from utils import futils, Visualizer
import glob
import os
visual_dir = '/home/xueyan/antialias-cnn/data/output/resnet101_ori/visual'
files = glob.glob(os.path.join(visual_dir, '*/*.*'))
class_dict = {}
for pth in files:
class_name = pth.split('/')[8]
if class_name not in class_dict.keys():
class_dict[class_name] = []
class_dict[class_name].append(pth)
for key in class_dict.keys():
visualizer = Visualizer(os.path.join(visual_dir, key),
demo_name='index.html')
for pth in class_dict[key]:
visual_pth = '/'.join(pth.split('/')[-1:])
visual_name = pth.split('/')[-1]
visualizer.insert(visual_pth, visual_name)
visualizer.write()
def validate_shift(val_loader, model, args, visual):
batch_time = AverageMeter()
consist = []
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
visual_folder, top_visualizer, bottom_visualizer = visual
# load json file for id to class
idx_to_cls = json.load(
open('../../data/imagenet-vid-robust/misc/imagenet_idx_to_name.json'))
# when validate shift, batch size must be 1.
assert args.batch_size == 1
# switch to evaluate mode
model.eval()
visualizer_dict = {}
# get evaluated class name
src_pth = '/home/xueyan/antialias-cnn/data/output/resnet101_ori/visual'
class_names = [x[0].split('/')[-1] for x in os.walk(src_pth)][1:]
with torch.no_grad():
end = time.time()
video_pred_var = {}
video_to_root_visual_pth = {}
for i, ((input, target), pth) in enumerate(val_loader):
hist_names = []
img_visual_names = []
img_visual_cap = []
class_name, img_name = pth[0].split('/')[5], pth[0].split('/')[6]
if i % 100 == 0:
print('process ',
'[' + str(i) + '|' + str(len(val_loader)) + ']')
if class_name in class_names:
continue
input_ = input.clone()
n, c, h, w = input.shape
h_c, w_c = h // 2, w // 2
half_s = args.size // 2
input = input[:, :, (h_c - half_s):(h_c + half_s),
(w_c - half_s):(w_c + half_s)]
# save input center
center_pth = os.path.join(args.out_dir, 'visual', class_name)
if class_name in visualizer_dict.keys():
visualizer = visualizer_dict[class_name]
else:
visualizer = Visualizer(os.path.join(args.out_dir, 'visual',
class_name),
demo_name='index.html')
visualizer_dict[class_name] = visualizer
if not os.path.exists(center_pth):
os.mkdir(center_pth)
cur_img_name = img_name[:-5] + '_a_h_0_w_0.png'
center_name = os.path.join(args.out_dir, 'visual', class_name,
cur_img_name)
hist_names.append(
os.path.join(args.out_dir, 'visual', class_name,
img_name[:-5] + '_bar_h_0_w_0.png'))
imsave(center_name,
((input[0, :, :, :].permute(1, 2, 0).cpu().numpy()) *
np.array(std) + np.array(mean).reshape(1, 1, 3)) * 255)
# append to html file
visual_pth = os.path.join('/',
args.out_dir.split('/')[-1], 'visual',
class_name, cur_img_name)
img_visual_names.append(visual_pth)
img_visual_cap.append(cur_img_name)
visualizer.insert(visual_pth, cur_img_name)
offset = [-16, -10, -5, -2, -1, 1, 2, 5, 10, 16]
for k in range(0, len(offset)):
# shift h
shift_h = input_[:, :, (h_c + offset[k] -
half_s):(h_c + offset[k] + half_s),
(w_c - half_s):(w_c + half_s)]
input = torch.cat((input, shift_h), dim=0)
cur_img_name = img_name[:-5] + '_a_' + 'h_' + str(
k) + '_w_0.png'
h_name = os.path.join(args.out_dir, 'visual', class_name,
cur_img_name)
hist_names.append(
os.path.join(
args.out_dir, 'visual', class_name,
img_name[:-5] + '_bar_h_' + str(k) + '_w_0.png'))
imsave(h_name,
((shift_h[0, :, :, :].permute(1, 2, 0).cpu().numpy()) *
np.array(std) + np.array(mean).reshape(1, 1, 3)) * 255)
# append to html file
visual_pth = os.path.join('/',
args.out_dir.split('/')[-1],
'visual', class_name, cur_img_name)
img_visual_names.append(visual_pth)
img_visual_cap.append(cur_img_name)
visualizer.insert(visual_pth, cur_img_name)
# shift w
shift_w = input_[:, :, (h_c - half_s):(h_c + half_s),
(w_c + offset[k] - half_s):(w_c + offset[k] +
half_s)]
input = torch.cat((input, shift_w), dim=0)
cur_img_name = img_name[:-5] + '_a_h_0_w_' + str(k) + '.png'
w_name = os.path.join(
args.out_dir, 'visual', class_name,
img_name[:-5] + '_a_h_0_w_' + str(k) + '.png')
hist_names.append(
os.path.join(
args.out_dir, 'visual', class_name,
img_name[:-5] + '_bar_h_0_w_' + str(k) + '.png'))
imsave(w_name,
((shift_w[0, :, :, :].permute(1, 2, 0).cpu().numpy()) *
np.array(std) + np.array(mean).reshape(1, 1, 3)) * 255)
# append to html file
visual_pth = os.path.join('/',
args.out_dir.split('/')[-1],
'visual', class_name, cur_img_name)
img_visual_names.append(visual_pth)
img_visual_cap.append(cur_img_name)
visualizer.insert(visual_pth, cur_img_name)
if args.gpu is not None:
input = input.cuda(args.gpu, non_blocking=True)
target = target.cuda(args.gpu, non_blocking=True)
output = torch.nn.Softmax(dim=1)(model(input))
consist.append(JS_Divergence(output))
output_sort, output_argsort = torch.sort(output,
dim=1,
descending=True)
for k in range(input.shape[0]):
top5_arg_name = [
idx_to_cls[str(x.item())] for x in output_argsort[k, 0:5]
]
top5_prob = output_sort[k, 0:5].cpu().numpy()
ax = sns.barplot(x=top5_arg_name, y=top5_prob)
plt.savefig(hist_names[k])
plt.clf()
cur_img_name = hist_names[k].split('/')[-1]
visual_pth = os.path.join('/',
args.out_dir.split('/')[-1],
'visual', class_name, cur_img_name)
img_visual_names.append(visual_pth)
img_visual_cap.append(cur_img_name)
visualizer.insert(visual_pth, cur_img_name)
video_to_root_visual_pth[pth[0]] = [
sorted(img_visual_names),
sorted(img_visual_cap)
]
video_pred_var[pth[0]] = JS_Divergence(output)
for key in visualizer_dict:
visualizer_dict[key].write()
def validate_shift(val_loader, model, args, visual):
batch_time = AverageMeter()
consist = []
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
visual_folder, top_visualizer, bottom_visualizer = visual
# load json file for id to class
idx_to_cls = json.load(
open('../../data/imagenet-vid-robust/misc/imagenet_idx_to_name.json'))
# when validate shift, batch size must be 1.
assert args.batch_size == 1
# switch to evaluate mode
model.eval()
with torch.no_grad():
end = time.time()
video_pred_var = {}
video_to_root_visual_pth = {}
for i, ((input, target), pth) in enumerate(val_loader):
if i % 25 == 0:
hist_names = []
img_visual_names = []
img_visual_cap = []
class_name, img_name = pth[0].split('/')[5], pth[0].split(
'/')[6]
input_ = input.clone()
n, c, h, w = input.shape
h_c, w_c = h // 2, w // 2
half_s = args.size // 2
input = input[:, :, (h_c - half_s):(h_c + half_s),
(w_c - half_s):(w_c + half_s)]
# save input center
center_pth = os.path.join(args.out_dir, 'visual', class_name)
visualizer = Visualizer(os.path.join(args.out_dir, 'visual',
class_name),
demo_name='index.html')
if not os.path.exists(center_pth):
os.mkdir(center_pth)
cur_img_name = 'a_' + img_name[:-5] + '_h_0_w_0.png'
center_name = os.path.join(args.out_dir, 'visual', class_name,
cur_img_name)
hist_names.append(
os.path.join(args.out_dir, 'visual', class_name,
'bar_' + img_name[:-5] + '_h_0_w_0.png'))
imsave(center_name,
((input[0, :, :, :].permute(1, 2, 0).cpu().numpy()) *
np.array(std) + np.array(mean).reshape(1, 1, 3)) * 255)
# append to html file
visual_pth = os.path.join('/',
args.out_dir.split('/')[-1],
'visual', class_name, cur_img_name)
img_visual_names.append(visual_pth)
img_visual_cap.append(cur_img_name)
visualizer.insert(visual_pth, cur_img_name)
offset = [-16, -10, -5, -2, -1, 1, 2, 5, 10, 16]
for k in range(0, len(offset)):
# shift h
shift_h = input_[:, :, (h_c + offset[k] -
half_s):(h_c + offset[k] + half_s),
(w_c - half_s):(w_c + half_s)]
input = torch.cat((input, shift_h), dim=0)
cur_img_name = 'a_' + img_name[:-5] + '_h_' + str(
k) + '_w_0.png'
h_name = os.path.join(args.out_dir, 'visual', class_name,
cur_img_name)
hist_names.append(
os.path.join(
args.out_dir, 'visual', class_name, 'bar_' +
img_name[:-5] + '_h_' + str(k) + '_w_0.png'))
imsave(
h_name,
((shift_h[0, :, :, :].permute(1, 2, 0).cpu().numpy()) *
np.array(std) + np.array(mean).reshape(1, 1, 3)) *
255)
# append to html file
visual_pth = os.path.join('/',
args.out_dir.split('/')[-1],
'visual', class_name,
cur_img_name)
img_visual_names.append(visual_pth)
img_visual_cap.append(cur_img_name)
visualizer.insert(visual_pth, cur_img_name)
# shift w
shift_w = input_[:, :, (h_c - half_s):(h_c + half_s),
(w_c + offset[k] -
half_s):(w_c + offset[k] + half_s)]
input = torch.cat((input, shift_w), dim=0)
cur_img_name = 'a_' + img_name[:-5] + '_h_0_w_' + str(
k) + '.png'
w_name = os.path.join(
args.out_dir, 'visual', class_name,
'a_' + img_name[:-5] + '_h_0_w_' + str(k) + '.png')
hist_names.append(
os.path.join(
args.out_dir, 'visual', class_name, 'bar_' +
img_name[:-5] + '_h_0_w_' + str(k) + '.png'))
imsave(
w_name,
((shift_w[0, :, :, :].permute(1, 2, 0).cpu().numpy()) *
np.array(std) + np.array(mean).reshape(1, 1, 3)) *
255)
# append to html file
visual_pth = os.path.join('/',
args.out_dir.split('/')[-1],
'visual', class_name,
cur_img_name)
img_visual_names.append(visual_pth)
img_visual_cap.append(cur_img_name)
visualizer.insert(visual_pth, cur_img_name)
if args.gpu is not None:
input = input.cuda(args.gpu, non_blocking=True)
target = target.cuda(args.gpu, non_blocking=True)
output = torch.nn.Softmax(dim=1)(model(input))
consist.append(JS_Divergence(output))
output_sort, output_argsort = torch.sort(output,
dim=1,
descending=True)
for k in range(input.shape[0]):
top5_arg_name = [
idx_to_cls[str(x.item())]
for x in output_argsort[k, 0:5]
]
top5_prob = output_sort[k, 0:5].cpu().numpy()
ax = sns.barplot(x=top5_arg_name, y=top5_prob)
plt.savefig(hist_names[k])
plt.clf()
cur_img_name = hist_names[k].split('/')[-1]
visual_pth = os.path.join('/',
args.out_dir.split('/')[-1],
'visual', class_name,
cur_img_name)
img_visual_names.append(visual_pth)
img_visual_cap.append(cur_img_name)
visualizer.insert(visual_pth, cur_img_name)
video_to_root_visual_pth[pth[0]] = [
sorted(img_visual_names),
sorted(img_visual_cap)
]
video_pred_var[pth[0]] = JS_Divergence(output)
visualizer.write()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
print('process ',
'[' + str(i) + '|' + str(len(val_loader)) + ']')
sorted_var_top = sorted(video_pred_var.items(),
key=lambda item: item[1],
reverse=True)
sorted_var_back = sorted(video_pred_var.items(),
key=lambda item: item[1],
reverse=False)
for i in range(0, 100):
vid_name, var = sorted_var_top[i]
video_visual_pths, video_captions = video_to_root_visual_pth[
vid_name]
for j in range(0, len(video_visual_pths)):
top_visualizer.insert(
video_visual_pths[j],
video_captions[j][:-4] + '_' + str(var)[0:6])
vid_name, var = sorted_var_back[i]
video_visual_pths, video_captions = video_to_root_visual_pth[
vid_name]
for j in range(0, len(video_visual_pths)):
bottom_visualizer.insert(
video_visual_pths[j],
video_captions[j][:-4] + '_' + str(var)[0:6])
top_visualizer.write(_sorted=False)
bottom_visualizer.write(_sorted=False)
print(' * Consistency ', sum(consist) / len(consist))
return consist
