app.py

import gradio as gr
import torch
import numpy as np
from transformers import AutoProcessor, AutoModel
from PIL import Image
from decord import VideoReader, cpu

# Use GPU if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torch.backends.cudnn.benchmark = True

def sample_uniform_frame_indices(clip_len, seg_len):
    if seg_len < clip_len:
        repeat_factor = np.ceil(clip_len / seg_len).astype(int)
        indices = np.arange(seg_len).tolist() * repeat_factor
        indices = indices[:clip_len]
    else:
        spacing = seg_len // clip_len
        indices = [i * spacing for i in range(clip_len)]
    return np.array(indices).astype(np.int64)

def read_video_decord(file_path, indices):
    vr = VideoReader(file_path, num_threads=1, ctx=cpu(0))
    video = vr.get_batch(indices).asnumpy()
    return video

def concatenate_frames(frames, clip_len):
    layout = {
        32: (4, 8),
        16: (4, 4),
        8:  (2, 4)
    }
    rows, cols = layout[clip_len]
    combined_image = Image.new('RGB', (frames[0].shape[1]*cols, frames[0].shape[0]*rows))
    frame_iter = iter(frames)
    y_offset = 0
    for i in range(rows):
        x_offset = 0
        for j in range(cols):
            img = Image.fromarray(next(frame_iter))
            combined_image.paste(img, (x_offset, y_offset))
            x_offset += frames[0].shape[1]
        y_offset += frames[0].shape[0]
    return combined_image

def model_interface(uploaded_video, model_choice, activity):
    clip_len = {
        "microsoft/xclip-base-patch16-zero-shot": 32,
        "microsoft/xclip-base-patch32-16-frames": 16,
        "microsoft/xclip-base-patch32": 8
    }.get(model_choice, 32)
    
    indices = sample_uniform_frame_indices(clip_len, seg_len=len(VideoReader(uploaded_video)))
    video = read_video_decord(uploaded_video, indices)
    concatenated_image = concatenate_frames(video, clip_len)

    activities_list = [activity, "other"]

    processor = AutoProcessor.from_pretrained(model_choice).to(device)
    model = AutoModel.from_pretrained(model_choice).to(device)
    
    # Convert the list of frames to a single numpy array for efficient conversion to a tensor
    video_np_array = np.array(video)
    
    inputs = processor(
        text=activities_list,
        videos=video_np_array,
        return_tensors="pt",
        padding=True,
    )

    inputs = {k: v.to(device) for k, v in inputs.items()}

    with torch.no_grad():
        outputs = model(**inputs)

    logits_per_video = outputs.logits_per_video
    probs = logits_per_video.softmax(dim=1)

    results_probs = []
    results_logits = []
    max_prob_index = torch.argmax(probs[0]).item()
    for i in range(len(activities_list)):
        current_activity = activities_list[i]
        prob = float(probs[0][i])
        logit = float(logits_per_video[0][i])
        results_probs.append((current_activity, f"Probability: {prob * 100:.2f}%"))
        results_logits.append((current_activity, f"Raw Score: {logit:.2f}"))

    likely_label = activities_list[max_prob_index]
    likely_probability = float(probs[0][max_prob_index]) * 100

    return concatenated_image, results_probs, results_logits, [ likely_label , likely_probability ]

iface = gr.Interface(
    fn=model_interface,
    inputs=[
        gr.components.Video(label="Upload a video file"),
        gr.components.Dropdown(choices=[
            "microsoft/xclip-base-patch16-zero-shot",
            "microsoft/xclip-base-patch32-16-frames",
            "microsoft/xclip-base-patch32"
        ], label="Model Choice"),
        gr.components.Textbox(default="dancing", label="Desired Activity to Recognize"),
    ],
    outputs=[
        gr.components.Image(type="pil", label="Sampled Frames"),
        gr.components.Textbox(type="text", label="Probabilities"),
        gr.components.Textbox(type="text", label="Raw Scores"),
        gr.components.Textbox(type="text", label="Top Prediction")
    ],
    live=False
)

iface.launch()