PyYel 0.1.0

Transformers release

PyYel

PyYel is a personnal library that aims at helping the deployement of strong data science tools, from data handling to deep learning.

Quick start

1. Install the library.

your_path> pip install PyYel

2. Import the library into you code.

import pyl

3. Import the relevant features.

from pyl.models.LLM import LLMDecodingPhi, LLMEncodingBARTLargeMNLI
from pyl.models.CNN import CNNClassificationResNet

Content

Data

A collection of features to manipulate the data. Can be used to implement pipelines, preprocessing, data augmentation...

• Augmentations: a compilation of classes featuring methods to augment a datapoint of various type.

  • ImageAugmentation : features a handfull of functions that can augment any type of data, as well as its labels.
  • TODO

• Reduction: acompilation of classes featuring methods to reduce datapoint of various type.

  • TODO/TO-REWORK

• Utils: a collection of powerful tools that permit an easy manipulation of the datapoints.

  • TODO/TO-REWORK

Models

The neural networks implementations. These are grouped by types and tasks.

• CNN (Convolutional Neural Networks)

Source model	PyYel model	Task	Status
ResNet	CNNCLassificationResNet	Classification	Implemented
FasterRCNN	CNNDetectionFasterRCNN	Detection	Implemented
SSD	CNNDetectionSSD	Detection	Implemented
RetinaNet	CNNDetectionRetinaNet	Detection	TODO
/	CNNKeypoint	Keypoint detection	TODO
FCN	CNNSegmentationFCN	Segmentation	Implemented/TODO
DeeplabV3	CNNSegmentationDeeplabV3	Segmentation	Implemented/TODO

Note: Traditionnal computer vision networks. Features a model builder to design custom small-sized networks.

• FCN (Fully Connected Networks)

Source model	PyYel model	Task	Status
/	FCNBuilder	/	TODO
Note: Dense models. Features a model builder to design custom small-sized networks.

• LLM (Large Language Models)

Source model	PyYel model	Task	Status
Mistral7B v0.1	LLMDecodingMistral7B	Decoding: text-to-text generation	Implemented
OPT 125M	LLMDecodingOPT125m	Decoding: text-to-text generation	Implemented
Phi 3.5 Mini Instruct	LLMDecodingPhi	Decoding: text-to-text generation	Implemented
Phi 3.5 MoE	LLMDecodingPhiMoE	Decoding: text-to-text generation	Implemented/TODO
BART Large	LLMEncodingBARTLargeMNLI	Encoding: zero-shoot classification	Implemented
DeBERTaV3 Base	LLMEncodingDeBERTaV3Base	Encoding: zero-shoot classification	Implemented
DeBERTaV3 Base	LLMEncodingDeBERTaV3BaseMNLI	Encoding: zero-shoot classification	Implemented
DeBERTaV3 Large	LLMEncodingDeBERTaV3Large	Encoding: zero-shoot classification	Implemented

Note: NLP transformers.

• LVM (Large Vision Models)

Source model	PyYel model	Task	Status
ViT	LVMVisionTransformerClassification	Classification	TODO

Note: Computer vision transformers.

Utils

A collection of higher-level tools, that simplifies the manipulation of the library

TODO/TO-REWORK

Notes

TODO