Skip to content

Latest commit

 

History

History
192 lines (156 loc) · 5.3 KB

facebookresearch_pytorchvideo_x3d.md

File metadata and controls

192 lines (156 loc) · 5.3 KB
Raw
layout background-class body-class category title summary image author tags github-link github-id featured_image_1 featured_image_2 accelerator demo-model-link
hub_detail
hub-background
hub
researchers
X3D
X3D networks pretrained on the Kinetics 400 dataset
x3d.png
FAIR PyTorchVideo
vision
facebookresearch/pytorchvideo
no-image
no-image
“cuda-optional”

Example Usage

Imports

Load the model:

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

Import remaining functions:

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
)

Setup

Set the model to eval mode and move to desired device.

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

Download the id to label mapping for the Kinetics 400 dataset on which the torch hub models were trained. This will be used to get the category label names from the predicted class ids.

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('"', "")

Define input transform

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

Run Inference

Download an example video.

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)

Load the video and transform it to the input format required by the model.

# 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)

Get Predictions

# 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))

Model Description

X3D model architectures are based on [1] pretrained on the Kinetics dataset.

| arch | depth | frame length x sample rate | top 1 | top 5 | Flops (G) | Params (M) | | --------------- | ----------- | ----------- | ----------- | ----------- | ----------- | ----------- | ----------- | | X3D | XS | 4x12 | 69.12 | 88.63 | 0.91 | 3.79 | | X3D | S | 13x6 | 73.33 | 91.27 | 2.96 | 3.79 | | X3D | M | 16x5 | 75.94 | 92.72 | 6.72 | 3.79 |

References

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