AutoMM Presets#
It is well-known that we usually need to set hyperparameters before the learning process begins. Deep learning models, e.g., pretrained foundation models, can have anywhere from a few hyperparameters to a few hundred. The hyperparameters can impact training speed, final model performance, and inference latency. However, choosing the proper hyperparameters may be challenging for many users with limited expertise.
In this tutorial, we will introduce the easy-to-use presets in AutoMM. Our presets can condense the complex hyperparameter setups into simple strings. More specifically, AutoMM supports three presets: medium_quality, high_quality, and best_quality.
import warnings
warnings.filterwarnings('ignore')
Dataset#
For demonstration, we use a subsampled Stanford Sentiment Treebank (SST) dataset, which consists of movie reviews and their associated sentiment. Given a new movie review, the goal is to predict the sentiment reflected in the text (in this case, a binary classification, where reviews are labeled as 1 if they conveyed a positive opinion and 0 otherwise). To get started, let’s download and prepare the dataset.
from autogluon.core.utils.loaders import load_pd
train_data = load_pd.load('https://autogluon-text.s3-accelerate.amazonaws.com/glue/sst/train.parquet')
test_data = load_pd.load('https://autogluon-text.s3-accelerate.amazonaws.com/glue/sst/dev.parquet')
subsample_size = 1000 # subsample data for faster demo, try setting this to larger values
train_data = train_data.sample(n=subsample_size, random_state=0)
train_data.head(10)
| sentence | label | |
|---|---|---|
| 43787 | very pleasing at its best moments | 1 |
| 16159 | , american chai is enough to make you put away... | 0 |
| 59015 | too much like an infomercial for ram dass 's l... | 0 |
| 5108 | a stirring visual sequence | 1 |
| 67052 | cool visual backmasking | 1 |
| 35938 | hard ground | 0 |
| 49879 | the striking , quietly vulnerable personality ... | 1 |
| 51591 | pan nalin 's exposition is beautiful and myste... | 1 |
| 56780 | wonderfully loopy | 1 |
| 28518 | most beautiful , evocative | 1 |
Medium Quality#
In some situations, we prefer fast training and inference to the prediction quality. medium_quality is designed for this purpose.
Among the three presets, medium_quality has the smallest model size. Now let’s fit the predictor using the medium_quality preset. Here we set a tight time budget for a quick demo.
from autogluon.multimodal import MultiModalPredictor
predictor = MultiModalPredictor(label='label', eval_metric='acc', presets="medium_quality")
predictor.fit(
train_data=train_data,
time_limit=30, # seconds
)
Global seed set to 123
No path specified. Models will be saved in: "AutogluonModels/ag-20230302_170215/"
AutoMM starts to create your model. ✨
- Model will be saved to "/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215".
- Validation metric is "acc".
- To track the learning progress, you can open a terminal and launch Tensorboard:
```shell
# Assume you have installed tensorboard
tensorboard --logdir /home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215
```
Enjoy your coffee, and let AutoMM do the job ☕☕☕ Learn more at https://auto.gluon.ai
Using 16bit None Automatic Mixed Precision (AMP)
GPU available: True (cuda), used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
| Name | Type | Params
-------------------------------------------------------------------
0 | model | HFAutoModelForTextPrediction | 13.5 M
1 | validation_metric | Accuracy | 0
2 | loss_func | CrossEntropyLoss | 0
-------------------------------------------------------------------
13.5 M Trainable params
0 Non-trainable params
13.5 M Total params
26.967 Total estimated model params size (MB)
Epoch 0, global step 3: 'val_acc' reached 0.47000 (best 0.47000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215/epoch=0-step=3.ckpt' as top 3
Epoch 0, global step 7: 'val_acc' reached 0.55000 (best 0.55000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215/epoch=0-step=7.ckpt' as top 3
Epoch 1, global step 10: 'val_acc' reached 0.59500 (best 0.59500), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215/epoch=1-step=10.ckpt' as top 3
Epoch 1, global step 14: 'val_acc' reached 0.64000 (best 0.64000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215/epoch=1-step=14.ckpt' as top 3
Epoch 2, global step 17: 'val_acc' reached 0.61000 (best 0.64000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215/epoch=2-step=17.ckpt' as top 3
Epoch 2, global step 21: 'val_acc' reached 0.71000 (best 0.71000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215/epoch=2-step=21.ckpt' as top 3
Epoch 3, global step 24: 'val_acc' reached 0.82000 (best 0.82000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215/epoch=3-step=24.ckpt' as top 3
Epoch 3, global step 28: 'val_acc' reached 0.84500 (best 0.84500), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215/epoch=3-step=28.ckpt' as top 3
Time limit reached. Elapsed time is 0:00:30. Signaling Trainer to stop.
Epoch 4, global step 30: 'val_acc' reached 0.84500 (best 0.84500), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215/epoch=4-step=30.ckpt' as top 3
Start to fuse 3 checkpoints via the greedy soup algorithm.
AutoMM has created your model 🎉🎉🎉
- To load the model, use the code below:
```python
from autogluon.multimodal import MultiModalPredictor
predictor = MultiModalPredictor.load("/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215")
```
- You can open a terminal and launch Tensorboard to visualize the training log:
```shell
# Assume you have installed tensorboard
tensorboard --logdir /home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170215
```
- If you are not satisfied with the model, try to increase the training time,
adjust the hyperparameters (https://auto.gluon.ai/stable/tutorials/multimodal/advanced_topics/customization.html),
or post issues on GitHub: https://github.com/autogluon/autogluon
<autogluon.multimodal.predictor.MultiModalPredictor at 0x7f20a9133a90>
Then we can evaluate the predictor on the test data.
scores = predictor.evaluate(test_data, metrics=["roc_auc"])
scores
{'roc_auc': 0.8920102930032836}
High Quality#
If you want to balance the prediction quality and training/inference speed, you can try the high_quality preset, which uses a larger model than medium_quality. Accordingly, we need to increase the time limit since larger models require more time to train.
from autogluon.multimodal import MultiModalPredictor
predictor = MultiModalPredictor(label='label', eval_metric='acc', presets="high_quality")
predictor.fit(
train_data=train_data,
time_limit=50, # seconds
)
Global seed set to 123
No path specified. Models will be saved in: "AutogluonModels/ag-20230302_170254/"
AutoMM starts to create your model. ✨
- Model will be saved to "/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170254".
- Validation metric is "acc".
- To track the learning progress, you can open a terminal and launch Tensorboard:
```shell
# Assume you have installed tensorboard
tensorboard --logdir /home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170254
```
Enjoy your coffee, and let AutoMM do the job ☕☕☕ Learn more at https://auto.gluon.ai
Using 16bit None Automatic Mixed Precision (AMP)
GPU available: True (cuda), used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
| Name | Type | Params
-------------------------------------------------------------------
0 | model | HFAutoModelForTextPrediction | 108 M
1 | validation_metric | Accuracy | 0
2 | loss_func | CrossEntropyLoss | 0
-------------------------------------------------------------------
108 M Trainable params
0 Non-trainable params
108 M Total params
217.786 Total estimated model params size (MB)
Epoch 0, global step 3: 'val_acc' reached 0.59500 (best 0.59500), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170254/epoch=0-step=3.ckpt' as top 3
Epoch 0, global step 7: 'val_acc' reached 0.68000 (best 0.68000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170254/epoch=0-step=7.ckpt' as top 3
Epoch 1, global step 10: 'val_acc' reached 0.65500 (best 0.68000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170254/epoch=1-step=10.ckpt' as top 3
Epoch 1, global step 14: 'val_acc' reached 0.84500 (best 0.84500), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170254/epoch=1-step=14.ckpt' as top 3
Epoch 2, global step 17: 'val_acc' reached 0.87000 (best 0.87000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170254/epoch=2-step=17.ckpt' as top 3
Time limit reached. Elapsed time is 0:00:50. Signaling Trainer to stop.
Epoch 2, global step 19: 'val_acc' reached 0.90000 (best 0.90000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170254/epoch=2-step=19.ckpt' as top 3
Start to fuse 3 checkpoints via the greedy soup algorithm.
AutoMM has created your model 🎉🎉🎉
- To load the model, use the code below:
```python
from autogluon.multimodal import MultiModalPredictor
predictor = MultiModalPredictor.load("/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170254")
```
- You can open a terminal and launch Tensorboard to visualize the training log:
```shell
# Assume you have installed tensorboard
tensorboard --logdir /home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170254
```
- If you are not satisfied with the model, try to increase the training time,
adjust the hyperparameters (https://auto.gluon.ai/stable/tutorials/multimodal/advanced_topics/customization.html),
or post issues on GitHub: https://github.com/autogluon/autogluon
<autogluon.multimodal.predictor.MultiModalPredictor at 0x7f211c56aa90>
Although high_quality requires more training time than medium_quality, it also brings performance gains.
scores = predictor.evaluate(test_data, metrics=["roc_auc"])
scores
{'roc_auc': 0.9499242232887092}
Best Quality#
If you want the best performance and don’t care about the training/inference cost, give it a try for the best_quality preset. High-end GPUs with large memory are preferred in this case. Compared to high_quality, it requires much longer training time.
from autogluon.multimodal import MultiModalPredictor
predictor = MultiModalPredictor(label='label', eval_metric='acc', presets="best_quality")
predictor.fit(train_data=train_data, time_limit=180)
Global seed set to 123
No path specified. Models will be saved in: "AutogluonModels/ag-20230302_170413/"
AutoMM starts to create your model. ✨
- Model will be saved to "/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170413".
- Validation metric is "acc".
- To track the learning progress, you can open a terminal and launch Tensorboard:
```shell
# Assume you have installed tensorboard
tensorboard --logdir /home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170413
```
Enjoy your coffee, and let AutoMM do the job ☕☕☕ Learn more at https://auto.gluon.ai
Using 16bit None Automatic Mixed Precision (AMP)
GPU available: True (cuda), used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
| Name | Type | Params
-------------------------------------------------------------------
0 | model | HFAutoModelForTextPrediction | 183 M
1 | validation_metric | Accuracy | 0
2 | loss_func | CrossEntropyLoss | 0
-------------------------------------------------------------------
183 M Trainable params
0 Non-trainable params
183 M Total params
367.666 Total estimated model params size (MB)
Epoch 0, global step 3: 'val_acc' reached 0.55000 (best 0.55000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170413/epoch=0-step=3.ckpt' as top 3
Epoch 0, global step 7: 'val_acc' reached 0.55000 (best 0.55000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170413/epoch=0-step=7.ckpt' as top 3
Epoch 1, global step 10: 'val_acc' reached 0.75000 (best 0.75000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170413/epoch=1-step=10.ckpt' as top 3
Epoch 1, global step 14: 'val_acc' reached 0.95000 (best 0.95000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170413/epoch=1-step=14.ckpt' as top 3
Time limit reached. Elapsed time is 0:03:00. Signaling Trainer to stop.
Epoch 2, global step 14: 'val_acc' reached 0.95000 (best 0.95000), saving model to '/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170413/epoch=2-step=14.ckpt' as top 3
Start to fuse 3 checkpoints via the greedy soup algorithm.
AutoMM has created your model 🎉🎉🎉
- To load the model, use the code below:
```python
from autogluon.multimodal import MultiModalPredictor
predictor = MultiModalPredictor.load("/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170413")
```
- You can open a terminal and launch Tensorboard to visualize the training log:
```shell
# Assume you have installed tensorboard
tensorboard --logdir /home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20230302_170413
```
- If you are not satisfied with the model, try to increase the training time,
adjust the hyperparameters (https://auto.gluon.ai/stable/tutorials/multimodal/advanced_topics/customization.html),
or post issues on GitHub: https://github.com/autogluon/autogluon
<autogluon.multimodal.predictor.MultiModalPredictor at 0x7f20c7f587c0>
We can see that best_quality achieves better performance than high_quality.
scores = predictor.evaluate(test_data, metrics=["roc_auc"])
scores
{'roc_auc': 0.9534104782352446}
HPO Presets#
The above three presets all use the default hyperparameters, which might not be optimal for your tasks. Fortunately, we also support hyperparameter optimization (HPO) with simple presets. To perform HPO, you can add a postfix _hpo in the three presets, resulting in medium_quality_hpo, high_quality_hpo, and best_quality_hpo.
Display Presets#
In case you want to see each preset’s inside details, we provide you with a util function to get the hyperparameter setups. For example, here are hyperparameters of preset high_quality.
import json
from autogluon.multimodal.presets import get_automm_presets
hyperparameters, hyperparameter_tune_kwargs = get_automm_presets(problem_type="default", presets="high_quality")
print(f"hyperparameters: {json.dumps(hyperparameters, sort_keys=True, indent=4)}")
print(f"hyperparameter_tune_kwargs: {json.dumps(hyperparameter_tune_kwargs, sort_keys=True, indent=4)}")
hyperparameters: {
"model.document_transformer.checkpoint_name": "microsoft/layoutlmv3-base",
"model.hf_text.checkpoint_name": "google/electra-base-discriminator",
"model.names": [
"categorical_mlp",
"numerical_mlp",
"timm_image",
"hf_text",
"document_transformer",
"fusion_mlp"
],
"model.timm_image.checkpoint_name": "swin_base_patch4_window7_224"
}
hyperparameter_tune_kwargs: {}
The HPO presets make several hyperparameters tunable such as model backbone, batch size, learning rate, max epoch, and optimizer type. Below are the details of preset high_quality_hpo.
import json
import yaml
from autogluon.multimodal.presets import get_automm_presets
hyperparameters, hyperparameter_tune_kwargs = get_automm_presets(problem_type="default", presets="high_quality_hpo")
print(f"hyperparameters: {yaml.dump(hyperparameters, allow_unicode=True, default_flow_style=False)}")
print(f"hyperparameter_tune_kwargs: {json.dumps(hyperparameter_tune_kwargs, sort_keys=True, indent=4)}")
hyperparameters: env.batch_size: !!python/object:ray.tune.search.sample.Categorical
categories:
- 16
- 32
- 64
- 128
- 256
sampler: !!python/object:ray.tune.search.sample._Uniform {}
model.document_transformer.checkpoint_name: microsoft/layoutlmv3-base
model.hf_text.checkpoint_name: !!python/object:ray.tune.search.sample.Categorical
categories:
- google/electra-base-discriminator
- google/flan-t5-base
- microsoft/deberta-v3-small
- roberta-base
- albert-xlarge-v2
sampler: !!python/object:ray.tune.search.sample._Uniform {}
model.names:
- categorical_mlp
- numerical_mlp
- timm_image
- hf_text
- document_transformer
- fusion_mlp
model.timm_image.checkpoint_name: !!python/object:ray.tune.search.sample.Categorical
categories:
- swin_base_patch4_window7_224
- convnext_base_in22ft1k
- vit_base_patch16_clip_224.laion2b_ft_in12k_in1k
sampler: !!python/object:ray.tune.search.sample._Uniform {}
optimization.learning_rate: !!python/object:ray.tune.search.sample.Float
lower: 1.0e-05
sampler: !!python/object:ray.tune.search.sample._LogUniform
base: 10
upper: 0.01
optimization.max_epochs: !!python/object:ray.tune.search.sample.Categorical
categories:
- 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
sampler: !!python/object:ray.tune.search.sample._Uniform {}
optimization.optim_type: !!python/object:ray.tune.search.sample.Categorical
categories:
- adamw
- sgd
sampler: !!python/object:ray.tune.search.sample._Uniform {}
hyperparameter_tune_kwargs: {
"num_trials": 512,
"scheduler": "ASHA",
"searcher": "bayes"
}
Other Examples#
You may go to AutoMM Examples to explore other examples about AutoMM.
Customization#
To learn how to customize AutoMM, please refer to Customize AutoMM.