Spaces:
Runtime error
Runtime error
File size: 6,567 Bytes
f0e0019 3b8674e f0e0019 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 |
from transformers import PaliGemmaProcessor, PaliGemmaForConditionalGeneration
from typing import List
import os
import supervision as sv
import uuid
from tqdm import tqdm
import gradio as gr
import torch
from PIL import Image
import spaces
import flax.linen as nn
import jax
import string
import functools
import jax.numpy as jnp
import numpy as np
import re
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_id = "google/paligemma-3b-mix-448"
model = PaliGemmaForConditionalGeneration.from_pretrained(model_id).eval().to(device)
processor = PaliGemmaProcessor.from_pretrained(model_id)
BOUNDING_BOX_ANNOTATOR = sv.BoundingBoxAnnotator()
MASK_ANNOTATOR = sv.MaskAnnotator()
LABEL_ANNOTATOR = sv.LabelAnnotator()
def calculate_end_frame_index(source_video_path):
video_info = sv.VideoInfo.from_video_path(source_video_path)
return min(
video_info.total_frames,
video_info.fps * 2
)
def annotate_image(
input_image,
detections,
labels
) -> np.ndarray:
output_image = MASK_ANNOTATOR.annotate(input_image, detections)
output_image = BOUNDING_BOX_ANNOTATOR.annotate(output_image, detections)
output_image = LABEL_ANNOTATOR.annotate(output_image, detections, labels=labels)
return output_image
@spaces.GPU
def process_video(
input_video,
labels,
progress=gr.Progress(track_tqdm=True)
):
video_info = sv.VideoInfo.from_video_path(input_video)
total = calculate_end_frame_index(input_video)
frame_generator = sv.get_video_frames_generator(
source_path=input_video,
end=total
)
result_file_name = f"{uuid.uuid4()}.mp4"
result_file_path = os.path.join("./", result_file_name)
with sv.VideoSink(result_file_path, video_info=video_info) as sink:
for _ in tqdm(range(total), desc="Processing video.."):
frame = next(frame_generator)
# list of dict of {"box": box, "mask":mask, "score":score, "label":label}
results, input_list = parse_detection(frame, labels)
detections = sv.Detections.from_transformers(results[0])
final_labels = []
for id in results[0]["labels"]:
final_labels.append(input_list[id])
frame = annotate_image(
input_image=frame,
detections=detections,
labels=final_labels,
)
sink.write_frame(frame)
return result_file_path
@spaces.GPU
def infer(
image: Image.Image,
text: str,
max_new_tokens: int
) -> str:
inputs = processor(text=text, images=image, return_tensors="pt").to(device)
with torch.inference_mode():
generated_ids = model.generate(
**inputs,
max_new_tokens=max_new_tokens,
do_sample=False
)
result = processor.batch_decode(generated_ids, skip_special_tokens=True)
return result[0][len(text):].lstrip("\n")
def parse_detection(input_image, input_text):
prompt = f"detect {input_text}"
out = infer(input_image, prompt, max_new_tokens=100)
objs = extract_objs(out.lstrip("\n"), input_image.shape[0], input_image.shape[1], unique_labels=True)
labels = list(obj.get('name') for obj in objs if obj.get('name'))
print("labels", labels)
input_list = input_text.split(";")
for ind, input in enumerate(input_list):
input_list[ind] = remove_special_characters(input).lstrip("\n").rstrip("\n")
label_indices = []
for label in labels:
label = remove_special_characters(label)
label_indices.append(input_list.index(label))
label_indices = torch.tensor(label_indices).to("cuda")
boxes = torch.tensor([list(obj["xyxy"]) for obj in objs])
return [{"boxes": boxes, "scores":torch.tensor([0.99 for _ in range(len(boxes))]).to("cuda"), "labels":label_indices}], input_list
_MODEL_PATH = 'vae-oid.npz'
_SEGMENT_DETECT_RE = re.compile(
r'(.*?)' +
r'<loc(\d{4})>' * 4 + r'\s*' +
'(?:%s)?' % (r'<seg(\d{3})>' * 16) +
r'\s*([^;<>]+)? ?(?:; )?',
)
def _quantized_values_from_codebook_indices(codebook_indices, embeddings):
batch_size, num_tokens = codebook_indices.shape
assert num_tokens == 16, codebook_indices.shape
unused_num_embeddings, embedding_dim = embeddings.shape
encodings = jnp.take(embeddings, codebook_indices.reshape((-1)), axis=0)
encodings = encodings.reshape((batch_size, 4, 4, embedding_dim))
return encodings
def remove_special_characters(word):
return re.sub(r'^[^a-zA-Z0-9]+|[^a-zA-Z0-9]+$', '', word)
def extract_objs(text, width, height, unique_labels=False):
"""Returns objs for a string with "<loc>" and "<seg>" tokens."""
objs = []
seen = set()
while text:
m = _SEGMENT_DETECT_RE.match(text)
if not m:
break
gs = list(m.groups())
before = gs.pop(0)
name = gs.pop()
y1, x1, y2, x2 = [int(x) / 1024 for x in gs[:4]]
y1, x1, y2, x2 = map(round, (y1*height, x1*width, y2*height, x2*width))
seg_indices = gs[4:20]
mask=None
content = m.group()
if before:
objs.append(dict(content=before))
content = content[len(before):]
while unique_labels and name in seen:
name = (name or '') + "'"
seen.add(name)
objs.append(dict(
content=content, xyxy=(x1, y1, x2, y2), mask=mask, name=name))
text = text[len(before) + len(content):]
if text:
objs.append(dict(content=text))
return objs
with gr.Blocks() as demo:
gr.Markdown("## Zero-shot Object Tracking with PaliGemma")
gr.Markdown("This is a demo for zero-shot object tracking using [PaliGemma](https://huggingface.co/google/paligemma-3b-mix-448) vision language model by Google.")
gr.Markdown("Simply upload a video and enter the candidate labels, or try the example below. Text input should be ; separated. 👇")
with gr.Tab(label="Video"):
with gr.Row():
input_video = gr.Video(
label='Input Video'
)
output_video = gr.Video(
label='Output Video'
)
with gr.Row():
candidate_labels = gr.Textbox(
label='Labels',
placeholder='Labels separated by a comma',
)
submit = gr.Button()
gr.Examples(
fn=process_video,
examples=[["./cats.mp4", "bird ; cat"]],
inputs=[
input_video,
candidate_labels,
],
outputs=output_video
)
submit.click(
fn=process_video,
inputs=[input_video, candidate_labels],
outputs=output_video
)
demo.launch(debug=False, show_error=True)
|