def main():
Model = 'four_flux'
R_km = 0
R_muf = 0
# load data
RealCompose_2 = np.load('RealCompose_2.npy')
RealIngredient_2 = np.load('RealIngredint_2.npy')
Composes = np.append(RealCompose, RealCompose_2, axis=0)
Ingredients = np.append(RealIngredient, RealIngredient_2, axis=0)
data_c = np.load('data_c.npy')
data_p = np.load('data_p.npy').T
#get K/S of base
base_f_km = mf.math_model(data_p[0] - R_km, 'km')
base_f_muf = mf.math_model(data_p[0] - R_muf, 'four_flux')
#get F
F_km = get_F(data_p[1:] - R_km, data_c, base_f_km, 'km')
F_muf = get_F(data_p[1:] - R_muf, data_c, base_f_muf, 'four_flux')
diffs_km = []
diffs_muf = []
for i in range(Composes.shape[0]):
PredictIngredint_km = Mix(F_km, base_f_km, Composes[i], 'km') + R_km
PredictIngredint_muf = Mix(F_muf, base_f_muf, Composes[i],
'four_flux') + R_muf
dif = mf.color_diff(PredictIngredint_km, Ingredients[i])
if dif < 6:
diffs_km.append(dif)
dif = mf.color_diff(PredictIngredint_muf, Ingredients[i])
if dif < 6:
diffs_muf.append(dif)
plt.figure()
X = np.arange(len(diffs_km))
plt.scatter(X, diffs_km, label='KM')
plt.plot(X, np.repeat(sum(diffs_km) / len(diffs_km), len(diffs_km)))
plt.text(1, sum(diffs_km) / len(diffs_km), sum(diffs_km) / len(diffs_km))
X = np.arange(len(diffs_muf))
plt.scatter(X, diffs_muf, label='four_flux')
plt.plot(X, np.repeat(sum(diffs_muf) / len(diffs_muf), len(diffs_muf)))
plt.text(1,
sum(diffs_muf) / len(diffs_muf),
sum(diffs_muf) / len(diffs_muf))
plt.xlabel('加料方案', fontproperties=font_set)
plt.ylabel('色差', fontproperties=font_set)
plt.legend()
print(matplotlib.get_backend())
plt.show()
def Mix(compose, data_c, data_p):
base_f = mf.math_model(data_p[0], 'four_flux')
F = np.zeros_like(base_f)
for i in range(compose.size):
df = (mf.math_model(data_p[i * 3 + 2], 'four_flux') -
base_f) / data_c[i][1]
F += df * compose[i]
return mf.i_math_model(F + base_f, 'four_flux')
def get_dfs_KM(w, data_c, data_p):
base_f = mf.math_model(data_p[0], 'four_flux')
dfs = []
for i in range(data_c.shape[0]):
df = (mf.math_model(data_p[i * 3 + 1], 'four_flux') - base_f -
correct_func(data_c[i][0], w)) / data_c[i][0]
df += (mf.math_model(data_p[i * 3 + 2], 'four_flux') - base_f -
correct_func(data_c[i][1], w)) / data_c[i][1]
df += (mf.math_model(data_p[i * 3 + 3], 'four_flux') - base_f -
correct_func(data_c[i][2], w)) / data_c[i][2]
dfs.append(df / 3)
return np.array(dfs)
def plot_ks(id, ax):
ax.set_title('id=%d' % id)
ks = mf.math_model(data_p[id * 3 + 1:id * 3 + 4],
"four_flux") - mf.math_model(data_p[0], "four_flux")
c = data_c[id][:, np.newaxis]
ks = ks / c
c = np.repeat(c, 31, axis=1).reshape(93)
#ks = ks.reshape(93)
ax.scatter3D(c, Y, ks)
ax.set_xlabel('C')
ax.set_ylabel('wavelength')
ax.set_zlabel('k/s per unit')
def get_total_loss(K1, K2):
R = (data_p - K1) / (1 - K1 - K2 + K2 * data_p)
F = mf.math_model(R, model)
ans = 0
for i in range(31):
for j in range(21):
ans += get_loss(F[i][3 * j + 1:3 * j + 4] - F[i][0], data_c[j])
return ans
def get_F(p,c,base,Model):
ff = mf.math_model(p,Model) - base
F = []
sample_size = c.shape[1]
for i in range(c.shape[0]):
F.append(ff[sample_size*i + int(sample_size/2)] / c[i][int(sample_size/2)])
return np.array(F)
def get_F(p, c, base, Model):
ff = mf.math_model(p, Model)
F = []
sample_size = c.shape[1]
for i in range(c.shape[0]):
k = 0.0
for j in range(sample_size):
k += ff[sample_size * i + j] / c[i][j]
F.append(k / sample_size)
return np.array(F)
def get_diffs(model):
base = mf.math_model(R[0], model)
F = get_F(R[1:], data_c, base, model)
diffs = []
for i in range(Composes.shape[0]):
PredictIngredint = Mix(F, base, Composes[i], model)
dif = mf.color_diff(i_Sanderson(PredictIngredint, k1, k2),
Ingredients[i])
if dif < 50:
dif = (i_Sanderson(PredictIngredint, k1, k2) -
Ingredients[i]) * diff_weight
diffs.append(dif)
else:
print(i)
return np.array(diffs)
def main():
# data about km
data_c = np.load('data_c.npy')
data_p = np.load('data_p.npy').T
# get K/S of base
base_f_km = mf.math_model(data_p[0], 'km')
F_km = get_F(data_p[1:], data_c, base_f_km, 'km')
floor = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 1000]
for x in range(10):
test_com = random_com(1)[0]
test_ref = Mix(F_km, base_f_km, test_com, 'km')
count = np.zeros(10).astype(int)
coms = random_com(65536)
for c in coms:
r = Mix(F_km, base_f_km, c, 'km')
dif = mf.color_diff(r, test_ref)
for i in range(10):
if floor[i] < dif and dif < floor[i + 1]:
count[i] += 1
print(count)
def PredictCompose(F, base_f, targetIngredint, Model):
target_f = mf.math_model(targetIngredint, Model) - base_f
return target_f.dot(F.T).dot(np.linalg.inv(F.dot(F.T)))
def get_F(p, c, base, Model):
ff = mf.math_model(p, Model) - base
F = []
sample_size = c.shape[1]
for i in range(c.shape[0]):
F.append(ff[sample_size * i + int(sample_size / 2)] /
c[i][int(sample_size / 2)])
return np.array(F)
data_c = np.load('data_c.npy')
data_p = np.load('data_p.npy').T
#get K/S of base
base_f_km = mf.math_model(data_p[0], 'km')
#get F
F_km = get_F(data_p[1:], data_c, base_f_km, 'km')
com0 = np.zeros(21).astype(float)
Noise = [0.00001, 0.0001, 0.0005, 0.001, 0.005, 0.01, 0.02, 0.03]
size = 50000
for noise in Noise:
diff = 0
for i in range(size):
R0 = np.random.rand(31)
diff += mf.color_diff(R0, R0 + noise * np.random.normal(0, 1, 31))
print('mean_diff_with_noise_', noise, ':', diff / size)
def main():
R0 = 0.048
fil = 3
scale = [0.515, 0.515, 0.52, 0.56]
w_names = [
'data_w_size6.npy',
'data_w_size12.npy',
'data_w_size18.npy',
'data_w_size21_v3_no_r0.npy',
]
# data_c 浓度数据，21种色浆 * 3次取点
# data_p 分光反射率数据 size: (1 + 21 * 3) * 31 , 1为基底
data_c = np.load('data_c.npy')
data_p = np.load('data_p.npy')
c, p = mf.data_filiter(filiters[fil], data_c, data_p)
p = p.T - R0
# 计算基底K/S
base_f = mf.math_model(p[0], 'four_flux')
# 获取w,计算单位k/s值
w = np.load(w_names[fil]).T
dfs = get_dfs_KM(w, c, p)
RealCompose_2 = np.load('RealCompose_2.npy')
RealIngredient_2 = np.load('RealIngredint_2.npy')
Composes = np.append(RealCompose, RealCompose_2, axis=0)
Ingredients = np.append(RealIngredient, RealIngredient_2, axis=0)
print('\n using : ', w_names[fil])
Y1 = []
Y2 = []
Y3 = []
#X_ = np.linspace(0,0.1,100)
max = 6
for com in range(Composes.shape[0]):
compose = mf.data_filiter_(filiters[fil], Composes[com])
if Counter(compose)[0] == 18:
y = mf.color_diff(Ingredients[com], Mix(compose, c, p + R0))
if y < max:
Y1.append(y)
y = mf.color_diff(
Ingredients[com],
corrected_Mix(np.array([compose]), w, dfs, base_f, scale[fil])
+ R0)
if y < max:
Y2.append(y)
y = mf.color_diff(Ingredients[com], Mix(compose, c, p) + R0)
if y < max:
Y3.append(y)
plt.figure()
X = np.arange(len(Y1))
plt.scatter(X, Y1, label='four_flux')
plt.plot(X, np.repeat(sum(Y1) / len(Y1), len(Y1)))
plt.text(1, sum(Y1) / len(Y1), sum(Y1) / len(Y1))
plt.scatter(X, Y3, marker='*', label='corrected_four_flux_1.0')
plt.plot(X, np.repeat(sum(Y3) / len(Y3), len(Y3)))
plt.text(1, sum(Y3) / len(Y3), sum(Y3) / len(Y3))
plt.scatter(X, Y2, marker='x', label='corrected_four_flux_2.0')
plt.plot(X, np.repeat(sum(Y2) / len(Y2), len(Y2)))
plt.text(1, sum(Y2) / len(Y2), sum(Y2) / len(Y2))
plt.xlabel('加料方案', fontproperties=font_set)
plt.ylabel('色差', fontproperties=font_set)
# plt.imshow(Y)
# ignore ticks
plt.xticks([])
plt.yticks([])
plt.legend()
plt.show()
def main():
# data about km
data_c = np.load('data_c.npy')
data_p = np.load('data_p.npy').T
# get K/S of base
base_f_km = mf.math_model(data_p[0], 'km')
F_km = get_F(data_p[1:], data_c, base_f_km, 'km')
#
# shu = np.zeros(base_color_num)
# shu[0] = 1
# shu[1] = 1
# shu[2] = 1
# np.random.shuffle(shu)
# random_com = np.random.rand(base_color_num) * shu
# test_ref = Mix(F_km, base_f_km, random_com, 'km')
data = np.load('data/data_01.npz')
concentrations = torch.from_numpy(data['concentrations']).float()
reflectance = torch.from_numpy(data['reflectance']).float()
random_com = concentrations[42].cpu().data.numpy()
test_ref = reflectance[42].cpu().data.numpy()
# data about inn
inn = torch.load('model_dir/model_02')
y_noise_scale = 1e-4
N_sample = 20000
N_class = 100
dim_x = base_color_num
dim_y = reflectance_dim
dim_z = 13
dim_total = max(dim_x, dim_y + dim_z)
sample,densi = generate_test_sample(test_ref, N_sample, y_noise_scale, dim_x, dim_y, dim_z, dim_total)
p_c = predict(0,0,sample,inn)
min = 100
index = -1
floor = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 1000]
count = np.zeros(10).astype(int)
for i in range(N_sample):
df = mf.color_diff(Mix(F_km,base_f_km,p_c[i],'km'),test_ref)
for i in range(10):
if floor[i] < df and df < floor[i + 1]:
count[i] += 1
if min>df:
min = df
index = i
print(min)
print(count)
N_sample = 10000
p_c = p_c[:N_sample][:]
# p_c = p_c*densi / densi.max()
kmeans_model = KMeans(N_class).fit(p_c)
labels = kmeans_model.labels_
# using center
for c in kmeans_model.cluster_centers_:
print('center_color_diff:',mf.color_diff(Mix(F_km,base_f_km,c,'km'),test_ref))
p_c = np.concatenate((p_c, [random_com],kmeans_model.cluster_centers_))
tsne = TSNE(n_components=2)
tsne.fit_transform(p_c)
res = np.array(tsne.embedding_)
d = []
for j in range(N_class):
l = []
for i in range(N_sample):
if j == labels[i]:
l.append(res[i])
d.append(np.array(l).T)
fig = plt.figure()
# ax = Axes3D(fig)
for j in range(N_class):
plt.scatter(d[j][0],d[j][1],s=1)
plt.scatter(res[index][0],res[index][1],marker='o',c='r',label ='min')
plt.scatter(res[N_sample][0], res[N_sample][1], marker='v', c='b',label = 'real')
# plot center
for i in range(N_class):
plt.scatter(res[N_sample+1 + i][0],res[N_sample+1 + i][1],marker='s',c='g')
plt.legend()
plt.show()