# pytorch-lr-finder
# 简介
该项目为 PyTorch 学习率查找器
学习率范围测试是一种提供有关最佳学习率的宝贵信息的测试。在预训练运行期间，学习率在两个边界之间线性或指数增加。较低的初始学习率允许网络开始收敛，随着学习率的增加，它最终会变得太大，网络会发散。
来自 fastai 的调整版本：以指数方式增加学习率并计算每个学习率的训练损失。
lr_finder.plot()绘制训练损失与对数学习率的关系。Leslie Smith 的方法：线性增加学习率并计算每个学习率的评估损失。
lr_finder.plot()绘制评估损失与学习率的关系图。这种方法通常会产生更精确的曲线，因为评估损失更容易发散，但执行测试所需的时间会更长，尤其是在评估数据集很大的情况下。# test_lr_finder.py 代码分析
# 该 test 原代码
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
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448import pytest
from torch.utils.data import DataLoader
from torch_lr_finder import LRFinder
from torch_lr_finder.lr_finder import (
DataLoaderIter, TrainDataLoaderIter, ValDataLoaderIter
)
import task as mod_task
import dataset as mod_dataset
import matplotlib.pyplot as plt
# Check available backends for mixed precision training
AVAILABLE_AMP_BACKENDS = []
try:
import apex.amp
AVAILABLE_AMP_BACKENDS.append("apex")
except ImportError:
pass
try:
import torch.amp
AVAILABLE_AMP_BACKENDS.append("torch")
except ImportError:
pass
def collect_task_classes():
names = [v for v in dir(mod_task) if v.endswith("Task") and v != "BaseTask"]
attrs = [getattr(mod_task, v) for v in names]
classes = [v for v in attrs if issubclass(v, mod_task.BaseTask)]
return classes
def prepare_lr_finder(task, **kwargs):
model = task.model
optimizer = task.optimizer
criterion = task.criterion
config = {
"device": kwargs.get("device", None),
"memory_cache": kwargs.get("memory_cache", True),
"cache_dir": kwargs.get("cache_dir", None),
"amp_backend": kwargs.get("amp_backend", None),
"amp_config": kwargs.get("amp_config", None),
"grad_scaler": kwargs.get("grad_scaler", None),
}
lr_finder = LRFinder(model, optimizer, criterion, **config)
return lr_finder
def get_optim_lr(optimizer):
return [grp["lr"] for grp in optimizer.param_groups]
def run_loader_iter(loader_iter, desired_runs=None):
"""Run a `DataLoaderIter` object for specific times.
Arguments:
loader_iter (torch_lr_finder.DataLoaderIter): the iterator to test.
desired_runs (int, optional): times that iterator should be iterated.
If it's not given, `len(loader_iter.data_loader)` will be used.
Returns:
is_achieved (bool): False if `loader_iter` cannot be iterated specific
times. It usually means `loader_iter` has raised `StopIteration`.
"""
assert isinstance(loader_iter, DataLoaderIter)
if desired_runs is None:
desired_runs = len(loader_iter.data_loader)
count = 0
try:
for i in range(desired_runs):
next(loader_iter)
count += 1
except StopIteration:
return False
return desired_runs == count
class TestRangeTest:
def test_run(self, cls_task):
task = cls_task()
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, end_lr=0.1)
# check whether lr is actually changed
assert max(lr_finder.history["lr"]) >= init_lrs[0]
def test_run_with_val_loader(self, cls_task):
task = cls_task(validate=True)
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, val_loader=task.val_loader, end_lr=0.1)
# check whether lr is actually changed
assert max(lr_finder.history["lr"]) >= init_lrs[0]
def test_run_non_tensor_dataset(self, cls_task):
task = cls_task()
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, end_lr=0.1)
# check whether lr is actually changed
assert max(lr_finder.history["lr"]) >= init_lrs[0]
def test_run_non_tensor_dataset_with_val_loader(self, cls_task):
task = cls_task(validate=True)
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, val_loader=task.val_loader, end_lr=0.1)
# check whether lr is actually changed
assert max(lr_finder.history["lr"]) >= init_lrs[0]
class TestReset:
)
def test_reset(self, cls_task):
task = cls_task()
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, val_loader=task.val_loader, end_lr=0.1)
lr_finder.reset()
restored_lrs = get_optim_lr(task.optimizer)
assert init_lrs == restored_lrs
class TestLRHistory:
def test_linear_lr_history(self):
task = mod_task.XORTask()
# prepare_lr_finder sets the starting lr to 1e-5
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(
task.train_loader, num_iter=5, step_mode="linear", end_lr=5e-5
)
assert len(lr_finder.history["lr"]) == 5
assert lr_finder.history["lr"] == pytest.approx([1e-5, 2e-5, 3e-5, 4e-5, 5e-5])
def test_exponential_lr_history(self):
task = mod_task.XORTask()
# prepare_lr_finder sets the starting lr to 1e-5
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, num_iter=5, step_mode="exp", end_lr=0.1)
assert len(lr_finder.history["lr"]) == 5
assert lr_finder.history["lr"] == pytest.approx([1e-5, 1e-4, 1e-3, 1e-2, 0.1])
class TestGradientAccumulation:
def test_gradient_accumulation(self, mocker):
desired_bs, accum_steps = 32, 4
real_bs = desired_bs // accum_steps
num_iter = 10
task = mod_task.XORTask(batch_size=real_bs)
lr_finder = prepare_lr_finder(task)
spy = mocker.spy(lr_finder, "criterion")
lr_finder.range_test(
task.train_loader, num_iter=num_iter, accumulation_steps=accum_steps
)
# NOTE: We are using smaller batch size to simulate a large batch.
# So that the actual times of model/criterion called should be
# `(desired_bs/real_bs) * num_iter` == `accum_steps * num_iter`
assert spy.call_count == accum_steps * num_iter
)
def test_gradient_accumulation_with_apex_amp(self, mocker):
desired_bs, accum_steps = 32, 4
real_bs = desired_bs // accum_steps
num_iter = 10
task = mod_task.XORTask(batch_size=real_bs)
# Wrap model and optimizer by `amp.initialize`. Beside, `amp` requires
# CUDA GPU. So we have to move model to GPU first.
model, optimizer, device = task.model, task.optimizer, task.device
model = model.to(device)
task.model, task.optimizer = apex.amp.initialize(model, optimizer)
lr_finder = prepare_lr_finder(task, amp_backend="apex")
spy = mocker.spy(apex.amp, "scale_loss")
lr_finder.range_test(
task.train_loader, num_iter=num_iter, accumulation_steps=accum_steps
)
assert spy.call_count == accum_steps * num_iter
)
def test_gradient_accumulation_with_torch_amp(self, mocker):
desired_bs, accum_steps = 32, 4
real_bs = desired_bs // accum_steps
num_iter = 10
task = mod_task.XORTask(batch_size=real_bs)
# Config for `torch.amp`. Though `torch.amp.autocast` supports various
# device types, we test it with CUDA only.
amp_config = {
"device_type": "cuda",
"dtype": torch.float16,
}
grad_scaler = torch.cuda.amp.GradScaler()
lr_finder = prepare_lr_finder(
task, amp_backend="torch", amp_config=amp_config, grad_scaler=grad_scaler
)
spy = mocker.spy(grad_scaler, "scale")
lr_finder.range_test(
task.train_loader, num_iter=num_iter, accumulation_steps=accum_steps
)
assert spy.call_count == accum_steps * num_iter
)
class TestMixedPrecision:
def test_mixed_precision_apex(self, mocker):
batch_size = 32
num_iter = 10
task = mod_task.XORTask(batch_size=batch_size)
# Wrap model and optimizer by `amp.initialize`. Beside, `amp` requires
# CUDA GPU. So we have to move model to GPU first.
model, optimizer, device = task.model, task.optimizer, task.device
model = model.to(device)
task.model, task.optimizer = apex.amp.initialize(model, optimizer)
assert hasattr(task.optimizer, "_amp_stash")
lr_finder = prepare_lr_finder(task, amp_backend="apex")
spy = mocker.spy(apex.amp, "scale_loss")
lr_finder.range_test(task.train_loader, num_iter=num_iter)
# NOTE: Here we did not perform gradient accumulation, so that call count
# of `amp.scale_loss` should equal to `num_iter`.
assert spy.call_count == num_iter
def test_mixed_precision_torch(self, mocker):
batch_size = 32
num_iter = 10
task = mod_task.XORTask(batch_size=batch_size)
amp_config = {
"device_type": "cuda",
"dtype": torch.float16,
}
grad_scaler = torch.cuda.amp.GradScaler()
lr_finder = prepare_lr_finder(
task, amp_backend="torch", amp_config=amp_config, grad_scaler=grad_scaler
)
spy = mocker.spy(grad_scaler, "scale")
lr_finder.range_test(task.train_loader, num_iter=num_iter)
# NOTE: Here we did not perform gradient accumulation, so that call count
# of `amp.scale_loss` should equal to `num_iter`.
assert spy.call_count == num_iter
class TestDataLoaderIter:
def test_traindataloaderiter(self):
batch_size, data_length = 32, 256
dataset = mod_dataset.RandomDataset(data_length)
dataloader = DataLoader(dataset, batch_size=batch_size)
loader_iter = TrainDataLoaderIter(dataloader)
assert run_loader_iter(loader_iter)
# `TrainDataLoaderIter` can reset itself, so that it's ok to reuse it
# directly and iterate it more than `len(dataloader)` times.
assert run_loader_iter(loader_iter, desired_runs=len(dataloader) + 1)
def test_valdataloaderiter(self):
batch_size, data_length = 32, 256
dataset = mod_dataset.RandomDataset(data_length)
dataloader = DataLoader(dataset, batch_size=batch_size)
loader_iter = ValDataLoaderIter(dataloader)
assert run_loader_iter(loader_iter)
# `ValDataLoaderIter` can't reset itself, so this should be False if
# we re-run it without resetting it.
assert not run_loader_iter(loader_iter)
# Reset it by `iter()`
loader_iter = iter(loader_iter)
assert run_loader_iter(loader_iter)
# `ValDataLoaderIter` can't be iterated more than `len(dataloader)` times
loader_iter = ValDataLoaderIter(dataloader)
assert not run_loader_iter(loader_iter, desired_runs=len(dataloader) + 1)
def test_run_range_test_with_traindataloaderiter(self, mocker):
task = mod_task.XORTask()
lr_finder = prepare_lr_finder(task)
num_iter = 5
loader_iter = TrainDataLoaderIter(task.train_loader)
spy = mocker.spy(loader_iter, "inputs_labels_from_batch")
lr_finder.range_test(loader_iter, num_iter=num_iter)
assert spy.call_count == num_iter
def test_run_range_test_with_valdataloaderiter(self, mocker):
task = mod_task.XORTask(validate=True)
lr_finder = prepare_lr_finder(task)
num_iter = 5
train_loader_iter = TrainDataLoaderIter(task.train_loader)
val_loader_iter = ValDataLoaderIter(task.val_loader)
spy_train = mocker.spy(train_loader_iter, "inputs_labels_from_batch")
spy_val = mocker.spy(val_loader_iter, "inputs_labels_from_batch")
lr_finder.range_test(
train_loader_iter, val_loader=val_loader_iter, num_iter=num_iter
)
assert spy_train.call_count == num_iter
assert spy_val.call_count == num_iter * len(task.val_loader)
def test_run_range_test_with_trainloaderiter_without_subclassing(self):
task = mod_task.XORTask()
lr_finder = prepare_lr_finder(task)
num_iter = 5
loader_iter = CustomLoaderIter(task.train_loader)
with pytest.raises(ValueError, match="`train_loader` has unsupported type"):
lr_finder.range_test(loader_iter, num_iter=num_iter)
def test_run_range_test_with_valloaderiter_without_subclassing(self):
task = mod_task.XORTask(validate=True)
lr_finder = prepare_lr_finder(task)
num_iter = 5
train_loader_iter = TrainDataLoaderIter(task.train_loader)
val_loader_iter = CustomLoaderIter(task.val_loader)
with pytest.raises(ValueError, match="`val_loader` has unsupported type"):
lr_finder.range_test(
train_loader_iter, val_loader=val_loader_iter, num_iter=num_iter
)
class CustomLoaderIter(object):
"""This class does not inherit from `DataLoaderIter`, should be used to
trigger exceptions related to type checking."""
def __init__(self, loader):
self.loader = loader
def __iter__(self):
return iter(self.loader)
def test_scheduler_and_num_iter(num_iter, scheduler):
task = mod_task.XORTask()
# prepare_lr_finder sets the starting lr to 1e-5
lr_finder = prepare_lr_finder(task)
with pytest.raises(ValueError, match="num_iter"):
lr_finder.range_test(
task.train_loader, num_iter=num_iter, step_mode=scheduler, end_lr=5e-5
)
def test_plot_with_skip_and_suggest_lr(suggest_lr, skip_start, skip_end):
task = mod_task.XORTask()
num_iter = 11
# prepare_lr_finder sets the starting lr to 1e-5
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(
task.train_loader, num_iter=num_iter, step_mode="exp", end_lr=0.1
)
fig, ax = plt.subplots()
results = lr_finder.plot(
skip_start=skip_start, skip_end=skip_end, suggest_lr=suggest_lr, ax=ax
)
if num_iter - skip_start - skip_end <= 1:
# handle data with one or zero lr
assert len(ax.lines) == 1
assert results is ax
else:
# handle different suggest_lr
# for 'steepest': the point with steepest gradient (minimal gradient)
assert len(ax.lines) == 1
assert len(ax.collections) == int(suggest_lr)
if results is not ax:
assert len(results) == 2
def test_suggest_lr():
task = mod_task.XORTask()
lr_finder = prepare_lr_finder(task)
lr_finder.history["loss"] = [10, 8, 4, 1, 4, 16]
lr_finder.history["lr"] = range(len(lr_finder.history["loss"]))
fig, ax = plt.subplots()
ax, lr = lr_finder.plot(skip_start=0, skip_end=0, suggest_lr=True, ax=ax)
assert lr == 2
# Loss with minimal gradient is the first element in history
lr_finder.history["loss"] = [1, 0, 1, 2, 3, 4]
lr_finder.history["lr"] = range(len(lr_finder.history["loss"]))
fig, ax = plt.subplots()
ax, lr = lr_finder.plot(skip_start=0, skip_end=0, suggest_lr=True, ax=ax)
assert lr == 0
# Loss with minimal gradient is the last element in history
lr_finder.history["loss"] = [0, 1, 2, 3, 4, 3]
lr_finder.history["lr"] = range(len(lr_finder.history["loss"]))
fig, ax = plt.subplots()
ax, lr = lr_finder.plot(skip_start=0, skip_end=0, suggest_lr=True, ax=ax)
assert lr == len(lr_finder.history["loss"]) - 1# 删除 assert 和必要提示后上传 LLM 的代码
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
392import pytest
from torch.utils.data import DataLoader
from torch_lr_finder import LRFinder
from torch_lr_finder.lr_finder import (
DataLoaderIter, TrainDataLoaderIter, ValDataLoaderIter
)
import task as mod_task
import dataset as mod_dataset
import matplotlib.pyplot as plt
# Check available backends for mixed precision training
AVAILABLE_AMP_BACKENDS = []
try:
import apex.amp
AVAILABLE_AMP_BACKENDS.append("apex")
except ImportError:
pass
try:
import torch.amp
AVAILABLE_AMP_BACKENDS.append("torch")
except ImportError:
pass
def collect_task_classes():
names = [v for v in dir(mod_task) if v.endswith("Task") and v != "BaseTask"]
attrs = [getattr(mod_task, v) for v in names]
classes = [v for v in attrs if issubclass(v, mod_task.BaseTask)]
return classes
def prepare_lr_finder(task, **kwargs):
model = task.model
optimizer = task.optimizer
criterion = task.criterion
config = {
"device": kwargs.get("device", None),
"memory_cache": kwargs.get("memory_cache", True),
"cache_dir": kwargs.get("cache_dir", None),
"amp_backend": kwargs.get("amp_backend", None),
"amp_config": kwargs.get("amp_config", None),
"grad_scaler": kwargs.get("grad_scaler", None),
}
lr_finder = LRFinder(model, optimizer, criterion, **config)
return lr_finder
def get_optim_lr(optimizer):
return [grp["lr"] for grp in optimizer.param_groups]
def run_loader_iter(loader_iter, desired_runs=None):
if desired_runs is None:
desired_runs = len(loader_iter.data_loader)
count = 0
try:
for i in range(desired_runs):
next(loader_iter)
count += 1
except StopIteration:
return False
return desired_runs == count
class TestRangeTest:
def test_run(self, cls_task):
task = cls_task()
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, end_lr=0.1)
def test_run_with_val_loader(self, cls_task):
task = cls_task(validate=True)
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, val_loader=task.val_loader, end_lr=0.1)
def test_run_non_tensor_dataset(self, cls_task):
task = cls_task()
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, end_lr=0.1)
def test_run_non_tensor_dataset_with_val_loader(self, cls_task):
task = cls_task(validate=True)
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, val_loader=task.val_loader, end_lr=0.1)
class TestReset:
)
def test_reset(self, cls_task):
task = cls_task()
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, val_loader=task.val_loader, end_lr=0.1)
lr_finder.reset()
restored_lrs = get_optim_lr(task.optimizer)
class TestLRHistory:
def test_linear_lr_history(self):
task = mod_task.XORTask()
# prepare_lr_finder sets the starting lr to 1e-5
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(
task.train_loader, num_iter=5, step_mode="linear", end_lr=5e-5
)
def test_exponential_lr_history(self):
task = mod_task.XORTask()
# prepare_lr_finder sets the starting lr to 1e-5
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, num_iter=5, step_mode="exp", end_lr=0.1)
class TestGradientAccumulation:
def test_gradient_accumulation(self, mocker):
desired_bs, accum_steps = 32, 4
real_bs = desired_bs // accum_steps
num_iter = 10
task = mod_task.XORTask(batch_size=real_bs)
lr_finder = prepare_lr_finder(task)
spy = mocker.spy(lr_finder, "criterion")
lr_finder.range_test(
task.train_loader, num_iter=num_iter, accumulation_steps=accum_steps
)
)
def test_gradient_accumulation_with_apex_amp(self, mocker):
desired_bs, accum_steps = 32, 4
real_bs = desired_bs // accum_steps
num_iter = 10
task = mod_task.XORTask(batch_size=real_bs)
# Wrap model and optimizer by `amp.initialize`. Beside, `amp` requires
# CUDA GPU. So we have to move model to GPU first.
model, optimizer, device = task.model, task.optimizer, task.device
model = model.to(device)
task.model, task.optimizer = apex.amp.initialize(model, optimizer)
lr_finder = prepare_lr_finder(task, amp_backend="apex")
spy = mocker.spy(apex.amp, "scale_loss")
lr_finder.range_test(
task.train_loader, num_iter=num_iter, accumulation_steps=accum_steps
)
)
def test_gradient_accumulation_with_torch_amp(self, mocker):
desired_bs, accum_steps = 32, 4
real_bs = desired_bs // accum_steps
num_iter = 10
task = mod_task.XORTask(batch_size=real_bs)
# Config for `torch.amp`. Though `torch.amp.autocast` supports various
# device types, we test it with CUDA only.
amp_config = {
"device_type": "cuda",
"dtype": torch.float16,
}
grad_scaler = torch.cuda.amp.GradScaler()
lr_finder = prepare_lr_finder(
task, amp_backend="torch", amp_config=amp_config, grad_scaler=grad_scaler
)
spy = mocker.spy(grad_scaler, "scale")
lr_finder.range_test(
task.train_loader, num_iter=num_iter, accumulation_steps=accum_steps
)
)
class TestMixedPrecision:
def test_mixed_precision_apex(self, mocker):
batch_size = 32
num_iter = 10
task = mod_task.XORTask(batch_size=batch_size)
# Wrap model and optimizer by `amp.initialize`. Beside, `amp` requires
# CUDA GPU. So we have to move model to GPU first.
model, optimizer, device = task.model, task.optimizer, task.device
model = model.to(device)
task.model, task.optimizer = apex.amp.initialize(model, optimizer)
lr_finder = prepare_lr_finder(task, amp_backend="apex")
spy = mocker.spy(apex.amp, "scale_loss")
lr_finder.range_test(task.train_loader, num_iter=num_iter)
# NOTE: Here we did not perform gradient accumulation, so that call count
# of `amp.scale_loss` should equal to `num_iter`.
def test_mixed_precision_torch(self, mocker):
batch_size = 32
num_iter = 10
task = mod_task.XORTask(batch_size=batch_size)
amp_config = {
"device_type": "cuda",
"dtype": torch.float16,
}
grad_scaler = torch.cuda.amp.GradScaler()
lr_finder = prepare_lr_finder(
task, amp_backend="torch", amp_config=amp_config, grad_scaler=grad_scaler
)
spy = mocker.spy(grad_scaler, "scale")
lr_finder.range_test(task.train_loader, num_iter=num_iter)
class TestDataLoaderIter:
def test_traindataloaderiter(self):
batch_size, data_length = 32, 256
dataset = mod_dataset.RandomDataset(data_length)
dataloader = DataLoader(dataset, batch_size=batch_size)
loader_iter = TrainDataLoaderIter(dataloader)
def test_valdataloaderiter(self):
batch_size, data_length = 32, 256
dataset = mod_dataset.RandomDataset(data_length)
dataloader = DataLoader(dataset, batch_size=batch_size)
loader_iter = ValDataLoaderIter(dataloader)
def test_run_range_test_with_traindataloaderiter(self, mocker):
task = mod_task.XORTask()
lr_finder = prepare_lr_finder(task)
num_iter = 5
loader_iter = TrainDataLoaderIter(task.train_loader)
spy = mocker.spy(loader_iter, "inputs_labels_from_batch")
lr_finder.range_test(loader_iter, num_iter=num_iter)
assert spy.call_count == num_iter
def test_run_range_test_with_valdataloaderiter(self, mocker):
task = mod_task.XORTask(validate=True)
lr_finder = prepare_lr_finder(task)
num_iter = 5
train_loader_iter = TrainDataLoaderIter(task.train_loader)
val_loader_iter = ValDataLoaderIter(task.val_loader)
spy_train = mocker.spy(train_loader_iter, "inputs_labels_from_batch")
spy_val = mocker.spy(val_loader_iter, "inputs_labels_from_batch")
lr_finder.range_test(
train_loader_iter, val_loader=val_loader_iter, num_iter=num_iter
)
assert spy_train.call_count == num_iter
assert spy_val.call_count == num_iter * len(task.val_loader)
def test_run_range_test_with_trainloaderiter_without_subclassing(self):
task = mod_task.XORTask()
lr_finder = prepare_lr_finder(task)
num_iter = 5
loader_iter = CustomLoaderIter(task.train_loader)
with pytest.raises(ValueError, match="`train_loader` has unsupported type"):
lr_finder.range_test(loader_iter, num_iter=num_iter)
def test_run_range_test_with_valloaderiter_without_subclassing(self):
task = mod_task.XORTask(validate=True)
lr_finder = prepare_lr_finder(task)
num_iter = 5
train_loader_iter = TrainDataLoaderIter(task.train_loader)
val_loader_iter = CustomLoaderIter(task.val_loader)
with pytest.raises(ValueError, match="`val_loader` has unsupported type"):
lr_finder.range_test(
train_loader_iter, val_loader=val_loader_iter, num_iter=num_iter
)
class CustomLoaderIter(object):
"""This class does not inherit from `DataLoaderIter`, should be used to
trigger exceptions related to type checking."""
def __init__(self, loader):
self.loader = loader
def __iter__(self):
return iter(self.loader)
def test_scheduler_and_num_iter(num_iter, scheduler):
task = mod_task.XORTask()
# prepare_lr_finder sets the starting lr to 1e-5
lr_finder = prepare_lr_finder(task)
with pytest.raises(ValueError, match="num_iter"):
lr_finder.range_test(
task.train_loader, num_iter=num_iter, step_mode=scheduler, end_lr=5e-5
)
def test_plot_with_skip_and_suggest_lr(suggest_lr, skip_start, skip_end):
task = mod_task.XORTask()
num_iter = 11
# prepare_lr_finder sets the starting lr to 1e-5
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(
task.train_loader, num_iter=num_iter, step_mode="exp", end_lr=0.1
)
fig, ax = plt.subplots()
results = lr_finder.plot(
skip_start=skip_start, skip_end=skip_end, suggest_lr=suggest_lr, ax=ax
)
if num_iter - skip_start - skip_end <= 1:
# handle data with one or zero lr
assert len(ax.lines) == 1
assert results is ax
else:
# handle different suggest_lr
# for 'steepest': the point with steepest gradient (minimal gradient)
assert len(ax.lines) == 1
assert len(ax.collections) == int(suggest_lr)
if results is not ax:
assert len(results) == 2
def test_suggest_lr():
task = mod_task.XORTask()
lr_finder = prepare_lr_finder(task)
lr_finder.history["loss"] = [10, 8, 4, 1, 4, 16]
lr_finder.history["lr"] = range(len(lr_finder.history["loss"]))
fig, ax = plt.subplots()
ax, lr = lr_finder.plot(skip_start=0, skip_end=0, suggest_lr=True, ax=ax)
assert lr == 2
# Loss with minimal gradient is the first element in history
lr_finder.history["loss"] = [1, 0, 1, 2, 3, 4]
lr_finder.history["lr"] = range(len(lr_finder.history["loss"]))
fig, ax = plt.subplots()
ax, lr = lr_finder.plot(skip_start=0, skip_end=0, suggest_lr=True, ax=ax)
assert lr == 0
# Loss with minimal gradient is the last element in history
lr_finder.history["loss"] = [0, 1, 2, 3, 4, 3]
lr_finder.history["lr"] = range(len(lr_finder.history["loss"]))
fig, ax = plt.subplots()
ax, lr = lr_finder.plot(skip_start=0, skip_end=0, suggest_lr=True, ax=ax)
assert lr == len(lr_finder.history["loss"]) - 1# LLM 自动填充 assert 后的代码
**base prompt 为：** 该 tests case 缺少 assert 断言 请你自动填充它
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
320import pytest
from torch.utils.data import DataLoader
from torch_lr_finder import LRFinder
from torch_lr_finder.lr_finder import (
DataLoaderIter, TrainDataLoaderIter, ValDataLoaderIter
)
import task as mod_task
import dataset as mod_dataset
import matplotlib.pyplot as plt
# Check available backends for mixed precision training
AVAILABLE_AMP_BACKENDS = []
try:
import apex.amp
AVAILABLE_AMP_BACKENDS.append("apex")
except ImportError:
pass
try:
import torch.amp
AVAILABLE_AMP_BACKENDS.append("torch")
except ImportError:
pass
def collect_task_classes():
names = [v for v in dir(mod_task) if v.endswith("Task") and v != "BaseTask"]
attrs = [getattr(mod_task, v) for v in names]
classes = [v for v in attrs if issubclass(v, mod_task.BaseTask)]
return classes
def prepare_lr_finder(task, **kwargs):
model = task.model
optimizer = task.optimizer
criterion = task.criterion
config = {
"device": kwargs.get("device", None),
"memory_cache": kwargs.get("memory_cache", True),
"cache_dir": kwargs.get("cache_dir", None),
"amp_backend": kwargs.get("amp_backend", None),
"amp_config": kwargs.get("amp_config", None),
"grad_scaler": kwargs.get("grad_scaler", None),
}
lr_finder = LRFinder(model, optimizer, criterion, **config)
return lr_finder
def get_optim_lr(optimizer):
return [grp["lr"] for grp in optimizer.param_groups]
def run_loader_iter(loader_iter, desired_runs=None):
if desired_runs is None:
desired_runs = len(loader_iter.data_loader)
count = 0
try:
for i in range(desired_runs):
next(loader_iter)
count += 1
except StopIteration:
return False
return desired_runs == count
class TestRangeTest:
def test_run(self, cls_task):
task = cls_task()
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, end_lr=0.1)
# Assert statements
assert len(lr_finder.history["lr"]) > 0
assert len(lr_finder.history["loss"]) > 0
def test_run_with_val_loader(self, cls_task):
task = cls_task(validate=True)
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, val_loader=task.val_loader, end_lr=0.1)
# Assert statements
assert len(lr_finder.history["lr"]) > 0
assert len(lr_finder.history["loss"]) > 0
def test_run_non_tensor_dataset(self, cls_task):
task = cls_task()
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, end_lr=0.1)
# Assert statements
assert len(lr_finder.history["lr"]) > 0
assert len(lr_finder.history["loss"]) > 0
def test_run_non_tensor_dataset_with_val_loader(self, cls_task):
task = cls_task(validate=True)
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, val_loader=task.val_loader, end_lr=0.1)
# Assert statements
assert len(lr_finder.history["lr"]) > 0
assert len(lr_finder.history["loss"]) > 0
class TestReset:
)
def test_reset(self, cls_task):
task = cls_task()
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, val_loader=task.val_loader, end_lr=0.1)
lr_finder.reset()
restored_lrs = get_optim_lr(task.optimizer)
# Assert statements
assert restored_lrs == init_lrs
class TestLRHistory:
def test_linear_lr_history(self):
task = mod_task.XORTask()
# prepare_lr_finder sets the starting lr to 1e-5
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(
task.train_loader, num_iter=5, step_mode="linear", end_lr=5e-5
)
# Assert statements
assert len(lr_finder.history["lr"]) > 0
assert len(lr_finder.history["loss"]) > 0
def test_exponential_lr_history(self):
task = mod_task.XORTask()
# prepare_lr_finder sets the starting lr to 1e-5
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, num_iter=5, step_mode="exp", end_lr=0.1)
# Assert statements
assert len(lr_finder.history["lr"]) > 0
assert len(lr_finder.history["loss"]) > 0
class TestGradientAccumulation:
def test_gradient_accumulation(self, mocker):
desired_bs, accum_steps = 32, 4
real_bs = desired_bs // accum_steps
num_iter = 10
task = mod_task.XORTask(batch_size=real_bs)
lr_finder = prepare_lr_finder(task)
spy = mocker.spy(lr_finder, "criterion")
lr_finder.range_test(
task.train_loader, num_iter=num_iter, accumulation_steps=accum_steps
)
# Assert statements
assert spy.call_count == num_iter
)
def test_gradient_accumulation_with_apex_amp(self, mocker):
desired_bs, accum_steps = 32, 4
real_bs = desired_bs // accum_steps
num_iter = 10
task = mod_task.XORTask(batch_size=real_bs)
# Wrap model and optimizer by `amp.initialize`. Beside, `amp` requires
# CUDA GPU. So we have to move model to GPU first.
model, optimizer, device = task.model, task.optimizer, task.device
model = model.to(device)
task.model, task.optimizer = apex.amp.initialize(model, optimizer)
lr_finder = prepare_lr_finder(task, amp_backend="apex")
spy = mocker.spy(apex.amp, "scale_loss")
lr_finder.range_test(
task.train_loader, num_iter=num_iter, accumulation_steps=accum_steps
)
# Assert statements
assert spy.call_count == num_iter
)
def test_gradient_accumulation_with_torch_amp(self, mocker):
desired_bs, accum_steps = 32, 4
real_bs = desired_bs // accum_steps
num_iter = 10
task = mod_task.XORTask(batch_size=real_bs)
# Config for `torch.amp`. Though `torch.amp.autocast` supports various
# device types, we test it with CUDA only.
amp_config = {
"device_type": "cuda",
"dtype": torch.float16,
}
grad_scaler = torch.cuda.amp.GradScaler()
lr_finder = prepare_lr_finder(
task, amp_backend="torch", amp_config=amp_config, grad_scaler=grad_scaler
)
spy = mocker.spy(grad_scaler, "scale")
lr_finder.range_test(
task.train_loader, num_iter=num_iter, accumulation_steps=accum_steps
)
# Assert statements
assert spy.call_count == num_iter
)
class TestMixedPrecision:
def test_mixed_precision_apex(self, mocker):
batch_size = 32
num_iter = 10
task = mod_task.XORTask(batch_size=batch_size)
# Wrap model and optimizer by `amp.initialize`. Beside, `amp` requires
# CUDA GPU. So we have to move model to GPU first.
model, optimizer, device = task.model, task.optimizer, task.device
model = model.to(device)
task.model, task.optimizer = apex.amp.initialize(model, optimizer)
lr_finder = prepare_lr_finder(task, amp_backend="apex")
spy = mocker.spy(apex.amp, "scale_loss")
lr_finder.range_test(task.train_loader, num_iter=num_iter)
# NOTE: Here we did not perform gradient accumulation, so that call count
# of `amp` could be only one in any situation.
# Assert statements
assert spy.call_count == 1
)
class TestMixedPrecisionWithTorchAMP:
def test_mixed_precision_torch_amp(self, mocker):
batch_size = 32
num_iter = 10
task = mod_task.XORTask(batch_size=batch_size)
# Config for `torch.amp`. Though `torch.amp.autocast` supports various
# device types, we test it with CUDA only.
amp_config = {
"device_type": "cuda",
"dtype": torch.float16,
}
grad_scaler = torch.cuda.amp.GradScaler()
lr_finder = prepare_lr_finder(
task, amp_backend="torch", amp_config=amp_config, grad_scaler=grad_scaler
)
spy = mocker.spy(grad_scaler, "scale")
lr_finder.range_test(task.train_loader, num_iter=num_iter)
# NOTE: Here we did not perform gradient accumulation, so that call count
# of `amp` could be only one in any situation.
# Assert statements
assert spy.call_count == 1
class TestMultiPhase:
)
def test_multi_phase(self, cls_task):
task = cls_task()
init_lrs = get_optim_lr(task.optimizer)
lr_finder = prepare_lr_finder(task)
lr_finder.range_test(task.train_loader, val_loader=task.val_loader, num_iter=10)
# Simulate a change in LR and perform another phase of LR range test
for param_group in task.optimizer.param_groups:
param_group["lr"] *= 0.1
lr_finder.range_test(
task.train_loader, val_loader=task.val_loader, num_iter=10, reset_lr=False
)
# Assert statements
assert len(lr_finder.history["lr"]) > 0
assert len(lr_finder.history["loss"]) > 0
