Skip to content

Instantly share code, notes, and snippets.

What would you like to do?
Experimenting with triplet loss embeddings.
import numpy as np
def build_rainbow(n, curve=None):
rgb = []
width = 2 * np.pi
for i in range(3):
offset = -i * width / 3
cur = np.cos(np.linspace(offset, offset + width, n))
rainbow = (1 + np.vstack(rgb)) / 2
if curve:
rainbow = curve(rainbow)
rainbow = np.minimum(rainbow * 256, 255).astype(int)
return rainbow.T
import numpy as np
def map_range(x, in_min, in_max, out_min, out_max):
return out_min + (out_max - out_min) * (x - in_min) / (in_max - in_min)
def plot_images(images, xy, blend=np.maximum, canvas_shape=(512,512), fill=0):
h,w = images.shape[1:3]
if images.ndim == 4:
canvas_shape = (canvas_shape[0], canvas_shape[1], images.shape[3])
min_xy = np.amin(xy, 0)
max_xy = np.amax(xy, 0)
min_canvas = np.array((0, 0))
max_canvas = np.array((canvas_shape[0] - h, canvas_shape[1] - w))
canvas = np.full(canvas_shape, fill)
for image, pos in zip(images, xy):
x_off, y_off = map_range(pos, min_xy, max_xy, min_canvas, max_canvas).astype(int)
sub_canvas = canvas[y_off:y_off+h, x_off:x_off+w]
sub_image = image[:h, :w]
canvas[y_off:y_off+h, x_off:x_off+w] = blend(sub_canvas, sub_image)
return canvas
try: # Python 2
from cStringIO import StringIO as BytesIO
except: # Python 3
from io import BytesIO
import numpy as np
import PIL.Image
import IPython.display
import shutil
from math import sqrt
def show_array(a, fmt='png', filename=None, retina=False, zoom=None):
if len(a.shape) == 1:
n = len(a)
side = int(sqrt(n))
if (side * side) == n:
a = a.reshape(side, side)
raise ValueError('input is one-dimensional', a.shape)
a = np.uint8(np.clip(a, 0, 255))
image_data = BytesIO()
PIL.Image.fromarray(a).save(image_data, fmt)
if filename is None:
height, width = a.shape[:2]
if zoom is not None:
width *= zoom
height *= zoom
with open(filename, 'wb') as f:
shutil.copyfileobj(image_data, f)
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment