Merge pull request #1223 from sebffischer/feat/ignite

Feat/ignite

DESCRIPTION

@@ -96,6 +96,14 @@ Collate:
     'gen-namespace.R'
     'generator.R'
     'help.R'
+    'utils-data.R'
+    'optim.R'
+    'optim-adam.R'
+    'optim-sgd.R'
+    'optim-rmsprop.R'
+    'optim-adagrad.R'
+    'optim-adamw.R'
+    'ignite.R'
     'indexing.R'
     'install.R'
     'ivalue.R'
@@ -107,7 +115,6 @@ Collate:
     'layout.R'
     'linalg.R'
     'memory_format.R'
-    'utils-data.R'
     'nn.R'
     'nn-activation.R'
     'nn-batchnorm.R'
@@ -145,17 +152,11 @@ Collate:
     'nnf-upsampling.R'
     'nnf-vision.R'
     'operators.R'
-    'optim.R'
     'optim-adadelta.R'
-    'optim-adagrad.R'
-    'optim-adam.R'
-    'optim-adamw.R'
     'optim-asgd.R'
     'optim-lbfgs.R'
     'optim-lr_scheduler.R'
-    'optim-rmsprop.R'
     'optim-rprop.R'
-    'optim-sgd.R'
     'positron.R'
     'package.R'
     'qscheme.R'

NAMESPACE

@@ -127,6 +127,7 @@ S3method(torch_save,nn_module)
 S3method(torch_save,torch_tensor)
 S3method(trunc,torch_tensor)
 export("%>%")
+export(OptimizerIgnite)
 export(as_array)
 export(as_iterator)
 export(autograd_backward)
@@ -467,11 +468,18 @@ export(optim_adagrad)
 export(optim_adam)
 export(optim_adamw)
 export(optim_asgd)
+export(optim_ignite_adagrad)
+export(optim_ignite_adam)
+export(optim_ignite_adamw)
+export(optim_ignite_rmsprop)
+export(optim_ignite_sgd)
 export(optim_lbfgs)
+export(optim_required)
 export(optim_rmsprop)
 export(optim_rprop)
 export(optim_sgd)
 export(optimizer)
+export(optimizer_ignite)
 export(sampler)
 export(set_autocast)
 export(slc)

NEWS.md

@@ -2,13 +2,14 @@
 
 ## Bug fixes
 
-- Fix french translation (#1176 @cregouby)
-
-## Bug fixes
-
 - `torch_iinfo()` now support all integer dtypes (#1190 @cregouby)
 - Fixed float key_padding_mask in `nnf_multi_head_attention_forward()` (#1205)
 - Updated to LibTorch v2.5.1 (#1204)
+- Fix french translation (#1176 @cregouby)
+- Feature: Faster optimizers (`optim_ignite_<name>()`) are available: Adam, AdamW, Adagrad, RMSprop,SGD.
+  These can be used as drop-in replacements for `optim_<name>` but are considerably
+  faster as they wrap the LibTorch implementation of the optimizer.
+  The biggest speed differences can be observed for complex optimizers such as `AdamW`.
 
 # torch 0.13.0
 

R/RcppExports.R

@@ -14889,6 +14889,150 @@ cpp_torch_cuda_set_rng_state <- function(device, state) {
     invisible(.Call(`_torch_cpp_torch_cuda_set_rng_state`, device, state))
 }
 
+rcpp_ignite_optim_get_param_groups <- function(opt) {
+    .Call(`_torch_rcpp_ignite_optim_get_param_groups`, opt)
+}
+
+rcpp_ignite_optim_param_groups_size <- function(groups) {
+    .Call(`_torch_rcpp_ignite_optim_param_groups_size`, groups)
+}
+
+rcpp_ignite_optim_get_param_group_params <- function(groups, i) {
+    .Call(`_torch_rcpp_ignite_optim_get_param_group_params`, groups, i)
+}
+
+rcpp_ignite_optim_step <- function(opt) {
+    invisible(.Call(`_torch_rcpp_ignite_optim_step`, opt))
+}
+
+rcpp_ignite_optim_zero_grad <- function(opt) {
+    invisible(.Call(`_torch_rcpp_ignite_optim_zero_grad`, opt))
+}
+
+rcpp_ignite_optim_parameters_with_state <- function(opt) {
+    .Call(`_torch_rcpp_ignite_optim_parameters_with_state`, opt)
+}
+
+rcpp_ignite_adamw <- function(params, lr, betas, eps, weight_decay, amsgrad) {
+    .Call(`_torch_rcpp_ignite_adamw`, params, lr, betas, eps, weight_decay, amsgrad)
+}
+
+rcpp_ignite_adamw_get_states <- function(opt) {
+    .Call(`_torch_rcpp_ignite_adamw_get_states`, opt)
+}
+
+rcpp_ignite_adamw_set_states <- function(opt, params, states) {
+    invisible(.Call(`_torch_rcpp_ignite_adamw_set_states`, opt, params, states))
+}
+
+rcpp_ignite_adamw_add_param_group <- function(opt, params, lr, betas, eps, weight_decay, amsgrad) {
+    invisible(.Call(`_torch_rcpp_ignite_adamw_add_param_group`, opt, params, lr, betas, eps, weight_decay, amsgrad))
+}
+
+rcpp_as_list_adamw_param_groups <- function(groups) {
+    .Call(`_torch_rcpp_as_list_adamw_param_groups`, groups)
+}
+
+rcpp_ignite_adamw_set_param_group_options <- function(opt, list) {
+    invisible(.Call(`_torch_rcpp_ignite_adamw_set_param_group_options`, opt, list))
+}
+
+rcpp_ignite_adam <- function(params, lr, betas, eps, weight_decay, amsgrad) {
+    .Call(`_torch_rcpp_ignite_adam`, params, lr, betas, eps, weight_decay, amsgrad)
+}
+
+rcpp_ignite_adam_get_states <- function(opt) {
+    .Call(`_torch_rcpp_ignite_adam_get_states`, opt)
+}
+
+rcpp_ignite_adam_set_states <- function(opt, params, states) {
+    invisible(.Call(`_torch_rcpp_ignite_adam_set_states`, opt, params, states))
+}
+
+rcpp_ignite_adam_add_param_group <- function(opt, params, lr, betas, eps, weight_decay, amsgrad) {
+    invisible(.Call(`_torch_rcpp_ignite_adam_add_param_group`, opt, params, lr, betas, eps, weight_decay, amsgrad))
+}
+
+rcpp_as_list_adam_param_groups <- function(groups) {
+    .Call(`_torch_rcpp_as_list_adam_param_groups`, groups)
+}
+
+rcpp_ignite_adam_set_param_group_options <- function(opt, list) {
+    invisible(.Call(`_torch_rcpp_ignite_adam_set_param_group_options`, opt, list))
+}
+
+rcpp_ignite_sgd <- function(params, lr, momentum, dampening, weight_decay, nesterov) {
+    .Call(`_torch_rcpp_ignite_sgd`, params, lr, momentum, dampening, weight_decay, nesterov)
+}
+
+rcpp_ignite_sgd_get_states <- function(opt) {
+    .Call(`_torch_rcpp_ignite_sgd_get_states`, opt)
+}
+
+rcpp_ignite_sgd_set_states <- function(opt, params, states) {
+    invisible(.Call(`_torch_rcpp_ignite_sgd_set_states`, opt, params, states))
+}
+
+rcpp_ignite_sgd_add_param_group <- function(opt, params, lr, momentum, dampening, weight_decay, nesterov) {
+    invisible(.Call(`_torch_rcpp_ignite_sgd_add_param_group`, opt, params, lr, momentum, dampening, weight_decay, nesterov))
+}
+
+rcpp_as_list_sgd_param_groups <- function(groups) {
+    .Call(`_torch_rcpp_as_list_sgd_param_groups`, groups)
+}
+
+rcpp_ignite_sgd_set_param_group_options <- function(opt, list) {
+    invisible(.Call(`_torch_rcpp_ignite_sgd_set_param_group_options`, opt, list))
+}
+
+rcpp_ignite_rmsprop <- function(params, lr, alpha, eps, weight_decay, momentum, centered) {
+    .Call(`_torch_rcpp_ignite_rmsprop`, params, lr, alpha, eps, weight_decay, momentum, centered)
+}
+
+rcpp_ignite_rmsprop_get_states <- function(opt) {
+    .Call(`_torch_rcpp_ignite_rmsprop_get_states`, opt)
+}
+
+rcpp_ignite_rmsprop_set_states <- function(opt, params, states) {
+    invisible(.Call(`_torch_rcpp_ignite_rmsprop_set_states`, opt, params, states))
+}
+
+rcpp_ignite_rmsprop_add_param_group <- function(opt, params, lr, alpha, eps, weight_decay, momentum, centered) {
+    invisible(.Call(`_torch_rcpp_ignite_rmsprop_add_param_group`, opt, params, lr, alpha, eps, weight_decay, momentum, centered))
+}
+
+rcpp_as_list_rmsprop_param_groups <- function(groups) {
+    .Call(`_torch_rcpp_as_list_rmsprop_param_groups`, groups)
+}
+
+rcpp_ignite_rmsprop_set_param_group_options <- function(opt, list) {
+    invisible(.Call(`_torch_rcpp_ignite_rmsprop_set_param_group_options`, opt, list))
+}
+
+rcpp_ignite_adagrad <- function(params, lr, lr_decay, weight_decay, eps, initial_accumulator_value) {
+    .Call(`_torch_rcpp_ignite_adagrad`, params, lr, lr_decay, weight_decay, eps, initial_accumulator_value)
+}
+
+rcpp_ignite_adagrad_get_states <- function(opt) {
+    .Call(`_torch_rcpp_ignite_adagrad_get_states`, opt)
+}
+
+rcpp_ignite_adagrad_set_states <- function(opt, params, states) {
+    invisible(.Call(`_torch_rcpp_ignite_adagrad_set_states`, opt, params, states))
+}
+
+rcpp_ignite_adagrad_add_param_group <- function(opt, params, lr, lr_decay, weight_decay, eps, initial_accumulator_value) {
+    invisible(.Call(`_torch_rcpp_ignite_adagrad_add_param_group`, opt, params, lr, lr_decay, weight_decay, eps, initial_accumulator_value))
+}
+
+rcpp_as_list_adagrad_param_groups <- function(groups) {
+    .Call(`_torch_rcpp_as_list_adagrad_param_groups`, groups)
+}
+
+rcpp_ignite_adagrad_set_param_group_options <- function(opt, list) {
+    invisible(.Call(`_torch_rcpp_ignite_adagrad_set_param_group_options`, opt, list))
+}
+
 enquos0 <- function(env) {
     .Call(`_torch_enquos0`, env)
 }