File size: 9,797 Bytes
ca3d977
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
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
94
95
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
144
145
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
"""Factorized, contrastively trained canonical state representation probes."""

from __future__ import annotations

from dataclasses import dataclass
import random

import torch
from torch import Tensor, nn
import torch.nn.functional as F


CANONICAL = 0
CONTRADICTED = 1
HISTORICAL = 2
INFERRED = 3


class FactorizedStateRepresentation(nn.Module):
    """Separate entity/relation/value/metadata fields projected to a P slot."""

    def __init__(
        self,
        entities: int,
        relations: int,
        values: int,
        metadata: int = 4,
        *,
        field_width: int = 128,
        slot_width: int = 512,
    ) -> None:
        super().__init__()
        self.entity = nn.Embedding(entities, field_width)
        self.relation = nn.Embedding(relations, field_width)
        self.value = nn.Embedding(values, field_width)
        self.metadata = nn.Embedding(metadata, field_width)
        factor_width = field_width * 4
        self.slot_projection = nn.Linear(factor_width, slot_width, bias=False)
        self.tuple_projection = nn.Linear(factor_width, slot_width, bias=False)
        self.query_projection = nn.Linear(field_width * 3, slot_width, bias=False)
        self.entity_decoder = nn.Linear(slot_width, entities)
        self.relation_decoder = nn.Linear(slot_width, relations)
        self.value_decoder = nn.Linear(slot_width, values)
        self.metadata_decoder = nn.Linear(slot_width, metadata)
        self.temperature = nn.Parameter(torch.tensor(0.07))

    def factors(self, entity: Tensor, relation: Tensor, value: Tensor, metadata: Tensor) -> Tensor:
        return torch.cat((
            self.entity(entity),
            self.relation(relation),
            self.value(value),
            self.metadata(metadata),
        ), dim=-1)

    def encode(self, entity: Tensor, relation: Tensor, value: Tensor, metadata: Tensor) -> Tensor:
        return F.normalize(self.slot_projection(self.factors(entity, relation, value, metadata)), dim=-1)

    def tuple_anchor(self, entity: Tensor, relation: Tensor, value: Tensor, metadata: Tensor) -> Tensor:
        return F.normalize(self.tuple_projection(self.factors(entity, relation, value, metadata)), dim=-1)

    def query(self, entity: Tensor, relation: Tensor, metadata: Tensor) -> Tensor:
        fields = torch.cat((self.entity(entity), self.relation(relation), self.metadata(metadata)), dim=-1)
        return F.normalize(self.query_projection(fields), dim=-1)

    def scores(self, query: Tensor, slots: Tensor) -> Tensor:
        temperature = self.temperature.clamp(0.02, 1.0)
        return torch.einsum("bd,bkd->bk", query, slots) / temperature

    def decode(self, slots: Tensor) -> tuple[Tensor, Tensor, Tensor, Tensor]:
        return (
            self.entity_decoder(slots),
            self.relation_decoder(slots),
            self.value_decoder(slots),
            self.metadata_decoder(slots),
        )


@dataclass(frozen=True)
class CompositionConfig:
    entities: int = 24
    relations: int = 3
    values: int = 36
    candidates: int = 4


def is_held_out(entity: int, relation: int, value: int) -> bool:
    return (entity * 31 + relation * 17 + value * 13) % 5 == 0


def composition_splits(config: CompositionConfig):
    train, held_out = [], []
    for entity in range(config.entities):
        for relation in range(config.relations):
            for value in range(config.values):
                target = held_out if is_held_out(entity, relation, value) else train
                target.append((entity, relation, value))
    return train, held_out


def _same_split_alternative(entity, relation, value, size, want_held_out, field):
    for offset in range(1, size):
        if field == "value":
            candidate = (entity, relation, (value + offset) % size)
        else:
            candidate = ((entity + offset) % size, relation, value)
        if is_held_out(*candidate) == want_held_out:
            return candidate
    raise RuntimeError("unable to construct composition-preserving negative")


def candidate_tuples(positive, config: CompositionConfig, *, held_out: bool):
    entity, relation, value = positive
    wrong_value = _same_split_alternative(
        entity, relation, value, config.values, held_out, "value"
    )
    wrong_entity = _same_split_alternative(
        entity, relation, value, config.entities, held_out, "entity"
    )
    return (
        (entity, relation, value, CANONICAL, "correct"),
        (*wrong_value, CONTRADICTED, "wrong_value"),
        (*wrong_entity, CANONICAL, "wrong_entity"),
        (entity, relation, value, HISTORICAL, "historical"),
    )


def make_batch(combinations, config, batch_size, rng, *, held_out, permute=True):
    selected = [combinations[rng.randrange(len(combinations))] for _ in range(batch_size)]
    candidates, targets, kinds = [], [], []
    for positive in selected:
        rows = list(candidate_tuples(positive, config, held_out=held_out))
        if permute:
            rng.shuffle(rows)
        candidates.append([row[:4] for row in rows])
        kinds.append([row[4] for row in rows])
        targets.append(next(index for index, row in enumerate(rows) if row[4] == "correct"))
    positive = torch.tensor(selected, dtype=torch.long)
    return positive, torch.tensor(candidates, dtype=torch.long), torch.tensor(targets), kinds


def representation_loss(model, positive, candidates, targets):
    entity, relation, value = positive.T
    flat = candidates.view(-1, 4)
    slots = model.encode(*flat.T).view(candidates.shape[0], candidates.shape[1], -1)
    query = model.query(entity, relation, torch.full_like(entity, CANONICAL))
    retrieval = F.cross_entropy(model.scores(query, slots), targets)
    anchor = model.tuple_anchor(entity, relation, value, torch.full_like(entity, CANONICAL))
    contrastive = F.cross_entropy(model.scores(anchor, slots), targets)
    decoded = model.decode(slots)
    canonical = sum(
        F.cross_entropy(logits.flatten(0, 1), flat[:, field])
        for field, logits in enumerate(decoded)
    )
    return retrieval + contrastive + 0.5 * canonical


def evaluate_representation(model, combinations, config, *, permutations=8, seed=101):
    model.eval()
    totals = {"correct": 0, "wrong_value": 0, "wrong_entity": 0, "historical": 0}
    count = 0
    decoded = torch.zeros(4)
    stable = 0
    rng = random.Random(seed)
    with torch.inference_mode():
        for positive in combinations:
            chosen_values = []
            for _ in range(permutations):
                pos, candidates, target, kinds = make_batch(
                    [positive], config, 1, rng, held_out=True, permute=True
                )
                flat = candidates.view(-1, 4)
                slots = model.encode(*flat.T).view(1, config.candidates, -1)
                query = model.query(pos[:, 0], pos[:, 1], torch.zeros(1, dtype=torch.long))
                scores = model.scores(query, slots)[0]
                correct_index = int(target[0])
                prediction = int(scores.argmax())
                selected_slot = slots[0, prediction]
                selected_value = int(model.value_decoder(selected_slot).argmax())
                chosen_values.append(selected_value)
                if prediction == correct_index and selected_value == positive[2]:
                    totals["correct"] += 1
                for index, kind in enumerate(kinds[0]):
                    if kind != "correct" and scores[correct_index] > scores[index]:
                        totals[kind] += 1
                decoded_logits = model.decode(slots[0, correct_index])
                truth = candidates[0, correct_index]
                decoded += torch.tensor([
                    int(logits.argmax() == truth[field])
                    for field, logits in enumerate(decoded_logits)
                ])
                count += 1
            stable += int(len(set(chosen_values)) == 1 and chosen_values[0] == positive[2])
    return {
        "p_only_state_recovery": totals["correct"] / count,
        "hard_negative_accuracy": {
            kind: totals[kind] / count for kind in ("wrong_value", "wrong_entity", "historical")
        },
        "canonical_decode_accuracy": {
            field: float(decoded[index] / count)
            for index, field in enumerate(("entity", "relation", "value", "metadata"))
        },
        "permutation_stability": stable / len(combinations),
        "held_out_combinations": len(combinations),
        "permutations_per_combination": permutations,
    }


def train_and_probe_representation(
    *, steps=600, batch_size=64, slot_width=512, seed=97, evaluation_limit=256
):
    torch.manual_seed(seed)
    config = CompositionConfig()
    train, held_out = composition_splits(config)
    model = FactorizedStateRepresentation(
        config.entities, config.relations, config.values, slot_width=slot_width
    )
    optimizer = torch.optim.AdamW(model.parameters(), lr=3e-3)
    rng = random.Random(seed)
    losses = []
    model.train()
    for _ in range(steps):
        positive, candidates, targets, _ = make_batch(
            train, config, batch_size, rng, held_out=False
        )
        loss = representation_loss(model, positive, candidates, targets)
        optimizer.zero_grad(set_to_none=True)
        loss.backward()
        optimizer.step()
        losses.append(float(loss.detach()))
    probe = evaluate_representation(
        model, held_out[:evaluation_limit], config, permutations=8, seed=seed + 1
    )
    probe.update({
        "training_loss_first": losses[0],
        "training_loss_last": losses[-1],
        "train_combinations": len(train),
        "total_held_out_combinations": len(held_out),
        "slot_width": slot_width,
    })
    return model, config, probe