def individual_param_quadratic(self, test_loader, layer, idxs):
"""
Method interpolates individual parameter of the model and evaluates the performance of the model when the
interpolated parameter replaces its original in the parameters of the model
:param test_loader: test dataset loader
:param layer: layer of parameter
:param idxs: position of parameter
"""
loss_res = Path("{}_{}_{}_q".format(paths.svloss_path, layer,
convert_list2str(idxs)))
loss_img = Path("{}_{}_{}_q".format(paths.svloss_img_path, layer,
convert_list2str(idxs)))
acc_res = Path("{}_{}_{}_q".format(paths.sacc_path, layer,
convert_list2str(idxs)))
acc_img = Path("{}_{}_{}_q".format(paths.sacc_img_path, layer,
convert_list2str(idxs)))
logger.debug(f"Result files:\n" f"{loss_res}\n" f"{acc_res}\n")
logger.debug(f"Img files:\n" f"{loss_img}\n" f"{acc_img}\n")
if not loss_res.exists() or not acc_res.exists():
logger.debug(
"Files with results not found - beginning interpolation.")
v_loss_list = []
acc_list = []
start_a = 0
mid_a = 0.5
end_a = 1
logger.debug(f"Start: {start_a}\n"
f"Mid: {mid_a}\n"
f"End: {end_a}")
mid_check = self.__get_mid_point(paths.checkpoints)
start_p = self.theta_i[layer + ".weight"][idxs].cpu()
mid_p = copy.deepcopy(
torch.load(Path(
os.path.join(paths.checkpoints,
mid_check))))[layer + ".weight"][idxs].cpu()
end_p = self.theta_f[layer + ".weight"][idxs].cpu()
logger.debug(f"Start loss: {start_p}\n"
f"Mid loss: {mid_p}\n"
f"End loss: {end_p}")
start = [start_a, start_p]
mid = [mid_a, mid_p]
end = [end_a, end_p]
logger.debug(f"Start: {start}\n" f"Mid: {mid}\n" f"End: {end}")
self.model.load_state_dict(self.theta_f)
for alpha_act in tqdm(
self.alpha,
desc=f"Parameter {layer}/{idxs} Level Quadratic",
dynamic_ncols=True):
self.__calc_theta_single_q(layer + ".weight", idxs, alpha_act,
start, mid, end)
self.model.load_state_dict(self.theta)
logger.debug(
f"Getting validation loss and accuracy for alpha = {alpha_act}"
)
val_loss, acc = net.test(self.model, test_loader, self.device)
acc_list.append(acc)
v_loss_list.append(val_loss)
logger.debug(f"Saving results to files ({loss_res}, {acc_res})")
np.savetxt(loss_res, v_loss_list)
np.savetxt(acc_res, acc_list)
self.model.load_state_dict(self.theta_f)
logger.debug(f"Saving results to figures {loss_img}, {acc_img} ...")
plot.plot_metric(self.alpha, np.loadtxt(loss_res), loss_img, "loss")
plot.plot_metric(self.alpha, np.loadtxt(acc_res), acc_img, "acc")
self.model.load_state_dict(self.theta_f)
return
def layers_quadratic(self, test_loader, layer):
"""
Method examines the parameters on the level of layers using the quadratic interpolation.
:param test_loader: test data set loader
:param layer: layer to be examined
"""
loss_res = Path("{}_{}_q".format(paths.vvloss_path, layer))
loss_img = Path("{}_{}_q".format(paths.vvloss_img_path, layer))
acc_res = Path("{}_{}_q".format(paths.vacc_path, layer))
acc_img = Path("{}_{}_q".format(paths.vacc_img_path, layer))
logger.debug(f"Result files:\n" f"{loss_res}\n" f"{acc_res}")
logger.debug(f"Img files:\n" f"{loss_img}\n" f"{acc_img}")
if not loss_res.exists() or not acc_res.exists():
logger.debug("Result files not found - beginning interpolation.")
v_loss_list = []
acc_list = []
start_a = 0
mid_a = 0.5
end_a = 1
logger.debug(f"Start: {start_a}\n"
f"Mid: {mid_a}\n"
f"End: {end_a}")
mid_check = self.__get_mid_point(paths.checkpoints)
start_p = self.theta_i[layer + ".weight"].cpu()
mid_p = copy.deepcopy(
torch.load(os.path.join(paths.checkpoints,
mid_check))[layer + ".weight"]).cpu()
end_p = self.theta_f[layer + ".weight"].cpu()
start_w = [start_a, start_p]
mid_w = [mid_a, mid_p]
end_w = [end_a, end_p]
start_pb = self.theta_i[layer + ".bias"].cpu()
mid_pb = copy.deepcopy(
torch.load(os.path.join(
paths.checkpoints,
mid_check))[layer + ".bias"]).cpu() # TODO AUTO MID
end_pb = self.theta_f[layer + ".bias"].cpu()
start_b = [start_a, start_pb]
mid_b = [mid_a, mid_pb]
end_b = [end_a, end_pb]
for alpha_act in tqdm(self.alpha,
desc=f"Layer {layer} Level Quadratic Path",
dynamic_ncols=True):
self.__calc_theta_vec_q(layer + ".weight", alpha_act, start_w,
mid_w, end_w)
self.__calc_theta_vec_q(layer + ".bias", alpha_act, start_b,
mid_b, end_b)
self.model.load_state_dict(self.theta)
logger.debug(
f"Getting validation loss and accuracy for alpha = {alpha_act}"
)
vloss, acc = net.test(self.model, test_loader, self.device)
v_loss_list.append(vloss)
acc_list.append(acc)
logger.debug(f"Saving results to files ({loss_res}, {acc_res})")
np.savetxt(loss_res, v_loss_list)
np.savetxt(acc_res, acc_list)
logger.debug(f"Saving results to figures {loss_img}, {acc_img} ...")
plot.plot_metric(self.alpha, np.loadtxt(loss_res), loss_img, "loss")
plot.plot_metric(self.alpha, np.loadtxt(acc_res), acc_img, "acc")
self.model.load_state_dict(self.theta_f)
return
def layers_linear(self, test_loader, layer):
"""
Method interpolates parameters of selected layer of the model and evaluates the model after each interpolation
step
:param test_loader: test loader
:param layer: layer to be interpolated
"""
loss_res = Path("{}_{}".format(paths.vvloss_path, layer))
loss_img = Path("{}_{}".format(paths.vvloss_img_path, layer))
acc_res = Path("{}_{}".format(paths.vacc_path, layer))
acc_img = Path("{}_{}".format(paths.vacc_img_path, layer))
dist = Path("{}_{}_{}".format(paths.vvloss_path, layer, "distance"))
logger.debug(f"Result files:\n" f"{loss_res}\n" f"{acc_res}")
logger.debug(f"Img files:\n" f"{loss_img}\n" f"{acc_img}")
logger.debug(f"Dist file:\n" f"{dist}")
if not loss_res.exists() or not acc_res.exists():
logger.debug("Result files not found - beginning interpolation.")
v_loss_list = []
acc_list = []
self.model.load_state_dict(self.theta_f)
for alpha_act in tqdm(self.alpha,
desc=f"Layer {layer} Level Linear",
dynamic_ncols=True):
self.__calc_theta_vec(layer + ".weight", alpha_act)
self.__calc_theta_vec(layer + ".bias", alpha_act)
self.model.load_state_dict(self.theta)
logger.debug(
f"Getting validation loss and accuracy for alpha = {alpha_act}"
)
vloss, acc = net.test(self.model, test_loader, self.device)
v_loss_list.append(vloss)
acc_list.append(acc)
logger.debug(f"Saving results to files ({loss_res}, {acc_res})")
np.savetxt(loss_res, v_loss_list)
np.savetxt(acc_res, acc_list)
if not dist.exists():
logger.info(f"Calculating distance for: {layer}")
distance = self.calc_distance(layer + ".weight")
logger.info(f"Distance: {distance}")
with open(dist, 'w') as fd:
fd.write("{}".format(distance))
logger.debug(f"Saving results to figures {loss_img}, {acc_img} ...")
plot.plot_metric(self.alpha, np.loadtxt(loss_res), loss_img, "loss")
plot.plot_metric(self.alpha, np.loadtxt(acc_res), acc_img, "acc")
self.model.load_state_dict(self.theta_f)
return