evaluator.py 9.99 KB
Newer Older
lintangsutawika's avatar
lintangsutawika committed
1
import random
Leo Gao's avatar
Leo Gao committed
2
import itertools
lintangsutawika's avatar
lintangsutawika committed
3
4
import collections

5
6
import torch

7
import numpy as np
lintangsutawika's avatar
lintangsutawika committed
8
9

import lm_eval.api
10
import lm_eval.tasks
lintangsutawika's avatar
lintangsutawika committed
11
import lm_eval.models
lintangsutawika's avatar
lintangsutawika committed
12
import lm_eval.api.metrics
lintangsutawika's avatar
lintangsutawika committed
13

lintangsutawika's avatar
lintangsutawika committed
14
15
16
17
from lm_eval.utils import (
    positional_deprecated,
    run_task_tests,
    make_table,
18
    create_iterator,
lintangsutawika's avatar
lintangsutawika committed
19
20
    get_git_commit_hash,
)
21

lintangsutawika's avatar
lintangsutawika committed
22
23
from lm_eval.logger import eval_logger

Fabrizio Milo's avatar
Fabrizio Milo committed
24

25
@positional_deprecated
Fabrizio Milo's avatar
Fabrizio Milo committed
26
27
28
29
30
31
32
33
34
35
36
37
38
def simple_evaluate(
    model,
    model_args=None,
    tasks=[],
    num_fewshot=0,
    batch_size=None,
    device=None,
    no_cache=False,
    limit=None,
    bootstrap_iters=100000,
    check_integrity=False,
    decontamination_ngrams_path=None,
):
39

40
    """Instantiate and evaluate a model on a list of tasks.
41

42
43
44
    :param model: Union[str, LM]
        Name of model or LM object, see lm_eval.models.get_model
    :param model_args: Optional[str]
Fabrizio Milo's avatar
Fabrizio Milo committed
45
        String arguments for each model class, see LM.create_from_arg_string.
46
47
        Ignored if `model` argument is a LM object.
    :param tasks: list[Union[str, Task]]
Leo Gao's avatar
Leo Gao committed
48
        List of task names or Task objects. Task objects will be taken to have name task.EVAL_HARNESS_NAME if defined and type(task).__name__ otherwise.
49
50
51
52
53
    :param num_fewshot: int
        Number of examples in few-shot context
    :param batch_size: int, optional
        Batch size for model
    :param device: str, optional
54
        PyTorch device (e.g. "cpu" or "cuda:0") for running models
55
    :param no_cache: bool
Leo Gao's avatar
Leo Gao committed
56
        Whether or not to cache
57
58
59
60
    :param limit: int, optional
        Limit the number of examples per task (only use this for testing)
    :param bootstrap_iters:
        Number of iterations for bootstrap statistics
Stephen Hogg's avatar
Stephen Hogg committed
61
62
    :param check_integrity: bool
        Whether to run the relevant part of the test suite for the tasks
63
    :return
64
        Dictionary of results
65
    """
66
67
68
    random.seed(1234)
    np.random.seed(1234)

69
70
71
    assert tasks != [], "No tasks specified"

    if isinstance(model, str):
Fabrizio Milo's avatar
Fabrizio Milo committed
72
73
        if model_args is None:
            model_args = ""
74
        lm = lm_eval.api.model.get_model(model).create_from_arg_string(
Fabrizio Milo's avatar
Fabrizio Milo committed
75
76
            model_args, {"batch_size": batch_size, "device": device}
        )
77
    else:
78
        assert isinstance(model, lm_eval.api.model.LM)
79
        lm = model
80

lintangsutawika's avatar
update  
lintangsutawika committed
81
    task_dict = lm_eval.tasks.get_task_dict(tasks, num_fewshot=num_fewshot)
Jonathan Tow's avatar
Merge  
Jonathan Tow committed
82

Stephen Hogg's avatar
Stephen Hogg committed
83
    if check_integrity:
84
        run_task_tests(task_list=tasks)
Stephen Hogg's avatar
Stephen Hogg committed
85

86
87
88
89
    results = evaluate(
        lm=lm,
        task_dict=task_dict,
        limit=limit,
Niklas Muennighoff's avatar
Niklas Muennighoff committed
90
        bootstrap_iters=bootstrap_iters,
Fabrizio Milo's avatar
Fabrizio Milo committed
91
        decontamination_ngrams_path=decontamination_ngrams_path,
92
    )
93

94
95
96
97
98
99
100
101
102
103
104
105
    if lm.rank == 0:
        # add info about the model and few shot config
        results["config"] = {
            "model": model,
            "model_args": model_args,
            "num_fewshot": num_fewshot,
            "batch_size": batch_size,
            "device": device,
            "no_cache": no_cache,
            "limit": limit,
            "bootstrap_iters": bootstrap_iters,
        }
106
        results["git_hash"] = get_git_commit_hash()
107
108
109
        return results
    else:
        return None
110

Leo Gao's avatar
Leo Gao committed
111

112
decontaminate_suffix = "_decontaminate"
Leo Gao's avatar
Leo Gao committed
113

Fabrizio Milo's avatar
Fabrizio Milo committed
114

115
@positional_deprecated
Fabrizio Milo's avatar
Fabrizio Milo committed
116
117
118
119
120
121
122
def evaluate(
    lm,
    task_dict,
    limit=None,
    bootstrap_iters=100000,
    decontamination_ngrams_path=None,
):
123
124
125
126
127
    """Instantiate and evaluate a model on a list of tasks.

    :param lm: obj
        Language Model
    :param task_dict: dict[str, Task]
Leo Gao's avatar
Leo Gao committed
128
        Dictionary of tasks. Tasks will be taken to have name task.EVAL_HARNESS_NAME if defined and type(task).__name__ otherwise.
129
130
131
132
133
134
135
136
137
    :param num_fewshot: int
        Number of examples in few-shot context
    :param limit: int, optional
        Limit the number of examples per task (only use this for testing)
    :param bootstrap_iters:
        Number of iterations for bootstrap statistics
    :return
        Dictionary of results
    """
138

lintangsutawika's avatar
lintangsutawika committed
139
    # decontaminate = decontamination_ngrams_path is not None
140

Leo Gao's avatar
Leo Gao committed
141
    results = collections.defaultdict(dict)
Leo Gao's avatar
Leo Gao committed
142
    versions = collections.defaultdict(dict)
Leo Gao's avatar
Leo Gao committed
143
144

    requests = collections.defaultdict(list)
lintangsutawika's avatar
lintangsutawika committed
145
    # requests_origin = collections.defaultdict(list)
Leo Gao's avatar
Leo Gao committed
146

lintangsutawika's avatar
lintangsutawika committed
147
    # docs = {}
Leo Gao's avatar
Leo Gao committed
148

149
    # get lists of each type of request
150
    for task_name, task in task_dict.items():
Leo Gao's avatar
Leo Gao committed
151
        versions[task_name] = task.VERSION
lintangsutawika's avatar
lintangsutawika committed
152

Leo Gao's avatar
Leo Gao committed
153
        # deterministically shuffle docs and chop off the first `limit` because sometimes docs are in some kind of order
154
155
156
157
158
        # task_docs = list(task_doc_func())
        # rnd = random.Random()
        # rnd.seed(42)
        # rnd.shuffle(task_docs)

159
160
        task.build_all_requests(limit=limit, rank=lm.rank, world_size=lm.world_size)

161
        # aggregate Instances by LM method requested to get output.
lintangsutawika's avatar
lintangsutawika committed
162
163
164
165
166
167
        reqtype = (
            "loglikelihood"
            if task.OUTPUT_TYPE == "multiple_choice"
            else task.OUTPUT_TYPE
        )  # TODO: this is hacky, fix in task.py
        requests[reqtype].extend(task.instances)
168
169

        if lm.world_size > 1:
170
171
172
173
            instances_rnk = torch.tensor(len(task._instances), device=lm.device)
            gathered_item = (
                lm.accelerator.gather(instances_rnk).cpu().detach().numpy().tolist()
            )
174

175
            # compute number of pseudobatches to pad with (FSDP/DDP require even batches among ranks)
176
            numpad = max(gathered_item) - gathered_item[lm.rank]
177

178
    ### Run LM on inputs, get all outputs ###
Leo Gao's avatar
Leo Gao committed
179
180
    # execute each type of request
    for reqtype, reqs in requests.items():
lintangsutawika's avatar
lintangsutawika committed
181
        eval_logger.info("Running {} requests".format(reqtype))
182
183
184
185
        # create `K` copies of each request `req` based off `K = req.repeats`
        cloned_reqs = []
        for req in reqs:
            cloned_reqs.extend([req] * req.repeats)
lintangsutawika's avatar
lintangsutawika committed
186

187
        if (lm.world_size > 1) and (numpad > 0):
188
189
190
            for _ in range(numpad):
                cloned_reqs.extend([req] * req.repeats)

191
192
193
194
195
196
197
        # run requests through model
        resps = getattr(lm, reqtype)(cloned_reqs)

        # put responses from model into a list of length K for each request.
        for x, req in zip(resps, cloned_reqs):
            req.resps.append(x)

198
199
200
    if lm.world_size > 1:
        lm.accelerator.wait_for_everyone()

201
202
203
204
205
206
    ### Postprocess outputs ###
    # TODO: del model here, maybe (idea: allow user to specify device of e.g. reward model separately)
    for task_name, task in task_dict.items():
        task.apply_filters()

    ### Collect values of metrics on all datapoints ###
lintangsutawika's avatar
lintangsutawika committed
207
    # TODO: make metric configurable, add metric registry
Leo Gao's avatar
Leo Gao committed
208
209
210
    vals = collections.defaultdict(list)

    # unpack results and sort back in order and return control to Task
211
212
213
214
    for task_name, task in task_dict.items():
        # calculate values for each filter setup (TODO: make getting list of keys cleaner)
        # TODO: make it possible to use a different metric per key
        for key in task.instances[0].filtered_resps.keys():
215
216
217
218
            doc_iterator = (
                itertools.islice(
                    enumerate(task.test_docs()), lm.rank, limit, lm.world_size
                )
lintangsutawika's avatar
lintangsutawika committed
219
                if task.has_test_docs()
220
221
222
223
                else itertools.islice(
                    enumerate(task.validation_docs()), lm.rank, limit, lm.world_size
                )
            )
224
            for doc_id, doc in doc_iterator:
225
226
                # subset instances to only this document id ; sort by idx
                requests = list(filter(lambda x: x.doc_id == doc_id, task.instances))
227
                requests.sort(key=lambda x: x.idx)
lintangsutawika's avatar
lintangsutawika committed
228
229
230
                metrics = task.process_results(
                    doc, [req.filtered_resps[key] for req in requests]
                )
231
232
233
                for metric, value in metrics.items():
                    vals[(task_name, key, metric)].append(value)

234
    if lm.world_size > 1:
235
        # if multigpu, then gather data across all ranks
236
237
        vals_torch = collections.defaultdict(list)
        for (task_name, key, metric), items in vals.items():
238
239

            numitem = 0
240
            if type(items[0]) == tuple:
241
242
                numitem = len(items[0])

243
244
            # distributed gather requires all ranks to have same dimensions
            # so we pad out with float32 min value
245
            pad_value = torch.finfo(torch.float32).min
246
247
248
249
250
251
            metrics_tensor = torch.tensor(items, device=lm.device)

            original_dtype = metrics_tensor.dtype  # store original dtype
            torch_device_tensor = lm.accelerator.pad_across_processes(
                metrics_tensor.to(torch.float32), pad_index=pad_value
            )
252
            gathered_item = lm.accelerator.gather(torch_device_tensor)
253

254
            if numitem > 0:
255
                gathered_filtered = gathered_item[gathered_item[:, 0] != pad_value]
256
257
            else:
                gathered_filtered = gathered_item[gathered_item != pad_value]
258
259
260
261

            gathered_item = (
                gathered_filtered.to(original_dtype).cpu().detach().numpy().tolist()
            )
262
263
264
            # reconvert if we were passed a tuple of values
            if numitem > 0:
                gathered_item = [tuple(g) for g in gathered_item]
265

266
267
            if lm.rank == 0:
                vals_torch[(task_name, key, metric)] = gathered_item
268

269
        vals = vals_torch
270

271
272
273
274
275
    if lm.rank == 0:
        ### Aggregate results over all datapoints ###
        # aggregate results ; run bootstrap CIs
        for (task_name, key, metric), items in vals.items():
            task = task_dict[task_name]
276
277
278
            results[task_name][metric + " - filter=" + key] = task.aggregation()[
                metric
            ](items)
Leo Gao's avatar
Leo Gao committed
279

280
281
            # hotfix: bleu, chrf, ter seem to be really expensive to bootstrap
            # so we run them less iterations. still looking for a cleaner way to do this
282

lintangsutawika's avatar
lintangsutawika committed
283
            stderr = lm_eval.api.metrics.stderr_for_metric(
284
285
286
287
288
289
290
                metric=task.aggregation()[metric],
                bootstrap_iters=min(bootstrap_iters, 1000)
                if metric in ["bleu", "chrf", "ter"]
                else bootstrap_iters,
            )

            if stderr is not None:
291
292
293
                results[task_name][metric + " - filter=" + key + "_stderr"] = stderr(
                    items
                )
Fabrizio Milo's avatar
Fabrizio Milo committed
294

295
        return {"results": dict(results), "versions": dict(versions)}
Fabrizio Milo's avatar
Fabrizio Milo committed
296

297
298
    else:
        return None