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------------')
 
 
 if __name__ == '__main__':

benchmark/fluid/run.sh

@@ -1,7 +1,9 @@
 #!/bin/bash
 # This script benchmarking the PaddlePaddle Fluid on
 # single thread single GPU.
-export CUDNN_PATH=/paddle/cudnn_v5/cuda/lib
+
+#export FLAGS_fraction_of_gpu_memory_to_use=0.0
+export CUDNN_PATH=/paddle/cudnn_v5
 
 # disable openmp and mkl parallel
 #https://github.com/PaddlePaddle/Paddle/issues/7199
@@ -25,25 +27,79 @@ export CUDA_VISIBLE_DEVICES=0
 export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
 export LD_LIBRARY_PATH=$CUDNN_PATH:$LD_LIBRARY_PATH
 
+# only query the gpu used
+nohup stdbuf -oL nvidia-smi \
+      --id=${CUDA_VISIBLE_DEVICES} \
+      --query-gpu=timestamp \
+      --query-compute-apps=pid,process_name,used_memory \
+      --format=csv \
+      --filename=mem.log  \
+      -l 1 &
+# mnist
+# mnist gpu mnist 128
+FLAGS_benchmark=true stdbuf -oL python fluid/mnist.py \
+               --device=GPU \
+               --batch_size=128 \
+               --skip_batch_num=5 \
+               --iterations=500 \
+               2>&1 | tee -a mnist_gpu_128.log
 
 # vgg16
-# cifar10 gpu cifar10 128
-FLAGS_benchmark=true python fluid/vgg.py \
+# gpu cifar10 128
+FLAGS_benchmark=true stdbuf -oL python fluid/vgg16.py \
                --device=GPU \
                --batch_size=128 \
                --skip_batch_num=5 \
-               --iterations=30  \
-               2>&1 > vgg16_gpu_128.log
+               --iterations=30 \
+               2>&1 | tee -a vgg16_gpu_128.log
+
+# flowers gpu  128
+FLAGS_benchmark=true stdbuf -oL python fluid/vgg16.py \
+               --device=GPU \
+               --batch_size=32 \
+               --data_set=flowers \
+               --skip_batch_num=5 \
+               --iterations=30 \
+               2>&1 | tee -a vgg16_gpu_flowers_32.log
 
 # resnet50
 # resnet50 gpu cifar10 128
-FLAGS_benchmark=true python fluid/resnet.py \
+FLAGS_benchmark=true stdbuf -oL python fluid/resnet50.py \
                --device=GPU \
                --batch_size=128 \
                --data_set=cifar10 \
                --model=resnet_cifar10 \
                --skip_batch_num=5 \
                --iterations=30 \
-               2>&1 > resnet50_gpu_128.log
+               2>&1 | tee -a resnet50_gpu_128.log
+
+# resnet50 gpu flowers 64
+FLAGS_benchmark=true stdbuf -oL python fluid/resnet50.py \
+               --device=GPU \
+               --batch_size=64 \
+               --data_set=flowers \
+               --model=resnet_imagenet \
+               --skip_batch_num=5 \
+               --iterations=30 \
+               2>&1 | tee -a resnet50_gpu_flowers_64.log
 
 # lstm
+# lstm gpu imdb 32 # tensorflow only support batch=32
+FLAGS_benchmark=true stdbuf -oL python fluid/stacked_dynamic_lstm.py \
+               --device=GPU \
+               --batch_size=32 \
+               --skip_batch_num=5 \
+               --iterations=30 \
+               --hidden_dim=512 \
+               --emb_dim=512 \
+               --crop_size=1500 \
+               2>&1 | tee -a lstm_gpu_32.log
+
+# seq2seq
+# seq2seq gpu wmb 128
+FLAGS_benchmark=true stdbuf -oL python fluid/machine_translation.py \
+               --device=GPU \
+               --batch_size=128 \
+               --skip_batch_num=5 \
+               --iterations=30 \
+               2>&1 | tee -a lstm_gpu_128.log