Shortcuts

A 20-Minute Guide to MMAction2 FrameWork

In this tutorial, we will demonstrate the overall architecture of our MMACTION2 1.0 through a step-by-step example of video action recognition.

The structure of this tutorial is as follows:

First, we need to initialize the scope for registry, to ensure that each module is registered under the scope of mmaction. For more detailed information about registry, please refer to MMEngine Tutorial.

from mmaction.utils import register_all_modules

register_all_modules(init_default_scope=True)

Step0: Prepare Data

Please download our self-made kinetics400_tiny dataset and extract it to the $MMACTION2/data directory. The directory structure after extraction should be as follows:

mmaction2
├── data
│   ├── kinetics400_tiny
│   │    ├── kinetics_tiny_train_video.txt
│   │    ├── kinetics_tiny_val_video.txt
│   │    ├── train
│   │    │   ├── 27_CSXByd3s.mp4
│   │    │   ├── 34XczvTaRiI.mp4
│   │    │   ├── A-wiliK50Zw.mp4
│   │    │   ├── ...
│   │    └── val
│   │       ├── 0pVGiAU6XEA.mp4
│   │       ├── AQrbRSnRt8M.mp4
│   │       ├── ...

Here are some examples from the annotation file kinetics_tiny_train_video.txt:

D32_1gwq35E.mp4 0
iRuyZSKhHRg.mp4 1
oXy-e_P_cAI.mp4 0
34XczvTaRiI.mp4 1
h2YqqUhnR34.mp4 0

Each line in the file represents the annotation of a video, where the first item denotes the video filename (e.g., D32_1gwq35E.mp4), and the second item represents the corresponding label (e.g., label 0 for D32_1gwq35E.mp4). In this dataset, there are only two categories.

Step1: Build a Pipeline

In order to decode, sample, resize, crop, format, and pack the input video and corresponding annotation, we need to design a pipeline to handle these processes. Specifically, we design seven Transform classes to build this video processing pipeline. Note that all Transform classes in OpenMMLab must inherit from the BaseTransform class in mmcv, implement the abstract method transform, and be registered to the TRANSFORMS registry. For more detailed information about data transform, please refer to MMEngine Tutorial.

import mmcv
import decord
import numpy as np
from mmcv.transforms import TRANSFORMS, BaseTransform, to_tensor
from mmaction.structures import ActionDataSample


@TRANSFORMS.register_module()
class VideoInit(BaseTransform):
    def transform(self, results):
        container = decord.VideoReader(results['filename'])
        results['total_frames'] = len(container)
        results['video_reader'] = container
        return results


@TRANSFORMS.register_module()
class VideoSample(BaseTransform):
    def __init__(self, clip_len, num_clips, test_mode=False):
        self.clip_len = clip_len
        self.num_clips = num_clips
        self.test_mode = test_mode

    def transform(self, results):
        total_frames = results['total_frames']
        interval = total_frames // self.clip_len

        if self.test_mode:
            # Make the sampling during testing deterministic
            np.random.seed(42)

        inds_of_all_clips = []
        for i in range(self.num_clips):
            bids = np.arange(self.clip_len) * interval
            offset = np.random.randint(interval, size=bids.shape)
            inds = bids + offset
            inds_of_all_clips.append(inds)

        results['frame_inds'] = np.concatenate(inds_of_all_clips)
        results['clip_len'] = self.clip_len
        results['num_clips'] = self.num_clips
        return results


@TRANSFORMS.register_module()
class VideoDecode(BaseTransform):
    def transform(self, results):
        frame_inds = results['frame_inds']
        container = results['video_reader']

        imgs = container.get_batch(frame_inds).asnumpy()
        imgs = list(imgs)

        results['video_reader'] = None
        del container

        results['imgs'] = imgs
        results['img_shape'] = imgs[0].shape[:2]
        return results


@TRANSFORMS.register_module()
class VideoResize(BaseTransform):
    def __init__(self, r_size):
        self.r_size = (np.inf, r_size)

    def transform(self, results):
        img_h, img_w = results['img_shape']
        new_w, new_h = mmcv.rescale_size((img_w, img_h), self.r_size)

        imgs = [mmcv.imresize(img, (new_w, new_h))
                for img in results['imgs']]
        results['imgs'] = imgs
        results['img_shape'] = imgs[0].shape[:2]
        return results


@TRANSFORMS.register_module()
class VideoCrop(BaseTransform):
    def __init__(self, c_size):
        self.c_size = c_size

    def transform(self, results):
        img_h, img_w = results['img_shape']
        center_x, center_y = img_w // 2, img_h // 2
        x1, x2 = center_x - self.c_size // 2, center_x + self.c_size // 2
        y1, y2 = center_y - self.c_size // 2, center_y + self.c_size // 2
        imgs = [img[y1:y2, x1:x2] for img in results['imgs']]
        results['imgs'] = imgs
        results['img_shape'] = imgs[0].shape[:2]
        return results


@TRANSFORMS.register_module()
class VideoFormat(BaseTransform):
    def transform(self, results):
        num_clips = results['num_clips']
        clip_len = results['clip_len']
        imgs = results['imgs']

        # [num_clips*clip_len, H, W, C]
        imgs = np.array(imgs)
        # [num_clips, clip_len, H, W, C]
        imgs = imgs.reshape((num_clips, clip_len) + imgs.shape[1:])
        # [num_clips, C, clip_len, H, W]
        imgs = imgs.transpose(0, 4, 1, 2, 3)

        results['imgs'] = imgs
        return results


@TRANSFORMS.register_module()
class VideoPack(BaseTransform):
    def __init__(self, meta_keys=('img_shape', 'num_clips', 'clip_len')):
        self.meta_keys = meta_keys

    def transform(self, results):
        packed_results = dict()
        inputs = to_tensor(results['imgs'])
        data_sample = ActionDataSample()
        data_sample.set_gt_label(results['label'])
        metainfo = {k: results[k] for k in self.meta_keys if k in results}
        data_sample.set_metainfo(metainfo)
        packed_results['inputs'] = inputs
        packed_results['data_samples'] = data_sample
        return packed_results

Below, we provide a code snippet (using D32_1gwq35E.mp4 0 from the annotation file) to demonstrate how to use the pipeline.

import os.path as osp
from mmengine.dataset import Compose

pipeline_cfg = [
    dict(type='VideoInit'),
    dict(type='VideoSample', clip_len=16, num_clips=1, test_mode=False),
    dict(type='VideoDecode'),
    dict(type='VideoResize', r_size=256),
    dict(type='VideoCrop', c_size=224),
    dict(type='VideoFormat'),
    dict(type='VideoPack')
]

pipeline = Compose(pipeline_cfg)
data_prefix = 'data/kinetics400_tiny/train'
results = dict(filename=osp.join(data_prefix, 'D32_1gwq35E.mp4'), label=0)
packed_results = pipeline(results)

inputs = packed_results['inputs']
data_sample = packed_results['data_samples']

print('shape of the inputs: ', inputs.shape)

# Get metainfo of the inputs
print('image_shape: ', data_sample.img_shape)
print('num_clips: ', data_sample.num_clips)
print('clip_len: ', data_sample.clip_len)

# Get label of the inputs
print('label: ', data_sample.gt_label)
shape of the inputs:  torch.Size([1, 3, 16, 224, 224])
image_shape:  (224, 224)
num_clips:  1
clip_len:  16
label:  tensor([0])

Step2: Build a Dataset and DataLoader

All Dataset classes in OpenMMLab must inherit from the BaseDataset class in mmengine. We can customize annotation loading process by overriding the load_data_list method. Additionally, we can add more information to the results dict that is passed as input to the pipeline by overriding the get_data_info method. For more detailed information about BaseDataset class, please refer to MMEngine Tutorial.

import os.path as osp
from mmengine.fileio import list_from_file
from mmengine.dataset import BaseDataset
from mmaction.registry import DATASETS


@DATASETS.register_module()
class DatasetZelda(BaseDataset):
    def __init__(self, ann_file, pipeline, data_root, data_prefix=dict(video=''),
                 test_mode=False, modality='RGB', **kwargs):
        self.modality = modality
        super(DatasetZelda, self).__init__(ann_file=ann_file, pipeline=pipeline, data_root=data_root,
                                           data_prefix=data_prefix, test_mode=test_mode,
                                           **kwargs)

    def load_data_list(self):
        data_list = []
        fin = list_from_file(self.ann_file)
        for line in fin:
            line_split = line.strip().split()
            filename, label = line_split
            label = int(label)
            filename = osp.join(self.data_prefix['video'], filename)
            data_list.append(dict(filename=filename, label=label))
        return data_list

    def get_data_info(self, idx: int) -> dict:
        data_info = super().get_data_info(idx)
        data_info['modality'] = self.modality
        return data_info

Next, we will demonstrate how to use dataset and dataloader to index data. We will use the Runner.build_dataloader method to construct the dataloader. For more detailed information about dataloader, please refer to MMEngine Tutorial.

from mmaction.registry import DATASETS

train_pipeline_cfg = [
    dict(type='VideoInit'),
    dict(type='VideoSample', clip_len=16, num_clips=1, test_mode=False),
    dict(type='VideoDecode'),
    dict(type='VideoResize', r_size=256),
    dict(type='VideoCrop', c_size=224),
    dict(type='VideoFormat'),
    dict(type='VideoPack')
]

val_pipeline_cfg = [
    dict(type='VideoInit'),
    dict(type='VideoSample', clip_len=16, num_clips=5, test_mode=True),
    dict(type='VideoDecode'),
    dict(type='VideoResize', r_size=256),
    dict(type='VideoCrop', c_size=224),
    dict(type='VideoFormat'),
    dict(type='VideoPack')
]

train_dataset_cfg = dict(
    type='DatasetZelda',
    ann_file='kinetics_tiny_train_video.txt',
    pipeline=train_pipeline_cfg,
    data_root='data/kinetics400_tiny/',
    data_prefix=dict(video='train'))

val_dataset_cfg = dict(
    type='DatasetZelda',
    ann_file='kinetics_tiny_val_video.txt',
    pipeline=val_pipeline_cfg,
    data_root='data/kinetics400_tiny/',
    data_prefix=dict(video='val'))

train_dataset = DATASETS.build(train_dataset_cfg)

packed_results = train_dataset[0]

inputs = packed_results['inputs']
data_sample = packed_results['data_samples']

print('shape of the inputs: ', inputs.shape)

# Get metainfo of the inputs
print('image_shape: ', data_sample.img_shape)
print('num_clips: ', data_sample.num_clips)
print('clip_len: ', data_sample.clip_len)

# Get label of the inputs
print('label: ', data_sample.gt_label)

from mmengine.runner import Runner

BATCH_SIZE = 2

train_dataloader_cfg = dict(
    batch_size=BATCH_SIZE,
    num_workers=0,
    persistent_workers=False,
    sampler=dict(type='DefaultSampler', shuffle=True),
    dataset=train_dataset_cfg)

val_dataloader_cfg = dict(
    batch_size=BATCH_SIZE,
    num_workers=0,
    persistent_workers=False,
    sampler=dict(type='DefaultSampler', shuffle=False),
    dataset=val_dataset_cfg)

train_data_loader = Runner.build_dataloader(dataloader=train_dataloader_cfg)
val_data_loader = Runner.build_dataloader(dataloader=val_dataloader_cfg)

batched_packed_results = next(iter(train_data_loader))

batched_inputs = batched_packed_results['inputs']
batched_data_sample = batched_packed_results['data_samples']

assert len(batched_inputs) == BATCH_SIZE
assert len(batched_data_sample) == BATCH_SIZE

The terminal output should be the same as the one shown in the Step1: Build a Pipeline.

Step3: Build a Recognizer

Next, we will construct the recognizer, which mainly consists of three parts: data preprocessor for batching and normalizing the data, backbone for feature extraction, and cls_head for classification.

The implementation of data_preprocessor is as follows:

import torch
from mmengine.model import BaseDataPreprocessor, stack_batch
from mmaction.registry import MODELS


@MODELS.register_module()
class DataPreprocessorZelda(BaseDataPreprocessor):
    def __init__(self, mean, std):
        super().__init__()

        self.register_buffer(
            'mean',
            torch.tensor(mean, dtype=torch.float32).view(-1, 1, 1, 1),
            False)
        self.register_buffer(
            'std',
            torch.tensor(std, dtype=torch.float32).view(-1, 1, 1, 1),
            False)

    def forward(self, data, training=False):
        data = self.cast_data(data)
        inputs = data['inputs']
        batch_inputs = stack_batch(inputs)  # Batching
        batch_inputs = (batch_inputs - self.mean) / self.std  # Normalization
        data['inputs'] = batch_inputs
        return data

Here is the usage of data_preprocessor: feed the batched_packed_results obtained from the Step2: Build a Dataset and DataLoader into the data_preprocessor for batching and normalization.

from mmaction.registry import MODELS

data_preprocessor_cfg = dict(
    type='DataPreprocessorZelda',
    mean=[123.675, 116.28, 103.53],
    std=[58.395, 57.12, 57.375])

data_preprocessor = MODELS.build(data_preprocessor_cfg)

preprocessed_inputs = data_preprocessor(batched_packed_results)
print(preprocessed_inputs['inputs'].shape)
torch.Size([2, 1, 3, 16, 224, 224])

The implementations of backbone, cls_head and recognizer are as follows:

import torch
import torch.nn as nn
import torch.nn.functional as F
from mmengine.model import BaseModel, BaseModule, Sequential
from mmengine.structures import LabelData
from mmaction.registry import MODELS


@MODELS.register_module()
class BackBoneZelda(BaseModule):
    def __init__(self, init_cfg=None):
        if init_cfg is None:
            init_cfg = [dict(type='Kaiming', layer='Conv3d', mode='fan_out', nonlinearity="relu"),
                        dict(type='Constant', layer='BatchNorm3d', val=1, bias=0)]

        super(BackBoneZelda, self).__init__(init_cfg=init_cfg)

        self.conv1 = Sequential(nn.Conv3d(3, 64, kernel_size=(3, 7, 7),
                                          stride=(1, 2, 2), padding=(1, 3, 3)),
                                nn.BatchNorm3d(64), nn.ReLU())
        self.maxpool = nn.MaxPool3d(kernel_size=(1, 3, 3), stride=(1, 2, 2),
                                    padding=(0, 1, 1))

        self.conv = Sequential(nn.Conv3d(64, 128, kernel_size=3, stride=2, padding=1),
                               nn.BatchNorm3d(128), nn.ReLU())

    def forward(self, imgs):
        # imgs: [batch_size*num_views, 3, T, H, W]
        # features: [batch_size*num_views, 128, T/2, H//8, W//8]
        features = self.conv(self.maxpool(self.conv1(imgs)))
        return features


@MODELS.register_module()
class ClsHeadZelda(BaseModule):
    def __init__(self, num_classes, in_channels, dropout=0.5, average_clips='prob', init_cfg=None):
        if init_cfg is None:
            init_cfg = dict(type='Normal', layer='Linear', std=0.01)

        super(ClsHeadZelda, self).__init__(init_cfg=init_cfg)

        self.num_classes = num_classes
        self.in_channels = in_channels
        self.average_clips = average_clips

        if dropout != 0:
            self.dropout = nn.Dropout(dropout)
        else:
            self.dropout = None

        self.fc = nn.Linear(self.in_channels, self.num_classes)
        self.pool = nn.AdaptiveAvgPool3d(1)
        self.loss_fn = nn.CrossEntropyLoss()

    def forward(self, x):
        N, C, T, H, W = x.shape
        x = self.pool(x)
        x = x.view(N, C)
        assert x.shape[1] == self.in_channels

        if self.dropout is not None:
            x = self.dropout(x)

        cls_scores = self.fc(x)
        return cls_scores

    def loss(self, feats, data_samples):
        cls_scores = self(feats)
        labels = torch.stack([x.gt_label for x in data_samples])
        labels = labels.squeeze()

        if labels.shape == torch.Size([]):
            labels = labels.unsqueeze(0)

        loss_cls = self.loss_fn(cls_scores, labels)
        return dict(loss_cls=loss_cls)

    def predict(self, feats, data_samples):
        cls_scores = self(feats)
        num_views = cls_scores.shape[0] // len(data_samples)
        # assert num_views == data_samples[0].num_clips
        cls_scores = self.average_clip(cls_scores, num_views)

        for ds, sc in zip(data_samples, cls_scores):
            pred = LabelData(item=sc)
            ds.pred_scores = pred
        return data_samples

    def average_clip(self, cls_scores, num_views):
          if self.average_clips not in ['score', 'prob', None]:
            raise ValueError(f'{self.average_clips} is not supported. '
                             f'Currently supported ones are '
                             f'["score", "prob", None]')

          total_views = cls_scores.shape[0]
          cls_scores = cls_scores.view(total_views // num_views, num_views, -1)

          if self.average_clips is None:
              return cls_scores
          elif self.average_clips == 'prob':
              cls_scores = F.softmax(cls_scores, dim=2).mean(dim=1)
          elif self.average_clips == 'score':
              cls_scores = cls_scores.mean(dim=1)

          return cls_scores


@MODELS.register_module()
class RecognizerZelda(BaseModel):
    def __init__(self, backbone, cls_head, data_preprocessor):
        super().__init__(data_preprocessor=data_preprocessor)

        self.backbone = MODELS.build(backbone)
        self.cls_head = MODELS.build(cls_head)

    def extract_feat(self, inputs):
        inputs = inputs.view((-1, ) + inputs.shape[2:])
        return self.backbone(inputs)

    def loss(self, inputs, data_samples):
        feats = self.extract_feat(inputs)
        loss = self.cls_head.loss(feats, data_samples)
        return loss

    def predict(self, inputs, data_samples):
        feats = self.extract_feat(inputs)
        predictions = self.cls_head.predict(feats, data_samples)
        return predictions

    def forward(self, inputs, data_samples=None, mode='tensor'):
        if mode == 'tensor':
            return self.extract_feat(inputs)
        elif mode == 'loss':
            return self.loss(inputs, data_samples)
        elif mode == 'predict':
            return self.predict(inputs, data_samples)
        else:
            raise RuntimeError(f'Invalid mode: {mode}')

The init_cfg is used for model weight initialization. For more information on model weight initialization, please refer to MMEngine Tutorial. The usage of the above modules is as follows:

import torch
import copy
from mmaction.registry import MODELS

model_cfg = dict(
    type='RecognizerZelda',
    backbone=dict(type='BackBoneZelda'),
    cls_head=dict(
        type='ClsHeadZelda',
        num_classes=2,
        in_channels=128,
        average_clips='prob'),
    data_preprocessor = dict(
        type='DataPreprocessorZelda',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375]))

model = MODELS.build(model_cfg)

# Train
model.train()
model.init_weights()
data_batch_train = copy.deepcopy(batched_packed_results)
data = model.data_preprocessor(data_batch_train, training=True)
loss = model(**data, mode='loss')
print('loss dict: ', loss)

# Test
with torch.no_grad():
    model.eval()
    data_batch_test = copy.deepcopy(batched_packed_results)
    data = model.data_preprocessor(data_batch_test, training=False)
    predictions = model(**data, mode='predict')
print('Label of Sample[0]', predictions[0].gt_label)
print('Scores of Sample[0]', predictions[0].pred_score)
04/03 23:28:01 - mmengine - INFO -
backbone.conv1.0.weight - torch.Size([64, 3, 3, 7, 7]):
KaimingInit: a=0, mode=fan_out, nonlinearity=relu, distribution =normal, bias=0

04/03 23:28:01 - mmengine - INFO -
backbone.conv1.0.bias - torch.Size([64]):
KaimingInit: a=0, mode=fan_out, nonlinearity=relu, distribution =normal, bias=0

04/03 23:28:01 - mmengine - INFO -
backbone.conv1.1.weight - torch.Size([64]):
The value is the same before and after calling `init_weights` of RecognizerZelda

04/03 23:28:01 - mmengine - INFO -
backbone.conv1.1.bias - torch.Size([64]):
The value is the same before and after calling `init_weights` of RecognizerZelda

04/03 23:28:01 - mmengine - INFO -
backbone.conv.0.weight - torch.Size([128, 64, 3, 3, 3]):
KaimingInit: a=0, mode=fan_out, nonlinearity=relu, distribution =normal, bias=0

04/03 23:28:01 - mmengine - INFO -
backbone.conv.0.bias - torch.Size([128]):
KaimingInit: a=0, mode=fan_out, nonlinearity=relu, distribution =normal, bias=0

04/03 23:28:01 - mmengine - INFO -
backbone.conv.1.weight - torch.Size([128]):
The value is the same before and after calling `init_weights` of RecognizerZelda

04/03 23:28:01 - mmengine - INFO -
backbone.conv.1.bias - torch.Size([128]):
The value is the same before and after calling `init_weights` of RecognizerZelda

04/03 23:28:01 - mmengine - INFO -
cls_head.fc.weight - torch.Size([2, 128]):
NormalInit: mean=0, std=0.01, bias=0

04/03 23:28:01 - mmengine - INFO -
cls_head.fc.bias - torch.Size([2]):
NormalInit: mean=0, std=0.01, bias=0

loss dict:  {'loss_cls': tensor(0.6853, grad_fn=<NllLossBackward0>)}
Label of Sample[0] tensor([0])
Scores of Sample[0] tensor([0.5240, 0.4760])

Step4: Build a Evaluation Metric

Note that all Metric classes in OpenMMLab must inherit from the BaseMetric class in mmengine and implement the abstract methods, process and compute_metrics. For more information on evaluation, please refer to MMEngine Tutorial.

import copy
from collections import OrderedDict
from mmengine.evaluator import BaseMetric
from mmaction.evaluation import top_k_accuracy
from mmaction.registry import METRICS


@METRICS.register_module()
class AccuracyMetric(BaseMetric):
    def __init__(self, topk=(1, 5), collect_device='cpu', prefix='acc'):
        super().__init__(collect_device=collect_device, prefix=prefix)
        self.topk = topk

    def process(self, data_batch, data_samples):
        data_samples = copy.deepcopy(data_samples)
        for data_sample in data_samples:
            result = dict()
            scores = data_sample['pred_score'].cpu().numpy()
            label = data_sample['gt_label'].item()
            result['scores'] = scores
            result['label'] = label
            self.results.append(result)

    def compute_metrics(self, results: list) -> dict:
        eval_results = OrderedDict()
        labels = [res['label'] for res in results]
        scores = [res['scores'] for res in results]
        topk_acc = top_k_accuracy(scores, labels, self.topk)
        for k, acc in zip(self.topk, topk_acc):
            eval_results[f'topk{k}'] = acc
        return eval_results
from mmaction.registry import METRICS

metric_cfg = dict(type='AccuracyMetric', topk=(1, 5))

metric = METRICS.build(metric_cfg)

data_samples = [d.to_dict() for d in predictions]

metric.process(batched_packed_results, data_samples)
acc = metric.compute_metrics(metric.results)
print(acc)
OrderedDict([('topk1', 0.5), ('topk5', 1.0)])

Step5: Train and Test with Native PyTorch

import torch.optim as optim
from mmengine import track_iter_progress


device = 'cuda' # or 'cpu'
max_epochs = 10

optimizer = optim.Adam(model.parameters(), lr=0.01)

for epoch in range(max_epochs):
    model.train()
    losses = []
    for data_batch in track_iter_progress(train_data_loader):
        data = model.data_preprocessor(data_batch, training=True)
        loss_dict = model(**data, mode='loss')
        loss = loss_dict['loss_cls']

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        losses.append(loss.item())

    print(f'Epoch[{epoch}]: loss ', sum(losses) / len(train_data_loader))

    with torch.no_grad():
        model.eval()
        for data_batch in track_iter_progress(val_data_loader):
            data = model.data_preprocessor(data_batch, training=False)
            predictions = model(**data, mode='predict')
            data_samples = [d.to_dict() for d in predictions]
            metric.process(data_batch, data_samples)

        acc = metric.acc = metric.compute_metrics(metric.results)
        for name, topk in acc.items():
            print(f'{name}: ', topk)