[onert] Support Checkpoint API

[jyoungyun]

What

Let's support Checkpoint API

Why

To save and retrieve intermediate steps in training, this issue implements Checkpoint API and related functions.

Details

API

/**
 * @brief Import circle checkpoint
 * @note  This function should be called on training mode
 *        This function should be called before {@link nnfw_train}
 *
 * @param[in] session The session to export a checkpoint
 * @param[in] path    The path to export a checkpoint
 * @return @c NNFW_STATUS_NO_ERROR if successful
 */
NNFW_STATUS nnfw_train_import_checkpoint(nnfw_session *session, const char *path);

/**
 * @brief Export circle checkpoint
 * @note  This function should be called on training mode
 *        This function should be called after {@link nnfw_train}
 *
 * @param[in] session The session to export a checkpoint
 * @param[in] path    The path to export a checkpoint
 * @return @c NNFW_STATUS_NO_ERROR if successful
 */
NNFW_STATUS nnfw_train_export_checkpoint(nnfw_session *session, const char *path);

These APIs are compatible with onert-micro APIs.

Checkpoint File Spec

• https://github.com/Samsung/ONE/discussions/13037#discussioncomment-9692320	0	2	3	4	8	12	16	20	...	20 + N * 4 - 4	20 + N * 4	...	Buffers_size + 20 + N * 4	m offset	...	m offset + N * 4 - 4	m offset + N * 4	...	m offset + Buffer_size + N * 4	v offset	...	v offset + N * 4 - 4	v offset + N * 4	...	v offset + Buffer_size + N * 4	other param offset	other param offset + 4
  MAGIC NUMBER	SCHEMA VERSION	RESERVED	m offset	v offset	other parameters offset	n_buffers (=N)	Offset 1	…	Offset N	Buffer 1 Data	…	Buffer N Data	Offset 1 for m	...	Offset N for m	Buffer 1 for m	...	Buffer N for m	Offset 1 for v	...	Offset N for v	Buffer 1 for v	...	Buffer N for v	cur step value	cur epoch value

• Details
  • Header, Tensor data(offset, buffer), Adam Optimizer data(offset, buffer), Footer(step, epoch)
• Issue
  • There is a limitation that it is difficult to add new data.
  • Only two optimizer buffer can be saved. It is not confirmed whether it is enough.

Draft: #13561

[jyoungyun]

Checkpoint File Spec

• The item to discuss about checkpoint file format
  • https://github.com/Samsung/ONE/discussions/13037#discussioncomment-10331312

[ragmani]

I'm sorry. There is something I missed related to optimizer. Optimizers have properties that should be written in circle checkpoint. Moreover, the space for properties to be written should be flexible because the kinds of properties may vary depending on the optimizer type. You can see properties for SGD, Adam in #11698

[jyoungyun]

@ragmani

Moreover, the space for properties to be written should be flexible because the kinds of properties may vary depending on the optimizer type. You can see properties for SGD, Adam in #11698

I think these properites can be saved to the circle+ file. Is there any change data that we need to save?

[ragmani]

I think these properites can be saved to the circle+ file. Is there any change data that we need to save?

As we talked offline, there seems to be no need to change data for properties.