Anki/rslib/src/scheduler/fsrs/memory_state.rs

494 lines
19 KiB
Rust

// Copyright: Ankitects Pty Ltd and contributors
// License: GNU AGPL, version 3 or later; http://www.gnu.org/licenses/agpl.html
use std::collections::HashMap;
use anki_proto::scheduler::ComputeMemoryStateResponse;
use fsrs::FSRSItem;
use fsrs::MemoryState;
use fsrs::FSRS;
use fsrs::FSRS5_DEFAULT_DECAY;
use fsrs::FSRS6_DEFAULT_DECAY;
use itertools::Itertools;
use super::params::ignore_revlogs_before_ms_from_config;
use super::rescheduler::Rescheduler;
use crate::card::CardType;
use crate::prelude::*;
use crate::revlog::RevlogEntry;
use crate::scheduler::answering::get_fuzz_seed;
use crate::scheduler::fsrs::params::reviews_for_fsrs;
use crate::scheduler::fsrs::params::Params;
use crate::scheduler::states::fuzz::with_review_fuzz;
use crate::search::Negated;
use crate::search::SearchNode;
use crate::search::StateKind;
#[derive(Debug, Clone, Copy, Default)]
pub struct ComputeMemoryProgress {
pub current_cards: u32,
pub total_cards: u32,
}
/// Helper function to determine the appropriate decay value based on FSRS
/// parameters
pub(crate) fn get_decay_from_params(params: &[f32]) -> f32 {
if params.is_empty() {
FSRS6_DEFAULT_DECAY // default decay for FSRS-6
} else if params.len() < 21 {
FSRS5_DEFAULT_DECAY // default decay for FSRS-4.5 and FSRS-5
} else {
params[20]
}
}
#[derive(Debug)]
pub(crate) struct UpdateMemoryStateRequest {
pub params: Params,
pub desired_retention: f32,
pub historical_retention: f32,
pub max_interval: u32,
pub reschedule: bool,
}
pub(crate) struct UpdateMemoryStateEntry {
pub req: Option<UpdateMemoryStateRequest>,
pub search: SearchNode,
pub ignore_before: TimestampMillis,
}
impl Collection {
/// For each provided set of params, locate cards with the provided search,
/// and update their memory state.
/// Should be called inside a transaction.
/// If Params are None, it means the user disabled FSRS, and the existing
/// memory state should be removed.
pub(crate) fn update_memory_state(
&mut self,
entries: Vec<UpdateMemoryStateEntry>,
) -> Result<()> {
let timing = self.timing_today()?;
let usn = self.usn()?;
for UpdateMemoryStateEntry {
req,
search,
ignore_before,
} in entries
{
let search =
SearchBuilder::all([search.into(), SearchNode::State(StateKind::New).negated()]);
let revlog = self.revlog_for_srs(search)?;
let reschedule = req.as_ref().map(|e| e.reschedule).unwrap_or_default();
let last_revlog_info = if reschedule {
Some(get_last_revlog_info(&revlog))
} else {
None
};
let mut rescheduler = self
.get_config_bool(BoolKey::LoadBalancerEnabled)
.then(|| Rescheduler::new(self))
.transpose()?;
let fsrs = FSRS::new(req.as_ref().map(|w| &w.params[..]).or(Some([].as_slice())))?;
let decay = req.as_ref().map(|w| get_decay_from_params(&w.params));
let historical_retention = req.as_ref().map(|w| w.historical_retention);
let items = fsrs_items_for_memory_states(
&fsrs,
revlog,
timing.next_day_at,
historical_retention.unwrap_or(0.9),
ignore_before,
)?;
let desired_retention = req.as_ref().map(|w| w.desired_retention);
let mut progress = self.new_progress_handler::<ComputeMemoryProgress>();
progress.update(false, |s| s.total_cards = items.len() as u32)?;
for (idx, (card_id, item)) in items.into_iter().enumerate() {
progress.update(true, |state| state.current_cards = idx as u32 + 1)?;
let mut card = self.storage.get_card(card_id)?.or_not_found(card_id)?;
let original = card.clone();
if let Some(req) = &req {
// Store decay and desired retention in the card so that add-ons, card info,
// stats and browser search/sorts don't need to access the deck config.
// Unlike memory states, scheduler doesn't use decay and dr stored in the card.
card.desired_retention = desired_retention;
card.decay = decay;
if let Some(item) = item {
card.set_memory_state(&fsrs, Some(item), historical_retention.unwrap())?;
// if rescheduling
if let Some(reviews) = &last_revlog_info {
// and we have a last review time for the card
if let Some(last_info) = reviews.get(&card.id) {
if let Some(last_review) = &last_info.last_reviewed_at {
let days_elapsed =
timing.next_day_at.elapsed_days_since(*last_review) as i32;
// and the card's not new
if let Some(state) = &card.memory_state {
// or in (re)learning
if card.ctype == CardType::Review {
let deck = self
.get_deck(card.original_or_current_deck_id())?
.or_not_found(card.original_or_current_deck_id())?;
let deckconfig_id = deck.config_id().unwrap();
// reschedule it
let original_interval = card.interval;
let interval = fsrs.next_interval(
Some(state.stability),
desired_retention.unwrap(),
0,
);
card.interval = rescheduler
.as_mut()
.and_then(|r| {
r.find_interval(
interval,
1,
req.max_interval,
days_elapsed as u32,
deckconfig_id,
get_fuzz_seed(&card, true),
)
})
.unwrap_or_else(|| {
with_review_fuzz(
card.get_fuzz_factor(true),
interval,
1,
req.max_interval,
)
});
let due = if card.original_due != 0 {
&mut card.original_due
} else {
&mut card.due
};
let new_due = (timing.days_elapsed as i32)
- days_elapsed
+ card.interval as i32;
if let Some(rescheduler) = &mut rescheduler {
rescheduler.update_due_cnt_per_day(
*due,
new_due,
deckconfig_id,
);
}
*due = new_due;
// Add a rescheduled revlog entry
self.log_rescheduled_review(
&card,
original_interval,
usn,
)?;
}
}
}
}
}
} else {
// clear memory states if item is None
card.memory_state = None;
}
} else {
// clear FSRS data if FSRS is disabled
card.clear_fsrs_data();
}
self.update_card_inner(&mut card, original, usn)?;
}
}
Ok(())
}
pub fn compute_memory_state(&mut self, card_id: CardId) -> Result<ComputeMemoryStateResponse> {
let mut card = self.storage.get_card(card_id)?.or_not_found(card_id)?;
let deck_id = card.original_deck_id.or(card.deck_id);
let deck = self.get_deck(deck_id)?.or_not_found(card.deck_id)?;
let conf_id = DeckConfigId(deck.normal()?.config_id);
let config = self
.storage
.get_deck_config(conf_id)?
.or_not_found(conf_id)?;
let desired_retention = config.inner.desired_retention;
let historical_retention = config.inner.historical_retention;
let params = config.fsrs_params();
let decay = get_decay_from_params(params);
let fsrs = FSRS::new(Some(params))?;
let revlog = self.revlog_for_srs(SearchNode::CardIds(card.id.to_string()))?;
let item = fsrs_item_for_memory_state(
&fsrs,
revlog,
self.timing_today()?.next_day_at,
historical_retention,
ignore_revlogs_before_ms_from_config(&config)?,
)?;
if item.is_some() {
card.set_memory_state(&fsrs, item, historical_retention)?;
Ok(ComputeMemoryStateResponse {
state: card.memory_state.map(Into::into),
desired_retention,
decay,
})
} else {
Ok(ComputeMemoryStateResponse {
state: None,
desired_retention,
decay,
})
}
}
}
impl Card {
pub(crate) fn set_memory_state(
&mut self,
fsrs: &FSRS,
item: Option<FsrsItemForMemoryState>,
historical_retention: f32,
) -> Result<()> {
let memory_state = if let Some(i) = item {
Some(fsrs.memory_state(i.item, i.starting_state)?)
} else if self.ctype == CardType::New || self.interval == 0 {
None
} else {
// no valid revlog entries; infer state from current card state
Some(fsrs.memory_state_from_sm2(
self.ease_factor(),
self.interval as f32,
historical_retention,
)?)
};
self.memory_state = memory_state.map(Into::into);
Ok(())
}
}
#[derive(Debug)]
pub(crate) struct FsrsItemForMemoryState {
pub item: FSRSItem,
/// When revlogs have been truncated, this stores the initial state at first
/// review
pub starting_state: Option<MemoryState>,
pub filtered_revlogs: Vec<RevlogEntry>,
}
/// Like [fsrs_item_for_memory_state], but for updating multiple cards at once.
pub(crate) fn fsrs_items_for_memory_states(
fsrs: &FSRS,
revlogs: Vec<RevlogEntry>,
next_day_at: TimestampSecs,
historical_retention: f32,
ignore_revlogs_before: TimestampMillis,
) -> Result<Vec<(CardId, Option<FsrsItemForMemoryState>)>> {
revlogs
.into_iter()
.chunk_by(|r| r.cid)
.into_iter()
.map(|(card_id, group)| {
Ok((
card_id,
fsrs_item_for_memory_state(
fsrs,
group.collect(),
next_day_at,
historical_retention,
ignore_revlogs_before,
)?,
))
})
.collect()
}
struct LastRevlogInfo {
/// Used to determine the actual elapsed time between the last time the user
/// reviewed the card and now, so that we can determine an accurate period
/// when the card has subsequently been rescheduled to a different day.
last_reviewed_at: Option<TimestampSecs>,
}
/// Return a map of cards to info about last review/reschedule.
fn get_last_revlog_info(revlogs: &[RevlogEntry]) -> HashMap<CardId, LastRevlogInfo> {
let mut out = HashMap::new();
revlogs
.iter()
.chunk_by(|r| r.cid)
.into_iter()
.for_each(|(card_id, group)| {
let mut last_reviewed_at = None;
for e in group.into_iter() {
if e.button_chosen >= 1 {
last_reviewed_at = Some(e.id.as_secs());
}
}
out.insert(card_id, LastRevlogInfo { last_reviewed_at });
});
out
}
/// When calculating memory state, only the last FSRSItem is required. If the
/// revlog is non-empty and no learning steps have been detected (indicative of
/// a truncated revlog), we return the starting state inferred from the first
/// revlog entry, so that the first review is not treated as if started from
/// scratch.
pub(crate) fn fsrs_item_for_memory_state(
fsrs: &FSRS,
entries: Vec<RevlogEntry>,
next_day_at: TimestampSecs,
historical_retention: f32,
ignore_revlogs_before: TimestampMillis,
) -> Result<Option<FsrsItemForMemoryState>> {
struct FirstReview {
interval: f32,
ease_factor: f32,
}
if let Some(mut output) = reviews_for_fsrs(entries, next_day_at, false, ignore_revlogs_before) {
let mut item = output.fsrs_items.pop().unwrap().1;
if output.revlogs_complete {
Ok(Some(FsrsItemForMemoryState {
item,
starting_state: None,
filtered_revlogs: output.filtered_revlogs,
}))
} else if let Some(first_user_grade) = output.filtered_revlogs.first() {
// the revlog has been truncated, but not fully
let first_review = FirstReview {
interval: first_user_grade.interval.max(1) as f32,
ease_factor: if first_user_grade.ease_factor == 0 {
2500
} else {
first_user_grade.ease_factor
} as f32
/ 1000.0,
};
let mut starting_state = fsrs.memory_state_from_sm2(
first_review.ease_factor,
first_review.interval,
historical_retention,
)?;
// if the ease factor is less than 1.1, the revlog entry is generated by FSRS
if first_review.ease_factor <= 1.1 {
starting_state.difficulty = (first_review.ease_factor - 0.1) * 9.0 + 1.0;
}
// remove the first review because it has been converted to the starting state
item.reviews.remove(0);
Ok(Some(FsrsItemForMemoryState {
item,
starting_state: Some(starting_state),
filtered_revlogs: output.filtered_revlogs,
}))
} else {
// only manual and rescheduled revlogs; treat like empty
Ok(None)
}
} else {
Ok(None)
}
}
#[cfg(test)]
mod tests {
use fsrs::MemoryState;
use super::*;
use crate::card::FsrsMemoryState;
use crate::revlog::RevlogReviewKind;
use crate::scheduler::fsrs::params::tests::convert;
use crate::scheduler::fsrs::params::tests::revlog;
/// Floating point precision can vary between platforms, and each FSRS
/// update tends to result in small changes to these numbers, so we
/// round them.
fn assert_int_eq(actual: Option<FsrsMemoryState>, expected: Option<FsrsMemoryState>) {
let actual = actual.unwrap();
let expected = expected.unwrap();
assert_eq!(actual.stability.round(), expected.stability.round());
assert_eq!(actual.difficulty.round(), expected.difficulty.round());
}
#[test]
fn bypassed_learning_is_handled() -> Result<()> {
// cards without any learning steps due to truncated history still have memory
// state calculated
let fsrs = FSRS::new(Some(&[])).unwrap();
let item = fsrs_item_for_memory_state(
&fsrs,
vec![
RevlogEntry {
ease_factor: 2500,
interval: 100,
..revlog(RevlogReviewKind::Review, 99)
},
revlog(RevlogReviewKind::Review, 0),
],
TimestampSecs::now(),
0.9,
0.into(),
)?
.unwrap();
assert_int_eq(
item.starting_state.map(Into::into),
Some(FsrsMemoryState {
stability: 100.0,
difficulty: 5.003576,
}),
);
let mut card = Card {
reps: 1,
..Default::default()
};
card.set_memory_state(&fsrs, Some(item), 0.9)?;
assert_int_eq(
card.memory_state,
Some(FsrsMemoryState {
stability: 248.9251,
difficulty: 4.9938006,
}),
);
// cards with a single review-type entry also get memory states from revlog
// rather than card states
let item = fsrs_item_for_memory_state(
&fsrs,
vec![RevlogEntry {
ease_factor: 2500,
interval: 100,
..revlog(RevlogReviewKind::Review, 100)
}],
TimestampSecs::now(),
0.9,
0.into(),
)?
.unwrap();
assert!(item.item.reviews.is_empty());
card.set_memory_state(&fsrs, Some(item), 0.9)?;
assert_int_eq(
card.memory_state,
Some(FsrsMemoryState {
stability: 100.0,
difficulty: 5.003576,
}),
);
Ok(())
}
#[test]
fn zero_history_is_handled() -> Result<()> {
// when the history is empty, no items are produced
assert_eq!(convert(&[], false), None);
// but memory state should still be inferred, by using the card's current state
let mut card = Card {
ctype: CardType::Review,
interval: 100,
ease_factor: 1300,
reps: 1,
..Default::default()
};
card.set_memory_state(&FSRS::new(Some(&[])).unwrap(), None, 0.9)?;
assert_int_eq(
card.memory_state,
Some(
MemoryState {
stability: 99.999954,
difficulty: 9.979899,
}
.into(),
),
);
Ok(())
}
}