Configuration¶
Main¶
-
class
config.
MAINARGS
[source]¶ Main simulation options. The default settings can be modified through command line arguments as follows
--<option-name> desired-value
- Variables
omp_cores (int:) – number of OpenMP cores. Default:
1
instate (str or Path) – input state file. Default:
None
instate_noise (float) – magnitude of noise applied to the input state, if any. Default:
0.0
ipeps_init_type (str) – initialization of the trial iPEPS state, if no
instate
is provided. Default:RANDOM
out_prefix (str) – output file prefix. Default:
output
bond_dim (int) – iPEPS auxiliary bond dimension. Default:
1
chi (int) – environment bond dimension. Default:
20
opt_max_iter (int) – maximal number of optimization steps. Default:
100
opt_resume (str or Path) – resume from checkpoint file. Default:
None
opt_resume_override_params (bool) – override optimizer parameters stored in checkpoint. Default:
False
seed (int) – PRNG seed. Default:
0
Global¶
-
class
config.
GLOBALARGS
[source]¶ Holds global configuration options. The default settings can be modified through command line arguments as follows
--GLOBALARGS_<variable-name> desired-value
- Variables
dtype (torch.dtype) – data type of all torch.tensor. Default:
torch.float64
device (str) – device on which all the torch.tensors are stored. Default:
'cpu'
gpu (str) – gpu used for optional acceleration. It might be desirable to store the model and all the intermediates of CTM on CPU and compute only the core parts of the expensive CTM step on GPU. Default:
''
Corner Transfer Matrix Algorithm¶
-
class
config.
CTMARGS
[source]¶ Holds configuration of the CTM algorithm. The default settings can be modified through command line arguments as follows
--CTMARGS_<variable-name> desired-value
- Variables
ctm_max_iter (int) – maximum iterations of directional CTM algorithm. Default:
50
ctm_env_init_type (str) – default initialization method for ENV objects. Default:
'CTMRG'
ctm_conv_tol (float) – threshold for convergence of CTM algorithm. Default:
'1.0e-10'
conv_check_cpu (bool) – execute CTM convergence check on cpu (if applicable). Default:
False
projector_method (str) –
method used to construct projectors which facilitate truncation of environment bond dimension \(\chi\) within CTM algorithm
4X4: Projectors are built from two halfs of 4x4 tensor network
4X2: Projectors are built from two enlarged corners (2x2) making up a 4x2 (or 2x4) tensor network
Default:
'4X4'
projector_svd_method (str) –
singular/eigen value decomposition algorithm used in the construction of the projectors:
'GESDD'
: pytorch wrapper of LAPACK’s gesdd'RSVD'
: randomized SVD'SYMEIG'
: pytorch wrapper of LAPACK’s dsyev for symmetric matrices'SYMARP'
: scipy wrapper of ARPACK’s dsaupd for symmetric matrices'ARP'
: scipy wrapper of ARPACK’s svds for general matrices
Default:
'SYMEIG'
for c4v-symmetric CTM, otherwise'GESDD'
projector_svd_reltol (float) – relative threshold on the magnitude of the smallest elements of singular value spectrum used in the construction of projectors. Default:
1.0e-8
ctm_move_sequence (list[tuple(int,int)]) –
sequence of directional moves within single CTM iteration. The possible directions are encoded as tuples(int,int)
up = (0,-1)
left = (-1,0)
down = (0,1)
right = (1,0)
Default:
[(0,-1), (-1,0), (0,1), (1,0)]
fwd_checkpoint_c2x2 (bool) – recompute forward pass of enlarged corner functions (c2x2_*) during backward pass within optimization to save memory. Default:
False
fwd_checkpoint_halves (bool) – recompute forward pass of halves functions (halves_of_4x4_*) during backward pass within optimization to save memory. Default:
False
fwd_checkpoint_projectors (bool) – recompute forward pass of projector construction (except SVD) during backward pass within optimization to save memory. Default:
False
fwd_checkpoint_absorb (bool) – recompute forward pass of absorp and truncate functions (absorb_truncate_*) during backward pass within optimization to save memory. Default:
False
fwd_checkpoint_move (bool) – recompute forward pass of whole
ctm_MOVE
during backward pass. Default:False
FPCM related options
- Variables
fpcm_init_iter (int) – minimal number of CTM steps before FPCM acceleration step is attempted. Default:
1
fpcm_freq (int) – frequency of FPCM steps per CTM steps. Default:
-1
fpcm_isogauge_tol (float) – tolerance on gauging the uniform MPS built from half-row/-column tensor T. Default:
1.0e-14
.fpcm_fpt_tol (float) – tolerance on convergence within FPCM step. Default:
1.0e-8
Logging and Debugging options
- Variables
ctm_logging – log debug statements into log file. Default:
False
verbosity_initialization (int) – verbosity of initialization method for ENV objects. Default:
0
verbosity_ctm_convergence (int) – verbosity of evaluation of CTM convergence criterion. Default:
0
verbosity_projectors (int) – verbosity of projector construction. Default:
0
verbosity_ctm_move (int) – verbosity of directional CTM moves. Default:
0
verbosity_rdm (int) – verbosity of reduced density matrix routines. Default:
0
step_core_gpu (bool) – assuming the default device is CPU, offload the core part of the CTM step to GPU. Together with CTM step checkpointing
fwd_checkpoint_move
allows to store all intermediates on CPU.
Optimization¶
-
class
config.
OPTARGS
[source]¶ Holds configuration of the optimization. The default settings can be modified through command line arguments as follows
--OPTARGS_<variable-name> desired-value
General options
- Variables
opt_ctm_reinit (bool) – reinitialize environment from scratch within every loss function evaluation. Default:
True
lr (float) – initial learning rate. Default:
1.0
line_search (str) – line search algorithm to use. L-BFGS supports
'strong_wolfe'
and'backtracking'
. SGD supports just'backtracking'
. Default:None
.line_search_ctm_reinit (bool) – recompute environment from scratch at each step within line search algorithm. Default:
True
.line_search_svd_method (str) – eigen decompostion method to use within line search environment computation. See options in
config.CTMARGS
. Default:'DEFAULT'
which depends on the particular CTM algorithm.
L-BFGS related options
- Variables
tolerance_grad (float) – stopping criterion wrt. norm of the gradient (which norm ? See
torch.optim.LBFGS
). Default:1.0e-5
tolerance_change (float) – stopping criterion wrt. change of the loss function. Default:
1.0e-9
max_iter_per_epoch (int) – maximum number of optimizer iterations per epoch. Default:
1
history_size (int) – number past of directions used to approximate inverse Hessian. Default:
100
.
SGD related options
- Variables
momentum (float) – momentum used in the SGD step
dampening (float) – dampening used in the SGD step
Gradients through finite differences
- Variables
fd_eps (float) – magnitude of displacement when computing the forward difference \(E(x_0 + \textrm{fd_eps})-E(x_0)/\textrm{fd_eps}\). Default:
1.0e-4
fd_ctm_reinit – recompute environment from scratch after applying the displacement. Default:
True
Logging
- Variables
opt_logging (bool) – turns on recording of additional data from optimization, such as CTM convergence, timings, gradients, etc. The information is logged in file
{out_prefix}_log.json
. Default:True
opt_log_grad (bool) – log values of gradient. Default:
False
verbosity_opt_epoch (int) – verbosity within optimization epoch. Default:
1