Merge branch 'speed_script' into 'master'

Speed benchmark script

See merge request open-mmlab/mmaction-lite!325

tools/benchmark.py

+import argparse
+import time
+
+import torch
+from mmcv import Config
+from mmcv.parallel import MMDataParallel
+from tools.fuse_conv_bn import fuse_module
+
+from mmaction.core import wrap_fp16_model
+from mmaction.datasets import build_dataloader, build_dataset
+from mmaction.models import build_model
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='MMAction benchmark a recognizer')
+    parser.add_argument('config', help='test config file path')
+    parser.add_argument(
+        '--log-interval', default=10, help='interval of logging')
+    parser.add_argument(
+        '--fuse-conv-bn',
+        action='store_true',
+        help='Whether to fuse conv and bn, this will slightly increase'
+        'the inference speed')
+    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
+    cfg.model.backbone.pretrained = None
+    cfg.data.test.test_mode = True
+
+    # build the dataloader
+    dataset = build_dataset(cfg.data.test, dict(test_mode=True))
+    data_loader = build_dataloader(
+        dataset,
+        videos_per_gpu=1,
+        workers_per_gpu=cfg.data.workers_per_gpu,
+        dist=False,
+        shuffle=False)
+
+    # build the model and load checkpoint
+    model = build_model(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
+    fp16_cfg = cfg.get('fp16', None)
+    if fp16_cfg is not None:
+        wrap_fp16_model(model)
+    if args.fuse_conv_bn:
+        model = fuse_module(model)
+
+    model = MMDataParallel(model, device_ids=[0])
+
+    model.eval()
+
+    # the first several iterations may be very slow so skip them
+    num_warmup = 5
+    pure_inf_time = 0
+
+    # benchmark with 2000 video 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:
+                fps = (i + 1 - num_warmup) / pure_inf_time
+                print(
+                    f'Done video [{i + 1:<3}/ 2000], fps: {fps:.1f} video / s')
+
+        if (i + 1) == 200:
+            pure_inf_time += elapsed
+            fps = (i + 1 - num_warmup) / pure_inf_time
+            print(f'Overall fps: {fps:.1f} video / s')
+            break
+
+
+if __name__ == '__main__':
+    main()

tools/fuse_conv_bn.py

+import argparse
+
+import torch
+import torch.nn as nn
+from mmcv.runner import save_checkpoint
+
+from mmaction.apis import init_recognizer
+
+
+def fuse_conv_bn(conv, bn):
+    """During inference, the functionary of batch norm layers is turned off but
+    only the mean and var alone channels are used, which exposes the chance to
+    fuse it with the preceding conv layers to save computations and simplify
+    network structures."""
+    conv_w = conv.weight
+    conv_b = conv.bias if conv.bias is not None else torch.zeros_like(
+        bn.running_mean)
+
+    factor = bn.weight / torch.sqrt(bn.running_var + bn.eps)
+    conv.weight = nn.Parameter(conv_w *
+                               factor.reshape([conv.out_channels, 1, 1, 1]))
+    conv.bias = nn.Parameter((conv_b - bn.running_mean) * factor + bn.bias)
+    return conv
+
+
+def fuse_module(m):
+    last_conv = None
+    last_conv_name = None
+
+    for name, child in m.named_children():
+        if isinstance(child, (nn.BatchNorm2d, nn.SyncBatchNorm)):
+            if last_conv is None:  # only fuse BN that is after Conv
+                continue
+            fused_conv = fuse_conv_bn(last_conv, child)
+            m._modules[last_conv_name] = fused_conv
+            # To reduce changes, set BN as Identity instead of deleting it.
+            m._modules[name] = nn.Identity()
+            last_conv = None
+        elif isinstance(child, nn.Conv2d):
+            last_conv = child
+            last_conv_name = name
+        else:
+            fuse_module(child)
+    return m
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='fuse Conv and BN layers in a model')
+    parser.add_argument('config', help='config file path')
+    parser.add_argument('checkpoint', help='checkpoint file path')
+    parser.add_argument('out', help='output path of the converted model')
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+    # build the model from a config file and a checkpoint file
+    model = init_recognizer(args.config, args.checkpoint)
+    # fuse conv and bn layers of the model
+    fused_model = fuse_module(model)
+    save_checkpoint(fused_model, args.out)
+
+
+if __name__ == '__main__':
+    main()