Multimodal Data Tables: Tabular, Text, and Image¶
Tip: Prior to reading this tutorial, it is recommended to have a basic understanding of the TabularPredictor API covered in Predicting Columns in a Table - Quick Start.
In this tutorial, we will train a multi-modal ensemble using data that contains image, text, and tabular features.
Note: A GPU is required for this tutorial in order to train the image and text models. Additionally, GPU installations are required for MXNet and Torch with appropriate CUDA versions.
The PetFinder Dataset¶
We will be using the PetFinder dataset. The PetFinder dataset provides information about shelter animals that appear on their adoption profile with the goal to predict the adoption rate of the animal. The end goal is for rescue shelters to use the predicted adoption rate to identify animals whose profiles could be improved so that they can find a home.
Each animal’s adoption profile contains a variety of information, such as pictures of the animal, a text description of the animal, and various tabular features such as age, breed, name, color, and more.
To get started, we first need to download the dataset. Datasets that contain images require more than a CSV file, so the dataset is packaged in a zip file in S3. We will first download it and unzip the contents:
download_dir = './ag_petfinder_tutorial'
zip_file = 'https://automl-mm-bench.s3.amazonaws.com/petfinder_kaggle.zip'
from autogluon.core.utils.loaders import load_zip
load_zip.unzip(zip_file, unzip_dir=download_dir)
Downloading ./ag_petfinder_tutorial/file.zip from https://automl-mm-bench.s3.amazonaws.com/petfinder_kaggle.zip...
100%|██████████| 2.00G/2.00G [00:31<00:00, 62.4MiB/s]
Now that the data is download and unzipped, let’s take a look at the contents:
import os
os.listdir(download_dir)
['file.zip', 'petfinder_processed']
‘file.zip’ is the original zip file we downloaded, and ‘petfinder_processed’ is a directory containing the dataset files.
dataset_path = download_dir + '/petfinder_processed'
os.listdir(dataset_path)
['train.csv', 'train_images', 'test.csv', 'test_images', 'dev.csv']
Here we can see the train, test, and dev CSV files, as well as two directories: ‘test_images’ and ‘train_images’ which contain the image JPG files.
Note: We will be using the dev data as testing data as dev contains the
ground truth labels for showing scores via predictor.leaderboard.
Let’s take a peek at the first 10 files inside of the ‘train_images’ directory:
os.listdir(dataset_path + '/train_images')[:10]
['d765ae877-1.jpg',
'756025f7c-2.jpg',
'e1a2d9477-4.jpg',
'6d18707ee-2.jpg',
'96607bca0-5.jpg',
'fde58f7fa-10.jpg',
'be7b65c23-3.jpg',
'dd36ab692-3.jpg',
'2d8db1c19-2.jpg',
'53037f091-2.jpg']
As expected, these are the images we will be training with alongside the other features.
Next, we will load the train and dev CSV files:
import pandas as pd
train_data = pd.read_csv(f'{dataset_path}/train.csv', index_col=0)
test_data = pd.read_csv(f'{dataset_path}/dev.csv', index_col=0)
train_data.head(3)
| Type | Name | Age | Breed1 | Breed2 | Gender | Color1 | Color2 | Color3 | MaturitySize | ... | Quantity | Fee | State | RescuerID | VideoAmt | Description | PetID | PhotoAmt | AdoptionSpeed | Images | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 10721 | 1 | Elbi | 2 | 307 | 307 | 2 | 5 | 0 | 0 | 3 | ... | 1 | 0 | 41336 | e9a86209c54f589ba72c345364cf01aa | 0 | I'm looking for people to adopt my dog | e4b90955c | 4.0 | 4 | train_images/e4b90955c-1.jpg;train_images/e4b9... |
| 13114 | 2 | Darling | 4 | 266 | 0 | 1 | 1 | 0 | 0 | 2 | ... | 1 | 0 | 41401 | 01f954cdf61526daf3fbeb8a074be742 | 0 | Darling was born at the back lane of Jalan Alo... | a0c1384d1 | 5.0 | 3 | train_images/a0c1384d1-1.jpg;train_images/a0c1... |
| 13194 | 1 | Wolf | 3 | 307 | 0 | 1 | 1 | 2 | 0 | 2 | ... | 1 | 0 | 41332 | 6e19409f2847326ce3b6d0cec7e42f81 | 0 | I found Wolf about a month ago stuck in a drai... | cf357f057 | 7.0 | 4 | train_images/cf357f057-1.jpg;train_images/cf35... |
3 rows × 25 columns
Looking at the first 3 examples, we can tell that there is a variety of tabular features, a text description (‘Description’), and an image path (‘Images’).
For the PetFinder dataset, we will try to predict the speed of adoption for the animal (‘AdoptionSpeed’), grouped into 5 categories. This means that we are dealing with a multi-class classification problem.
label = 'AdoptionSpeed'
image_col = 'Images'
Preparing the image column¶
Let’s take a look at what a value in the image column looks like:
train_data[image_col].iloc[0]
'train_images/e4b90955c-1.jpg;train_images/e4b90955c-2.jpg;train_images/e4b90955c-3.jpg;train_images/e4b90955c-4.jpg'
Currently, AutoGluon only supports one image per row. Since the PetFinder dataset contains one or more images per row, we first need to preprocess the image column to only contain the first image of each row.
train_data[image_col] = train_data[image_col].apply(lambda ele: ele.split(';')[0])
test_data[image_col] = test_data[image_col].apply(lambda ele: ele.split(';')[0])
train_data[image_col].iloc[0]
'train_images/e4b90955c-1.jpg'
AutoGluon loads images based on the file path provided by the image column.
Here we update the path to point to the correct location on disk:
def path_expander(path, base_folder):
path_l = path.split(';')
return ';'.join([os.path.abspath(os.path.join(base_folder, path)) for path in path_l])
train_data[image_col] = train_data[image_col].apply(lambda ele: path_expander(ele, base_folder=dataset_path))
test_data[image_col] = test_data[image_col].apply(lambda ele: path_expander(ele, base_folder=dataset_path))
train_data[image_col].iloc[0]
'/home/ci/autogluon/docs/_build/eval/tutorials/tabular_prediction/ag_petfinder_tutorial/petfinder_processed/train_images/e4b90955c-1.jpg'
train_data.head(3)
| Type | Name | Age | Breed1 | Breed2 | Gender | Color1 | Color2 | Color3 | MaturitySize | ... | Quantity | Fee | State | RescuerID | VideoAmt | Description | PetID | PhotoAmt | AdoptionSpeed | Images | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 10721 | 1 | Elbi | 2 | 307 | 307 | 2 | 5 | 0 | 0 | 3 | ... | 1 | 0 | 41336 | e9a86209c54f589ba72c345364cf01aa | 0 | I'm looking for people to adopt my dog | e4b90955c | 4.0 | 4 | /home/ci/autogluon/docs/_build/eval/tutorials/... |
| 13114 | 2 | Darling | 4 | 266 | 0 | 1 | 1 | 0 | 0 | 2 | ... | 1 | 0 | 41401 | 01f954cdf61526daf3fbeb8a074be742 | 0 | Darling was born at the back lane of Jalan Alo... | a0c1384d1 | 5.0 | 3 | /home/ci/autogluon/docs/_build/eval/tutorials/... |
| 13194 | 1 | Wolf | 3 | 307 | 0 | 1 | 1 | 2 | 0 | 2 | ... | 1 | 0 | 41332 | 6e19409f2847326ce3b6d0cec7e42f81 | 0 | I found Wolf about a month ago stuck in a drai... | cf357f057 | 7.0 | 4 | /home/ci/autogluon/docs/_build/eval/tutorials/... |
3 rows × 25 columns
Analyzing an example row¶
Now that we have preprocessed the image column, let’s take a look at an example row of data and display the text description and the picture.
example_row = train_data.iloc[1]
example_row
Type 2
Name Darling
Age 4
Breed1 266
Breed2 0
Gender 1
Color1 1
Color2 0
Color3 0
MaturitySize 2
FurLength 1
Vaccinated 2
Dewormed 2
Sterilized 2
Health 1
Quantity 1
Fee 0
State 41401
RescuerID 01f954cdf61526daf3fbeb8a074be742
VideoAmt 0
Description Darling was born at the back lane of Jalan Alo...
PetID a0c1384d1
PhotoAmt 5.0
AdoptionSpeed 3
Images /home/ci/autogluon/docs/_build/eval/tutorials/...
Name: 13114, dtype: object
example_row['Description']
'Darling was born at the back lane of Jalan Alor and was foster by a feeder. All his siblings had died of accident. His mother and grandmother had just been spayed. Darling make a great condo/apartment cat. He love to play a lot. He would make a great companion for someone looking for a cat to love.'
example_image = example_row['Images']
from IPython.display import Image, display
pil_img = Image(filename=example_image)
display(pil_img)
The PetFinder dataset is fairly large. For the purposes of the tutorial, we will sample 300 rows for training.
Training on large multi-modal datasets can be very computationally
intensive, especially if using the best_quality preset in AutoGluon.
When prototyping, it is recommended to sample your data to get an idea
of which models are worth training, then gradually train with larger
amounts of data and longer time limits as you would with any other
machine learning algorithm.
train_data = train_data.sample(300, random_state=0)
Constructing the FeatureMetadata¶
Next, let’s see what AutoGluon infers the feature types to be by constructing a FeatureMetadata object from the training data:
from autogluon.tabular import FeatureMetadata
feature_metadata = FeatureMetadata.from_df(train_data)
print(feature_metadata)
('float', []) : 1 | ['PhotoAmt']
('int', []) : 19 | ['Type', 'Age', 'Breed1', 'Breed2', 'Gender', ...]
('object', []) : 4 | ['Name', 'RescuerID', 'PetID', 'Images']
('object', ['text']) : 1 | ['Description']
Notice that FeatureMetadata automatically identified the column ‘Description’ as text, so we don’t need to manually specify that it is text.
In order to leverage images, we need to tell AutoGluon which column contains the image path. We can do this by specifying a FeatureMetadata object and adding the ‘image_path’ special type to the image column. We later pass this custom FeatureMetadata to TabularPredictor.fit.
feature_metadata = feature_metadata.add_special_types({image_col: ['image_path']})
print(feature_metadata)
('float', []) : 1 | ['PhotoAmt']
('int', []) : 19 | ['Type', 'Age', 'Breed1', 'Breed2', 'Gender', ...]
('object', []) : 3 | ['Name', 'RescuerID', 'PetID']
('object', ['image_path']) : 1 | ['Images']
('object', ['text']) : 1 | ['Description']
Specifying the hyperparameters¶
Next, we need to specify the models we want to train with. This is done
via the hyperparameters argument to TabularPredictor.fit.
AutoGluon has a predefined config that works well for multimodal datasets called ‘multimodal’. We can access it via:
from autogluon.tabular.configs.hyperparameter_configs import get_hyperparameter_config
hyperparameters = get_hyperparameter_config('multimodal')
hyperparameters
{'NN_TORCH': {},
'GBM': [{},
{'extra_trees': True, 'ag_args': {'name_suffix': 'XT'}},
'GBMLarge'],
'CAT': {},
'XGB': {},
'AG_TEXT_NN': {'presets': 'medium_quality_faster_train'},
'AG_IMAGE_NN': {},
'VW': {}}
This hyperparameter config will train a variety of Tabular models as well as finetune an Electra BERT text model, and a ResNet image model.
Fitting with TabularPredictor¶
Now we will train a TabularPredictor on the dataset, using the feature metadata and hyperparameters we defined prior. This TabularPredictor will leverage tabular, text, and image features all at once.
from autogluon.tabular import TabularPredictor
predictor = TabularPredictor(label=label).fit(
train_data=train_data,
hyperparameters=hyperparameters,
feature_metadata=feature_metadata,
time_limit=600,
)
No path specified. Models will be saved in: "AutogluonModels/ag-20230204_010942/"
Beginning AutoGluon training ... Time limit = 600s
AutoGluon will save models to "AutogluonModels/ag-20230204_010942/"
AutoGluon Version: 0.6.3b20230204
Python Version: 3.8.13
Operating System: Linux
Platform Machine: x86_64
Platform Version: #1 SMP Tue Nov 30 00:17:50 UTC 2021
Train Data Rows: 300
Train Data Columns: 24
Label Column: AdoptionSpeed
Preprocessing data ...
AutoGluon infers your prediction problem is: 'multiclass' (because dtype of label-column == int, but few unique label-values observed).
5 unique label values: [2, 3, 4, 0, 1]
If 'multiclass' is not the correct problem_type, please manually specify the problem_type parameter during predictor init (You may specify problem_type as one of: ['binary', 'multiclass', 'regression'])
Warning: Updated label_count_threshold from 10 to 8 to avoid cutting too many classes.
Train Data Class Count: 5
Using Feature Generators to preprocess the data ...
Fitting AutoMLPipelineFeatureGenerator...
Available Memory: 31409.67 MB
Train Data (Original) Memory Usage: 0.3 MB (0.0% of available memory)
Stage 1 Generators:
Fitting AsTypeFeatureGenerator...
Note: Converting 2 features to boolean dtype as they only contain 2 unique values.
Stage 2 Generators:
Fitting FillNaFeatureGenerator...
Stage 3 Generators:
Fitting IdentityFeatureGenerator...
Fitting IdentityFeatureGenerator...
Fitting RenameFeatureGenerator...
Fitting CategoryFeatureGenerator...
Fitting CategoryMemoryMinimizeFeatureGenerator...
Fitting TextSpecialFeatureGenerator...
Fitting BinnedFeatureGenerator...
Fitting DropDuplicatesFeatureGenerator...
Fitting TextNgramFeatureGenerator...
Fitting CountVectorizer for text features: ['Description']
CountVectorizer fit with vocabulary size = 96
Fitting IdentityFeatureGenerator...
Fitting IsNanFeatureGenerator...
Stage 4 Generators:
Fitting DropUniqueFeatureGenerator...
Unused Original Features (Count: 1): ['PetID']
These features were not used to generate any of the output features. Add a feature generator compatible with these features to utilize them.
Features can also be unused if they carry very little information, such as being categorical but having almost entirely unique values or being duplicates of other features.
These features do not need to be present at inference time.
('object', []) : 1 | ['PetID']
Types of features in original data (raw dtype, special dtypes):
('float', []) : 1 | ['PhotoAmt']
('int', []) : 18 | ['Type', 'Age', 'Breed1', 'Breed2', 'Gender', ...]
('object', []) : 2 | ['Name', 'RescuerID']
('object', ['image_path']) : 1 | ['Images']
('object', ['text']) : 1 | ['Description']
Types of features in processed data (raw dtype, special dtypes):
('category', []) : 2 | ['Name', 'RescuerID']
('category', ['text_as_category']) : 1 | ['Description']
('float', []) : 1 | ['PhotoAmt']
('int', []) : 16 | ['Age', 'Breed1', 'Breed2', 'Gender', 'Color1', ...]
('int', ['binned', 'text_special']) : 20 | ['Description.char_count', 'Description.word_count', 'Description.capital_ratio', 'Description.lower_ratio', 'Description.digit_ratio', ...]
('int', ['bool']) : 2 | ['Type', 'VideoAmt']
('int', ['text_ngram']) : 97 | ['__nlp__.active', '__nlp__.adopt', '__nlp__.adoption', '__nlp__.adorable', '__nlp__.all', ...]
('object', ['image_path']) : 1 | ['Images']
('object', ['text']) : 1 | ['Description_raw_text']
0.3s = Fit runtime
23 features in original data used to generate 141 features in processed data.
Train Data (Processed) Memory Usage: 0.29 MB (0.0% of available memory)
Data preprocessing and feature engineering runtime = 0.35s ...
AutoGluon will gauge predictive performance using evaluation metric: 'accuracy'
To change this, specify the eval_metric parameter of Predictor()
Automatically generating train/validation split with holdout_frac=0.2, Train Rows: 240, Val Rows: 60
Fitting 9 L1 models ...
Fitting model: LightGBM ... Training model for up to 599.65s of the 599.64s of remaining time.
0.4 = Validation score (accuracy)
1.16s = Training runtime
0.01s = Validation runtime
Fitting model: LightGBMXT ... Training model for up to 598.48s of the 598.47s of remaining time.
0.4 = Validation score (accuracy)
1.33s = Training runtime
0.01s = Validation runtime
Fitting model: CatBoost ... Training model for up to 597.08s of the 597.08s of remaining time.
0.4167 = Validation score (accuracy)
3.83s = Training runtime
0.01s = Validation runtime
Fitting model: XGBoost ... Training model for up to 593.24s of the 593.24s of remaining time.
0.3667 = Validation score (accuracy)
1.83s = Training runtime
0.01s = Validation runtime
Fitting model: NeuralNetTorch ... Training model for up to 591.38s of the 591.38s of remaining time.
0.3167 = Validation score (accuracy)
1.64s = Training runtime
0.02s = Validation runtime
Fitting model: VowpalWabbit ... Training model for up to 589.71s of the 589.71s of remaining time.
0.25 = Validation score (accuracy)
0.36s = Training runtime
0.02s = Validation runtime
Fitting model: LightGBMLarge ... Training model for up to 589.14s of the 589.14s of remaining time.
0.3833 = Validation score (accuracy)
1.1s = Training runtime
0.01s = Validation runtime
Fitting model: TextPredictor ... Training model for up to 588.03s of the 588.02s of remaining time.
Warning: Exception caused TextPredictor to fail during training... Skipping this model.
Unknown preset type: medium_quality_faster_train
Detailed Traceback:
Traceback (most recent call last):
File "/home/ci/autogluon/core/src/autogluon/core/trainer/abstract_trainer.py", line 1502, in _train_and_save
model = self._train_single(X, y, model, X_val, y_val, total_resources=total_resources, **model_fit_kwargs)
File "/home/ci/autogluon/core/src/autogluon/core/trainer/abstract_trainer.py", line 1447, in _train_single
model = model.fit(X=X, y=y, X_val=X_val, y_val=y_val, total_resources=total_resources, **model_fit_kwargs)
File "/home/ci/autogluon/core/src/autogluon/core/models/abstract/abstract_model.py", line 703, in fit
out = self._fit(**kwargs)
File "/home/ci/autogluon/tabular/src/autogluon/tabular/models/automm/automm_model.py", line 193, in _fit
self.model.fit(
File "/home/ci/autogluon/multimodal/src/autogluon/multimodal/predictor.py", line 843, in fit
self._fit(**_fit_args)
File "/home/ci/autogluon/multimodal/src/autogluon/multimodal/predictor.py", line 1004, in _fit
config = get_config(
File "/home/ci/autogluon/multimodal/src/autogluon/multimodal/utils/config.py", line 143, in get_config
preset_overrides, _ = get_automm_presets(problem_type=problem_type, presets=presets)
File "/home/ci/autogluon/multimodal/src/autogluon/multimodal/presets.py", line 617, in get_automm_presets
hyperparameters, hyperparameter_tune_kwargs = automm_presets.create(problem_type, presets)
File "/home/ci/autogluon/multimodal/src/autogluon/multimodal/registry.py", line 92, in create
return self.get(name)(*args, **kwargs)
File "/home/ci/autogluon/multimodal/src/autogluon/multimodal/presets.py", line 89, in default
raise ValueError(f"Unknown preset type: {presets}")
ValueError: Unknown preset type: medium_quality_faster_train
Fitting model: ImagePredictor ... Training model for up to 584.32s of the 584.32s of remaining time.
/home/ci/opt/venv/lib/python3.8/site-packages/torch/functional.py:504: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at ../aten/src/ATen/native/TensorShape.cpp:3190.)
return _VF.meshgrid(tensors, **kwargs) # type: ignore[attr-defined]
Downloading: "https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window7_224_22kto1k.pth" to /home/ci/.cache/torch/hub/checkpoints/swin_base_patch4_window7_224_22kto1k.pth
/home/ci/opt/venv/lib/python3.8/site-packages/pytorch_lightning/utilities/cloud_io.py:33: LightningDeprecationWarning: pytorch_lightning.utilities.cloud_io.get_filesystem has been deprecated in v1.8.0 and will be removed in v1.10.0. Please use lightning_lite.utilities.cloud_io.get_filesystem instead.
rank_zero_deprecation(
/home/ci/opt/venv/lib/python3.8/site-packages/pytorch_lightning/utilities/cloud_io.py:25: LightningDeprecationWarning: pytorch_lightning.utilities.cloud_io.atomic_save has been deprecated in v1.8.0 and will be removed in v1.10.0. This function is internal but you can copy over its implementation.
rank_zero_deprecation(
0.3667 = Validation score (accuracy)
210.76s = Training runtime
1.59s = Validation runtime
Fitting model: WeightedEnsemble_L2 ... Training model for up to 360.0s of the 371.94s of remaining time.
0.4167 = Validation score (accuracy)
0.19s = Training runtime
0.0s = Validation runtime
AutoGluon training complete, total runtime = 238.28s ... Best model: "WeightedEnsemble_L2"
TabularPredictor saved. To load, use: predictor = TabularPredictor.load("AutogluonModels/ag-20230204_010942/")
After the predictor is fit, we can take a look at the leaderboard and see the performance of the various models:
leaderboard = predictor.leaderboard(test_data)
model score_test score_val pred_time_test pred_time_val fit_time pred_time_test_marginal pred_time_val_marginal fit_time_marginal stack_level can_infer fit_order
0 ImagePredictor 0.304101 0.366667 16.983891 1.585901 210.755301 16.983891 1.585901 210.755301 1 True 8
1 XGBoost 0.299433 0.366667 0.122267 0.005663 1.825729 0.122267 0.005663 1.825729 1 True 4
2 LightGBM 0.296099 0.400000 0.032434 0.005447 1.155379 0.032434 0.005447 1.155379 1 True 1
3 NeuralNetTorch 0.292431 0.316667 0.041564 0.019151 1.643744 0.041564 0.019151 1.643744 1 True 5
4 CatBoost 0.291764 0.416667 0.019191 0.008565 3.827535 0.019191 0.008565 3.827535 1 True 3
5 WeightedEnsemble_L2 0.291764 0.416667 0.023420 0.009218 4.014848 0.004229 0.000653 0.187313 2 True 9
6 LightGBMXT 0.291430 0.400000 0.178973 0.007930 1.326265 0.178973 0.007930 1.326265 1 True 2
7 LightGBMLarge 0.275425 0.383333 0.346112 0.005080 1.098513 0.346112 0.005080 1.098513 1 True 7
8 VowpalWabbit 0.260754 0.250000 0.700607 0.024119 0.364768 0.700607 0.024119 0.364768 1 True 6
That’s all it takes to train with image, text, and tabular data (at the same time) using AutoGluon!
For an in-depth tutorial on text + tabular multimodal functionality, refer to Multimodal Data Tables: Combining BERT/Transformers and Classical Tabular Models.
For more tutorials, refer to Predicting Columns in a Table - Quick Start and Predicting Columns in a Table - In Depth.