Skip to content

stateful

PriorityMap

A map from keys to ordered priorities, supporting efficient pop and peek operations returning the key with the lowest priority.

It probably wan't reasonable to write our own implementation of this, but it was easier than finding a robust existing implementation.

Source code in genlm/control/potential/stateful.py 
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
class PriorityMap:
    """A map from keys to ordered priorities, supporting efficient
    pop and peek operations returning the key with the lowest priority.

    It probably wan't reasonable to write our own implementation of this,
    but it was easier than finding a robust existing implementation.
    """

    def __init__(self):
        self.__priorities = {}
        self.__priorities_to_keys = defaultdict(set)
        self.__heap = []

    def __repr__(self):
        return f"PriorityMap({self.__priorities})"

    def __clear_dead_priorities(self):
        while self.__heap and self.__heap[0] not in self.__priorities_to_keys:
            heapq.heappop(self.__heap)

    def __setitem__(self, key, value):
        if key in self.__priorities:
            old_value = self.__priorities[key]
            if old_value == value:
                return
            else:
                self.__priorities_to_keys[old_value].remove(key)
                if not self.__priorities_to_keys[old_value]:
                    del self.__priorities_to_keys[old_value]

        self.__priorities[key] = value
        if value not in self.__priorities_to_keys:
            heapq.heappush(self.__heap, value)
        self.__priorities_to_keys[value].add(key)

    def __getitem__(self, key):
        return self.__priorities[key]

    def __delitem__(self, key):
        value = self.__priorities.pop(key)
        self.__priorities_to_keys[value].remove(key)
        if not self.__priorities_to_keys[value]:
            del self.__priorities_to_keys[value]

    def peek(self):
        if not self.__heap:
            raise ValueError("Peek on empty PriorityMap")
        self.__clear_dead_priorities()
        min_priority = self.__heap[0]
        assert self.__priorities_to_keys[min_priority]
        key = next(iter(self.__priorities_to_keys[min_priority]))
        return key, min_priority

    def pop(self):
        if not self.__heap:
            raise ValueError("Pop on empty PriorityMap")
        self.__clear_dead_priorities()
        min_priority = self.__heap[0]
        assert self.__priorities_to_keys[min_priority]
        key = next(iter(self.__priorities_to_keys[min_priority]))
        del self.__priorities[key]
        self.__priorities_to_keys[min_priority].remove(key)
        if not self.__priorities_to_keys[min_priority]:
            del self.__priorities_to_keys[min_priority]
            heapq.heappop(self.__heap)
        return key, min_priority

    def clear(self):
        self.__priorities.clear()
        self.__priorities_to_keys.clear()
        self.__heap.clear()

    def __len__(self):
        return len(self.__priorities)

ParticleState

Bases: ABC

Source code in genlm/control/potential/stateful.py 
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
class ParticleState(ABC):
    def __init__(self, owner):
        self.owner = owner
        self.finished = False
        self.context = []

    async def update_context(self, incremental_context):
        """Update the context with more data that has come in."""
        if self.finished:
            return
        self.context.extend(incremental_context)
        await self.impl_update_context(incremental_context)

    async def start(self):
        """May be implemented by subclasses to perform initialization logic
        that needs to be async."""
        pass

    async def finish(self):
        """Mark this state as finished, clearing up any associated
        state, and updating the current score to reflect whether
        this is a valid string in the associated language."""
        if self.finished:
            return
        self.finished = True
        await self.impl_finish()

    @abstractproperty
    def current_score(self):
        """The current score associated with this potential, which
        will reflect whether the current context is a suitable member
        of the language if this has been finished, or whether it is a
        suitable prefix if it has not."""

    @abstractmethod
    async def impl_update_context(self, incremental_context): ...

    @abstractmethod
    async def impl_finish(self): ...

    async def clone(self):
        if self.finished:
            return self
        result = await self.owner.new_state()
        await result.update_context(self.context)
        assert self.context == result.context
        assert self.current_score == result.current_score
        return result

update_context(incremental_context) async

Update the context with more data that has come in.

Source code in genlm/control/potential/stateful.py 
102
103
104
105
106
107
async def update_context(self, incremental_context):
    """Update the context with more data that has come in."""
    if self.finished:
        return
    self.context.extend(incremental_context)
    await self.impl_update_context(incremental_context)

start() async

May be implemented by subclasses to perform initialization logic that needs to be async.

Source code in genlm/control/potential/stateful.py 
109
110
111
112
async def start(self):
    """May be implemented by subclasses to perform initialization logic
    that needs to be async."""
    pass

finish() async

Mark this state as finished, clearing up any associated state, and updating the current score to reflect whether this is a valid string in the associated language.

Source code in genlm/control/potential/stateful.py 
114
115
116
117
118
119
120
121
async def finish(self):
    """Mark this state as finished, clearing up any associated
    state, and updating the current score to reflect whether
    this is a valid string in the associated language."""
    if self.finished:
        return
    self.finished = True
    await self.impl_finish()

current_score()

The current score associated with this potential, which will reflect whether the current context is a suitable member of the language if this has been finished, or whether it is a suitable prefix if it has not.

Source code in genlm/control/potential/stateful.py 
123
124
125
126
127
128
@abstractproperty
def current_score(self):
    """The current score associated with this potential, which
    will reflect whether the current context is a suitable member
    of the language if this has been finished, or whether it is a
    suitable prefix if it has not."""

StatefulPotential

Bases: Potential

Source code in genlm/control/potential/stateful.py 
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
class StatefulPotential(Potential):
    def __init__(
        self, vocabulary, token_type=None, eos=None, state_class=None, cache_size=100
    ):
        super().__init__(vocabulary=vocabulary, token_type=token_type, eos=eos)
        self.__state_class = state_class

        self.__cache_size = cache_size
        self.__state_count = 0

        self.__state_pool = defaultdict(list)
        self.__known_contexts = []

        self.__eviction_heap = PriorityMap()

        self.__epoch = 0

    def __tick(self):
        self.__epoch += 1
        return self.__epoch

    async def new_state(self) -> ParticleState:
        if self.__state_class is None:
            raise NotImplementedError()
        result = self.__state_class(self)
        await result.start()
        return result

    async def cleanup(self):
        await asyncio.gather(
            *[state.finish() for pool in self.__state_pool.values() for state in pool]
        )
        self.__state_pool.clear()
        self.__known_contexts.clear()
        self.__eviction_heap.clear()

    async def __look_up_state(self, context):
        context = make_immutable(context)

        state = None

        i = bisect.bisect_left(self.__known_contexts, context)
        if i < len(self.__known_contexts):
            existing = self.__known_contexts[i]
            if context[: len(existing)] == existing:
                pool = self.__state_pool[existing]
                if pool:
                    state = pool.pop()
                    if not pool:
                        del self.__known_contexts[i]
                        del self.__state_pool[existing]
                        del self.__eviction_heap[existing]
                    self.__state_count -= 1
                    self.__check_soundness()
        if state is None:
            state = await self.new_state()
        if len(context) > len(state.context):
            await state.update_context(context[len(state.context) :])
        assert len(state.context) == len(context)
        assert list(state.context) == list(context)
        return state

    def __check_soundness(self):
        assert self.__state_count >= 0
        assert len(self.__eviction_heap) == len(self.__state_pool)
        assert self.__state_count >= len(self.__state_pool)
        assert len(self.__known_contexts) == len(self.__state_pool)

    def __return_state(self, state):
        self.__check_soundness()
        assert not state.finished
        context = make_immutable(state.context)
        i = bisect.bisect_left(self.__known_contexts, context)
        if i >= len(self.__known_contexts):
            self.__known_contexts.append(context)
        elif self.__known_contexts[i] != context:
            self.__known_contexts.insert(i, context)
        self.__state_pool[context].append(state)
        self.__state_count += 1
        age = self.__tick()
        self.__eviction_heap[context] = age
        self.__check_soundness()
        assert len(self.__eviction_heap) > 0
        assert self.__eviction_heap[context] == age
        while self.__state_count > self.__cache_size:
            self.__check_soundness()
            to_evict, _ = self.__eviction_heap.peek()
            i = bisect.bisect_left(self.__known_contexts, to_evict)
            assert self.__known_contexts[i] == to_evict
            pool = self.__state_pool[to_evict]
            pool.pop()
            self.__state_count -= 1
            if not pool:
                del self.__state_pool[to_evict]
                check, _ = self.__eviction_heap.pop()
                assert check == to_evict
                del self.__known_contexts[i]
            self.__check_soundness()

    async def prefix(self, context):
        state = await self.__look_up_state(context)
        result = state.current_score
        self.__return_state(state)
        return result

    async def complete(self, context):
        state = await self.__look_up_state(context)
        await state.finish()
        return state.current_score

    async def logw_next(self, context):
        """Compute the next-token weights of each token in `self.vocab_eos` given `context`.

        Args:
            context (list): Sequence of tokens.

        Returns:
            (LazyWeights): Weights of each token in the vocabulary and EOS.
        """
        state = await self.__look_up_state(context)
        assert not state.finished
        ctx_log_w = state.current_score

        if ctx_log_w == float("-inf"):
            raise ValueError(f"Context {context!r} has weight zero under `prefix`.")

        async def step_score(x):
            local_state = await state.clone()
            await local_state.update_context([x])

            if x == self.eos:
                await local_state.finish()
                return local_state.current_score
            else:
                result = local_state.current_score
                await local_state.finish()
                return result

        scores = np.array(
            await asyncio.gather(*[step_score(x) for x in self.vocab_eos])
        )

        logws = scores - ctx_log_w

        return self.make_lazy_weights(logws)

logw_next(context) async

Compute the next-token weights of each token in self.vocab_eos given context.

Parameters:

Name Type Description Default
context list

Sequence of tokens.

 required

Returns:

Type Description
LazyWeights

Weights of each token in the vocabulary and EOS.
Source code in genlm/control/potential/stateful.py 
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
async def logw_next(self, context):
    """Compute the next-token weights of each token in `self.vocab_eos` given `context`.

    Args:
        context (list): Sequence of tokens.

    Returns:
        (LazyWeights): Weights of each token in the vocabulary and EOS.
    """
    state = await self.__look_up_state(context)
    assert not state.finished
    ctx_log_w = state.current_score

    if ctx_log_w == float("-inf"):
        raise ValueError(f"Context {context!r} has weight zero under `prefix`.")

    async def step_score(x):
        local_state = await state.clone()
        await local_state.update_context([x])

        if x == self.eos:
            await local_state.finish()
            return local_state.current_score
        else:
            result = local_state.current_score
            await local_state.finish()
            return result

    scores = np.array(
        await asyncio.gather(*[step_score(x) for x in self.vocab_eos])
    )

    logws = scores - ctx_log_w

    return self.make_lazy_weights(logws)