--- a
+++ b/tests/test_runtime/test_inference.py
@@ -0,0 +1,149 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import mmcv
+import numpy as np
+import pytest
+import torch
+import torch.nn as nn
+
+from mmaction.apis import inference_recognizer, init_recognizer
+
+video_config_file = 'configs/recognition/tsn/tsn_r50_video_inference_1x1x3_100e_kinetics400_rgb.py'  # noqa: E501
+frame_config_file = 'configs/recognition/tsn/tsn_r50_inference_1x1x3_100e_kinetics400_rgb.py'  # noqa: E501
+flow_frame_config_file = 'configs/recognition/tsn/tsn_r50_320p_1x1x3_110e_kinetics400_flow.py'  # noqa: E501
+video_path = 'demo/demo.mp4'
+frames_path = 'tests/data/imgs'
+
+
+def test_init_recognizer():
+    with pytest.raises(TypeError):
+        # config must be a filename or Config object
+        init_recognizer(dict(config_file=None))
+
+    if torch.cuda.is_available():
+        device = 'cuda:0'
+    else:
+        device = 'cpu'
+
+    model = init_recognizer(video_config_file, None, device)
+
+    config = mmcv.Config.fromfile(video_config_file)
+    config.model.backbone.pretrained = None
+
+    isinstance(model, nn.Module)
+    if torch.cuda.is_available():
+        assert next(model.parameters()).is_cuda is True
+    else:
+        assert next(model.parameters()).is_cuda is False
+    assert model.cfg.model.backbone.pretrained is None
+
+
+def test_video_inference_recognizer():
+    if torch.cuda.is_available():
+        device = 'cuda:0'
+    else:
+        device = 'cpu'
+    model = init_recognizer(video_config_file, None, device)
+
+    with pytest.raises(RuntimeError):
+        # video path doesn't exist
+        inference_recognizer(model, 'missing.mp4')
+
+    for ops in model.cfg.data.test.pipeline:
+        if ops['type'] in ('TenCrop', 'ThreeCrop'):
+            # Use CenterCrop to reduce memory in order to pass CI
+            ops['type'] = 'CenterCrop'
+
+    top5_label = inference_recognizer(model, video_path)
+    scores = [item[1] for item in top5_label]
+    assert len(top5_label) == 5
+    assert scores == sorted(scores, reverse=True)
+
+    _, feat = inference_recognizer(
+        model, video_path, outputs=('backbone', 'cls_head'), as_tensor=False)
+    assert isinstance(feat, dict)
+    assert 'backbone' in feat and 'cls_head' in feat
+    assert isinstance(feat['backbone'], np.ndarray)
+    assert isinstance(feat['cls_head'], np.ndarray)
+    assert feat['backbone'].shape == (25, 2048, 7, 7)
+    assert feat['cls_head'].shape == (1, 400)
+
+    _, feat = inference_recognizer(
+        model,
+        video_path,
+        outputs=('backbone.layer3', 'backbone.layer3.1.conv1'))
+    assert 'backbone.layer3.1.conv1' in feat and 'backbone.layer3' in feat
+    assert isinstance(feat['backbone.layer3.1.conv1'], torch.Tensor)
+    assert isinstance(feat['backbone.layer3'], torch.Tensor)
+    assert feat['backbone.layer3'].size() == (25, 1024, 14, 14)
+    assert feat['backbone.layer3.1.conv1'].size() == (25, 256, 14, 14)
+
+    cfg_file = 'configs/recognition/slowfast/slowfast_r50_video_inference_4x16x1_256e_kinetics400_rgb.py'  # noqa: E501
+    sf_model = init_recognizer(cfg_file, None, device)
+    for ops in sf_model.cfg.data.test.pipeline:
+        # Changes to reduce memory in order to pass CI
+        if ops['type'] in ('TenCrop', 'ThreeCrop'):
+            ops['type'] = 'CenterCrop'
+        if ops['type'] == 'SampleFrames':
+            ops['num_clips'] = 1
+    _, feat = inference_recognizer(
+        sf_model, video_path, outputs=('backbone', 'cls_head'))
+    assert isinstance(feat, dict) and isinstance(feat['backbone'], tuple)
+    assert 'backbone' in feat and 'cls_head' in feat
+    assert len(feat['backbone']) == 2
+    assert isinstance(feat['backbone'][0], torch.Tensor)
+    assert isinstance(feat['backbone'][1], torch.Tensor)
+    assert feat['backbone'][0].size() == (1, 2048, 4, 8, 8)
+    assert feat['backbone'][1].size() == (1, 256, 32, 8, 8)
+    assert feat['cls_head'].size() == (1, 400)
+
+
+def test_frames_inference_recognizer():
+    if torch.cuda.is_available():
+        device = 'cuda:0'
+    else:
+        device = 'cpu'
+    rgb_model = init_recognizer(frame_config_file, None, device)
+    flow_model = init_recognizer(flow_frame_config_file, None, device)
+
+    with pytest.raises(RuntimeError):
+        # video path doesn't exist
+        inference_recognizer(rgb_model, 'missing_path')
+
+    for ops in rgb_model.cfg.data.test.pipeline:
+        if ops['type'] in ('TenCrop', 'ThreeCrop'):
+            # Use CenterCrop to reduce memory in order to pass CI
+            ops['type'] = 'CenterCrop'
+            ops['crop_size'] = 224
+    for ops in flow_model.cfg.data.test.pipeline:
+        if ops['type'] in ('TenCrop', 'ThreeCrop'):
+            # Use CenterCrop to reduce memory in order to pass CI
+            ops['type'] = 'CenterCrop'
+            ops['crop_size'] = 224
+
+    top5_label = inference_recognizer(rgb_model, frames_path)
+    scores = [item[1] for item in top5_label]
+    assert len(top5_label) == 5
+    assert scores == sorted(scores, reverse=True)
+
+    _, feat = inference_recognizer(
+        flow_model,
+        frames_path,
+        outputs=('backbone', 'cls_head'),
+        as_tensor=False)
+    assert isinstance(feat, dict)
+    assert 'backbone' in feat and 'cls_head' in feat
+    assert isinstance(feat['backbone'], np.ndarray)
+    assert isinstance(feat['cls_head'], np.ndarray)
+    assert feat['backbone'].shape == (25, 2048, 7, 7)
+    assert feat['cls_head'].shape == (1, 400)
+
+    _, feat = inference_recognizer(
+        rgb_model,
+        frames_path,
+        outputs=('backbone.layer3', 'backbone.layer3.1.conv1'))
+
+    assert 'backbone.layer3.1.conv1' in feat and 'backbone.layer3' in feat
+    assert isinstance(feat['backbone.layer3.1.conv1'], torch.Tensor)
+    assert isinstance(feat['backbone.layer3'], torch.Tensor)
+    assert feat['backbone.layer3'].size() == (25, 1024, 14, 14)
+    assert feat['backbone.layer3.1.conv1'].size() == (25, 256, 14, 14)