Spaces:
Running
Running
File size: 5,487 Bytes
256a58e 5f988b9 144548a 5f988b9 0ddfda6 c1dbdfa 0ddfda6 ecfe0fd 0ddfda6 1b84ca8 0ddfda6 256a58e 1435787 256a58e 941a695 5f988b9 941a695 256a58e 5f988b9 256a58e 5f988b9 256a58e 5f988b9 cc6ed45 5f988b9 cc6ed45 5f988b9 cc6ed45 256a58e cc6ed45 256a58e 5f988b9 b4760ab b758449 5f988b9 cc6ed45 5f988b9 256a58e 0ddfda6 256a58e cc6ed45 256a58e cc6ed45 256a58e cc6ed45 256a58e 697a6d7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 |
import gradio as gr
import os
import numpy as np
import pandas as pd
from IPython import display
import faiss
import torch
from transformers import CLIPTokenizer, CLIPTextModelWithProjection
HTML="""
<!DOCTYPE html>
<html>
<style>
.container {
align-items: center;
justify-content: center;
}
img {
max-width: 10%;
max-height:10%;
float: left;
}
.text {
font-size: 20px;
padding-top: 10%;
padding-left: 20px;
padding-bottom: 5%;
float: left;
}
</style>
<body>
<div class="container">
<div class="image">
<img src="https://huggingface.co/spaces/Diangle/Clip4Clip-webvid/resolve/main/Searchium.png" width="333" height="216">
</div>
<div class="text">
<h1 style="font-size: 48px;"> Large Scale Video Search </h1>
</div>
</div>
</body>
</html>
"""
DESCRIPTION="""This is a video retrieval demo using [Diangle/clip4clip-webvid](https://huggingface.co/Diangle/clip4clip-webvid)."""
DATA_PATH = './data'
ft_visual_features_file = DATA_PATH + '/dataset_v1_visual_features_database.npy'
#load database features:
ft_visual_features_database = np.load(ft_visual_features_file)
database_csv_path = os.path.join(DATA_PATH, 'dataset_v1.csv')
database_df = pd.read_csv(database_csv_path)
class NearestNeighbors:
"""
Class for NearestNeighbors.
"""
def __init__(self, n_neighbors=10, metric='cosine', rerank_from=-1):
"""
metric = 'cosine' / 'binary'
if metric ~= 'cosine' and rerank_from > n_neighbors then a cosine rerank will be performed
"""
self.n_neighbors = n_neighbors
self.metric = metric
self.rerank_from = rerank_from
def normalize(self, a):
return a / np.sum(a**2, axis=1, keepdims=True)
def fit(self, data, o_data=None):
if self.metric == 'cosine':
data = self.normalize(data)
self.index = faiss.IndexFlatIP(data.shape[1])
elif self.metric == 'binary':
self.o_data = data if o_data is None else o_data
#assuming data already packed
self.index = faiss.IndexBinaryFlat(data.shape[1]*8)
self.index.add(np.ascontiguousarray(data))
def kneighbors(self, q_data):
if self.metric == 'cosine':
q_data = self.normalize(q_data)
sim, idx = self.index.search(q_data, self.n_neighbors)
else:
if self.metric == 'binary':
print('This is binary search.')
bq_data = np.packbits((q_data > 0.0).astype(bool), axis=1)
sim, idx = self.index.search(bq_data, max(self.rerank_from, self.n_neighbors))
if self.rerank_from > self.n_neighbors:
re_sims = np.zeros([len(q_data), self.n_neighbors], dtype=float)
re_idxs = np.zeros([len(q_data), self.n_neighbors], dtype=float)
for i, q in enumerate(q_data):
rerank_data = self.o_data[idx[i]]
rerank_search = NearestNeighbors(n_neighbors=self.n_neighbors, metric='cosine')
rerank_search.fit(rerank_data)
re_sim, re_idx = rerank_search.kneighbors(np.asarray([q]))
re_sims[i, :] = re_sim
re_idxs[i, :] = idx[i][re_idx]
idx = re_idxs
sim = re_sims
return sim, idx
model = CLIPTextModelWithProjection.from_pretrained("Diangle/clip4clip-webvid")
tokenizer = CLIPTokenizer.from_pretrained("Diangle/clip4clip-webvid")
def search(search_sentence):
inputs = tokenizer(text=search_sentence , return_tensors="pt")
outputs = model(input_ids=inputs["input_ids"], attention_mask=inputs["attention_mask"])
# Normalizing the embeddings:
final_output = outputs[0] / outputs[0].norm(dim=-1, keepdim=True)
sequence_output = final_output.cpu().detach().numpy()
nn_search = NearestNeighbors(n_neighbors=5, metric='binary', rerank_from=100)
nn_search.fit(np.packbits((ft_visual_features_database > 0.0).astype(bool), axis=1), o_data=ft_visual_features_database)
sims, idxs = nn_search.kneighbors(sequence_output)
# print(database_df.iloc[idxs[0]]['contentUrl'])
urls = database_df.iloc[idxs[0]]['contentUrl'].to_list()
AUTOPLAY_VIDEOS = []
for url in urls:
AUTOPLAY_VIDEOS.append("""<video controls muted autoplay>
<source src={} type="video/mp4">
</video>""".format(url))
return AUTOPLAY_VIDEOS
with gr.Blocks() as demo:
gr.HTML(HTML)
gr.Markdown(DESCRIPTION)
gr.Markdown("Retrieval of top 5 videos relevant to the input sentence: ")
with gr.Row():
with gr.Column():
inp = gr.Textbox(placeholder="Write a sentence.")
btn = gr.Button(value="Retrieve")
ex = [["mind-blowing magic tricks"],["baking chocolate cake"],
["birds fly in the sky"], ["natural wonders of the world"]]
gr.Examples(examples=ex,
inputs=[inp]
)
with gr.Column():
out = [gr.HTML() for _ in range(5)]
btn.click(search, inputs=inp, outputs=out)
demo.launch() |