def thickness_sim(model, data, labels, ref_thickness):
opt_r = 0
diff = 2
current_acc = 0
current_tau = 0
acc = []
tau = []
r = 1
curr_data = data
while abs(ref_thickness - current_tau) > diff:
downscaled = np.array(scale_down(curr_data))
current_acc = test_model(model, downscaled, labels)
current_tau = utils.calc_linewidth(curr_data)
tau.append(current_tau)
acc.append(current_acc)
print("{Current acc: %s, Current tau: %s}" %
(current_acc, current_tau))
r = int(np.sign(ref_thickness - current_tau))
if abs(ref_thickness - current_tau) > diff:
opt_r += r
curr_data = change_thickness(curr_data, r)
return opt_r, acc, tau, curr_data
def thickness_sim(model, data, labels, ref_thickness):
opt_r = 0
epsilon = 2
current_acc = test_out_digits(model, data, labels)
current_tau = utils.calc_linewidth(data)
metrics = [current_acc]
tau = [current_tau]
curr_data = data
r = int(np.sign(ref_thickness - current_tau))
new_r = r
while abs(ref_thickness - current_tau) > epsilon:
opt_r += r
pre_acc = current_acc
new_data = change_thickness(curr_data, r)
current_acc = test_out_digits(model, new_data, labels)
current_tau = utils.calc_linewidth(new_data)
metrics.append(current_acc)
tau.append(current_tau)
curr_data = new_data
new_r = int(np.sign(ref_thickness - current_tau))
print(current_acc)
return opt_r, metrics, tau, curr_data
def normalize_digit_thickness(digits):
global pwa
tau = utils.calc_linewidth(digits)
datapoints = digits.shape[0]
pb = ProgressBar(total=datapoints,
prefix="processing digits:",
length=20,
fill="=",
zfill="_")
pwa.reset(tau, pb)
print("Normalizing %s digits to tau = %s" % (datapoints, tau))
with Pool(6) as p:
d_normalized = p.map(wrapper_pwa_linewidth, digits)
return np.array(d_normalized)
def change_thickness(data, ref_thickness):
print("Adjusting the thicknesses to %s" % ref_thickness)
diff = 0.5
curr_data = data
while abs(ref_thickness - tau) > diff:
tau = utils.calc_linewidth(curr_data)
print("current linethickness : %s" % tau)
dt = ref_thickness - tau
r = int(round(dt))
r = r if r != 0 else int(
np.sign(dt)) # Ensures that r is at least 1 or -1
if abs(ref_thickness - tau) > diff:
print("Adjusting Linewidth with r = %s" % r)
curr_data = change_thickness_helper(curr_data, r)
return curr_data
},
{
"type" : "Dense",
"units" : 10,
"activation" : "softmax"
}
]
}
"""
def resize_helper(img):
sides = 280
return cv2.resize(img, (sides,sides))
xtrain, ytrain, xtest, ytest = utils.load_mnist()
chunksize = 1000
mnist_total_size = xtest.shape[0]
partitions = int(mnist_total_size/chunksize)
taus = []
for i in range(partitions):
print("Starting Loop")
with Pool(6) as p:
mnist_batch = xtest[i*chunksize:(i+1)*chunksize]
mnist_batch_big = np.array(p.map(resize_helper, mnist_batch))
taus.append(utils.calc_linewidth(mnist_batch_big))
print("Calculated Line width of %s Digits of %s" % ((i+1) * chunksize, mnist_total_size))
print(np.mean(taus))