X3D

使用示例

匯入

載入模型

import torch
# Choose the `x3d_s` model
model_name = 'x3d_s'
model = torch.hub.load('facebookresearch/pytorchvideo', model_name, pretrained=True)

匯入剩餘函式

import json
import urllib
from pytorchvideo.data.encoded_video import EncodedVideo

from torchvision.transforms import Compose, Lambda
from torchvision.transforms._transforms_video import (
    CenterCropVideo,
    NormalizeVideo,
)
from pytorchvideo.transforms import (
    ApplyTransformToKey,
    ShortSideScale,
    UniformTemporalSubsample
)

設定

將模型設定為評估模式並移動到所需裝置。

# Set to GPU or CPU
device = "cpu"
model = model.eval()
model = model.to(device)

下載 Kinetics 400 資料集的 ID 到標籤對映，該資料集用於訓練 torch hub 模型。這將用於從預測的類別 ID 獲取類別標籤名稱。

json_url = "https://dl.fbaipublicfiles.com/pyslowfast/dataset/class_names/kinetics_classnames.json"
json_filename = "kinetics_classnames.json"
try: urllib.URLopener().retrieve(json_url, json_filename)
except: urllib.request.urlretrieve(json_url, json_filename)

with open(json_filename, "r") as f:
    kinetics_classnames = json.load(f)

# Create an id to label name mapping
kinetics_id_to_classname = {}
for k, v in kinetics_classnames.items():
    kinetics_id_to_classname[v] = str(k).replace('"', "")

定義輸入變換

mean = [0.45, 0.45, 0.45]
std = [0.225, 0.225, 0.225]
frames_per_second = 30
model_transform_params  = {
    "x3d_xs": {
        "side_size": 182,
        "crop_size": 182,
        "num_frames": 4,
        "sampling_rate": 12,
    },
    "x3d_s": {
        "side_size": 182,
        "crop_size": 182,
        "num_frames": 13,
        "sampling_rate": 6,
    },
    "x3d_m": {
        "side_size": 256,
        "crop_size": 256,
        "num_frames": 16,
        "sampling_rate": 5,
    }
}

# Get transform parameters based on model
transform_params = model_transform_params[model_name]

# Note that this transform is specific to the slow_R50 model.
transform =  ApplyTransformToKey(
    key="video",
    transform=Compose(
        [
            UniformTemporalSubsample(transform_params["num_frames"]),
            Lambda(lambda x: x/255.0),
            NormalizeVideo(mean, std),
            ShortSideScale(size=transform_params["side_size"]),
            CenterCropVideo(
                crop_size=(transform_params["crop_size"], transform_params["crop_size"])
            )
        ]
    ),
)

# The duration of the input clip is also specific to the model.
clip_duration = (transform_params["num_frames"] * transform_params["sampling_rate"])/frames_per_second

執行推理

下載示例影片。

url_link = "https://dl.fbaipublicfiles.com/pytorchvideo/projects/archery.mp4"
video_path = 'archery.mp4'
try: urllib.URLopener().retrieve(url_link, video_path)
except: urllib.request.urlretrieve(url_link, video_path)

載入影片並將其轉換為模型所需的輸入格式。

# Select the duration of the clip to load by specifying the start and end duration
# The start_sec should correspond to where the action occurs in the video
start_sec = 0
end_sec = start_sec + clip_duration

# Initialize an EncodedVideo helper class and load the video
video = EncodedVideo.from_path(video_path)

# Load the desired clip
video_data = video.get_clip(start_sec=start_sec, end_sec=end_sec)

# Apply a transform to normalize the video input
video_data = transform(video_data)

# Move the inputs to the desired device
inputs = video_data["video"]
inputs = inputs.to(device)

獲取預測

# Pass the input clip through the model
preds = model(inputs[None, ...])

# Get the predicted classes
post_act = torch.nn.Softmax(dim=1)
preds = post_act(preds)
pred_classes = preds.topk(k=5).indices[0]

# Map the predicted classes to the label names
pred_class_names = [kinetics_id_to_classname[int(i)] for i in pred_classes]
print("Top 5 predicted labels: %s" % ", ".join(pred_class_names))

模型描述

X3D 模型架構基於 [1]，並在 Kinetics 資料集上進行了預訓練。

架構	深度	幀長度 x 取樣率	Top 1	Top 5	浮點運算 (G)	引數 (M)
X3D	XS	4×12	69.12	88.63	0.91	3.79
X3D	S	13×6	73.33	91.27	2.96	3.79
X3D	M	16×5	75.94	92.72	6.72	3.79

參考文獻

[1] Christoph Feichtenhofer, “X3D: Expanding Architectures for Efficient Video Recognition.” https://arxiv.org/abs/2004.04730