README.md

metadata

library_name: tensorflow-lite
tags:
  - litert
  - tflite
  - android
  - on-device-training
  - computer-vision
  - model-collection

LiteRT · general-purpose Android conversions

25 current variants, including 15 trainable heads/output adapters. Visual backbones stay frozen; the exact contract defines adaptation scope.

Quick card

Field	Reference
Use	Select a conversion by output task and documented adaptation contract.
Expected input	The exact image / auxiliary tensor signature, preprocessing and labels of the chosen variant.
Example	Inspect the runtime contract, check inference, then adapt reviewed targets only when genuine train/save/restore signatures exist.
Limits	Conversion and synthetic-learning checks do not establish common task accuracy, full ARM qualification or application integration for every task.
Contract	Per-variant contracts · Retained qualification matrix

Choose a variant · Resource navigation · Update log

Contract task	Variants
classification	caformer_s18_learning, convformer_s18_learning, dinov2_learning, edgenext_small_usi_learning, edgenext_x_small_learning, edgenext_xx_small_learning, efficientformer_l1_learning, hgnetv2_b0_learning, repvit_m1_learning, tinyclip_learning
classification + visual embeddings	convformer_s18, hgnetv2_b0, repvit_m1
image embeddings	dinov2
prompted segmentation	efficientvit_sam
vision-language generation	florence2
object detection	rfdetr, rtmdet_tiny
object_detection	rfdetr_learning, rtmdet_tiny_learning, table_transformer_detection_learning, table_transformer_structure_learning
image / text embeddings	tinyclip
pose keypoints	vitpose
pointing	vitpose_learning

Read the exact evidence

The retained matrix separates synthetic host learning, standalone SDK and application checks. Five recorded application passes concern EdgeNeXt Small/X-Small/XX-Small learning, RepViT M1 inference and RTMDet Tiny learning. RTMDet Tiny inference retains its failure; other variants remain pending according to that matrix. The reported RepViT learning SDK pass is separate. No general physical ARM or representative-domain accuracy qualification is claimed.

Specialized Cadryl models and recorded comparisons

Source weights/software keep their original terms. Read each variant’s notices, preserve original weights and human annotations, and keep adapted checkpoints separate. Existing artifact paths are retained.

Likes are always welcome! ❤️ If these resources help you, a little heart makes our day and helps others discover the work. Thank you!