158 lines
5.6 KiB
Python
158 lines
5.6 KiB
Python
# --------------------------------------------------------
|
|
# X-Decoder -- Generalized Decoding for Pixel, Image, and Language
|
|
# Copyright (c) 2022 Microsoft
|
|
# Licensed under The MIT License [see LICENSE for details]
|
|
# Written by Xueyan Zou (xueyan@cs.wisc.edu)
|
|
# --------------------------------------------------------
|
|
|
|
import os
|
|
import sys
|
|
import logging
|
|
|
|
pth = '/'.join(sys.path[0].split('/')[:-1])
|
|
sys.path.insert(0, pth)
|
|
|
|
from PIL import Image
|
|
import numpy as np
|
|
np.random.seed(0)
|
|
import cv2
|
|
|
|
import torch
|
|
from torchvision import transforms
|
|
|
|
from utils.arguments import load_opt_command
|
|
|
|
from detectron2.data import MetadataCatalog
|
|
from modeling.language.loss import vl_similarity
|
|
from modeling.BaseModel import BaseModel
|
|
from modeling import build_model
|
|
from utils.visualizer import Visualizer
|
|
from utils.distributed import init_distributed
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
def main(args=None):
|
|
'''
|
|
Main execution point for PyLearn.
|
|
'''
|
|
opt, cmdline_args = load_opt_command(args)
|
|
if cmdline_args.user_dir:
|
|
absolute_user_dir = os.path.abspath(cmdline_args.user_dir)
|
|
opt['base_path'] = absolute_user_dir
|
|
opt = init_distributed(opt)
|
|
|
|
# META DATA
|
|
pretrained_pth = os.path.join(opt['RESUME_FROM'])
|
|
output_root = './output'
|
|
|
|
image_list = ['inference/images/coco/000.jpg', 'inference/images/coco/001.jpg', 'inference/images/coco/002.jpg', 'inference/images/coco/003.jpg']
|
|
text = ['pizza on the plate']
|
|
|
|
model = BaseModel(opt, build_model(opt)).from_pretrained(pretrained_pth).eval().cuda()
|
|
model.model.sem_seg_head.predictor.lang_encoder.get_text_embeddings(["background", "background"], is_eval=False)
|
|
|
|
t = []
|
|
t.append(transforms.Resize(224, interpolation=Image.BICUBIC))
|
|
transform_ret = transforms.Compose(t)
|
|
t = []
|
|
t.append(transforms.Resize(512, interpolation=Image.BICUBIC))
|
|
transform_grd = transforms.Compose(t)
|
|
|
|
metadata = MetadataCatalog.get('ade20k_panoptic_train')
|
|
color = [0/255, 255/255, 0/255]
|
|
|
|
with torch.no_grad():
|
|
batch_inputs = []
|
|
candidate_list = []
|
|
for j in range(len(image_list)):
|
|
image_ori = Image.open(image_list[j])
|
|
width = image_ori.size[0]
|
|
height = image_ori.size[1]
|
|
|
|
image = transform_ret(image_ori)
|
|
image = np.asarray(image)
|
|
candidate_list += [image]
|
|
image_list += [image]
|
|
image_ori = np.asarray(image_ori)
|
|
images = torch.from_numpy(image.copy()).permute(2,0,1).cuda()
|
|
batch_inputs += [{'image': images, 'height': height, 'width': width}]
|
|
|
|
outputs = model.model.evaluate(batch_inputs)
|
|
v_emb = torch.cat([x['captions'][-1:] for x in outputs])
|
|
v_emb = v_emb / (v_emb.norm(dim=-1, keepdim=True) + 1e-7)
|
|
|
|
model.model.sem_seg_head.predictor.lang_encoder.get_text_embeddings(text, is_eval=False, name='caption', prompt=False)
|
|
t_emb = getattr(model.model.sem_seg_head.predictor.lang_encoder, '{}_text_embeddings'.format('caption'))
|
|
temperature = model.model.sem_seg_head.predictor.lang_encoder.logit_scale
|
|
logits = vl_similarity(v_emb, t_emb, temperature)
|
|
max_prob, max_id = logits.softmax(0).max(dim=0)
|
|
|
|
frame_pth = image_list[max_id.item()]
|
|
image_ori = Image.open(frame_pth)
|
|
width = image_ori.size[0]
|
|
height = image_ori.size[1]
|
|
|
|
image = transform_grd(image_ori)
|
|
image = np.asarray(image)
|
|
image_ori = np.asarray(image_ori)
|
|
images = torch.from_numpy(image.copy()).permute(2,0,1).cuda()
|
|
batch_inputs = [{'image': images, 'height': height, 'width': width, 'groundings': {'texts': [text]}}]
|
|
outputs = model.model.evaluate_grounding(batch_inputs, None)
|
|
|
|
visual = Visualizer(image_ori, metadata=metadata)
|
|
grd_masks = (outputs[0]['grounding_mask'] > 0).float().cpu().numpy()
|
|
for text_, mask in zip(text, grd_masks):
|
|
demo = visual.draw_binary_mask(mask, color=color, text='', alpha=0.5)
|
|
region_img = demo.get_image()
|
|
candidate_list[max_id.item()] = region_img
|
|
out_image = np.zeros((224*4+60, 448*4, 3))
|
|
for ii, img in enumerate(candidate_list):
|
|
img = cv2.resize(img, (448, 224))
|
|
if ii != max_id.item():
|
|
img = img * 0.4
|
|
hs, ws = 60+(ii//4)*224, (ii%4)*448
|
|
out_image[hs:hs+224,ws:ws+448,:] = img[:,:,::-1]
|
|
|
|
font = cv2.FONT_HERSHEY_DUPLEX
|
|
fontScale = 1.2
|
|
thickness = 3
|
|
lineType = 2
|
|
bottomLeftCornerOfText = (10, 40)
|
|
fontColor = [255,255,255]
|
|
cv2.putText(out_image, text[0],
|
|
bottomLeftCornerOfText,
|
|
font,
|
|
fontScale,
|
|
fontColor,
|
|
thickness,
|
|
lineType)
|
|
|
|
x1 = (max_id.item()%4) * 448
|
|
y1 = (max_id.item()//4) * 224 + 60
|
|
cv2.rectangle(out_image, (x1, y1), (x1+448, y1+224), (0,0,255), 3)
|
|
|
|
x1 = x1
|
|
y1 = y1 + 224 - 30
|
|
cv2.rectangle(out_image, (x1+2, y1), (x1+60, y1+28), (0,0,0), -1)
|
|
|
|
fontScale = 1.0
|
|
thickness = 2
|
|
|
|
bottomLeftCornerOfText = (x1, y1+21)
|
|
cv2.putText(out_image, str(max_prob.item())[0:4],
|
|
bottomLeftCornerOfText,
|
|
font,
|
|
0.8,
|
|
[0,0,255],
|
|
thickness,
|
|
lineType)
|
|
|
|
if not os.path.exists(output_root):
|
|
os.makedirs(output_root)
|
|
cv2.imwrite(os.path.join(output_root, 'region_retrieval.png'), out_image)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|
|
sys.exit(0) |