Self-supervised-Learning Demo

This repository deploys the pre-trained neural networks from self-supervised pretraining for the variety downstream applications.

---

Section 1 Image Retrieval (Text-to-Image retrieval, Image-to-Image Retrieval)

1. Text-to-Image Retrieval

• Transformer model to extract the text-queries embedding representation

• ViTs model to extract Image embedding representation

• Retrieving Top-K similarity between the text-queries & data images-embedding

2. Image-to-Image Retrieval

• ViTs model extract Image-Queries embedding representation

• ViTs model extract all other images embedding representation

• Retrieving Top-K similarity between the Image-queries & data images-embedding

---

Section 2 Patch-Level Retrieval

• ViTs model extract Patch-Queries embedding representation

• ViTs model extract all other images Patches embedding representation

• Retrieving Top-K similarity between the Patch-queries & other Patches-embedding

---

Section 3 Image Segmentation

1. CoCo dataset segmentation

• ResNet50 pretraining with Heuristic Attention Representation Learning for Self-Supervised Pretraining
• Fine-tune MaskRCNN with ResNet50 backbone using coco dataset

• Segment objects in an image

2. Model feature attention part

• ResNet50 pretraining with Heuristic Attention Representation Learning for Self-Supervised Pretraining

• Get the attention map from ResNet50.

---

Section 4 Natural Image Classification Tasks

1. Linear evaluation for 12 datasets

2. Linear Evaluation with Sweeping the Hyper-parameters

3. Finetuning Pretrained model for 12 datasets

4. Finetuning with Sweeping the Hyper-parameters

---

Section 5 ImageNet Linear Evaluation and Semi-Supervised Learning

1. ImageNet Linear evaluation

2. Linear Evaluation with Sweeping the Hyper-parameters

3. ImageNet Semi-Supervised Learning

4. Finetuning with 1% & 10% plus Sweeping the Hyper-parameters