def visualize_batch(x0, x, step_i):
vis0 = np.hstack(common.to_rgb(x0).numpy())
vis1 = np.hstack(common.to_rgb(x).numpy())
vis = np.vstack([vis0, vis1])
common.imwrite('train_log/batches_%04d.jpg' % step_i, vis)
print('batch (before/after):')
common.imshow(vis)
def main(
draw_boxes=True,
draw_ourboxes=True,
equalize_color_dist=False,
show=False,
dataset='val', # 'val' or 'test'
skip=0,
):
predictions = {}
ground_truth = {}
image_sets = get_image_sets(sets=100, filter=dataset)
for i in range(skip):
next(image_sets)
for images, paths, boxes in tqdm(image_sets):
if equalize_color_dist:
images = equalize_color_distribution(images)
print("Images color distribution equalized")
# check that equalization worked
# print(images[5,6])
# print(images[5,6].min())
# print(images[5,6].mean())
# print(images[5,6].max())
# print(np.percentile(images[0,0], 25, axis=(0,1))) # should be three 0s
# print(np.percentile(images[0,0], 75, axis=(0,1))) # should be three 1s
p = Paths.output / paths[3, 4]
p.parent.mkdir(parents=True, exist_ok=True)
heatmaps = get_temporal_diff_heatmaps(images, boxes, blur_color_diff=lambda x: blor(x))
detection_map, ourboxes = get_detection_map(heatmaps, images, boxes=boxes, thresh_quantile=0.98)
im = images[3, 4].copy()
im[np.where(detection_map > 0)] = [0, 0, 255]
if draw_boxes:
draw_bounding_boxes(im, boxes)
if draw_ourboxes:
draw_bounding_boxes(im, ourboxes, box_color=(255, 255, 0))
cv2.imwrite(str(p), im)
print(f'Image exported: {p}')
if show:
imshow(im)
predictions[p.parent.name] = ourboxes.tolist()
ground_truth[p.parent.name] = boxes.tolist()
p = Paths.output / f'{dataset}.json'
with open(p, 'w') as outfile:
json.dump(predictions, outfile)
print(f'JSON created: {p}')
if len(ground_truth):
ap = compute_AP(predictions, ground_truth)
print(f"Average precision={ap:.5f} on {dataset} set.")
def vgg_eval_model(dataset_root_dir, restore_model: str):
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("DEVICE WILL BE USED: ", device)
net = VggNet(torchvision.models.vgg16_bn(True))
net = net.to(device)
classes = ('not hotdog', 'hotdog')
if restore_model is not None and len(restore_model) > 0:
# original saved file with DataParallel
state_dict = torch.load(restore_model, map_location={'cuda:0': 'cpu'})
net.load_state_dict(state_dict)
print("Model {} restored".format(restore_model))
else:
print("ERROR: no restore model file found!")
return
#from torch.autograd import Variable
#dummy_input = Variable(torch.randn(1, 3, 224, 224), requires_grad=True)
# input_names = ["actual_input_1"] + ["learned_%d" % i for i in range(12)]
# output_names = ["output1"]
#torch.onnx.export(net, dummy_input, "vgg_hot_dog.onnx", export_params=True, verbose=True)
#print("SUCCESS")
hot_dog_dataset_test = HotDogsDatasetEval(
root_dir=dataset_root_dir,
transform=transforms.Compose([
Rescale((224, 224)), #normalize,
ToTensor(),
]))
test_dataloader = DataLoader(hot_dog_dataset_test,
batch_size=4,
shuffle=True,
num_workers=4)
for dl, type in zip([test_dataloader], ['test']):
correct = 0
total = 0
with torch.no_grad():
for data in dl:
images, names = data['image'].float(), data['name']
images = images.to(device)
outputs = net(images)
total += len(names)
for id, prediction in enumerate(outputs.data):
res = torch.nn.functional.softmax(prediction, dim=0)
_, rid = torch.max(res, 0)
print('{} is {}'.format(names[id], classes[rid]))
imshow(torchvision.utils.make_grid(images[id]))
def get_detection_map(
heatmaps: np.ndarray,
images: np.ndarray,
kernel_size=19,
thresh_quantile=0.95,
thresh_factor=2,
min_detections=7,
boxes=[],
debug=False,
) -> Tuple[np.ndarray, np.ndarray]:
kernel = np.ones((kernel_size, kernel_size)) / kernel_size ** 2
detection_maps = []
for heatmap in heatmaps.reshape(-1, *heatmaps.shape[2:]):
conv_heatmap = convolve2d(heatmap, kernel, mode='same')
thresh = thresh_factor * np.nanquantile(conv_heatmap, thresh_quantile)
detection = conv_heatmap > thresh
detection_maps.append(detection)
detection_maps = np.reshape(detection_maps, heatmaps.shape)
# detection_map = detection_maps.sum(axis=(0, 1)) >= min_detections
detection_map = detection_maps.sum(axis=(0, 1))
if debug:
detection_map_bb = detection_map.copy()
draw_bounding_boxes(np.moveaxis([detection_map_bb] * 3, 0, 2), boxes, box_color=(0, 170, 255))
imshow(detection_map_bb * 255 / 20)
cluster_positions, cluster_intensities = find_clusters(detection_map)
for i in reversed(range(len(cluster_intensities))):
if cluster_intensities[i] < min_detections:
cluster_positions.pop(i)
cluster_intensities.pop(i)
label = create_label_clusters(detection_map, cluster_positions, cluster_intensities, min_detections)
ourboxes = find_boxes(label)
return label, ourboxes
def run(args):
face1_file = args.face1
face2_file = args.face2
#face1_file, face2_file = face2_file, face1_file
print("loading face 1: {}".format(face1_file))
face1 = cv2.imread(face1_file)
if face1 is None:
print("can't read image ", face1_file)
return
print("loading face 2: {}".format(face2_file))
face2 = cv2.imread(face2_file)
if face2 is None:
print("can't read image ", face2_file)
return
print("loading models...")
models = load_models(args)
result1, result2 = swap_faces(models, face1, face2)
if args.output1 is not None:
print("writing output result for image1...")
cv2.imwrite(args.output1, result1)
if args.output2 is not None:
print("writing output result for image2...")
cv2.imwrite(args.output2, result2)
if args.show:
imshow("result 1", result1)
imshow("result 2", result2)
cv2.waitKey()
def swap_faces(models,
image1,
image2,
params=build_params(),
debug=False):
if params['precomputed1'] is not None:
pc = params['precomputed1']
face1 = pc['face']
landmarks1 = pc['landmarks']
lp1 = pc['landmark_points']
else:
face1, landmarks1 = get_landmarks(models, image1)
lp1 = get_landmarks_points(landmarks1)
if params['precomputed2'] is not None:
pc = params['precomputed2']
face2 = pc['face']
landmarks2 = pc['landmarks']
lp2 = pc['landmark_points']
else:
face2, landmarks2 = get_landmarks(models, image2)
lp2 = get_landmarks_points(landmarks2)
if params['triplets'] is None:
_, triplets = delunay(lp1)
else:
triplets = params['triplets']
if params['smooth1']:
if 'hist1' not in params:
params['hist1'] = []
params['hist1'].append(lp1)
if len(params['hist1']) > 2:
params['hist1'].pop(0)
h1 = np.array(params['hist1'])
h1 = np.round(np.median(h1, axis=0)).astype(np.int32)
lp1 = list(h1)
#warped1 = image1.copy()
#warped2 = image2.copy()
warped1 = np.zeros_like(image1)
warped2 = np.zeros_like(image2)
two_ways = params['two_ways']
for triplet in triplets:
warp_triangle(
image1,
image2,
warped1,
warped2,
triplet,
lp1,
lp2,
two_ways=two_ways
)
out1 = blend(image1, warped1, lp1, clone_mode=params['clone'])
if two_ways:
out2 = blend(image2, warped2, lp2, clone_mode=params['clone'])
else:
out2 = None
if debug:
imshow("landmarks 1", draw_landmarks(image1, landmarks1, face1))
imshow("landmarks 2", draw_landmarks(image2, landmarks2, face2))
imshow("triangles 1", draw_triangles(image1, triplets, lp1))
imshow("triangles 2", draw_triangles(image2, triplets, lp2))
imshow("result 1", warped1)
imshow("result 2", warped2)
imshow("out 1", out1)
imshow("out 2", out2)
cv2.waitKey()
return out1, out2
# -*- coding: utf-8 -*-
"""
Created on Wed Aug 22 09:37:17 2018
@author: Lovro
"""
import cv2
import numpy as np
import common
# read image
img = cv2.imread('figures\\box.jpg')
common.imshow(img, 800, 600)
# convert to gray
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# find corners
corners = cv2.goodFeaturesToTrack(gray, 100, 0.01, 10)
corners = np.int0(corners)
#mark corners on image
for i in corners:
x, y = i.ravel()
cv2.circle(img, (x, y), 3, 255, -1)
common.imshow(img)
import cv2
import common
img1 = common.imread('dog.jpg',0)
img2 = cv2.blur(img1,(5,5))
common.imshow('blur',img1,img2)
import cv2
import common
img1 = common.imread('dog.jpg')
img2 = img1.copy()
v = img2.item(10, 20)
print(v)
for y in range(10, 20):
for x in range(20, 40):
img2.itemset((y, x), (x + y) * 4)
v = img2.item(10, 20)
print(v)
v = img2.item(19, 39)
print(v)
common.imshow('pixel', img1, img2)
import cv2
import common
img1 = common.imread('dog.jpg')
img2 = img1.copy()
v = img2.item(10,20)
print(v)
for y in range(10,20):
for x in range(20,40):
img2.itemset((y,x),(x+y)*4)
v = img2.item(10,20)
print(v)
v = img2.item(19,39)
print(v)
common.imshow('pixel',img1,img2)
import cv2
import numpy
import common
img1 = numpy.zeros((60,360,3),numpy.uint8)
img2 = img1
h,w,_ = img1.shape
for y in range(0,h):
for x in range(0,w):
img1.itemset((y,x,0),x/2)
img1.itemset((y,x,1),255)
img1.itemset((y,x,2),255)
img2 = cv2.cvtColor(img1,cv2.COLOR_HSV2BGR)
common.imshow('colors',img1,img2)
import cv2
import common
import numpy
img1 = common.imread('shapes.jpg',0)
ret,img1cp = cv2.threshold(img1,127,255,cv2.THRESH_BINARY)
contours, hierarchy = cv2.findContours(img1cp,cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE)
img2 = numpy.zeros(img1.shape,numpy.uint8)
cv2.drawContours(img2, contours, -1, 255, 1)
print "number of contours = %d" % (len(contours))
for i in range(0,len(contours)):
c = contours[i]
f = cv2.FONT_HERSHEY_SIMPLEX
a = cv2.contourArea(c)
l = cv2.arcLength(c,True)
m = "%d" % (i)
print "%d: len=%d, arcLen=%d, area=%d" % (i,len(c),l,a)
p = tuple(c[0][0])
cv2.putText(img2,m,p,f,0.4,(255),1,cv2.CV_AA)
common.imshow('contours',img1,img2)
# See below for more details:
# http://docs.opencv.org/3.0-beta/doc/py_tutorials/py_imgproc/py_contours/py_contour_features/py_contour_features.html#contour-features
import cv2
import common
img1 = common.imread('dog.jpg',0)
ret,img2 = cv2.threshold(img1,127,255,cv2.THRESH_BINARY)
common.imshow('thresh',img1,img2)
vis1 = np.hstack(common.to_rgb(x).numpy())
vis = np.vstack([vis0, vis1])
common.imwrite('train_log/batches_%04d.jpg' % step_i, vis)
print('batch (before/after):')
common.imshow(vis)
def plot_loss(loss_log):
pl.figure(figsize=(10, 4))
pl.title('Loss history (log10)')
pl.plot(np.log10(loss_log), '.', alpha=0.1)
pl.show()
target_img = common.load_emoji(url)
common.imshow(common.zoom(common.to_rgb(target_img), 2), fmt='png')
p = common.TARGET_PADDING
pad_target = tf.pad(target_img, [(p, p), (p, p), (0, 0)])
h, w = pad_target.shape[:2]
seed = np.zeros([h, w, common.CHANNEL_N], np.float32)
seed[h // 2, w // 2, 3:] = 1.0
def loss_f(x):
return tf.reduce_mean(tf.square(common.to_rgba(x) - pad_target),
[-2, -3, -1])
ca = common.CAModel()
import cv2
import common
img1 = common.imread('dog.jpg',0)
img2 = img1[100:180, 90:150]
cv2.rectangle(img2,(0,0),(59,79),255,3)
common.imshow('roi',img1,img2)
import cv2
import common
img1 = common.imread('dog.jpg',0)
# do something.
common.imshow('noname',img1,img2)
import cv2
import common
img1 = common.imread('dog.jpg',1)
img2 = img1.copy()
img2[:,:,2] = 0 # Make all red pixels to zero.
# b,g,r = cv2.split(img1)
# img2 = cv2.merge((b,g,r)) # Same image
# img2 = b # Grayscale image using red pixels
common.imshow('split',img1,img2)
# See below for details:
# http://docs.opencv.org/3.0-beta/doc/py_tutorials/py_core/py_basic_ops/py_basic_ops.html#splitting-and-merging-image-channels
Created on Wed Aug 22 08:28:36 2018
@author: Lovro
"""
#pro packages
import cv2
#my packages
import common
#read image
img = cv2.imread('figures\\box.jpg')
#display image
common.imshow(img, 800, 600)
#rgb2gray
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
#display
common.imshow(gray, 800, 600)
#corner detection
corner = cv2.cornerHarris(gray, 2, 3, 0.04)
#make corner more visible
corner = cv2.dilate(corner, None)
#add corner to original image
img_corner = img
img_corner[corner > 0.01 * corner.max()] = [0, 255, 0]
import cv2
import common
img1 = common.imread('dog.jpg', 1)
img2 = img1.copy()
img2[:, :, 2] = 0 # Make all red pixels to zero.
# b,g,r = cv2.split(img1)
# img2 = cv2.merge((b,g,r)) # Same image
# img2 = b # Grayscale image using red pixels
common.imshow('split', img1, img2)
# See below for details:
# http://docs.opencv.org/3.0-beta/doc/py_tutorials/py_core/py_basic_ops/py_basic_ops.html#splitting-and-merging-image-channels
import cv2
import common
img1 = common.imread('dog.jpg', 0)
img2 = cv2.blur(img1, (5, 5))
common.imshow('blur', img1, img2)
import cv2
import common
img1 = common.imread('dog.jpg', 0)
ret, img2 = cv2.threshold(img1, 127, 255, cv2.THRESH_BINARY)
common.imshow('thresh', img1, img2)
