def generate_visualization(self, subset=False, num_output_imgs=10):
# grab all dirs to grab all images
mode = self.mode.name
for i in range(8):
cols = []
imgs = self.visualize_set[i]
if subset:
step = len(imgs) // num_output_imgs
indices = np.arange(0, step * num_output_imgs + 1, step)
imgs = [imgs[i] for i in range(len(imgs)) if i in indices]
gt_dir = [
os.path.join(self.root_img_dir, im_ext) for im_ext in imgs
]
cols += [Col('img', "GT", gt_dir)]
for k, v in self.dirs.items():
viz_set_list = [
os.path.join(v, f'viz/{i+1}/generated/{im_ext[15:]}')
for im_ext in imgs
]
cols += [Col('img', k, viz_set_list)]
# output html
outfile = self.exp_root_dir + f"testset_viz_{i}.html"
imagetable(cols,
out_file=outfile,
title=f"Results for Video {i+1}",
pathrep=('/home/jl5/data/data-meta/', '../../'))
def to_html(imfiles_keep, imfiles_go):
scores_keep = [str(sc) for _, sc in imfiles_keep]
scores_go = [str(sc) for _, sc in imfiles_go]
finfo = imfiles_keep[0][0] if len(imfiles_keep) else imfiles_go[0][0]
dout = finfo.replace("/".join(finfo.split("/")[-2:]), "")
cols = [
Col("id1", "ID"), # 1-based indexing
Col(
"img",
"keep",
[f.replace(dout, "") for f, _ in imfiles_keep],
None,
"background: #28aade",
href=[f.replace(dout, "") for f, _ in imfiles_keep],
),
Col("text", "keep (score)", scores_keep),
Col(
"img",
"dump",
[f.replace(dout, "") for f, _ in imfiles_go],
None,
"background: #db8a0f",
href=[f.replace(dout, "") for f, _ in imfiles_go],
),
Col("text", "dump (score)", scores_go, style="text {width: 2in;}"),
]
summary_row = [
"ID",
len(imfiles_keep),
np.mean([float(sc) for _, sc in imfiles_keep]),
len(imfiles_go),
np.mean([float(sc) for _, sc in imfiles_go]),
]
imagetable(
cols,
dout + "face-cleanup.html",
"Sample Decision and Score",
summary_row=
summary_row, # add a summary row showing overall statistics of the dataset
summary_color="#fff9b7", # highlight the summary row
sortable=True, # enable interactive sorting
sticky_header=True, # keep the header on the top
style="img {border: 1px solid black;};",
sort_style=
"materialize", # use the theme "materialize" from jquery.tablesorter
zebra=True, # use zebra-striped table
)
def create_html_page_face_montage(p1, p2, labels):
neg_ids = np.where(labels == 0)[0][-100:]
pos_ids = np.where(labels == 1)[0][:100]
p1_pos = p1[pos_ids]
p1_neg = p1[neg_ids]
p2_pos = p2[pos_ids]
p2_neg = p2[neg_ids]
hard_pos_pairs = np.array(list(zip(p1_pos, p2_pos)))
hard_neg_pairs = np.array(list(zip(p1_neg, p2_neg)))
cols = [
Col('text', 'FID1', t1[pos_ids]),
Col('img', 'P1', [
CONFIGS.path.dfid + f + '.jpg' for f in list(hard_pos_pairs[:, 0])
]),
Col('img', 'P2', [
CONFIGS.path.dfid + f + '.jpg' for f in list(hard_pos_pairs[:, 1])
]),
Col('text', 'Scores',
["{0:0.5}".format(sc * 100) for sc in scores[pos_ids]])
]
# cols2 = []
imagetable(
cols,
imscale=0.75, # scale all images to 50%
sticky_header=True, # keep the header on the top
out_file=dir_fold + 'hard_positives.html',
style=
'img {border: 1px solid black;-webkit-box-shadow: 2px 2px 1px #ccc; box-shadow: 2px 2px 1px #ccc;}',
)
cols = [
Col('text', 'FID1', [
"{}\n{}".format(tt1, tt2)
for (tt1, tt2) in zip(t1[neg_ids], t2[neg_ids])
]),
Col('img', 'P1', [
CONFIGS.path.dfid + f + '.jpg' for f in list(hard_neg_pairs[:, 0])
]),
Col('img', 'P2', [
CONFIGS.path.dfid + f + '.jpg' for f in list(hard_neg_pairs[:, 1])
]),
# Col('text', 'FID2', t2[neg_ids]),
Col('text', 'Scores',
["{0:0.5}".format(sc * 100) for sc in scores[neg_ids]])
]
imagetable(
cols,
imscale=0.75, # scale all images to 50%
sticky_header=True, # keep the header on the top
out_file=dir_fold + 'hard_negatives.html',
style=
'img {border: 1px solid black;-webkit-box-shadow: 2px 2px 1px #ccc; box-shadow: 2px 2px 1px #ccc;}',
)
def generate_html(G_net,
D_net,
device,
data_loaders,
html_save_path,
max_rows=64,
outpaint=True):
'''
Visualizes one batch from both the training and validation sets.
Images are stored in the specified HTML file path.
'''
G_net.eval()
D_net.eval()
torch.set_grad_enabled(False)
if os.path.exists(html_save_path):
shutil.rmtree(html_save_path)
os.makedirs(html_save_path + '/images')
# Evaluate examples
for phase in ['train', 'val']:
imgs, masked_imgs, masked_parts = next(iter(data_loaders[phase]))
masked_imgs = masked_imgs.to(device)
outputs = G_net(masked_imgs)
masked_imgs = masked_imgs.cpu()
if not (outpaint):
results = finish_inpaint(imgs, outputs.cpu())
else:
results = outputs.cpu()
# Store images
for i in range(min(imgs.shape[0], max_rows)):
save_image(
masked_imgs[i], html_save_path + '/images/' + phase + '_' +
str(i) + '_masked.jpg')
save_image(
results[i], html_save_path + '/images/' + phase + '_' +
str(i) + '_result.jpg')
save_image(
imgs[i], html_save_path + '/images/' + phase + '_' + str(i) +
'_truth.jpg')
# Generate table
cols = [
Col('id1', 'ID'),
Col('img', 'Training set masked',
html_save_path + '/images/train_*_masked.jpg'),
Col('img', 'Training set result',
html_save_path + '/images/train_*_result.jpg'),
Col('img', 'Training set truth',
html_save_path + '/images/train_*_truth.jpg'),
Col('img', 'Validation set masked',
html_save_path + '/images/val_*_masked.jpg'),
Col('img', 'Validation set result',
html_save_path + '/images/val_*_result.jpg'),
Col('img', 'Validation set truth',
html_save_path + '/images/val_*_truth.jpg'),
]
imagetable(cols,
out_file=html_save_path + '/index.html',
pathrep=(html_save_path + '/images', 'images'))
print('Generated image table at: ' + html_save_path + '/index.html')
def generate_html(pairs):
image1_list = []
image2_list = []
labels_list = []
for index in pairs:
info = pairs[index]
image1_list.append(get_relative_image_path(info['image1']))
image2_list.append(get_relative_image_path(info['image2']))
labels_list.append(str(info['label']))
cols = [
Col('id1', 'ID'),
Col('img', 'Image 1', image1_list),
Col('img', 'Image 2', image2_list),
Col('text', 'Label', labels_list)
]
imagetable(cols, out_file='index.html', imsize=(320, 240))
def generate_file(out_file_name, authors, links, titles, poster_id,
select_ids):
ls = []
# import pdb; pdb.set_trace()
for i in range(len(authors)):
if int(poster_id[i]) in select_ids:
ls.append(i)
print(ls)
authors = [authors[x] for x in ls]
titles = [titles[x] for x in ls]
poster_id = [poster_id[x] for x in ls]
links = [links[x] for x in ls]
cols = [
Col('text', 'ID', poster_id),
Col('text', 'Title', titles, href=links),
Col('text', 'Authors', authors)
]
imagetable(
cols,
out_file_name,
sortable=True,
style=
"td {text-align: left} table, th, td { border: 1px solid black; border-collapse: collapse; padding: 5px;}"
)
consecs = np.arange(n_consec)
sel = [n_consec * i + consecs for i in sel]
sel = np.array(sel).flatten().tolist()
# sel = list(range(n_show))
img_paths = []
vid_paths = []
for idx in sel:
img = imgs[idx]
seq = seqs[img['sid']]
img_paths.append(join(seq, img['name'][:-3] + 'jpg'))
vid_paths.append(seq + '.mp4')
cols = [Col('id1', 'ID')]
summary_row = [metric]
for i, name in enumerate(names):
paths = [join(dirs[i], p) for p in img_paths]
if link_video:
hrefs = [join(dirs[i], p) for p in vid_paths]
else:
hrefs = paths
cols.append(Col('img', name, paths, href=hrefs))
if 'Exp' in dirs[i]:
eval_path = join(dirs[i].replace('/vis', '/output'),
'eval_summary.pkl')
eval_path = eval_path.replace('-c3-eta0', '')
if isfile(eval_path):
from html4vision import Col, imagetable
import os
baseline_tpfp_chair = [
Col('img', 'CHAIR', 'vis_top_detections_baseline/chair/*'),
]
baseline_tpfp_bed = [
Col('img', 'BED', 'vis_top_detections_baseline/bed/*'),
]
baseline_tpfp_toilet = [
Col('img', 'TOILET', 'vis_top_detections_baseline/toilet/*'),
]
baseline_tpfp_tv = [
Col('img', 'TV', 'vis_top_detections_baseline/tv/*'),
]
baseline_tpfp_couch = [
Col('img', 'COUCH', 'vis_top_detections_baseline/couch/*'),
]
baseline_tpfp_pottedplant = [
Col('img', 'POTTED PLANT', 'vis_top_detections_baseline/potted_plant/*'),
]
# html table generation
imagetable(baseline_tpfp_chair, 'baseline_tpfp_chair.html', imscale=0.7)
from html4vision import Col, imagetable
cols = [
Col('img', 'Image', 'images/sketch_img_*'),
Col('img', 'Sketch 1', 'images/sketch_*_1.svg'),
Col('img', 'Sketch 2', 'images/sketch_*_2.svg'),
Col('img', 'Sketch 3', 'images/sketch_*_3.svg',
None, 'background: #e9f9fe'), # set background to light blue for this column
]
imagetable(
cols,
'formatting.html',
imsize=0, # resize sketch svg to match corresponding image size (column 0)
imscale=0.5, # scale all images to 50%
# adding image border and box shadows to the entire table
style='img {border: 1px solid black;-webkit-box-shadow: 2px 2px 1px #ccc; box-shadow: 2px 2px 1px #ccc;}',
)
summary_per_class = shorten(
np.loadtxt(join(predict_dir, 'eval-val-summary-per-class.txt')))
per_inst = shorten(np.loadtxt(join(predict_dir, 'eval-val-per-inst.txt')))
per_inst_per_class = shorten(
np.loadtxt(join(predict_dir, 'eval-val-per-inst-per-class.txt')))
train_class_name = [
label.category + '.' + label.name for label in labels
if label.trainId != 255 and label.trainId >= 0
]
if opts.class_subset is not None:
train_class_name = [train_class_name[i] for i in opts.class_subset[:-1]]
sel = None # all
# sel = 6
overlay_col = Col('overlay', '', inputs, sel, 'opacity: 0.15')
cols = [
Col('id1', 'ID'),
Col('text', 'mIoU', per_inst[:, 0], sel),
Col('img', 'Input Image', inputs, sel),
Col('img', 'Output Segmentation', outputs, sel),
overlay_col,
Col('img', 'Ground Truth (standard, 20)', targets, sel),
overlay_col,
Col('img', 'Ground Truth (original, 35)', targets_org, sel),
overlay_col,
Col('text', 'mAccu', per_inst[:, 1], sel),
Col('text', 'Accu', per_inst[:, 2], sel),
]
for i in range(len(train_class_name)):
from html4vision import Col, imagetable
cols = [
# href can even be used on id columns, and uses the same glob pattern as for the content in img column
Col('id1', 'ID', subset=1, href='images/road_*_image.jpg'),
Col('img', 'Image + Label', 'images/road_*_image.jpg', 1, href='images/road_*_image.jpg'),
# for overlay columns, href should be attached to the preceding img column
Col('overlay', '', 'images/road_*_label.png', 1, 'opacity: 0.4'),
# href supports a list of URLs, each one for each row
Col('img', 'Amodel Road Mask', 'images/road_*_amodelroad.png', 1, href=['https://github.com/mtli/HTML4Vision']),
]
imagetable(cols, 'href.html', 'Hyperlink Example')
# Note for image overaly, href conflicts with overlay_toggle=True, since they are both triggered by the mouse click
# detection_image_path = os.path.join('full/vis_detect', detection_image)
# detection_images.append(detection_image_path)
# print(tp_fp_images[0])
# print(detection_images[0])
# cols = [
# # Col('id1', 'ID'), # make a column of 1-based indice
# # Col('img', 'Images with TP/FP', 'full/vis_gt/*.png'), # specify image content for column 2
# Col('img', 'Images with TP/FP', tp_fp_images),
# Col('img', 'Images with Detections', detection_images), # specify image content for column 3
# ]
cols_all = [
# Col('id1', 'ID'), # make a column of 1-based indice
# Col('img', 'Images with TP/FP', 'full/vis_gt/*.png'), # specify image content for column 2
Col('img', 'CHAIR', 'vis_top_detections/chair/*'),
Col('img', 'BED', 'vis_top_detections/bed/*'),
# Col('img', 'BED ONLY FALSE POSITIVES', 'vis_top_detections_bed_fp_only/bed/*'),
Col('img', 'TOILET', 'vis_top_detections/toilet/*'),
Col('img', 'TV', 'vis_top_detections/tv/*'),
Col('img', 'COUCH', 'vis_top_detections/couch/*'),
Col('img', 'POTTED PLANT', 'vis_top_detections/potted_plant/*'),
#Col('img', 'Images with Detections', detection_images), # specify image content for column 3
]
cols_chair = [
Col('img', 'CHAIR', 'vis_top_detections/chair/*'),
]
cols_bed = [
Col('img', 'BED', 'vis_top_detections/bed/*'),
title = 'Debug'
metric = 'AP'
link_video = True
sel = 100
img_dir = '/data2/mengtial/Exp/ArgoVerse1.1-debug/pps_mrcnn50_nm_ds_s0.75_fba_iou_lin_pkt/5ab2697b-6e3e-3454-a36a-aba2c6f27818'
srv_dir = data_dir
srv_port = 40001
# host_name = socket.gethostname()
host_name = 'trinity.vision.cs.cmu.edu'
##
# print(f'python vis/make_videos.py "{img_dir}" --fps 30')
cols = [
Col('id1', 'ID'),
Col('img', 'Image', img_dir + '/*.jpg', sel),
]
imagetable(
cols,
join(out_dir, out_name),
title,
imscale=0.5,
pathrep=srv_dir,
)
url = f'http://{host_name}:{srv_port}{out_dir.replace(srv_dir, "")}/{out_name}'
print(url)
mport os.path as osp
from html4vision import Col, imagetable
# table description
cols = [
Col('id1', 'ID'), # make a column of 1-based indices
Col('img', 'vgg16', 'det_vgg16_Flickr_vis-0.70/*.jpg'), # specify image content for column 2
]
imagetable(cols, outfile='det_vgg16_Flickr_vis-0.70.html', title='Flickr conf pert-adapt',
style=None)
print(len(col_dict[0]))
print(len(col_dict[1]))
print(len(col_dict[2]))
print(len(col_dict[3]))
print(len(col_dict[4]))
print(len(col_dict[5]))
print(len(col_dict_local[0]))
print(len(col_dict_local[1]))
print(len(col_dict_local[2]))
print(len(col_dict_local[3]))
print(len(col_dict_local[4]))
print(len(col_dict_local[5]))
baseline_html = [
Col('img', '1', col_dict[0]),
Col('img', '2', col_dict[1]),
Col('img', '3', col_dict[2]),
Col('img', '4', col_dict[3]),
Col('img', '5', col_dict[4]),
Col('img', '6', col_dict[5]),
]
local_html = [
Col('img', '1', col_dict_local[0]),
Col('img', '2', col_dict_local[1]),
Col('img', '3', col_dict_local[2]),
Col('img', '4', col_dict_local[3]),
Col('img', '5', col_dict_local[4]),
Col('img', '6', col_dict_local[5]),
]
seg_colors = join(seq_dir, 'SegColor/*.png')
source_imgs = join(seq_dir, 'RGB/*.png')
matched = join(exp_dir, 'SegMatch-256-Vis/*.jpg')
canvas = join(exp_dir, 'Canvas-512-Vis/*.jpg')
refined = join(exp_dir, 'Refined-512/*.png')
road1_source = join(exp_dir, 'SegMatch-256-VisEx/*_road1.jpg')
car1_source = join(exp_dir, 'SegMatch-256-VisEx/*_car1.jpg')
car2_source = join(exp_dir, 'SegMatch-256-VisEx/*_car2.jpg')
car3_source = join(exp_dir, 'SegMatch-256-VisEx/*_car3.jpg')
building1_source = join(exp_dir, 'SegMatch-256-VisEx/*_building1.jpg')
sel = None # all
overlay_col = Col('overlay', '', seg_colors, sel, 'opacity: 0.15')
cols = [
Col('id1', 'ID'), # 1-based indexing
Col('img', 'CARLA RGB', source_imgs, sel),
overlay_col,
Col('img', 'Input Label', seg_colors, sel),
Col('img', 'Output Image', refined, sel),
overlay_col,
Col('img', 'Segment Matching', matched, sel),
overlay_col,
Col('img', 'Canvas', canvas, sel),
overlay_col,
Col('img', 'Road 1 Source', road1_source, sel),
Col('img', 'Car 1 Source', car1_source, sel),
Col('img', 'Car 2 Source', car2_source, sel),
Col('img', 'Car 3 Source', car3_source, sel),
from html4vision import Col, imagetable
cols = [
Col('img', 'Image', 'images/road_*_image.jpg'),
Col('img', 'Label', 'images/road_*_label.png',
1), # 1 is used to select only the first item
Col('img', 'Image + Label', 'images/road_*_image.jpg'),
Col('overlay', '', 'images/road_*_label.png', 1, 'opacity: 0.4'),
]
imagetable(cols,
'overlay.html',
'Image Overlay',
imscale=1.5,
overlay_toggle=True)
consecs = np.arange(n_consec)
sel = [n_consec*i + consecs for i in sel]
sel = np.array(sel).flatten().tolist()
# sel = list(range(n_show))
img_paths = []
vid_paths = []
for idx in sel:
img_paths.append(join(seq, img_names[idx] + '.jpg'))
vid_paths.append(seq + '.mp4')
cols = [Col('id1', 'ID')]
summary_row = [metric]
for i, name in enumerate(names):
paths = [join(dirs[i], p) for p in img_paths]
if link_video:
hrefs = [join(dirs[i], p) for p in vid_paths]
else:
hrefs = paths
cols.append(
Col('img', name, paths, href=hrefs)
)
if 'Exp' in dirs[i]:
eval_path = join(dirs[i].replace('/visf-th0.5/', '/output/'), 'eval_summary_mask.pkl')
print(eval_path)
# sel_idx = np.random.choice(len(frames))
# print(sel_idx)
frame = frames[sel_idx]
img_paths[i].append(join(seq, frame))
vid_paths[i].append(seq + '.mp4')
cols = []
summary_row = []
for i, name in enumerate(names):
paths = [join(dirs[i], p) for p in img_paths[i]]
if link_video:
hrefs = [join(dirs[i], p) for p in vid_paths[i]]
else:
hrefs = paths
cols.append(Col('img', name, paths, href=hrefs))
method_names = basename(dirname(dirs[i]))
idx = method_names.find('-vs-')
assert idx >= 0
method_A = method_names[:idx]
method_B = method_names[idx + 4:]
if 'Exp' in dirs[i]:
eval_A = join(data_dir, 'Exp', data_cfg, 'output', method_A, split,
'eval_summary.pkl')
eval_B = join(data_dir, 'Exp', data_cfg, 'output', method_B, split,
'eval_summary.pkl')
if isfile(eval_A):
eval_summary = pickle.load(open(eval_A, 'rb'))
summary_A = '%.1f' % (100 * eval_summary['stats'][0])
from html4vision import Col, imagetable
cols = [
Col('id1', 'ID'), # 1-based indexing
Col('img', 'Label Map', 'images/road_*_label.png'),
Col('img', 'Road Object Map', 'images/road_*_roadobj.png'),
Col('img', 'Amodel Road Mask', 'images/road_*_amodelroad.png'),
]
imagetable(cols)
ave_perf = n * [None]
class_perf = [n * [None] for i in range(n_class)]
summary_info = (n_class + 1) * [None]
for i, s in enumerate(pref):
items = s.split(' ')
if i == 0: # the header contains dataset summary
for j in range(n_class + 1):
summary_info[j] = float(items[j])
else: # each row contains result for each data point
ave_perf[i - 1] = float(items[0])
for j in range(n_class):
class_perf[j][i - 1] = float(items[j + 1])
cols = [
Col('id1', 'ID'), # 1-based indexing
Col('img', 'Detection', 'images/det_*.jpg'),
Col('text', 'Class Average', ave_perf),
]
for i in range(n_class):
cols.append(Col('text', class_names[i], class_perf[i]))
summary_row = ['S', '', *summary_info]
imagetable(
cols,
'sort.html',
'Sorting Example',
summary_row=
summary_row, # add a summary row showing overall statistics of the dataset
summary_color='#fff9b7', # highlight the summary row
imscale=0.4, # scale images to 0.4 of the original size
