import numpy as np
import tensorflow as tf
from [Link] import layers
import cv2
import [Link] as plt
(X_num, y_num), _ = [Link].load_data()
X_num = np.expand_dims(X_num, axis= - 1).astype(np.float32) / 255.0
grid_size = 16
def make_numbers(X, y):
for _ in range(3):
idx = [Link](len(X_num))
number = X_num[idx] @ ([Link](1, 3) + 0.1)
number[number > 0.1] = [Link](number[number > 0.1], 0.5, 0.8)
kls = y_num[idx]
px, py = [Link] .randint(0, 100), [Link](0, 100)
mx, my = (px+14) // grid_size, (py+14) // grid_size
channels = y[my][mx]
if channels[0] > 0:
continue
channels[0] = 1.0
channels[1] = px - (mx * grid_size)
channels[2] = py - (my * grid_size)
channels[3] = 28.0
channels[4] = 28.0
channels[5 + kls] = 1.0
# put digit in X
X[py:py+28, px:px+28] += number
def make_data(size=64):
X = [Link]((size, 128, 128, 3), dtype=np.float32)
y = [Link]((size, 8, 8, 15), dtype=np.float32)
for i in range(size):
make_numbers(X[i], y[i])
X = [Link](X, 0.0, 1.0)
return X, y
def get_color_by_probability(p):
if p < 0.3:
return (1., 0., 0.)
if p < 0.7:
return (1., 1., 0.)
return (0., 1., 0.)
def show_predict(X, y, threshold=0.1):
X = [Link]()
for mx in range(8):
for my in range(8):
channels = y[my][mx]
prob, x1, y1, x2, y2 = channels[:5]
if prob < threshold:
continue
color = get_color_by_probability(prob)
px, py = (mx * grid_size) + x1, (my * grid_size) + y1
[Link](X, (int(px), int(py)), (int(px + x2), int(py
+ y2)), color, 1)
# la bel
[Link](X, (int(px), int(py - 10)), (int(px + 12),
int(py)), color, - 1)
kls = [Link](channels[5:])
[Link](X, f'{kls}', (int(px + 2), int(py - 2)),
cv2.FONT_HERSHEY_PLAIN, 0.7, (0.0, 0.0, 0.0))
[Link](X)
X, y = make_data(size=1)
show_predict(X[0], y[0])
