Add benchmark.py (#35)

docs/getting_started.md

@@ -306,6 +306,12 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} con
 CUDA_VISIBLE_DEVICES=4,5,6,7 ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} config2.py ${WORK_DIR} 4
 ```
 
+## Benchmark
+You can get average inference speed using the following script. Not that it does not includes the IO time and pre-processing time.
+```shell
+python tools/benchmark_inference.py ${MMPOSE_CONFIG_FILE}
+```
+
 ## Tutorials
 
 Currently, we provide some tutorials for users to [finetune model](tutorials/finetune.md),

tools/benchmark_inference.py

@@ -0,0 +1,72 @@
+import argparse
+import time
+
+import torch
+from mmcv import Config
+from mmcv.parallel import MMDataParallel
+
+from mmpose.core import wrap_fp16_model
+from mmpose.datasets import build_dataloader, build_dataset
+from mmpose.models import build_posenet
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='MMPose benchmark a recognizer')
+    parser.add_argument('config', help='test config file path')
+    parser.add_argument(
+        '--log-interval', default=10, help='interval of logging')
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+
+    cfg = Config.fromfile(args.config)
+    # set cudnn_benchmark
+    if cfg.get('cudnn_benchmark', False):
+        torch.backends.cudnn.benchmark = True
+
+    # build the dataloader
+    dataset = build_dataset(cfg.data.val)
+    data_loader = build_dataloader(
+        dataset,
+        samples_per_gpu=1,
+        workers_per_gpu=cfg.data.workers_per_gpu,
+        dist=False,
+        shuffle=False)
+
+    # build the model and load checkpoint
+    model = build_posenet(cfg.model)
+    fp16_cfg = cfg.get('fp16', None)
+    if fp16_cfg is not None:
+        wrap_fp16_model(model)
+    model = MMDataParallel(model, device_ids=[0])
+
+    # the first several iterations may be very slow so skip them
+    num_warmup = 5
+    pure_inf_time = 0
+
+    # benchmark with total batch and take the average
+    for i, data in enumerate(data_loader):
+
+        torch.cuda.synchronize()
+        start_time = time.perf_counter()
+        with torch.no_grad():
+            model(return_loss=False, **data)
+
+        torch.cuda.synchronize()
+        elapsed = time.perf_counter() - start_time
+
+        if i >= num_warmup:
+            pure_inf_time += elapsed
+            if (i + 1) % args.log_interval == 0:
+                its = (i + 1 - num_warmup) / pure_inf_time
+                print(f'Done item [{i + 1:<3}],  {its:.2f} items / s')
+    print(f'Overall average: {its:.2f} items / s')
+    print(f'Total time: {pure_inf_time:.2f} s')
+
+
+if __name__ == '__main__':
+    main()