def diff(clean, error):
d1 = read_data(clean)
d2 = read_data(error)
mse = (np.square(d1 - d2)).mean(axis=None)
diff_count = np.sum(np.abs(d1 - d2) > 10e-5) + np.sum(np.isnan(d2))
diff_max = np.max(np.abs(d1 - d2))
diff_rel_max = np.max(np.abs((d1 - d2) / d1))
return mse, diff_count, diff_max, diff_rel_max
def diff(clean, error):
d1 = read_data(clean)
d2 = read_data(error)
bit, x, y, z = error.split("_")[-5:-1]
bit, x, y, z = int(bit), int(x), int(y), int(z)
print("bit: %s, x: %s, y: %s, z: %s" %(bit, x, y ,z))
print("clean: %s, error: %s" %(d1[x,y,z], d2[x,y,z]))
sum_error = np.sum(np.abs(d1 - d2))
diff_count = np.sum(d1 != d2)
diff_max = np.max(np.abs(d1 - d2))
return sum_error, diff_count, diff_max
def restart(checkpoint_file, postfix=0, var_name="dens"):
timestep = int(checkpoint_file[-4:])
start_checkpoint_file = checkpoint_file
os.system("cp ./clean/" + checkpoint_file + " ./" + start_checkpoint_file)
end_checkpoint_file = start_checkpoint_file[0:-4] + (
"0000" + str(timestep + 1))[-4:]
print(start_checkpoint_file, end_checkpoint_file)
# 1. insert an error into the checkpoint file
# remeber the error bit and error position (x,y,z)
f = h5py.File(start_checkpoint_file, "r+")
x, y, z = get_random_indices()
error, bit = get_flip_error(f[var_name][0, z, y, x], 0, 30)
print(bit, x, y, z, f[var_name][0, z, y, x], error)
f[var_name][0, z, y, x] = error
f.close()
# 2. read flash.par and change it so we could restart from the corrupted checkpoint
# need to specify restart from which iteration
# and also the end iteration
# sed can do this easily
os.system("cp ./flash.par ./flash.par_" + postfix)
os.system("sed -i \'s/.*checkpointFileNumber.*/checkpointFileNumber = " +
str(timestep) + "/g\' ./flash.par_" + postfix)
#os.system("sed -i \'s/.*basenm.*/basenm = " + postfix+"_sod_" + "/g\' ./flash.par_"+postfix)
#os.system("sed -i \'s/.*nend.*/nend = " +str(timestep+delay) + "/g\' ./flash.par_"+postfix)
# 3. restart the flash program
os.system("mpirun -np 4 ./flash4 -par_file ./flash.par_" + postfix)
# 4. read the corrupted checkpoint file
new_start_name = "error_%s_%s_%s_%s_%s_start" % (bit, x, y, z, timestep)
new_end_name = "error_%s_%s_%s_%s_%s_end" % (bit, x, y, z, timestep)
np.save(new_start_name + ".npy", read_data(start_checkpoint_file,
var_name))
np.save(new_end_name + ".npy", read_data(end_checkpoint_file, var_name))
os.system("rm " + start_checkpoint_file)
os.system("rm " + end_checkpoint_file)
# 5. delete unnecessary files
os.system("rm *.dat *.log *_plt_cnt_*")
elif args.test:
data_gen = FlashDatasetGenerator(args.data, BATCH_SIZE, args.k == 0,
args.bits)
model.load_weights(model_file)
scores = model.predict_generator(generator=data_gen,
use_multiprocessing=False,
workers=1,
verbose=1)
for threshold in [
0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95, 0.99
]:
print("Threshold =", threshold)
pred = (scores >= threshold)
compute_metrics(pred, data_gen.labels)
elif args.detect:
model.load_weights(model_file)
error = 0
# match both for clean checkpoint files and k-delay corrupted files
files = glob.glob(args.detect + "/*chk_*") + glob.glob(args.detect +
"/*error*")
files.sort()
files = files[0:200]
for filename in files:
dens = read_data(filename)
dens_blocks = np.expand_dims(np.squeeze(split_to_windows(dens)),
-1)
if detection(model, dens_blocks):
error += 1
print("detected %s error samples, total: %s, recall: %s" %
(error, len(files), error * 1.0 / len(files)))