-
Notifications
You must be signed in to change notification settings - Fork 24
/
retomaton.py
250 lines (207 loc) · 11 KB
/
retomaton.py
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
242
243
244
245
246
247
248
249
250
from collections import defaultdict
import os
import logging
import pickle
import time
import numpy as np
import torch
from torch import nn
from enum import Enum, auto
from pathlib import Path
import glob
from itertools import zip_longest
from tqdm import tqdm
from knnlm import KNNWrapper, get_dstore_path
import faiss
import faiss.contrib.torch_utils
from faiss import IndexFlatL2
import scipy.sparse as sp
logger = logging.getLogger(__name__)
logger.setLevel(20)
class RetomatonWrapper(KNNWrapper):
def __init__(self, no_pointer=False, min_knns=1, max_knns=1024, members=None, **kwargs):
super().__init__(**kwargs)
self.no_pointer = no_pointer
self.min_knns = min_knns
self.max_knns = max_knns
if members is None:
available_member_files = glob.glob(f'{self.dstore_dir}/members*')
if len(available_member_files) == 0:
logger.info(f'No member files found in {self.dstore_dir}, not using clustering')
else:
members = available_member_files[0]
logger.info(f'Found the following cluster members files: {available_member_files}')
logger.info(f'Using members file {members}')
if members is None:
self.extend_pointers_using_clusters = lambda pointers: pointers
else:
with open(members, 'rb') as file:
self.members = pickle.load(file)
members_for_indices = np.nonzero(self.members[np.arange(self.members.shape[0])])
self.cluster = torch.zeros((self.dstore_size, ), dtype=torch.int32).to(self.device)
self.cluster[members_for_indices[1]] = torch.from_numpy(members_for_indices[0]).to(self.device)
self.generate_cur_knns = torch.tensor([], dtype=torch.int64)
self.generate_cur_dists = torch.tensor([], dtype=torch.float32)
self.no_lookup_counter_history = []
def post_forward_hook(self, module, input, output):
shift = 0 if self.is_encoder_decoder else 1
if self.labels is None:
# In "generate" mode, we don't support yet tracking of the beam search hypotheses across time,
# which we need to track in order to implement RetoMaton correctly.
# In the meantime, use kNN-LM's generate
return super().post_forward_hook(module, input, output)
lm_logits = output
lm_logits = torch.nn.functional.log_softmax(lm_logits, dim=-1) # (batch, time, vocab)
queries = self.activation_capturer.captured # (batch, time, dim)
shifted_labels = self.labels[:, shift:]
nonpad_mask = torch.cat([
shifted_labels != -100,
torch.zeros([self.labels.shape[0], shift], dtype=torch.bool).to(self.device)
], axis=-1)
captured_labels = shifted_labels[shifted_labels != -100] # (nonpad)
queries = queries[nonpad_mask] # (nonpad, dim)
lm_logits = lm_logits[nonpad_mask] # (nonpad, vocab)
all_knn_probs = []
cur_knns = torch.tensor([], dtype=torch.int64)
cur_dists = torch.tensor([], dtype=torch.float32)
no_lookup_counter = 0
for timestep_query, label in zip_longest(queries, captured_labels):
perform_search = False
extended_pointers = None
pointers = cur_knns + 1
if self.no_pointer or cur_knns.numel() < self.min_knns:
perform_search = True
self.no_lookup_counter_history.append(no_lookup_counter)
no_lookup_counter = 0
else:
no_lookup_counter += 1
if self.no_pointer:
extended_pointers = None
elif pointers.numel() >= self.max_knns:
extended_pointers = pointers[:self.max_knns]
else:
extended_pointers = self.extend_pointers_using_clusters(pointers)
# (vocab_size, ) , (k, ), (k, ), (k, )
cur_knn_log_prob, knns, dists, vals_at_knns = self.get_knn_log_prob(
timestep_query,
pointers=extended_pointers,
perform_search=perform_search)
all_knn_probs.append(cur_knn_log_prob)
if not self.no_pointer and label is not None:
vals_are_correct_and_pointer_available = (vals_at_knns == label) & (knns < self.dstore_size - 1)
cur_knns = knns[vals_are_correct_and_pointer_available]
cur_dists = dists[vals_are_correct_and_pointer_available]
cur_knns = cur_knns[cur_dists.argsort(descending=True)]
interpolated_scores = KNNWrapper.interpolate(torch.stack(all_knn_probs), lm_logits, self.lmbda) # (nonpad, vocab)
output[nonpad_mask] = interpolated_scores
return output
def get_knn_log_prob(self, query, pointers, perform_search):
pointer_dists = torch.tensor([[]]).to(self.device)
if pointers is not None and pointers.numel() > 0 and not self.recompute_dists:
pointer_vectors = self.reconstruct_ids(pointers)
pointer_dists = self.dist_func(query, pointer_vectors)
if perform_search:
dists, knns = self.get_knns(query.unsqueeze(0)) # (1, k)
dists, knns = dists.squeeze(0), knns.squeeze(0) # (k, )
if pointers is not None and pointers.numel() > 0:
knns = torch.cat([knns, pointers], axis=-1)
dists = torch.cat([dists, pointer_dists], axis=-1)
else:
knns = pointers
dists = pointer_dists
if self.recompute_dists:
knns_vecs = torch.from_numpy(self.keys[knns]).to(self.device)
dists = self.dist_func(query, knns_vecs)
neg_dists = -dists
knn_log_probs, vals_at_knns = self.knns_to_log_prob(knns, neg_dists)
return knn_log_probs, knns, neg_dists, vals_at_knns
def extend_pointers_using_clusters(self, pointers):
if pointers.numel() == 0:
return pointers
# Don't take the same cluster twice
# pointers = pointers.numpy()
clusters, cluster_counts = torch.unique(self.cluster[pointers], return_counts=True)
# Take smaller clusters first
clusters = clusters[torch.argsort(-cluster_counts)]
members = torch.from_numpy(np.nonzero(self.members[clusters.cpu().numpy()])[1]).to(self.device)
# Prefer datastore entries that were directly pointed to by the previous time step's
# datastore entries, over other members of their cluster
extended_pointers = torch.cat([pointers, members])
if len(extended_pointers) > self.max_knns:
extended_pointers = extended_pointers[:self.max_knns]
return extended_pointers
def reconstruct_ids(self, ids):
# Converting to numpy only because GPU indexes do not support reconstructing vectors:
# https://github.com/facebookresearch/faiss/issues/2181
ids = ids.cpu().numpy()
# faiss's index.reconstruct supports a single ID at a time, so batching is performed
# using numpy.vectorize:
# https://github.com/facebookresearch/faiss/issues/1163
reconstruct_func = np.vectorize(lambda x: self.reconstruct_index.reconstruct(int(x)), otypes=[object])
vectors = reconstruct_func(ids)
vectors = np.stack(vectors).reshape(ids.shape + (self.dimension, ))
vectors_t = torch.from_numpy(vectors).to(self.device)
return vectors_t
def get_metrics(self):
return {'lookups_saved': np.sum(self.no_lookup_counter_history)/
(np.sum(self.no_lookup_counter_history) + len(self.no_lookup_counter_history)),
}
def break_out(self):
super().break_out()
self.print_stats()
def print_stats(self):
if len(self.no_lookup_counter_history) > 0:
saved = np.sum(self.no_lookup_counter_history) / \
(np.sum(self.no_lookup_counter_history) + len(self.no_lookup_counter_history))
logger.info(f'Lookups saved: {saved*100}%')
def cluster_dstore(self, num_clusters, sample_size, model, batch_size=500000):
keys_vals_prefix = get_dstore_path(self.dstore_dir, model.config.model_type, self.dstore_size, self.dimension)
keys = np.memmap(f'{keys_vals_prefix}_keys.npy', dtype=np.float16, mode='r', shape=(self.dstore_size, self.dimension))
if sample_size > self.dstore_size:
logger.info('Taking all data for training')
to_cluster = keys[:]
else:
idx = np.random.RandomState(1).choice(np.arange(self.dstore_size), size=sample_size, replace=False)
to_cluster = keys[idx]
to_cluster = to_cluster.astype(np.float32)
logger.info(f'Starting to cluster {sample_size} examples into {num_clusters} clusters')
kmeans = faiss.Kmeans(self.dimension, num_clusters, niter=20, verbose=True, gpu=True, seed=1)
kmeans.train(to_cluster)
logger.info(f'Done training, assigning {self.dstore_size} examples to the clusters')
index = IndexFlatL2(self.dimension)
index.add(kmeans.centroids)
logger.info('Index created, added centroids')
if torch.cuda.is_available() and torch.cuda.device_count() > 0:
logger.info('Moving index to GPU')
co = faiss.GpuClonerOptions()
co.useFloat16 = True
index = faiss.index_cpu_to_gpu(faiss.StandardGpuResources(), 0, index, co)
logger.info('Moved index to GPU')
start = 0
centroid_ids = []
logger.info('Starting to add tokens')
while start < self.dstore_size:
end = min(self.dstore_size, start + batch_size)
to_search = keys[start:end].copy()
_, key_i = index.search(torch.from_numpy(to_search).float(), 1)
centroid_ids.append(key_i.squeeze())
start += batch_size
if (start % 1000000) == 0:
print('Assigned %d tokens so far' % start)
centroid_ids = np.concatenate(centroid_ids)
logger.info('Processing the mapping of cluster->members')
parent_cluster = centroid_ids
cluster_to_members = defaultdict(set)
for key_i, cluster in tqdm(enumerate(parent_cluster), total=self.dstore_size):
cluster_to_members[cluster.item()].add(key_i)
row_ind = [k for k, v in cluster_to_members.items() for _ in range(len(v))]
col_ind = [i for ids in cluster_to_members.values() for i in ids]
members_sp = sp.csr_matrix(([1]*len(row_ind), (row_ind, col_ind)))
members_filename = get_members_path(self.dstore_dir,
model.config.model_type, self.dstore_size, self.dimension,
sample_size, num_clusters)
with open(members_filename, 'wb') as f:
pickle.dump(members_sp, f)
logger.info(f'Done, found {len(cluster_to_members)} clusters, written to {members_filename}')
def get_members_path(dstore_dir, model_type, dstore_size, dimension, sample_size, num_clusters):
return f'{dstore_dir}/members_{model_type}_{dstore_size}_{dimension}_{sample_size}_{num_clusters}.pkl'