def imagine2(dim, code, save=False, visualize=True):
# set the shape
shape = (1, dim, dim, zed)
code1 = code[0:zed]
code2 = code[zed:zed * 2]
# overall tone
means = code1 - 0.5 # -0.5 or 0.5
variances = code2 * .3 + .2 # .2 or .5
# final latent
z = np.random.normal(loc=means, scale=variances, size=shape)
gened = gm.infer(z)
gened *= 0.5 # normalize to 0..1
gened += 0.5
im = gened[0]
if visualize:
# visualize
vis.show_autoscaled(im, name='imagined', limit=600.)
fmap = np.transpose(z, (3, 1, 2, 0)) + .5
vis.show_batch_autoscaled(fmap, name='z_map', limit=600.)
# outputs the image
return im
def show(threshold=.5):
from cv2tools import vis, filt
bs = 16
j = np.random.choice(len(xt) - 16)
minibatch = xt[j:j + bs]
code, rec, code2, rec2, noisy_x = test(minibatch, threshold)
code = np.transpose(code[0:1], axes=(3, 1, 2, 0))
code2 = np.transpose(code2[0:1], axes=(3, 1, 2, 0))
vis.show_batch_autoscaled(code, name='code(quant)', limit=600.)
vis.show_batch_autoscaled(code2, name='code2(no quant)', limit=600.)
vis.show_batch_autoscaled(noisy_x, name='input')
vis.show_batch_autoscaled(rec, name='recon(quant)')
vis.show_batch_autoscaled(rec2, name='recon(no quant)')
#..................
import csv
import numpy
np.set_printoptions(threshold=np.inf)
f_code = open("code.csv", "w")
w_code = csv.writer(f_code)
w_code.writerow(code2)
f_code = open("input.csv", "w")
w_code = csv.writer(f_code)
w_code.writerow(noisy_x)
f_code = open("rec2.csv", "w")
w_code = csv.writer(f_code)
w_code.writerow(rec2)
def testPour():
u, q = pour(16)
vis.show_batch_autoscaled(u, name='uv')
vis.show_batch_autoscaled(q, name='qr')
print('deque len:', len(dek))
def show(threshold=.5):
from cv2tools import vis,filt
import math
bs = 16
j = np.random.choice(len(xt)-16)
minibatch = xt[j:j+bs]
code, rec, code2, rec2, noisy_x = test(minibatch,threshold)
code = np.transpose(code[0:1],axes=(3,1,2,0))
code2 = np.transpose(code2[0:1],axes=(3,1,2,0))
vis.show_batch_autoscaled(code, name='code(quant)', limit=600.)
vis.show_batch_autoscaled(code2, name='code2(no quant)', limit=600.)
vis.show_batch_autoscaled(noisy_x,name='input')
vis.show_batch_autoscaled(rec,name='recon(quant)')
vis.show_batch_autoscaled(rec2,name='recon(no quant)')
#..................
#import csv
#import numpy
#np.set_printoptions(threshold=np.inf)
#f_code = open("code.csv","w")
#w_code = csv.writer(f_code)
#w_code.writerow(code2)
#f_code = open("input.csv","w")
# w_code = csv.writer(f_code)
#w_code.writerow(noisy_x)
# f_code = open("rec2.csv","w")
# w_code = csv.writer(f_code)
# w_code.writerow(rec2)
#compute the Mean Sqare Error
mse=noisy_x-rec2
#print(mse)
sq_mse=np.power(mse,2)
#abs_mse=abs(mse)
print("******************")
#print(sq_mse)
sum_sq_error=sq_mse.sum()
#print (sum_sq_error)
#print("legth")
#print(len(noisy_x[0][0]))
print("MSE=")
mean=sum_sq_error/(32*32*16)
print(mean)
temp=1/mean
PSNR=10*math.log10(temp)
print("PSNR")
print(PSNR)
def test():
vf2t = getvf2t()
# print(vf2t.minibatch(5))
# for i in range(10):
# print(vf2t.randone())
for i in range(20):
a, b = vf2t.minibatch(20, side=128)
# cv2.imshow('yolo', b)
vis.show_batch_autoscaled(b)
cv2.waitKey(0)
def show(threshold=.5):
from cv2tools import vis,filt
bs = 16
j = np.random.choice(len(xt)-16)
minibatch = xt[j:j+bs]
code, rec, code2, rec2, noisy_x = test(minibatch,threshold)
code = np.transpose(code[0:1],axes=(3,1,2,0))
code2 = np.transpose(code2[0:1],axes=(3,1,2,0))
vis.show_batch_autoscaled(code, name='code(quant)', limit=600.)
vis.show_batch_autoscaled(code2, name='code2(no quant)', limit=600.)
vis.show_batch_autoscaled(noisy_x,name='input')
vis.show_batch_autoscaled(rec,name='recon(quant)')
vis.show_batch_autoscaled(rec2,name='recon(no quant)')
def jeffdemo():
jeff = load_image('../jeff.jpg')
h, w = jeff.shape[0:2]
jeff = vis.resize_perfect(jeff, 192, 192)
print(jeff.shape)
vis.show_autoscaled(jeff)
jefff = np.divide(jeff, 255., dtype=np.float32)
jeffff = forenet_infer(jefff)
print(jeffff.shape)
jeffff = np.transpose(jeffff, [2, 0, 1])
jeffff.shape += (1, )
vis.show_batch_autoscaled(jeffff * 0.5 + 0.5)
cv2.waitKey(0)
def show(): # evaluate result on validation set
from cv2tools import filt, vis
import cv2
# generate our own set of samples from scratch
xt, yt = needsamples(1)
index = np.random.choice(len(xt))
mbx = xt[index:index + 1]
mby = yt[index:index + 1]
gru_state = None
resarr = []
for i in range(len(mbx[0])): # timesteps
resy, state = stateful_predict(gru_state, mbx[0:1, i:i + 1])
resarr.append(resy) # [1,1,h,w,1]
gru_state = state
resarr = np.concatenate(resarr, axis=1)
print(resarr.shape)
vis.show_batch_autoscaled(mbx[0], name='input image')
vis.show_batch_autoscaled(resarr[0], name='inference')
vis.show_batch_autoscaled(mby[0], name='ground truth')
def show():
import cv2
from cv2tools import vis, filt
index = np.random.choice(len(xt))
img = xt[index:index + 1]
tiledx, y_softmaxed, shifted_means, variances = test(img)
for idxt, dist in enumerate(y_softmaxed[0]):
print('step', idxt, 'guess:', np.argmax(dist))
tiledx += 0.5
tiledx_copy = tiledx.copy()
tiledx = (tiledx * 255.).astype('uint16') # 16-bit-ify
if tiledx.shape[-1] == 1:
tiledx = np.tile(tiledx, (1, 1, 1, 1, 3)) # colorify
shifted_means += np.array([img.shape[1] - 1, img.shape[2] - 1],
dtype='float32') / 2
# shift from image center to image coordinates
# draw the circles...
for idxt, receptors in enumerate(shifted_means[0]):
tmp = tiledx[0, idxt] * 0 # [HWC]
for idxr, receptor in enumerate(receptors):
# additively paint circles
tmp += cv2.circle(np.zeros_like(tmp, dtype='uint8'),
(int(receptor[1] * 16), int(receptor[0] * 16)),
radius=int(np.sqrt(variances[idxr, 0]) * 16),
color=(80, 140, 180),
thickness=-1,
lineType=cv2.LINE_AA,
shift=4)
tiledx[0,
idxt] = tiledx[0, idxt] * 0.5 + tmp * 0.5 # blend with original
tiledx = tiledx.clip(min=0, max=255).astype('uint8')
vis.show_batch_autoscaled(tiledx_copy[0], name='input sequence')
vis.show_batch_autoscaled(tiledx[0], name='attention over time')
def show(threshold=.5):
from cv2tools import vis,filt
import math
bs = 16
j = np.random.choice(len(xt)-16)
minibatch = xt[j:j+bs]
code, rec, code2, rec2, noisy_x = test(minibatch,threshold)
code = np.transpose(code[0:1],axes=(3,1,2,0))
code2 = np.transpose(code2[0:1],axes=(3,1,2,0))
vis.show_batch_autoscaled(code, name='code(quant)', limit=600.)
vis.show_batch_autoscaled(code2, name='code2(no quant)', limit=600.)
vis.show_batch_autoscaled(noisy_x,name='input')
vis.show_batch_autoscaled(rec,name='recon(quant)')
vis.show_batch_autoscaled(rec2,name='recon(no quant)')
mse=noisy_x-rec2
#print(mse)
sq_mse=np.power(mse,2)
#abs_mse=abs(mse)
print("******************")
#print(sq_mse)
sum_sq_error=sq_mse.sum()
#print (sum_sq_error)
#print("legth")
#print(len(noisy_x[0][0]))
print("MSE=")
mean=sum_sq_error/(32*32*16)
print(mean)
temp=1/mean
PSNR=10*math.log10(temp)
print("PSNR")
print(PSNR)
def t():
uv, qr = pour(10)
y, sigmred = test(qr)
y[:, :, :, 2:3] = sigmred
vis.show_batch_autoscaled(y, name='output')
vis.show_batch_autoscaled(qr, name='input')
vis.show_batch_autoscaled(uv, name='gt')
def show(): # evaluate result on validation set
from drone_samples_queue import needsamples
from cv2tools import filt, vis
import cv2
# generate our own set of samples from scratch
mbx, mby = needsamples(10)
mbx.shape = (10, ) + mbx.shape[2:]
mby.shape = (10, ) + mby.shape[2:]
res = predict(mbx)
print(res.shape)
vis.show_batch_autoscaled(mbx, name='input image')
vis.show_batch_autoscaled(res, name='inference')
vis.show_batch_autoscaled(mby, name='ground truth')