This repository has been archived by the owner on Dec 16, 2022. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 2.2k
/
predictor.py
414 lines (348 loc) · 17.5 KB
/
predictor.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
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
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
from typing import List, Iterator, Dict, Tuple, Any, Type, Union, Optional
import logging
from os import PathLike
import json
import re
from contextlib import contextmanager
import numpy
import torch
from torch.utils.hooks import RemovableHandle
from torch import Tensor
from torch import backends
from allennlp.common import Registrable, plugins
from allennlp.common.util import JsonDict, sanitize
from allennlp.data import DatasetReader, Instance
from allennlp.data.batch import Batch
from allennlp.models import Model
from allennlp.models.archival import Archive, load_archive
from allennlp.nn import util
logger = logging.getLogger(__name__)
class Predictor(Registrable):
"""
a `Predictor` is a thin wrapper around an AllenNLP model that handles JSON -> JSON predictions
that can be used for serving models through the web API or making predictions in bulk.
"""
def __init__(self, model: Model, dataset_reader: DatasetReader, frozen: bool = True) -> None:
if frozen:
model.eval()
self._model = model
self._dataset_reader = dataset_reader
self.cuda_device = next(self._model.named_parameters())[1].get_device()
self._token_offsets: List[Tensor] = []
def load_line(self, line: str) -> JsonDict:
"""
If your inputs are not in JSON-lines format (e.g. you have a CSV)
you can override this function to parse them correctly.
"""
return json.loads(line)
def dump_line(self, outputs: JsonDict) -> str:
"""
If you don't want your outputs in JSON-lines format
you can override this function to output them differently.
"""
return json.dumps(outputs) + "\n"
def predict_json(self, inputs: JsonDict) -> JsonDict:
instance = self._json_to_instance(inputs)
return self.predict_instance(instance)
def json_to_labeled_instances(self, inputs: JsonDict) -> List[Instance]:
"""
Converts incoming json to a [`Instance`](../data/instance.md),
runs the model on the newly created instance, and adds labels to the
`Instance`s given by the model's output.
# Returns
`List[instance]`
A list of `Instance`'s.
"""
instance = self._json_to_instance(inputs)
self._dataset_reader.apply_token_indexers(instance)
outputs = self._model.forward_on_instance(instance)
new_instances = self.predictions_to_labeled_instances(instance, outputs)
return new_instances
def get_gradients(self, instances: List[Instance]) -> Tuple[Dict[str, Any], Dict[str, Any]]:
"""
Gets the gradients of the loss with respect to the model inputs.
# Parameters
instances : `List[Instance]`
# Returns
`Tuple[Dict[str, Any], Dict[str, Any]]`
The first item is a Dict of gradient entries for each input.
The keys have the form `{grad_input_1: ..., grad_input_2: ... }`
up to the number of inputs given. The second item is the model's output.
# Notes
Takes a `JsonDict` representing the inputs of the model and converts
them to [`Instances`](../data/instance.md)), sends these through
the model [`forward`](../models/model.md#forward) function after registering hooks on the embedding
layer of the model. Calls `backward` on the loss and then removes the
hooks.
"""
# set requires_grad to true for all parameters, but save original values to
# restore them later
original_param_name_to_requires_grad_dict = {}
for param_name, param in self._model.named_parameters():
original_param_name_to_requires_grad_dict[param_name] = param.requires_grad
param.requires_grad = True
embedding_gradients: List[Tensor] = []
hooks: List[RemovableHandle] = self._register_embedding_gradient_hooks(embedding_gradients)
for instance in instances:
self._dataset_reader.apply_token_indexers(instance)
dataset = Batch(instances)
dataset.index_instances(self._model.vocab)
dataset_tensor_dict = util.move_to_device(dataset.as_tensor_dict(), self.cuda_device)
# To bypass "RuntimeError: cudnn RNN backward can only be called in training mode"
with backends.cudnn.flags(enabled=False):
outputs = self._model.make_output_human_readable(
self._model.forward(**dataset_tensor_dict) # type: ignore
)
loss = outputs["loss"]
# Zero gradients.
# NOTE: this is actually more efficient than calling `self._model.zero_grad()`
# because it avoids a read op when the gradients are first updated below.
for p in self._model.parameters():
p.grad = None
loss.backward()
for hook in hooks:
hook.remove()
grad_dict = dict()
for idx, grad in enumerate(embedding_gradients):
key = "grad_input_" + str(idx + 1)
grad_dict[key] = grad.detach().cpu().numpy()
# restore the original requires_grad values of the parameters
for param_name, param in self._model.named_parameters():
param.requires_grad = original_param_name_to_requires_grad_dict[param_name]
return grad_dict, outputs
def get_interpretable_layer(self) -> torch.nn.Module:
"""
Returns the input/embedding layer of the model.
If the predictor wraps around a non-AllenNLP model,
this function should be overridden to specify the correct input/embedding layer.
For the cases where the input layer _is_ an embedding layer, this should be the
layer 0 of the embedder.
"""
try:
return util.find_embedding_layer(self._model)
except RuntimeError:
raise RuntimeError(
"If the model does not use `TextFieldEmbedder`, please override "
"`get_interpretable_layer` in your predictor to specify the embedding layer."
)
def get_interpretable_text_field_embedder(self) -> torch.nn.Module:
"""
Returns the first `TextFieldEmbedder` of the model.
If the predictor wraps around a non-AllenNLP model,
this function should be overridden to specify the correct embedder.
"""
try:
return util.find_text_field_embedder(self._model)
except RuntimeError:
raise RuntimeError(
"If the model does not use `TextFieldEmbedder`, please override "
"`get_interpretable_text_field_embedder` in your predictor to specify "
"the embedding layer."
)
def _register_embedding_gradient_hooks(self, embedding_gradients):
"""
Registers a backward hook on the embedding layer of the model. Used to save the gradients
of the embeddings for use in get_gradients()
When there are multiple inputs (e.g., a passage and question), the hook
will be called multiple times. We append all the embeddings gradients
to a list.
We additionally add a hook on the _forward_ pass of the model's `TextFieldEmbedder` to save
token offsets, if there are any. Having token offsets means that you're using a mismatched
token indexer, so we need to aggregate the gradients across wordpieces in a token. We do
that with a simple sum.
"""
def hook_layers(module, grad_in, grad_out):
grads = grad_out[0]
if self._token_offsets:
# If you have a mismatched indexer with multiple TextFields, it's quite possible
# that the order we deal with the gradients is wrong. We'll just take items from
# the list one at a time, and try to aggregate the gradients. If we got the order
# wrong, we should crash, so you'll know about it. If you get an error because of
# that, open an issue on github, and we'll see what we can do. The intersection of
# multiple TextFields and mismatched indexers is pretty small (currently empty, that
# I know of), so we'll ignore this corner case until it's needed.
offsets = self._token_offsets.pop(0)
span_grads, span_mask = util.batched_span_select(grads.contiguous(), offsets)
span_mask = span_mask.unsqueeze(-1)
span_grads *= span_mask # zero out paddings
span_grads_sum = span_grads.sum(2)
span_grads_len = span_mask.sum(2)
# Shape: (batch_size, num_orig_tokens, embedding_size)
grads = span_grads_sum / torch.clamp_min(span_grads_len, 1)
# All the places where the span length is zero, write in zeros.
grads[(span_grads_len == 0).expand(grads.shape)] = 0
embedding_gradients.append(grads)
def get_token_offsets(module, inputs, outputs):
offsets = util.get_token_offsets_from_text_field_inputs(inputs)
if offsets is not None:
self._token_offsets.append(offsets)
hooks = []
text_field_embedder = self.get_interpretable_text_field_embedder()
hooks.append(text_field_embedder.register_forward_hook(get_token_offsets))
embedding_layer = self.get_interpretable_layer()
hooks.append(embedding_layer.register_backward_hook(hook_layers))
return hooks
@contextmanager
def capture_model_internals(self, module_regex: str = ".*") -> Iterator[dict]:
"""
Context manager that captures the internal-module outputs of
this predictor's model. The idea is that you could use it as follows:
```
with predictor.capture_model_internals() as internals:
outputs = predictor.predict_json(inputs)
return {**outputs, "model_internals": internals}
```
"""
results = {}
hooks = []
# First we'll register hooks to add the outputs of each module to the results dict.
def add_output(idx: int):
def _add_output(mod, _, outputs):
results[idx] = {"name": str(mod), "output": sanitize(outputs)}
return _add_output
regex = re.compile(module_regex)
for idx, (name, module) in enumerate(self._model.named_modules()):
if regex.fullmatch(name) and module != self._model:
hook = module.register_forward_hook(add_output(idx))
hooks.append(hook)
# If you capture the return value of the context manager, you get the results dict.
yield results
# And then when you exit the context we remove all the hooks.
for hook in hooks:
hook.remove()
def predict_instance(self, instance: Instance) -> JsonDict:
self._dataset_reader.apply_token_indexers(instance)
outputs = self._model.forward_on_instance(instance)
return sanitize(outputs)
def predictions_to_labeled_instances(
self, instance: Instance, outputs: Dict[str, numpy.ndarray]
) -> List[Instance]:
"""
This function takes a model's outputs for an Instance, and it labels that instance according
to the `outputs`. This function is used to (1) compute gradients of what the model predicted;
(2) label the instance for the attack. For example, (a) for the untargeted attack for classification
this function labels the instance according to the class with the highest probability; (b) for
targeted attack, it directly constructs fields from the given target.
The return type is a list because in some tasks there are multiple predictions in the output
(e.g., in NER a model predicts multiple spans). In this case, each instance in the returned list of
Instances contains an individual entity prediction as the label.
"""
raise RuntimeError("implement this method for model interpretations or attacks")
def _json_to_instance(self, json_dict: JsonDict) -> Instance:
"""
Converts a JSON object into an [`Instance`](../data/instance.md)
and a `JsonDict` of information which the `Predictor` should pass through,
such as tokenized inputs.
"""
raise NotImplementedError
def predict_batch_json(self, inputs: List[JsonDict]) -> List[JsonDict]:
instances = self._batch_json_to_instances(inputs)
return self.predict_batch_instance(instances)
def predict_batch_instance(self, instances: List[Instance]) -> List[JsonDict]:
for instance in instances:
self._dataset_reader.apply_token_indexers(instance)
outputs = self._model.forward_on_instances(instances)
return sanitize(outputs)
def _batch_json_to_instances(self, json_dicts: List[JsonDict]) -> List[Instance]:
"""
Converts a list of JSON objects into a list of `Instance`s.
By default, this expects that a "batch" consists of a list of JSON blobs which would
individually be predicted by `predict_json`. In order to use this method for
batch prediction, `_json_to_instance` should be implemented by the subclass, or
if the instances have some dependency on each other, this method should be overridden
directly.
"""
instances = []
for json_dict in json_dicts:
instances.append(self._json_to_instance(json_dict))
return instances
@classmethod
def from_path(
cls,
archive_path: Union[str, PathLike],
predictor_name: str = None,
cuda_device: int = -1,
dataset_reader_to_load: str = "validation",
frozen: bool = True,
import_plugins: bool = True,
overrides: Union[str, Dict[str, Any]] = "",
**kwargs,
) -> "Predictor":
"""
Instantiate a `Predictor` from an archive path.
If you need more detailed configuration options, such as overrides,
please use `from_archive`.
# Parameters
archive_path : `Union[str, PathLike]`
The path to the archive.
predictor_name : `str`, optional (default=`None`)
Name that the predictor is registered as, or None to use the
predictor associated with the model.
cuda_device : `int`, optional (default=`-1`)
If `cuda_device` is >= 0, the model will be loaded onto the
corresponding GPU. Otherwise it will be loaded onto the CPU.
dataset_reader_to_load : `str`, optional (default=`"validation"`)
Which dataset reader to load from the archive, either "train" or
"validation".
frozen : `bool`, optional (default=`True`)
If we should call `model.eval()` when building the predictor.
import_plugins : `bool`, optional (default=`True`)
If `True`, we attempt to import plugins before loading the predictor.
This comes with additional overhead, but means you don't need to explicitly
import the modules that your predictor depends on as long as those modules
can be found by `allennlp.common.plugins.import_plugins()`.
overrides : `Union[str, Dict[str, Any]]`, optional (default = `""`)
JSON overrides to apply to the unarchived `Params` object.
**kwargs : `Any`
Additional key-word arguments that will be passed to the `Predictor`'s
`__init__()` method.
# Returns
`Predictor`
A Predictor instance.
"""
if import_plugins:
plugins.import_plugins()
return Predictor.from_archive(
load_archive(archive_path, cuda_device=cuda_device, overrides=overrides),
predictor_name,
dataset_reader_to_load=dataset_reader_to_load,
frozen=frozen,
extra_args=kwargs,
)
@classmethod
def from_archive(
cls,
archive: Archive,
predictor_name: str = None,
dataset_reader_to_load: str = "validation",
frozen: bool = True,
extra_args: Optional[Dict[str, Any]] = None,
) -> "Predictor":
"""
Instantiate a `Predictor` from an [`Archive`](../models/archival.md);
that is, from the result of training a model. Optionally specify which `Predictor`
subclass; otherwise, we try to find a corresponding predictor in `DEFAULT_PREDICTORS`, or if
one is not found, the base class (i.e. `Predictor`) will be used. Optionally specify
which [`DatasetReader`](../data/dataset_readers/dataset_reader.md) should be loaded;
otherwise, the validation one will be used if it exists followed by the training dataset reader.
Optionally specify if the loaded model should be frozen, meaning `model.eval()` will be called.
"""
# Duplicate the config so that the config inside the archive doesn't get consumed
config = archive.config.duplicate()
if not predictor_name:
model_type = config.get("model").get("type")
model_class, _ = Model.resolve_class_name(model_type)
predictor_name = model_class.default_predictor
predictor_class: Type[Predictor] = (
Predictor.by_name(predictor_name) if predictor_name is not None else cls # type: ignore
)
if dataset_reader_to_load == "validation":
dataset_reader = archive.validation_dataset_reader
else:
dataset_reader = archive.dataset_reader
model = archive.model
if frozen:
model.eval()
if extra_args is None:
extra_args = {}
return predictor_class(model, dataset_reader, **extra_args)