DS - Scaler | ex MLE - dataX.ai | Amazon ML Summer School '23 | Research (NLP & RL) BigCode | Applied ML Instructor

github

devto

hashnode

Swayam Singh  • Peerlist

Gravitational Waves Sonification

L-WAVE is an innovative project that seamlessly integrates the power of a cutting-edge Llama2 model with the vast resources of the web. With the ability to access unknown information through intelligent Google searches, L-WAVE empowers users with instant knowledge retrieval. Not only does L-WAVE provide text-based answers, but it also boasts an impressive image rendering feature that brings visual context to your inquiries. Harnessing the strengths of both the Llama model and the web, L-WAVE is your ultimate companion for uncovering insights and visualizing information. Experience the future of information access with L-WAVE.

L-WAVE: LlaMA Web Access and Visual Enhancement

•Using XLM-RoBERTa finetuned over {German, French and Italian} from xtreme dataset
•Can perform Zero-Shot Classification over other languages up to decent f1-score
•Processed and analyzed thoroughly using custom Error Analysis pipeline and mitigating the possible failures.

Multilingual Name Entity Recognition

PythonCoder is a code generation model only trained on Python dataset (codeparrot/codeparrot-clean) . It is a custom model with a context window of 1024 tokens and its architecture is based on OpenAI's GPT2 with MultiQuery Attention and FlashAttention (MultiHead Attention is also available as an option).

It is not a commercial code-gen model instead made for educative purposes to demonstrate how each aspect is implemented and combined using PyTorch to build and train a GPT-like code-gen model.

PythonCoder

- Gradio interface to assess the reasoning capabilities of StarCoder via 3-level of strategies: PAL, TA, and MathPrompter.
- Released along with StarCoder under BigCode organisation at Hugging Face.

Reasoning with StarCoder

recursion_trace is a Python package that provides a decorator to trace recursive function calls. It generates a visual recursion tree using Graphviz, making it easier to understand and debug recursive algorithms.

Recurion-Trace

•It's a custom Gym environment built to train RL models on the Edwin Lander game to land via following the rules
•Current model is trained on PPO algorithm and attained the perfect landing score by training up to 1M ste

Reinforced Edwin Lander

•Allow user to try on different clothes virtually, without going to any trial room.
•Used ResNet101 and UNet to segment the cloth and model respectively and pose estimation using openpose then final predictions are done using the PyTorch implementation of VITON.
•With SSIM (Structural Similarity Index Metric) of 0.895

Virtual Clothing Assistant (100 + ⭐️ on GitHub)

- It can impose the texture/style of one image on the other without disturbing the content of the other image.
- Used pre-trained VGG-19 network to extract the features from multiple convolutional layers for style and content then merged them while maintaining the content and style tradeoff.