def main(_):
FLAGS = tf.compat.v1.app.flags.FLAGS.flag_values_dict()
FLAGS = preprocess_flags(FLAGS)
globals().update(FLAGS)
print(partial_kernel_n_proc)
number_completed = 0
files = find_partial_kernel_filenames(FLAGS)
for f in files:
cnt = int(f.split("_")[-2])
if cnt > partial_kernel_n_proc:
number_completed = cnt - partial_kernel_n_proc
if f == kernel_filename(FLAGS):
number_completed = partial_kernel_n_proc
#break
while number_completed < partial_kernel_n_proc:
files = find_partial_kernel_filenames(FLAGS)
print(files)
if len(files) > 1:
for i, f in enumerate(files):
print(f)
if i == 0:
cov = load_kernel_by_filename(f)
if number_completed == 0:
number_completed += 1
else:
cov += load_kernel_by_filename(f)
number_completed += 1
if number_completed >= partial_kernel_n_proc:
save_kernel(cov / number_completed, FLAGS)
else:
save_kernel_partial(cov, FLAGS,
partial_kernel_n_proc + number_completed)
for f in files:
os.remove(f)
sleep(100)
Xfull = np.concatenate([X, X_test])
ys2 = [[y] for y in ys]
ysfull = ys2 + [[y] for y in test_ys]
Yfull = np.array(ysfull)
Y = np.array(ys2)
#%%
# filename = net+"_KMNIST_1000_0.0_0.0_True_False_True_4_3.0_0.0_None_0000_max_kernel.npy"
filename = "newer_arch_sweep_ce_sgd__" + str(
net
) + "_EMNIST_1000_0.0_0.0_True_False_True_4_1.414_0.0_None_0000_max_kernel.npy"
from utils import load_kernel_by_filename
Kfull = load_kernel_by_filename("kernels/" + filename)
# m = 1000
# Kfull.max()
# K = Kfull/Kfull.max()
# K *= 1000
eigs = np.linalg.eigh(K)[0]
eigs
m = 1000
epsilon = 1
sigma = 1e-1
for i in range(1000):
epsilon = np.sum(eigs / (1 + eigs * m / (sigma**2 + epsilon)))
ys2 = [[y] for y in ys]
ysfull = ys2 + [[y] for y in test_ys]
Yfull = np.array(ysfull)
Y = np.array(ys2)
#%%
# del K,theta
# filename = net+"_KMNIST_1000_0.0_0.0_True_False_True_4_3.0_0.0_None_0000_max_kernel.npy"
filename = prefix+"_"+str(net)+"_EMNIST_1000_0.0_0.0_True_False_True_4_1.414_0.0_None_0000_max_kernel.npy"
# filename = "newer_arch_sweep_ce_sgd__"+str(net)+"_EMNIST_1000_0.0_0.0_True_False_True_4_1.414_0.0_None_0000_max_NTK_kernel.npy"
##NNGP kernel
from utils import load_kernel_by_filename
try:
theta = load_kernel_by_filename("kernels/"+filename)
except FileNotFoundError:
print("File not found :P")
continue
##NTK kernel
# filename = "newer_arch_sweep_ce_sgd__"+str(net)+"_EMNIST_1000_0.0_0.0_True_False_True_4_1.414_0.0_None_0000_max_NTK_kernel.npy"
# from utils import load_kernel_by_filename
# try:
# Kfull = load_kernel_by_filename("kernels/"+filename)
# except FileNotFoundError:
# print("File not found :P")
# continue
# K = Kfull
# K = 1*Kfull
from scipy.special import comb
# comb(3,2)
def num_per_degree(S):
return comb(n,S)
Ss = range(n)
reduced_eigs = np.array(list(map(eig_for_degree, Ss)))
num_eigs = np.array(list(map(num_per_degree, Ss)))
# m=1000
from utils import load_kernel_by_filename
K = load_kernel_by_filename("kernels/mnist_msweep_kernels__fc_mnist_"+str(8192)+"_0.0_0.0_True_False_True_2_1.41_0.1_None_00_max_kernel.npy")
eigs = np.linalg.eigh(K)[0]
n=13
2**13
ms = list(map(lambda x: 2**x, range(n)))
epsilons = []
for m in ms:
epsilon = 1
sigma=1e-1
for i in range(1000):
epsilon = np.sum(eigs/(1+eigs*m/(sigma**2+epsilon)))
# epsilon = np.sum(num_eigs*reduced_eigs/(1+reduced_eigs*m/(sigma**2+epsilon)))
epsilons.append(epsilon)
import matplotlib.pyplot as plt