Source code for optim.ad_optim_lbfgs_mod

import copy
import warnings
import time
import json
import logging
log = logging.getLogger(__name__)
import torch
from optim import lbfgs_modified
import config as cfg

[docs]def store_checkpoint(checkpoint_file, state, optimizer, current_epoch, current_loss,\ verbosity=0): r""" :param checkpoint_file: target file :param state: ipeps wavefunction :param optimizer: Optimizer :param current_epoch: current epoch :param current_loss: current value of a loss function :param verbosity: verbosity :type checkpoint_file: str or Path :type state: IPEPS :type optimizer: torch.optim.Optimizer :type current_epoch: int :type current_loss: float :type verbosity: int Store the current state of the optimization in ``checkpoint_file``. """ torch.save({ 'epoch': current_epoch, 'loss': current_loss, 'parameters': state.get_checkpoint(), 'optimizer_state_dict': optimizer.state_dict()}, checkpoint_file) if verbosity>0: print(checkpoint_file)
[docs]def optimize_state(state, ctm_env_init, loss_fn, obs_fn=None, post_proc=None, main_args=cfg.main_args, opt_args=cfg.opt_args,ctm_args=cfg.ctm_args, global_args=cfg.global_args): r""" :param state: initial wavefunction :param ctm_env_init: initial environment corresponding to ``state`` :param loss_fn: loss function :param model: model with definition of observables :param local_args: parsed command line arguments :param opt_args: optimization configuration :param ctm_args: CTM algorithm configuration :param global_args: global configuration :type state: IPEPS :type ctm_env_init: ENV :type loss_fn: function(IPEPS,ENV,CTMARGS,OPTARGS,GLOBALARGS)->torch.tensor :type model: TODO Model base class :type local_args: argparse.Namespace :type opt_args: OPTARGS :type ctm_args: CTMARGS :type global_args: GLOBALARGS Optimizes initial wavefunction ``state`` with respect to ``loss_fn`` using :class:`optim.lbfgs_modified.LBFGS_MOD` optimizer. The main parameters influencing the optimization process are given in :class:`config.OPTARGS`. Calls to functions ``loss_fn``, ``obs_fn``, and ``post_proc`` pass the current configuration as dictionary ``{"ctm_args":ctm_args, "opt_args":opt_args}`` The optimizer saves the best energy state into file ``main_args.out_prefix+"_state.json"`` and checkpoints the optimization at every step to ``main_args.out_prefix+"_state.json"``. """ verbosity = opt_args.verbosity_opt_epoch checkpoint_file = main_args.out_prefix+"_checkpoint.p" outputstatefile= main_args.out_prefix+"_state.json" t_data = dict({"loss": [], "min_loss": 1.0e+16, "loss_ls": [], "min_loss_ls": 1.0e+16}) current_env= [ctm_env_init] context= dict({"ctm_args":ctm_args, "opt_args":opt_args, "loss_history": t_data}) epoch=0 parameters= state.get_parameters() for A in parameters: A.requires_grad_(True) optimizer = lbfgs_modified.LBFGS_MOD(parameters, max_iter=opt_args.max_iter_per_epoch, lr=opt_args.lr, \ tolerance_grad=opt_args.tolerance_grad, tolerance_change=opt_args.tolerance_change, \ history_size=opt_args.history_size, line_search_fn=opt_args.line_search, \ line_search_eps=opt_args.line_search_tol) # load and/or modify optimizer state from checkpoint if main_args.opt_resume is not None: print(f"INFO: resuming from check point. resume = {main_args.opt_resume}") if not str(global_args.device)==str(state.device): warnings.warn(f"Device mismatch: state.device {state.device}"\ +f" global_args.device {global_args.device}",RuntimeWarning) checkpoint = torch.load(main_args.opt_resume,map_location=state.device) epoch0 = checkpoint["epoch"] loss0 = checkpoint["loss"] cp_state_dict= checkpoint["optimizer_state_dict"] cp_opt_params= cp_state_dict["param_groups"][0] cp_opt_history= cp_state_dict["state"][cp_opt_params["params"][0]] if main_args.opt_resume_override_params: cp_opt_params["lr"] = opt_args.lr cp_opt_params["max_iter"] = opt_args.max_iter_per_epoch cp_opt_params["tolerance_grad"] = opt_args.tolerance_grad cp_opt_params["tolerance_change"] = opt_args.tolerance_change cp_opt_params["line_search_fn"] = opt_args.line_search cp_opt_params["line_search_eps"] = opt_args.line_search_tol # resize stored old_dirs, old_stps, ro, al to new history size cp_history_size= cp_opt_params["history_size"] cp_opt_params["history_size"] = opt_args.history_size if opt_args.history_size < cp_history_size: if len(cp_opt_history["old_dirs"]) > opt_args.history_size: assert len(cp_opt_history["old_dirs"])==len(cp_opt_history["old_stps"])\ ==len(cp_opt_history["ro"]), "Inconsistent L-BFGS history" cp_opt_history["old_dirs"]= cp_opt_history["old_dirs"][-opt_args.history_size:] cp_opt_history["old_stps"]= cp_opt_history["old_stps"][-opt_args.history_size:] cp_opt_history["ro"]= cp_opt_history["ro"][-opt_args.history_size:] cp_al_filtered= list(filter(None,cp_opt_history["al"])) if len(cp_al_filtered) > opt_args.history_size: cp_opt_history["al"]= cp_al_filtered[-opt_args.history_size:] else: cp_opt_history["al"]= cp_al_filtered + [None for i in range(opt_args.history_size-len(cp_al_filtered))] cp_state_dict["param_groups"][0]= cp_opt_params cp_state_dict["state"][cp_opt_params["params"][0]]= cp_opt_history optimizer.load_state_dict(cp_state_dict) print(f"checkpoint.loss = {loss0}") #@profile def closure(linesearching=False): context["line_search"]= linesearching # 0) evaluate loss optimizer.zero_grad() loss, ctm_env, history, t_ctm, t_check = loss_fn(state, current_env[0], context) # 4) evaluate gradient t_grad0= time.perf_counter() loss.backward() t_grad1= time.perf_counter() # 6) detach current environment from autograd graph ctm_env.detach_() current_env[0]= ctm_env # current_env[0]= ctm_env.detach().clone() # 1) record loss and store current state if the loss improves if linesearching: t_data["loss_ls"].append(loss.item()) if t_data["min_loss_ls"] > t_data["loss_ls"][-1]: t_data["min_loss_ls"]= t_data["loss_ls"][-1] else: t_data["loss"].append(loss.item()) if t_data["min_loss"] > t_data["loss"][-1]: t_data["min_loss"]= t_data["loss"][-1] state.write_to_file(outputstatefile, normalize=True) # 2) log CTM metrics for debugging if opt_args.opt_logging: log_entry=dict({"id": epoch, "loss": t_data["loss"][-1], "t_ctm": t_ctm, \ "t_check": t_check}) if linesearching: log_entry["LS"]=len(t_data["loss_ls"]) log_entry["loss"]=t_data["loss_ls"] log.info(json.dumps(log_entry)) # 3) compute desired observables if obs_fn is not None: obs_fn(state, current_env[0], context) # 5) log grad metrics if opt_args.opt_logging: log_entry=dict({"id": epoch}) if linesearching: log_entry["LS"]=len(t_data["loss_ls"]) else: log_entry["t_grad"]=t_grad1-t_grad0 # log just l2 and l\infty norm of the full grad # log_entry["grad_mag"]= [p.grad.norm().item() for p in parameters] flat_grad= torch.cat(tuple(p.grad.view(-1) for p in parameters)) log_entry["grad_mag"]= [flat_grad.norm().item(), flat_grad.norm(p=float('inf')).item()] log_entry["grad_mags"]= [p.grad.norm().item() for p in parameters] if opt_args.opt_log_grad: log_entry["grad"]= [p.grad.tolist() for p in parameters] log.info(json.dumps(log_entry)) return loss # closure for derivative-free line search. This closure # is to be called within torch.no_grad context @torch.no_grad() def closure_linesearch(linesearching): context["line_search"]=linesearching # 1) evaluate loss loc_opt_args= copy.deepcopy(opt_args) loc_opt_args.opt_ctm_reinit= opt_args.line_search_ctm_reinit loc_ctm_args= copy.deepcopy(ctm_args) if opt_args.line_search_svd_method != 'DEFAULT': loc_ctm_args.projector_svd_method= opt_args.line_search_svd_method ls_context= dict({"ctm_args":loc_ctm_args, "opt_args":loc_opt_args, "loss_history": t_data, "line_search": linesearching}) loss, ctm_env, history, t_ctm, t_check = loss_fn(state, current_env[0],\ ls_context) current_env[0]= ctm_env # 2) store current state if the loss improves t_data["loss_ls"].append(loss.item()) if t_data["min_loss_ls"] > t_data["loss_ls"][-1]: t_data["min_loss_ls"]= t_data["loss_ls"][-1] # 3) log metrics for debugging if opt_args.opt_logging: log_entry=dict({"id": epoch, "LS": len(t_data["loss_ls"]), \ "loss": t_data["loss_ls"], "t_ctm": t_ctm, "t_check": t_check}) log.info(json.dumps(log_entry)) # 4) compute desired observables if obs_fn is not None: obs_fn(state, current_env[0], context) return loss for epoch in range(main_args.opt_max_iter): # checkpoint the optimizer # checkpointing before step, guarantees the correspondence between the wavefunction # and the last computed value of loss t_data["loss"][-1] if epoch>0: store_checkpoint(checkpoint_file, state, optimizer, epoch, t_data["loss"][-1]) # After execution closure ``current_env`` **IS NOT** corresponding to ``state``, since # the ``state`` on-site tensors have been modified by gradient. optimizer.step_2c(closure, closure_linesearch) # reset line search history t_data["loss_ls"]=[] t_data["min_loss_ls"]=1.0e+16 if post_proc is not None: post_proc(state, current_env[0], context) # terminate condition if len(t_data["loss"])>1 and \ abs(t_data["loss"][-1]-t_data["loss"][-2])<opt_args.tolerance_change: break if (opt_args.line_search not in ["default", None]) and \ "STATUS" in context and context["STATUS"]=="ENV_ANTIVAR": raise RuntimeError("Over-optimized environment, see log file for "\ +" env_sensitivity and loss_diff.") # optimization is over, store the last checkpoint if at least a single step was made if len(t_data["loss"])>0: store_checkpoint(checkpoint_file, state, optimizer, \ main_args.opt_max_iter, t_data["loss"][-1])