About Me

I am Deep Learning Programmer specialising in training Deep Learning model on computer vision and also visualising neural network attentions to see how model is doing. It includes Grad-CAM, neurons activations, layers visualisation etc. I have trained several models on object detection (large as satellite-images and small as cashews nuts), gaze detection, facial attributes (like gender, emotion, age, pose on single model).

Now, I have been investing effective timings in fine-tuning LLMs. I have done couple of PoCs and hosted website for question-answering with BERT/GPT model and vector database (pincecone;redis). I have been fine-tuning model like FLAN-T5, GPT with PEFT using LoRA, Adapter techniques which reduces the parameters to be trained hence most weights remain unchanged. Also worked on BLEU and ROUGE score for model performance and evaluation. I have also worked on models hosted on hugging face (like Llama, GPT etc) and benchmarks like HELM and GLUE.

I have done research on temporal data for determining fight, vandalism, bullying etc. I have trained these models on (with modified) MobileNets, Squeezenets, FireNets etc. to run on embedded devices like RockPro, Nvidia Nano etc. Where device is not concern, i trained models on ResNet-101 also with little bit modification as related to features extractions.

Summary of my DL/ML proficiency skills are:

• PyTorch (training on clusters, optimising neural networks, Grad-CAM, statistical modeling, 2D, 3D, convolution, depth convolution, SK-Net optimised, trained with depth information for Medical science (CT-scan, MRI etc))
• Python (Threading and multiprocessing; Flask; scikit-learn, bayesian modelling, matplotlib, ggplot, visualising techniques)
• Shell scripting (running jobs from kernel, shell scripting, AWK, ruby, well versed with regex for fast parsing while making training datasets, cronjob initiation)
• Data Analysis (extensive analysis on data before going for training, visualisation in all respects to check the data forming which normalisation, plotting, features extraction and normalising decision factor)
• Deployment (deployment of product using docker-image with pip requirements file, Jira, cluster Jupiter notebook management, kubernetes scalable.
• Optimization (quantisation, model running by hyper parameters, prunning network layers, int8,int16,FP16 conversion while training)
• Model testing (test benchmark on trained model to check how good is model, Normalise Mutual Information, Adjust Mutual Information, R-precision, ROC/AUC curve, Precision, Recall, specificity (confusion matrix), F1-score)
• Squeezing model if size and accuracy is a BIG concern to run on embedded devices.