lsmpp's picture
Add files using upload-large-folder tool
4cef5ec verified
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# 콜백 [[callbacks]]
μ½œλ°±μ€ PyTorch [`Trainer`]의 반볡 ν•™μŠ΅ λ™μž‘μ„ μ‚¬μš©μž μ •μ˜ν•  수 μžˆλŠ” κ°μ²΄μž…λ‹ˆλ‹€
(이 κΈ°λŠ₯은 TensorFlowμ—μ„œλŠ” 아직 κ΅¬ν˜„λ˜μ§€ μ•Šμ•˜μŠ΅λ‹ˆλ‹€). μ½œλ°±μ€ 반볡 ν•™μŠ΅μ˜ μƒνƒœλ₯Ό
κ²€μ‚¬ν•˜μ—¬ (μ§„ν–‰ 상황 보고, TensorBoard λ˜λŠ” 기타 λ¨Έμ‹  λŸ¬λ‹ ν”Œλž«νΌμ— 둜그 남기기 λ“±)
κ²°μ •(예: μ‘°κΈ° μ’…λ£Œ)을 내릴 수 μžˆμŠ΅λ‹ˆλ‹€.
μ½œλ°±μ€ [`TrainerControl`] 객체λ₯Ό λ°˜ν™˜ν•˜λŠ” 것 μ™Έμ—λŠ” 반볡 ν•™μŠ΅μ—μ„œ μ–΄λ–€ 것도 λ³€κ²½ν•  수 μ—†λŠ”
"읽기 μ „μš©" μ½”λ“œ μ‘°κ°μž…λ‹ˆλ‹€. 반볡 ν•™μŠ΅μ— 변경이 ν•„μš”ν•œ μ‚¬μš©μž μ •μ˜ μž‘μ—…μ΄ ν•„μš”ν•œ 경우,
[`Trainer`]λ₯Ό μ„œλΈŒν΄λž˜μŠ€λ‘œ λ§Œλ“€μ–΄ ν•„μš”ν•œ λ©”μ†Œλ“œλ“€μ„ μ˜€λ²„λΌμ΄λ“œν•΄μ•Ό ν•©λ‹ˆλ‹€ (μ˜ˆμ‹œλŠ” [trainer](trainer)λ₯Ό μ°Έμ‘°ν•˜μ„Έμš”).
기본적으둜 `TrainingArguments.report_to`λŠ” `"all"`둜 μ„€μ •λ˜μ–΄ μžˆμœΌλ―€λ‘œ, [`Trainer`]λŠ” λ‹€μŒ μ½œλ°±μ„ μ‚¬μš©ν•©λ‹ˆλ‹€.
- [`DefaultFlowCallback`]λŠ” 둜그, μ €μž₯, 평가에 λŒ€ν•œ κΈ°λ³Έ λ™μž‘μ„ μ²˜λ¦¬ν•©λ‹ˆλ‹€.
- [`PrinterCallback`] λ˜λŠ” [`ProgressCallback`]λŠ” μ§„ν–‰ 상황을 ν‘œμ‹œν•˜κ³  둜그λ₯Ό 좜λ ₯ν•©λ‹ˆλ‹€
([`TrainingArguments`]λ₯Ό 톡해 tqdm을 λΉ„ν™œμ„±ν™”ν•˜λ©΄ 첫 번째 콜백이 μ‚¬μš©λ˜κ³ , κ·Έλ ‡μ§€ μ•ŠμœΌλ©΄ 두 λ²ˆμ§Έκ°€ μ‚¬μš©λ©λ‹ˆλ‹€).
- [`~integrations.TensorBoardCallback`]λŠ” TensorBoardκ°€ (PyTorch >= 1.4
λ˜λŠ” tensorboardXλ₯Ό 톡해) μ ‘κ·Ό κ°€λŠ₯ν•˜λ©΄ μ‚¬μš©λ©λ‹ˆλ‹€.
- [`~integrations.WandbCallback`]λŠ” [wandb](https://www.wandb.com/)κ°€ μ„€μΉ˜λ˜μ–΄ 있으면
μ‚¬μš©λ©λ‹ˆλ‹€.
- [`~integrations.CometCallback`]λŠ” [comet_ml](https://www.comet.com/site/)이 μ„€μΉ˜λ˜μ–΄ 있으면 μ‚¬μš©λ©λ‹ˆλ‹€.
- [`~integrations.MLflowCallback`]λŠ” [mlflow](https://www.mlflow.org/)κ°€ μ„€μΉ˜λ˜μ–΄ 있으면 μ‚¬μš©λ©λ‹ˆλ‹€.
- [`~integrations.NeptuneCallback`]λŠ” [neptune](https://neptune.ai/)이 μ„€μΉ˜λ˜μ–΄ 있으면 μ‚¬μš©λ©λ‹ˆλ‹€.
- [`~integrations.AzureMLCallback`]λŠ” [azureml-sdk](https://pypi.org/project/azureml-sdk/)κ°€ μ„€μΉ˜λ˜μ–΄
있으면 μ‚¬μš©λ©λ‹ˆλ‹€.
- [`~integrations.CodeCarbonCallback`]λŠ” [codecarbon](https://pypi.org/project/codecarbon/)이 μ„€μΉ˜λ˜μ–΄
있으면 μ‚¬μš©λ©λ‹ˆλ‹€.
- [`~integrations.ClearMLCallback`]λŠ” [clearml](https://github.com/allegroai/clearml)이 μ„€μΉ˜λ˜μ–΄ 있으면 μ‚¬μš©λ©λ‹ˆλ‹€.
- [`~integrations.DagsHubCallback`]λŠ” [dagshub](https://dagshub.com/)이 μ„€μΉ˜λ˜μ–΄ 있으면 μ‚¬μš©λ©λ‹ˆλ‹€.
- [`~integrations.FlyteCallback`]λŠ” [flyte](https://flyte.org/)κ°€ μ„€μΉ˜λ˜μ–΄ 있으면 μ‚¬μš©λ©λ‹ˆλ‹€.
- [`~integrations.DVCLiveCallback`]λŠ” [dvclive](https://dvc.org/doc/dvclive)κ°€ μ„€μΉ˜λ˜μ–΄ 있으면 μ‚¬μš©λ©λ‹ˆλ‹€.
- [`~integrations.SwanLabCallback`]λŠ” [swanlab](https://swanlab.cn)κ°€ μ„€μΉ˜λ˜μ–΄ 있으면 μ‚¬μš©λ©λ‹ˆλ‹€.
νŒ¨ν‚€μ§€κ°€ μ„€μΉ˜λ˜μ–΄ μžˆμ§€λ§Œ ν•΄λ‹Ή 톡합 κΈ°λŠ₯을 μ‚¬μš©ν•˜κ³  μ‹Άμ§€ μ•Šλ‹€λ©΄, `TrainingArguments.report_to`λ₯Ό μ‚¬μš©ν•˜κ³ μž ν•˜λŠ” 톡합 κΈ°λŠ₯ λͺ©λ‘μœΌλ‘œ λ³€κ²½ν•  수 μžˆμŠ΅λ‹ˆλ‹€ (예: `["azure_ml", "wandb"]`).
μ½œλ°±μ„ κ΅¬ν˜„ν•˜λŠ” μ£Όμš” ν΄λž˜μŠ€λŠ” [`TrainerCallback`]μž…λ‹ˆλ‹€. 이 ν΄λž˜μŠ€λŠ” [`Trainer`]λ₯Ό
μΈμŠ€ν„΄μŠ€ν™”ν•˜λŠ” 데 μ‚¬μš©λœ [`TrainingArguments`]λ₯Ό κ°€μ Έμ˜€κ³ , ν•΄λ‹Ή Trainer의 λ‚΄λΆ€ μƒνƒœλ₯Ό
[`TrainerState`]λ₯Ό 톡해 μ ‘κ·Όν•  수 있으며, [`TrainerControl`]을 톡해 반볡 ν•™μŠ΅μ—μ„œ 일뢀
μž‘μ—…μ„ μˆ˜ν–‰ν•  수 μžˆμŠ΅λ‹ˆλ‹€.
## μ‚¬μš© κ°€λŠ₯ν•œ 콜백 [[available-callbacks]]
λΌμ΄λΈŒλŸ¬λ¦¬μ—μ„œ μ‚¬μš© κ°€λŠ₯ν•œ [`TrainerCallback`] λͺ©λ‘μ€ λ‹€μŒκ³Ό κ°™μŠ΅λ‹ˆλ‹€:
[[autodoc]] integrations.CometCallback
- setup
[[autodoc]] DefaultFlowCallback
[[autodoc]] PrinterCallback
[[autodoc]] ProgressCallback
[[autodoc]] EarlyStoppingCallback
[[autodoc]] integrations.TensorBoardCallback
[[autodoc]] integrations.WandbCallback
- setup
[[autodoc]] integrations.MLflowCallback
- setup
[[autodoc]] integrations.AzureMLCallback
[[autodoc]] integrations.CodeCarbonCallback
[[autodoc]] integrations.NeptuneCallback
[[autodoc]] integrations.ClearMLCallback
[[autodoc]] integrations.DagsHubCallback
[[autodoc]] integrations.FlyteCallback
[[autodoc]] integrations.DVCLiveCallback
- setup
[[autodoc]] integrations.SwanLabCallback
- setup
## TrainerCallback [[trainercallback]]
[[autodoc]] TrainerCallback
μ—¬κΈ° PyTorch [`Trainer`]와 ν•¨κ»˜ μ‚¬μš©μž μ •μ˜ μ½œλ°±μ„ λ“±λ‘ν•˜λŠ” μ˜ˆμ‹œκ°€ μžˆμŠ΅λ‹ˆλ‹€:
```python
class MyCallback(TrainerCallback):
"A callback that prints a message at the beginning of training"
def on_train_begin(self, args, state, control, **kwargs):
print("Starting training")
trainer = Trainer(
model,
args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
callbacks=[MyCallback], # μš°λ¦¬λŠ” 콜백 클래슀λ₯Ό 이 λ°©μ‹μœΌλ‘œ μ „λ‹¬ν•˜κ±°λ‚˜ κ·Έκ²ƒμ˜ μΈμŠ€ν„΄μŠ€(MyCallback())λ₯Ό 전달할 수 μžˆμŠ΅λ‹ˆλ‹€
)
```
또 λ‹€λ₯Έ μ½œλ°±μ„ λ“±λ‘ν•˜λŠ” 방법은 `trainer.add_callback()`을 ν˜ΈμΆœν•˜λŠ” κ²ƒμž…λ‹ˆλ‹€:
```python
trainer = Trainer(...)
trainer.add_callback(MyCallback)
# λ‹€λ₯Έ λ°©λ²•μœΌλ‘œλŠ” 콜백 클래슀의 μΈμŠ€ν„΄μŠ€λ₯Ό 전달할 수 μžˆμŠ΅λ‹ˆλ‹€
trainer.add_callback(MyCallback())
```
## TrainerState [[trainerstate]]
[[autodoc]] TrainerState
## TrainerControl [[trainercontrol]]
[[autodoc]] TrainerControl