merge

19 files changed, 727 insertions, 19 deletions

diff --git a/.canny-single.py.swp b/.canny-single.py.swp
new file mode 100644
index 0000000..9c12044
--- /dev/null
+++ b/.canny-single.py.swp
diff --git a/.gitignore b/.gitignore
index 4ef3174..6c811ec 100644
--- a/.gitignore
+++ b/.gitignore
@@ -43,3 +43,5 @@ test/data/legacy_serialized.pt
 .idea
 recursive/
 mov/
+*.swp
+
diff --git a/canny-cv.py b/canny-cv.py
new file mode 100644
index 0000000..f8b5dc3
--- /dev/null
+++ b/canny-cv.py
@@ -0,0 +1,132 @@
+import os
+from options.test_options import TestOptions
+from data import CreateRecursiveDataLoader
+from models import create_model
+from util.visualizer import Visualizer
+from util.util import mkdirs, crop_image
+from util import html
+from shutil import move, copyfile
+from PIL import Image, ImageOps
+from skimage.transform import resize
+from scipy.misc import imresize
+from shutil import copyfile, rmtree
+import numpy as np
+import cv2
+import time
+
+import subprocess
+from time import sleep
+
+blur = 3
+sigma = 0
+canny_lo = 10
+canny_hi = 220
+frac_a = 0.99
+frac_b = 1 - frac_a
+
+if __name__ == '__main__':
+    opt = TestOptions().parse()
+    opt.nThreads = 1   # test code only supports nThreads = 1
+    opt.batchSize = 1  # test code only supports batchSize = 1
+    opt.serial_batches = True  # no shuffle
+    opt.no_flip = True  # no flip
+    opt.experiment = opt.start_img.split("/")[-1].split(".")[0]
+
+    render_dir = opt.results_dir + opt.name + "/exp:" + opt.experiment + "/"
+
+    if os.path.exists(render_dir):
+        rmtree(render_dir)
+    mkdirs(render_dir)
+
+    cmd = ("convert", opt.start_img, '-canny', '0x1+10%+30%', render_dir + "frame_00000.png")
+    process = subprocess.Popen(cmd, stdout=subprocess.PIPE)
+    output, error = process.communicate()
+
+    #copyfile(opt.start_img, render_dir + "frame_00000.png")
+
+    data_loader = CreateRecursiveDataLoader(opt)
+    dataset = data_loader.load_data()
+    ds = dataset.dataset
+    model = create_model(opt)
+    visualizer = Visualizer(opt)
+    # create website
+    web_dir = os.path.join(opt.results_dir, opt.name, '%s_%s' % (opt.phase, opt.which_epoch))
+    webpage = html.HTML(web_dir, 'Experiment = %s, Phase = %s, Epoch = %s' % (opt.name, opt.phase, opt.which_epoch))
+    # test
+    last_im = None
+    for i, data in enumerate(data_loader):
+        if i >= opt.how_many:
+            break
+        model.set_input(data)
+        model.test()
+        visuals = model.get_current_visuals()
+        img_path = model.get_image_paths()
+        print('%04d: process image... %s' % (i, img_path))
+        ims = visualizer.save_images(webpage, visuals, img_path, aspect_ratio=opt.aspect_ratio)
+
+        im = visuals['fake_B']
+        tmp_path = render_dir + "frame_{:05d}_tmp.png".format(i+1)
+        edges_path = render_dir + "frame_{:05d}.png".format(i+1)
+        render_path = render_dir + "ren_{:05d}.png".format(i+1)
+
+        if dataset.name() == 'RecursiveDatasetDataLoader':
+            # print(visuals.keys())
+            # s = 256
+            # p = 8
+            # im = imresize(im, (s-p, s-p), interp='bicubic')
+            # image_pil = Image.fromarray(im)
+            # image_pil = ImageOps.expand(image_pil, p)
+            # image_pil.save(save_path)
+            # copyfile(save_path, final_path)
+            if last_im is not None:
+                tmp_im = im.copy()
+                #array_a = np.multiply(im.astype('float64'), frac_a)
+                #array_b = np.multiply(last_im.astype('float64'), frac_b)
+                #im = np.add(array_a, array_b).astype('uint8')
+                # print(im.shape, im.dtype)
+                last_im = np.roll(tmp_im, 1, axis=1)
+            else:
+                last_im = im.copy().astype('uint8')
+                tmp_im = im.copy().astype('uint8')
+                #print(im.shape, im.dtype)
+
+            image_pil = Image.fromarray(tmp_im, mode='RGB')
+            image_pil.save(tmp_path)
+            os.rename(tmp_path, render_path)
+
+        image_pil = Image.fromarray(im, mode='RGB')
+        image_pil = crop_image(image_pil, (0.50, 0.50), 0.5)
+        im = np.asarray(image_pil).astype('uint8')
+        #print(im.shape, im.dtype)
+        opencv_image = im[:, :, ::-1].copy()
+        opencv_image = cv2.GaussianBlur(opencv_image, (blur,blur), sigma)
+        opencv_image = cv2.Canny(opencv_image, canny_lo, canny_hi)
+        cv2.imwrite(tmp_path, opencv_image)
+        os.rename(tmp_path, edges_path)
+
+
+    webpage.save()
+
+    os.remove(render_dir + "frame_00000.png")
+
+    t = time.time()
+    t /= 60
+    t %= 525600
+    video_fn = "{}_{}_canmix_{}frame_{}mix_{}blur_{}sigma_{}lo_{}hi_{}.mp4".format(
+        opt.name, opt.experiment,
+        opt.how_many, frac_a,
+        blur, sigma, canny_lo, canny_hi,
+        int(t))
+
+    cmd = ("/usr/bin/ffmpeg", "-i", render_dir + "ren_%05d.png", "-y", "-c:v", "libx264", "-vf", "fps=30", "-pix_fmt", "yuv420p", render_dir + video_fn)
+    process = subprocess.Popen(cmd, stdout=subprocess.PIPE)
+    output, error = process.communicate()
+
+    print("________")
+
+    cmd = ("scp", render_dir + video_fn, "jules@asdf.us:asdf/neural/")
+    process = subprocess.Popen(cmd, stdout=subprocess.PIPE)
+    output, error = process.communicate()
+
+    print("https://asdf.us/neural/" + video_fn)
+
diff --git a/canny-dir.py b/canny-dir.py
new file mode 100644
index 0000000..8a671dd
--- /dev/null
+++ b/canny-dir.py
@@ -0,0 +1,31 @@
+import os
+from options.test_options import TestOptions
+from shutil import move, copyfile
+from PIL import Image, ImageOps
+from shutil import copyfile, rmtree
+import numpy as np
+import cv2
+
+import subprocess
+from time import sleep
+
+if __name__ == '__main__':
+    opt = TestOptions().parse()
+    opt.nThreads = 1   # test code only supports nThreads = 1
+    opt.batchSize = 1  # test code only supports batchSize = 1
+    opt.serial_batches = True  # no shuffle
+    opt.no_flip = True  # no flip
+    opt.experiment = opt.start_img.split("/")[-1].split(".")[0]
+
+    render_dir = opt.results_dir + opt.name + "/exp:" + opt.experiment + "/"
+
+    if os.path.exists(render_dir):
+        rmtree(render_dir)
+    mkdirs(render_dir)
+    pil_image = Image.open(opt.start_img).convert('RGB')
+    opencv_image = np.array(pil_image)
+    opencv_image = opencv_image[:, :, ::-1].copy()
+    opencv_image = cv2.GaussianBlur(opencv_image, (3,3), 1)
+    opencv_image = cv2.Canny(opencv_image, 100, 200)
+    cv2.imwrite(render_dir + "frame_00000.png", opencv_image)
+
diff --git a/canny-single.py b/canny-single.py
new file mode 100644
index 0000000..749067a
--- /dev/null
+++ b/canny-single.py
@@ -0,0 +1,36 @@
+import os
+from options.test_options import TestOptions
+from shutil import move, copyfile
+from PIL import Image, ImageOps
+from shutil import copyfile, rmtree
+import numpy as np
+import cv2
+
+import subprocess
+from time import sleep
+
+if __name__ == '__main__':
+    opt = TestOptions().parse()
+    opt.nThreads = 1   # test code only supports nThreads = 1
+    opt.batchSize = 1  # test code only supports batchSize = 1
+    opt.serial_batches = True  # no shuffle
+    opt.no_flip = True  # no flip
+    opt.experiment = opt.start_img
+
+    render_dir = opt.results_dir + opt.experiment + "/"
+
+    if os.path.exists(render_dir):
+        rmtree(render_dir)
+    mkdirs(render_dir)
+    i = 0
+    for f in sorted(os.listdir(opt.start_img)):
+        if not os.path.isfile(f):
+            continue
+        pil_image = Image.open(f).convert('RGB')
+        opencv_image = np.array(pil_image)
+        opencv_image = opencv_image[:, :, ::-1].copy()
+        opencv_image = cv2.GaussianBlur(opencv_image, (3,3), 1)
+        opencv_image = cv2.Canny(opencv_image, 100, 200)
+        cv2.imwrite(render_dir + "frame_{:04d}.png".format(i), opencv_image)
+        i += 1
+
diff --git a/canny-single.sh b/canny-single.sh
new file mode 100755
index 0000000..7ad010f
--- /dev/null
+++ b/canny-single.sh
@@ -0,0 +1,14 @@
+python canny-single.py \
+  --dataroot /home/ubuntu/Desktop/thumbs/wood/B/train/ \
+  --name wood \
+  --start_img /home/ubuntu/Desktop/thumbs/wood/B/train/frame_00100.png \
+  --loadSize 256 \
+  --fineSize 256 \
+  --how_many 30 \
+  --model test  \
+  --which_model_netG unet_256  \
+  --which_direction AtoB  \
+  --dataset_mode recursive  \
+  --norm batch
+
+#  --aspect_ratio 1.777777 \
diff --git a/canny.py b/canny.py
new file mode 100644
index 0000000..b8dc6e8
--- /dev/null
+++ b/canny.py
@@ -0,0 +1,111 @@
+import os
+from options.test_options import TestOptions
+from data import CreateRecursiveDataLoader
+from models import create_model
+from util.visualizer import Visualizer
+from util.util import mkdirs
+from util import html
+from shutil import move, copyfile
+from PIL import Image, ImageOps
+from skimage.transform import resize
+from scipy.misc import imresize
+from shutil import copyfile, rmtree
+import time as time
+import numpy as np
+import cv2
+
+import subprocess
+from time import sleep
+
+if __name__ == '__main__':
+    opt = TestOptions().parse()
+    opt.nThreads = 1   # test code only supports nThreads = 1
+    opt.batchSize = 1  # test code only supports batchSize = 1
+    opt.serial_batches = True  # no shuffle
+    opt.no_flip = True  # no flip
+    opt.experiment = opt.start_img.split("/")[-1].split(".")[0]
+
+    render_dir = opt.results_dir + opt.name + "/exp:" + opt.experiment + "/"
+
+    if os.path.exists(render_dir):
+        rmtree(render_dir)
+    mkdirs(render_dir)
+
+    cmd = ("convert", opt.start_img, '-canny', '0x1+10%+30%', render_dir + "frame_00000.png")
+    process = subprocess.Popen(cmd, stdout=subprocess.PIPE)
+    output, error = process.communicate()
+
+    data_loader = CreateRecursiveDataLoader(opt)
+    dataset = data_loader.load_data()
+    ds = dataset.dataset
+    model = create_model(opt)
+    visualizer = Visualizer(opt)
+    # create website
+    web_dir = os.path.join(opt.results_dir, opt.name, '%s_%s' % (opt.phase, opt.which_epoch))
+    webpage = html.HTML(web_dir, 'Experiment = %s, Phase = %s, Epoch = %s' % (opt.name, opt.phase, opt.which_epoch))
+    # test
+    last_im = None
+    for i, data in enumerate(data_loader):
+        if i >= opt.how_many:
+            break
+        model.set_input(data)
+        model.test()
+        visuals = model.get_current_visuals()
+        img_path = model.get_image_paths()
+        print('%04d: process image... %s' % (i, img_path))
+        ims = visualizer.save_images(webpage, visuals, img_path, aspect_ratio=opt.aspect_ratio)
+        if dataset.name() == 'RecursiveDatasetDataLoader':
+            # print(visuals.keys())
+            im = visuals['fake_B']
+            tmp_path = render_dir + "frame_{:05d}_tmp.png".format(i+1)
+            edges_path = render_dir + "frame_{:05d}.png".format(i+1)
+            render_path = render_dir + "ren_{:05d}.png".format(i+1)
+            # s = 256
+            # p = 8
+            # im = imresize(im, (s-p, s-p), interp='bicubic')
+            # image_pil = Image.fromarray(im)
+            # image_pil = ImageOps.expand(image_pil, p)
+            # image_pil.save(save_path)
+            # copyfile(save_path, final_path)
+            #if last_im is not None:
+            #    frac_a = 999/1000
+            #    frac_b = 1/1000
+            #    tmp_im = im.copy()
+            #    array_a = np.multiply(im, frac_a)
+            #    array_b = np.multiply(last_im, frac_b)
+            #    # im = np.add(array_a, array_b).astype('int8')
+            #    # print(im.shape, im.dtype)
+            #    last_im = np.roll(tmp_im, 1, axis=1)
+            #else:
+            #    last_im = im.copy()
+            #    print(im.shape, im.dtype)
+            image_pil = Image.fromarray(im, mode='RGB')
+            image_pil.save(tmp_path)
+            os.rename(tmp_path, render_path)
+
+            cmd = ("convert", render_path, '-canny', '0x1+10%+30%', tmp_path)
+            process = subprocess.Popen(cmd, stdout=subprocess.PIPE)
+            output, error = process.communicate()
+
+            os.rename(tmp_path, edges_path)
+
+    webpage.save()
+
+    os.remove(render_dir + "frame_00000.png")
+
+    t = time.time()
+    t /= 60
+    t %= 525600
+    video_fn = opt.name + "_" + opt.experiment + "_canny_" + str(opt.how_many) + "_" + str(int(t)) + ".mp4"
+
+    cmd = ("/usr/bin/ffmpeg", "-i", render_dir + "ren_%05d.png", "-y", "-c:v", "libx264", "-vf", "fps=30", "-pix_fmt", "yuv420p", render_dir + video_fn)
+    process = subprocess.Popen(cmd, stdout=subprocess.PIPE)
+    output, error = process.communicate()
+
+    print("________")
+
+    cmd = ("scp", render_dir + video_fn, "jules@asdf.us:asdf/neural/")
+    process = subprocess.Popen(cmd, stdout=subprocess.PIPE)
+    output, error = process.communicate()
+
+    print("https://asdf.us/neural/" + video_fn)
diff --git a/recursive-canny.sh b/recursive-canny.sh
new file mode 100755
index 0000000..0150438
--- /dev/null
+++ b/recursive-canny.sh
@@ -0,0 +1,14 @@
+python canny-cv.py \
+  --dataroot /home/ubuntu/Desktop/thumbs/wood/B/train/ \
+  --name woodcanny \
+  --start_img /home/ubuntu/Desktop/thumbs/wood/B/train/frame_00100.png \
+  --loadSize 256 \
+  --fineSize 256 \
+  --how_many 200 \
+  --model test  \
+  --which_model_netG unet_256  \
+  --which_direction AtoB  \
+  --dataset_mode recursive  \
+  --norm batch
+
+#  --aspect_ratio 1.777777 \
diff --git a/recursive.sh b/recursive.sh
index daffba1..6ccc60d 100755
--- a/recursive.sh
+++ b/recursive.sh
@@ -1,7 +1,7 @@
 python test.py \
-  --dataroot /home/lens/Desktop/thumbs/woodscaled/A/train/ \
-  --name woodscaled_pix2pix \
-  --start_img /home/lens/Desktop/thumbs/woodscaled/A/train/frame_1002.png \
+  --dataroot /home/ubuntu/Desktop/thumbs/wood/A/train/ \
+  --name wood \
+  --start_img /home/ubuntu/Desktop/thumbs/wood/A/train/frame_1002.png \
   --how_many 2000 \
   --model test  \
   --which_model_netG unet_256  \
diff --git a/run-canny.sh b/run-canny.sh
new file mode 100755
index 0000000..aef1118
--- /dev/null
+++ b/run-canny.sh
@@ -0,0 +1,91 @@
+dataset="woodcanny"
+
+#/home/lens/Desktop/dataset-random.pl
+
+#python datasets/combine_A_and_B.py \
+#  --fold_A "$HOME/Desktop/thumbs/$dataset/A" \
+#  --fold_B "$HOME/Desktop/thumbs/$dataset/B" \
+#  --fold_AB "./datasets/$dataset/"
+
+python train.py \
+   --dataroot "./datasets/$dataset" \
+   --name "$dataset" \
+   --model pix2pix  \
+   --loadSize 276 \
+   --fineSize 256 \
+   --which_model_netG unet_256  \
+   --which_direction AtoB  \
+   --lambda_B 100  \
+   --dataset_mode aligned  \
+   --which_epoch latest \
+   --continue_train \
+   --no_lsgan --norm batch --pool_size 0
+
+python canny-cv.py \
+  --dataroot /home/ubuntu/Desktop/thumbs/wood/B/train/ \
+  --name woodcanny \
+  --start_img /home/ubuntu/Desktop/thumbs/wood/B/train/frame_00100.png \
+  --loadSize 256 \
+  --fineSize 256 \
+  --how_many 200 \
+  --model test  \
+  --which_model_netG unet_256  \
+  --which_direction AtoB  \
+  --dataset_mode recursive  \
+  --norm batch
+
+python train.py \
+   --dataroot "./datasets/$dataset" \
+   --name "$dataset" \
+   --model pix2pix  \
+   --loadSize 276 \
+   --fineSize 256 \
+   --which_model_netG unet_256  \
+   --which_direction AtoB  \
+   --lambda_B 100  \
+   --dataset_mode aligned  \
+   --which_epoch latest \
+   --continue_train \
+   --no_lsgan --norm batch --pool_size 0
+
+python canny-cv.py \
+  --dataroot /home/ubuntu/Desktop/thumbs/wood/B/train/ \
+  --name woodcanny \
+  --start_img /home/ubuntu/Desktop/thumbs/wood/B/train/frame_00100.png \
+  --loadSize 256 \
+  --fineSize 256 \
+  --how_many 200 \
+  --model test  \
+  --which_model_netG unet_256  \
+  --which_direction AtoB  \
+  --dataset_mode recursive  \
+  --norm batch
+
+python train.py \
+   --dataroot "./datasets/$dataset" \
+   --name "$dataset" \
+   --model pix2pix  \
+   --loadSize 276 \
+   --fineSize 256 \
+   --which_model_netG unet_256  \
+   --which_direction AtoB  \
+   --lambda_B 100  \
+   --dataset_mode aligned  \
+   --which_epoch latest \
+   --continue_train \
+   --no_lsgan --norm batch --pool_size 0
+
+python canny-cv.py \
+  --dataroot /home/ubuntu/Desktop/thumbs/wood/B/train/ \
+  --name woodcanny \
+  --start_img /home/ubuntu/Desktop/thumbs/wood/B/train/frame_00100.png \
+  --loadSize 256 \
+  --fineSize 256 \
+  --how_many 200 \
+  --model test  \
+  --which_model_netG unet_256  \
+  --which_direction AtoB  \
+  --dataset_mode recursive  \
+  --norm batch
+
+#  --aspect_ratio 1.777777 \
diff --git a/run-flow.sh b/run-flow.sh
new file mode 100755
index 0000000..3f650dc
--- /dev/null
+++ b/run-flow.sh
@@ -0,0 +1,36 @@
+dataset="woodflowfixed"
+
+#/home/lens/Desktop/dataset-random.pl
+
+#python datasets/combine_A_and_B.py \
+#  --fold_A "$HOME/Desktop/thumbs/$dataset/A" \
+#  --fold_B "$HOME/Desktop/thumbs/$dataset/B" \
+#  --fold_AB "./datasets/$dataset/"
+
+python train.py \
+   --dataroot "./datasets/$dataset" \
+   --name "$dataset" \
+   --model pix2pix  \
+   --loadSize 276 \
+   --fineSize 256 \
+   --which_model_netG unet_256  \
+   --which_direction AtoB  \
+   --lambda_B 100  \
+   --dataset_mode aligned  \
+   --continue_train \
+   --which_epoch latest \
+   --no_lsgan --norm batch --pool_size 0
+
+#python canny-cv.py \
+#  --dataroot /home/ubuntu/Desktop/thumbs/wood/B/train/ \
+#  --name woodcanny \
+#  --start_img /home/ubuntu/Desktop/thumbs/wood/B/train/frame_00100.png \
+#  --loadSize 256 \
+#  --fineSize 256 \
+#  --how_many 200 \
+#  --model test  \
+#  --which_model_netG unet_256  \
+#  --which_direction AtoB  \
+#  --dataset_mode recursive  \
+#  --norm batch
+
diff --git a/run.sh b/run.sh
index 75ada20..5f49b7d 100755
--- a/run.sh
+++ b/run.sh
@@ -1,24 +1,25 @@
-dataset="randomcrops5k"
+dataset="woodcanny"
 
 #/home/lens/Desktop/dataset-random.pl
-#
+
 #python datasets/combine_A_and_B.py \
-#  --fold_A "/home/lens/Desktop/thumbs/$dataset/A" \
-#  --fold_B "/home/lens/Desktop/thumbs/$dataset/B" \
+#  --fold_A "$HOME/Desktop/thumbs/$dataset/A" \
+#  --fold_B "$HOME/Desktop/thumbs/$dataset/B" \
 #  --fold_AB "./datasets/$dataset/"
 
-# python train.py \
-#   --dataroot "./datasets/$dataset" \
-#   --name "$dataset" \
-#   --model pix2pix  \
-#   --loadSize 256 \
-#   --fineSize 256 \
-#   --which_model_netG unet_256  \
-#   --which_direction AtoB  \
-#   --lambda_B 100  \
-#   --dataset_mode aligned  \
-#   --no_lsgan --norm batch --pool_size 0 \
-#   --continue_train
+python train.py \
+   --dataroot "./datasets/$dataset" \
+   --name "$dataset" \
+   --model pix2pix  \
+   --loadSize 276 \
+   --fineSize 256 \
+   --which_model_netG unet_256  \
+   --which_direction AtoB  \
+   --lambda_B 100  \
+   --dataset_mode aligned  \
+   --which_epoch latest \
+   --continue_train \
+   --no_lsgan --norm batch --pool_size 0
 
 python test.py \
   --dataroot "/home/lens/Desktop/thumbs/$dataset/A/train/" \
diff --git a/scripts/builders/canny-dir.py b/scripts/builders/canny-dir.py
new file mode 100644
index 0000000..ae1bd62
--- /dev/null
+++ b/scripts/builders/canny-dir.py
@@ -0,0 +1,40 @@
+import os
+from shutil import move, copyfile
+from PIL import Image, ImageOps
+from shutil import copyfile, rmtree
+import numpy as np
+import cv2
+
+work_dir = "woodsized/"
+render_dir = "woodcanny/"
+
+if os.path.exists(render_dir):
+    rmtree(render_dir)
+
+os.makedirs(render_dir)
+os.makedirs(render_dir + "A/")
+os.makedirs(render_dir + "A/train/")
+os.makedirs(render_dir + "A/test/")
+os.makedirs(render_dir + "A/val/")
+os.makedirs(render_dir + "B/")
+os.makedirs(render_dir + "B/train/")
+os.makedirs(render_dir + "B/test/")
+os.makedirs(render_dir + "B/val/")
+
+for i,fn in enumerate(sorted(os.listdir(work_dir))):
+    pil_image = Image.open(work_dir + fn).convert('RGB')
+    opencv_image = np.array(pil_image)
+    opencv_image = opencv_image[:, :, ::-1].copy()
+    opencv_image = cv2.GaussianBlur(opencv_image, (3,3), 1)
+    opencv_image = cv2.Canny(opencv_image, 100, 200)
+    ren = "frame_{:05d}.png".format(i)
+    if (i % 10) == 3:
+      wd = "test/"
+    elif (i % 10) == 6:
+      wd = "val/"
+    else:
+      wd = "train/"
+    cv2.imwrite(render_dir + "A/" + wd + ren, opencv_image)
+    copyfile(work_dir + fn, render_dir + "B/" + wd + ren)
+
+
diff --git a/scripts/builders/flow-dir.py b/scripts/builders/flow-dir.py
new file mode 100644
index 0000000..620ba32
--- /dev/null
+++ b/scripts/builders/flow-dir.py
@@ -0,0 +1,75 @@
+import os
+import sys
+from shutil import move, copyfile
+from PIL import Image, ImageOps
+from shutil import copyfile, rmtree
+import numpy as np
+import cv2
+
+work_dir = "woodsized/"
+render_dir = "woodflow/"
+
+if os.path.exists(render_dir):
+    rmtree(render_dir)
+
+os.makedirs(render_dir)
+os.makedirs(render_dir + "A/")
+os.makedirs(render_dir + "A/train/")
+os.makedirs(render_dir + "A/test/")
+os.makedirs(render_dir + "A/val/")
+os.makedirs(render_dir + "B/")
+os.makedirs(render_dir + "B/train/")
+os.makedirs(render_dir + "B/test/")
+os.makedirs(render_dir + "B/val/")
+
+hsv = []
+
+for i,fn in enumerate(sorted(os.listdir(work_dir))):
+  # load image and convert to grayscale
+  pil_image = Image.open(work_dir + fn).convert('RGB')
+  im = np.array(pil_image)
+  im = im[:, :, ::-1].copy()
+  im = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
+
+  # store first frame
+  if i == 0:
+    prev = im
+    hsv = np.zeros((256,256,3))
+    hsv[...,1] = 255
+    continue
+
+  # compute optical flow
+  flow = cv2.calcOpticalFlowFarneback(prev, im, None, 0.5, 3, 15, 3, 5, 1.2, 0)
+
+  # turn into magnitude/angle
+  mag, ang = cv2.cartToPolar(flow[...,0], flow[...,1])
+
+  # store angle as hue
+  hsv[...,0] = ang * 180 / np.pi / 2
+
+  # store magnitude as lum
+  hsv[...,2] = cv2.normalize(mag, None, 0, 255, cv2.NORM_MINMAX)
+
+  # convert this HSL to BGR
+  bgr = cv2.cvtColor(hsv.astype('uint8'), cv2.COLOR_HSV2BGR)
+
+  ren = "frame_{:05d}.png".format(i)
+  if (i % 10) == 3:
+    wd = "test/"
+  elif (i % 10) == 6:
+    wd = "val/"
+  else:
+    wd = "train/"
+
+  if i == 1:
+    prev_hsv = np.copy(hsv)
+    prev_bgr = np.copy(bgr)
+    continue
+
+  cv2.imwrite(render_dir + "A/" + wd + ren, prev_bgr)
+  cv2.imwrite(render_dir + "B/" + wd + ren, bgr)
+  # copyfile(work_dir + fn, render_dir + "B/" + wd + ren)
+  prev = im
+  prev_hsv = np.copy(hsv)
+  prev_bgr = np.copy(bgr)
+
diff --git a/scripts/builders/pair-dataset.pl b/scripts/builders/pair-dataset.pl
new file mode 100644
index 0000000..06c40c0
--- /dev/null
+++ b/scripts/builders/pair-dataset.pl
@@ -0,0 +1,55 @@
+#!/usr/bin/perl
+
+use strict;
+
+our $images_dir = "/home/ubuntu/Desktop/wood/";
+our $thumbs_dir = "/home/ubuntu/Desktop/thumbs/wood/";
+
+our $dt = 1;
+
+mkdir($thumbs_dir);
+mkdir($thumbs_dir . "A/");
+mkdir($thumbs_dir . "B/");
+mkdir($thumbs_dir . "A/train/");
+mkdir($thumbs_dir . "B/train/");
+mkdir($thumbs_dir . "A/val/");
+mkdir($thumbs_dir . "B/val/");
+mkdir($thumbs_dir . "A/test/");
+mkdir($thumbs_dir . "B/test/");
+
+our @files = ();
+
+opendir DIR, $images_dir;
+while (readdir DIR) {
+  next if /^\./;
+  push(@files, $_);
+}
+closedir DIR;
+
+our @images = sort @files;
+
+my $count = scalar(@images) - $dt;
+
+my $i; my $x; my $y; my $dir;
+for ($i = 0; $i < $count; $i++) {
+  my $id = $i;
+  if ( ($id % 7) == 3) {
+    $dir = "test/";
+  } elsif ( ($id % 7) == 6) {
+    $dir = "val/";
+  } else {
+    $dir = "train/";
+  }
+  if ( $i && ($i % 1000) == 0) {
+    print($id . "...\n")
+  }
+
+  my $a_frame = $images_dir . $images[$id];
+  my $b_frame = $images_dir . $images[$id+$dt];
+  my $fn = sprintf("frame_%05d.png", $i);
+  #print "$fn $x $y $a_frame\n";
+  system("convert", $a_frame, "-resize", '256x256!', '-canny', '0x1+10%+30%', $thumbs_dir . "A/" . $dir . $fn);
+  system("convert", $b_frame, "-resize", '256x256!', $thumbs_dir . "B/" . $dir . $fn);
+}
+print $count . "\n";
+
diff --git a/test-canny.sh b/test-canny.sh
new file mode 100644
index 0000000..57b56b6
--- /dev/null
+++ b/test-canny.sh
@@ -0,0 +1,11 @@
+python canny-cv.py \
+  --dataroot /home/ubuntu/Desktop/thumbs/wood/B/train/ \
+  --name woodcanny \
+  --loadSize 256 \
+  --fineSize 256 \
+  --how_many 200 \
+  --model test  \
+  --which_model_netG unet_256  \
+  --which_direction AtoB  \
+  --dataset_mode aligned  \
+  --norm batch
diff --git a/test.py b/test.py
index 6a5113c..beac2c0 100644
--- a/test.py
+++ b/test.py
@@ -103,3 +103,4 @@ if __name__ == '__main__':
 
     print("https://asdf.us/neural/" + filename)
     print("\n")
+
diff --git a/test.sh b/test.sh
new file mode 100755
index 0000000..c25c987
--- /dev/null
+++ b/test.sh
@@ -0,0 +1,38 @@
+dataset="wood"
+
+#/home/lens/Desktop/dataset-random.pl
+#
+#python datasets/combine_A_and_B.py \
+#  --fold_A "$HOME/Desktop/thumbs/$dataset/A" \
+#  --fold_B "$HOME/Desktop/thumbs/$dataset/B" \
+#  --fold_AB "./datasets/$dataset/"
+
+# python train.py \
+#   --dataroot "./datasets/$dataset" \
+#   --name "$dataset" \
+#   --model pix2pix  \
+#   --loadSize 256 \
+#   --fineSize 256 \
+#   --which_model_netG unet_256  \
+#   --which_direction AtoB  \
+#   --lambda_B 100  \
+#   --dataset_mode aligned  \
+#   --which_epoch latest \
+#   --no_lsgan --norm batch --pool_size 0
+#   --continue_train \
+
+python test.py \
+  --dataroot "/home/lens/Desktop/thumbs/$dataset/A/train/" \
+  --name "$dataset" \
+  --start_img "/home/lens/Desktop/thumbs/$dataset/A/train/frame_1008.png" \
+  --how_many 1000 \
+  --model test  \
+  --aspect_ratio 1.777777 \
+  --which_model_netG unet_256  \
+  --which_direction AtoB  \
+  --dataset_mode test \
+  --loadSize 256 \
+  --fineSize 256 \
+  --norm batch
+
+
diff --git a/util/util.py b/util/util.py
index 7a452a6..1c03715 100644
--- a/util/util.py
+++ b/util/util.py
@@ -54,3 +54,23 @@ def mkdirs(paths):
 def mkdir(path):
     if not os.path.exists(path):
         os.makedirs(path)
+
+def crop_image(img, xy, scale_factor):
+    '''Crop the image around the tuple xy
+
+    Inputs:
+    -------
+    img: Image opened with PIL.Image
+    xy: tuple with relative (x,y) position of the center of the cropped image
+        x and y shall be between 0 and 1
+    scale_factor: the ratio between the original image's size and the cropped image's size
+    '''
+    center = (img.size[0] * xy[0], img.size[1] * xy[1])
+    new_size = (img.size[0] / scale_factor, img.size[1] / scale_factor)
+    left = max (0, (int) (center[0] - new_size[0] / 2))
+    right = min (img.size[0], (int) (center[0] + new_size[0] / 2))
+    upper = max (0, (int) (center[1] - new_size[1] / 2))
+    lower = min (img.size[1], (int) (center[1] + new_size[1] / 2))
+    cropped_img = img.crop((left, upper, right, lower))
+    return cropped_img
+