def downloadImageNet():
print("downloading ImageNet dataset.")
classes, labels = classesOfClassifier(True)
cls_urls = {}
for cls, _, label in nested(labels, depth=2):
csf_idx, name = label
wnid = wnidOfName(name)
urls = fetchImageUrlsOfWnid(wnid)
if cls not in cls_urls:
cls_urls[cls] = []
cls_urls[cls] += [(url, csf_idx) for url in urls]
counter = {cls: 0 for cls in classes}
for cls, data in cls_urls.items():
shuffle(data)
_data = data[:100]
data = data[100:]
for url, csf_idx in _data:
try:
img = imread(url)
except:
_data.append(data[0])
data = data[1:]
continue
if img is None:
continue
idx = counter[cls]
counter[cls] += 1
file_name = "{}_{}_{}.jpg".format(getClassName(cls), csf_idx, idx)
file_path = join(PATH.DATA.IMAGENET.IMGS, file_name)
saveImage(img, file_path, plugin='skimage')
def test_mask_image(self):
self.log()
img = self.getImage("2008_001979.jpg")
mask = np.zeros(img.shape[:-1])
center = np.asarray(mask.shape) // 2
size = (100, 100)
indices = [np.arange(c - s, c + s) for c, s in zip(center, size)]
for r in indices[0]:
for c in indices[1]:
mask[r][c] = 1
masks = [mask]
mask = np.zeros(img.shape[:-1])
center = (0, 0)
size = (200, 200)
indices = [np.arange(c, c + s) for c, s in zip(center, size)]
for r in indices[0]:
for c in indices[1]:
mask[r][c] = 1
masks.append(mask)
img = maskImage(img, masks, alpha=0.6)
# cv2.imshow("test plotter", img)
# cv2.waitKey(0)
saveImage(img, os.path.join(root_path, "test_plotter.png"))
def test_reveal_mask(self):
self.log()
bl = BatchLoader(amount=1)
batch = bl.nextBatch()
img = batch[1][0]
annos = batch[2][0]
for anno in annos:
aid, mask = anno
output = revealMask(img, mask)
path = os.path.join(PATH.TEST.ROOT,
"{}.png".format(getClassName(aid, full=True)))
saveImage(output, path)
def test_patch(self):
self.log()
bl = BatchLoader(amount=10)
batch = bl.nextBatch()
imgs = batch[1]
annos = batch[2]
imgs, aids = patch(imgs, annos)
anames = getClassNames(aids)
path = os.path.join(PATH.TEST.ROOT, "patch")
idx = 0
for img, aname in zip(imgs, anames):
idx += 1
file_name = "{}_{}.jpg".format(aname, idx)
file_path = os.path.join(path, file_name)
saveImage(img, file_path)
def test_visual_reflect(self):
self.log()
bl = BatchLoader(amount=2)
model = ModelAgent(input_size=1)
batch = bl.nextBatch()
imgs = batch[1][-1:]
annos = batch[2][-1:]
activ_maps = model.getActivMaps(imgs, ["conv3_1", "conv4_1", "conv5_1"])
field_maps = model.getFieldmaps()
reflected = reflect(activ_maps, field_maps)
img = imgs[0]
path = os.path.join(PATH.TEST.ROOT, "output/")
saveImage(img, os.path.join(path, "raw_img.jpg"))
for unit, ref in reflected.items():
ref = np.asarray(ref[0])
saved = np.zeros(shape=ref.shape+(3,))
indices = np.argwhere(ref>0)
saved[indices[:,0], indices[:,1]] = [255, 0, 0]
saveImage(saved, os.path.join(path, unit+".jpg"))
def test_deconv_activ(self):
bl = BatchLoader(amount=1)
model = ModelAgent(input_size=1, deconv=True)
batch = bl.nextBatch()
img = batch[1]
activ_maps, switches = model.getActivMaps(img, ["pool5"])
unit = "pool5_511"
activ_maps = {unit : activ_maps[unit]}
deconv = model.getDeconvMaps(activ_maps, switches)
deconv = deconv[unit][0]
saveImage(deconv, os.path.join(PATH.TEST.ROOT, "deconv.jpg"))
deconv[deconv<0] = 0
deconv[deconv>0] = 255
for ri, row in enumerate(deconv):
for ci, col in enumerate(row):
if np.sum(col>0) > 0:
deconv[ri][ci] = [255, 255, 255]
saveImage(deconv, os.path.join(PATH.TEST.ROOT, "filtered.jpg"))