Auto-Deep-Learning (Auto Deep Learning)
auto_deep_learning
: with this package, you will be able to create, train and deploy neural networks automatically based on the input that you provide.
Alert
This package is still on development, but Start the Project to know further updates in next days!
For the moment, would be for computer vision classification tasks on images (multi-modal included).
Installation
Use the package manager pip to install auto_deep_learning.
To install the package:
pip install auto_deep_learning
If using an old version of the package, update it:
pip install --upgrade auto_deep_learning
Project Structure
The project structure is the following one:
├── auto_deep_learning
│ ├── cloud
│ │ ├── aws
│ │ └── gcp
│ ├── enum
│ ├── exceptions
│ │ ├── model
│ │ └── utils
│ │ └── data_handler
│ ├── model
│ │ └── arch
│ │ └── convolution
│ ├── schemas
│ └── utils
│ ├── data_handler
│ │ ├── creator
│ │ └── transform
│ └── model
├── examples
└── tests
Basic Usage
How easy can be to create and train a deep learning model:
from auto_deep_learning import Model
from auto_deep_learning.utils import DataCreator, DataSampler, image_folder_convertion
df = image_folder_convertion()
data = DataCreator(df)
data_sampled = DatasetSampler(data)
model = Model(data_sampled)
model.fit()
model.predict('image.jpg')
We provide also with a configuration object, where it centralizes some of the most important configurations that you might want to do:
ConfigurationObject(
n_epochs: int = 10,
batch_size_train: int = 64,
batch_size_valid: int = 128,
batch_size_test: int = 128,
valid_size: float = 0.1,
test_size: float = 0.05,
image_size: int = 224,
num_workers: int = 6,
objective: ModelObjective = ModelObjective.THROUGHPUT,
img_transformers: Dict[str, ImageTransformer] = {
'train': ImageTransformer(
rotation=3.0,
color_jitter_brightness=3.0,
color_jitter_contrast=3.0,
color_jitter_hue=3.0,
color_jitter_saturation=3.0,
color_jitter_enabled=True,
resized_crop_enabled=True
),
'valid': ImageTransformer(),
'test': ImageTransformer()
}
)
So by default, it is going to do image augmentation on the training data.
Note that if for example we did not want to make a validation split because our dataset is too small, we would change this value as:
conf_obj = ConfigurationObject()
conf_obj.valid_size = 0.0
Dataset
The data that it expects is a pd.DataFrame(), where the columns are the following:
- image_path: the path to the image
- class1: the classification of the class nr. 1. For example: {t-shirt, glasses, ...}
- class2: the classification of the class nr. 2. For example: {summer, winter, ...}
- ...
- split_type: whether it is for train/valid/test
For better performance, it is suggested that the classes and the type are of dtype category in the pandas DataFrame.
If the type is not provided in the dataframe, you should use the utils function of data_split_types (in utils.dataset.sampler file).
If instead you have the images ordered in the structure of ImageFolder, which is the following structure:
train/
class1_value/
1.jgp
2.jpg
...
class2_value/
3.jpg
4.jpg
...
test/
class1_value/
1.jgp
2.jpg
...
class2_value/
3.jpg
4.jpg
...
For simplifying logic, we have provided a logic that gives you the expected dataframe that we wanted, with the function of image_folder_convertion (in utils.functions), where it is expecting a path to the parent folder where the train/ and /test folders are.