File size: 8,253 Bytes
c985ba4 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 |
import argparse
import os
import sys
import numpy as np
import torch
from PIL import Image, ImageDraw, ImageFont
import groundingdino.datasets.transforms as T
from groundingdino.models import build_model
from groundingdino.util import box_ops
from groundingdino.util.slconfig import SLConfig
from groundingdino.util.utils import clean_state_dict, get_phrases_from_posmap
from groundingdino.util.vl_utils import create_positive_map_from_span
def plot_boxes_to_image(image_pil, tgt):
H, W = tgt["size"]
boxes = tgt["boxes"]
labels = tgt["labels"]
assert len(boxes) == len(labels), "boxes and labels must have same length"
draw = ImageDraw.Draw(image_pil)
mask = Image.new("L", image_pil.size, 0)
mask_draw = ImageDraw.Draw(mask)
# draw boxes and masks
for box, label in zip(boxes, labels):
# from 0..1 to 0..W, 0..H
box = box * torch.Tensor([W, H, W, H])
# from xywh to xyxy
box[:2] -= box[2:] / 2
box[2:] += box[:2]
# random color
color = tuple(np.random.randint(0, 255, size=3).tolist())
# draw
x0, y0, x1, y1 = box
x0, y0, x1, y1 = int(x0), int(y0), int(x1), int(y1)
draw.rectangle([x0, y0, x1, y1], outline=color, width=6)
# draw.text((x0, y0), str(label), fill=color)
font = ImageFont.load_default()
if hasattr(font, "getbbox"):
bbox = draw.textbbox((x0, y0), str(label), font)
else:
w, h = draw.textsize(str(label), font)
bbox = (x0, y0, w + x0, y0 + h)
# bbox = draw.textbbox((x0, y0), str(label))
draw.rectangle(bbox, fill=color)
draw.text((x0, y0), str(label), fill="white")
mask_draw.rectangle([x0, y0, x1, y1], fill=255, width=6)
return image_pil, mask
def load_image(image_path):
# load image
image_pil = Image.open(image_path).convert("RGB") # load image
transform = T.Compose(
[
T.RandomResize([800], max_size=1333),
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
]
)
image, _ = transform(image_pil, None) # 3, h, w
return image_pil, image
def load_model(model_config_path, model_checkpoint_path, cpu_only=False):
args = SLConfig.fromfile(model_config_path)
args.device = "cuda" if not cpu_only else "cpu"
model = build_model(args)
checkpoint = torch.load(model_checkpoint_path, map_location="cpu")
load_res = model.load_state_dict(clean_state_dict(checkpoint["model"]), strict=False)
print(load_res)
_ = model.eval()
return model
def get_grounding_output(model, image, caption, box_threshold, text_threshold=None, with_logits=True, cpu_only=False, token_spans=None):
assert text_threshold is not None or token_spans is not None, "text_threshould and token_spans should not be None at the same time!"
caption = caption.lower()
caption = caption.strip()
if not caption.endswith("."):
caption = caption + "."
device = "cuda" if not cpu_only else "cpu"
model = model.to(device)
image = image.to(device)
with torch.no_grad():
outputs = model(image[None], captions=[caption])
logits = outputs["pred_logits"].sigmoid()[0] # (nq, 256)
boxes = outputs["pred_boxes"][0] # (nq, 4)
# filter output
if token_spans is None:
logits_filt = logits.cpu().clone()
boxes_filt = boxes.cpu().clone()
filt_mask = logits_filt.max(dim=1)[0] > box_threshold
logits_filt = logits_filt[filt_mask] # num_filt, 256
boxes_filt = boxes_filt[filt_mask] # num_filt, 4
# get phrase
tokenlizer = model.tokenizer
tokenized = tokenlizer(caption)
# build pred
pred_phrases = []
for logit, box in zip(logits_filt, boxes_filt):
pred_phrase = get_phrases_from_posmap(logit > text_threshold, tokenized, tokenlizer)
if with_logits:
pred_phrases.append(pred_phrase + f"({str(logit.max().item())[:4]})")
else:
pred_phrases.append(pred_phrase)
else:
# given-phrase mode
positive_maps = create_positive_map_from_span(
model.tokenizer(text_prompt),
token_span=token_spans
).to(image.device) # n_phrase, 256
logits_for_phrases = positive_maps @ logits.T # n_phrase, nq
all_logits = []
all_phrases = []
all_boxes = []
for (token_span, logit_phr) in zip(token_spans, logits_for_phrases):
# get phrase
phrase = ' '.join([caption[_s:_e] for (_s, _e) in token_span])
# get mask
filt_mask = logit_phr > box_threshold
# filt box
all_boxes.append(boxes[filt_mask])
# filt logits
all_logits.append(logit_phr[filt_mask])
if with_logits:
logit_phr_num = logit_phr[filt_mask]
all_phrases.extend([phrase + f"({str(logit.item())[:4]})" for logit in logit_phr_num])
else:
all_phrases.extend([phrase for _ in range(len(filt_mask))])
boxes_filt = torch.cat(all_boxes, dim=0).cpu()
pred_phrases = all_phrases
return boxes_filt, pred_phrases
if __name__ == "__main__":
parser = argparse.ArgumentParser("Grounding DINO example", add_help=True)
parser.add_argument("--config_file", "-c", type=str, required=True, help="path to config file")
parser.add_argument(
"--checkpoint_path", "-p", type=str, required=True, help="path to checkpoint file"
)
parser.add_argument("--image_path", "-i", type=str, required=True, help="path to image file")
parser.add_argument("--text_prompt", "-t", type=str, required=True, help="text prompt")
parser.add_argument(
"--output_dir", "-o", type=str, default="outputs", required=True, help="output directory"
)
parser.add_argument("--box_threshold", type=float, default=0.3, help="box threshold")
parser.add_argument("--text_threshold", type=float, default=0.25, help="text threshold")
parser.add_argument("--token_spans", type=str, default=None, help=
"The positions of start and end positions of phrases of interest. \
For example, a caption is 'a cat and a dog', \
if you would like to detect 'cat', the token_spans should be '[[[2, 5]], ]', since 'a cat and a dog'[2:5] is 'cat'. \
if you would like to detect 'a cat', the token_spans should be '[[[0, 1], [2, 5]], ]', since 'a cat and a dog'[0:1] is 'a', and 'a cat and a dog'[2:5] is 'cat'. \
")
parser.add_argument("--cpu-only", action="store_true", help="running on cpu only!, default=False")
args = parser.parse_args()
# cfg
config_file = args.config_file # change the path of the model config file
checkpoint_path = args.checkpoint_path # change the path of the model
image_path = args.image_path
text_prompt = args.text_prompt
output_dir = args.output_dir
box_threshold = args.box_threshold
text_threshold = args.text_threshold
token_spans = args.token_spans
# make dir
os.makedirs(output_dir, exist_ok=True)
# load image
image_pil, image = load_image(image_path)
# load model
model = load_model(config_file, checkpoint_path, cpu_only=args.cpu_only)
# visualize raw image
image_pil.save(os.path.join(output_dir, "raw_image.jpg"))
# set the text_threshold to None if token_spans is set.
if token_spans is not None:
text_threshold = None
print("Using token_spans. Set the text_threshold to None.")
# run model
boxes_filt, pred_phrases = get_grounding_output(
model, image, text_prompt, box_threshold, text_threshold, cpu_only=args.cpu_only, token_spans=eval(f"{token_spans}")
)
# visualize pred
size = image_pil.size
pred_dict = {
"boxes": boxes_filt,
"size": [size[1], size[0]], # H,W
"labels": pred_phrases,
}
# import ipdb; ipdb.set_trace()
image_with_box = plot_boxes_to_image(image_pil, pred_dict)[0]
image_with_box.save(os.path.join(output_dir, "pred.jpg"))
|