NoShade.Vision

Summarized Videos

Abhishek-Thakur

Kaggle's 30 Days Of ML (Day 1): Getting Started With Kaggle

The Kaggle's 30 Days of ML challenge is introduced in this video, which involves learning something new about machine learning each day for the first 15 days and then working on a Kaggle competition for the next 15 days. The speaker provides an overview of setting up a Kaggle account and navigating the site, including the four categories, how to earn medals, and how points are earned in competitions. The basics of Kaggle are discussed using the Titanic competition as an example. The video covers importing packages, loading data, exploratory data analysis (EDA), and building a random forest machine learning model. The process of preparing data, submitting an entry to a Kaggle competition, and checking the leaderboard is also demonstrated. Finally, the speaker discusses other available features on Kaggle, the importance of feature engineering, and becoming a contributor.

Finetune LLMs (llama, vicuna, gptneo, pythia) without any code!

The video demonstrates how to fine-tune various LLM models such as llama, vicuna, GPT Neo, and pythia without writing any code, with the help of Hugging Face's Auto Train Advanced. The process involves selecting a natural language processing task, choosing a dataset for fine-tuning, selecting the appropriate base model, and setting the desired number of models for training. Once the project is created, the Auto Train Advanced takes care of the rest of the training process. The speaker also notes that Amazon SageMaker's Autopilot feature can be used to train and deploy models easily.

Stable Diffusion Inpainting with Segment Anything Model (SAM)

The video introduces the Stable Diffusion Inpainting with Segment Anything Model (SAM), which combines the segment editing model with the stable diffusion model to create a new and exciting model. The presenter demonstrates how to combine the two models and build a simple demo where an input image can have parts replaced with prompts or have the background replaced instead of the object. SAM is also used to inpaint an image using a segment mask created by clicking on the original image or to create a new image with a different background. The video concludes by encouraging viewers to like, subscribe, and share the video.

🤗 AutoTrain: Train state of-the-art image classification models (no code)

The video showcases the process of training a state-of-the-art image classification model without any coding using Hugging Face's AutoTrain Advanced. To do so, users need to duplicate the AutoTrain Advanced space, add their zip file-formatted dataset, select the number of models to tune, and choose between Auto Train or a model from the Hugging Face Hub. The presenter illustrates the process by uploading a flower classification dataset, which is automatically divided into training and validation sets. The free tier of AutoTrain allows users to train models up to 1,500 images, after which there is a cost. Once trained, the model is stored in the user's private repository on the Hugging Face Hub and tested using the hosted inference API. The video concludes with a request to like, subscribe, and share the content.

Custom object detection in Python using YOLOv8

The video covers the process of training and using YOLOv8, a state-of-the-art object detection model, on a custom dataset using Python. The presenter explains the process of converting the data set into the format expected by YOLOv8 and demonstrates how to create the necessary directories and files for training. They also show how to create a yaml file with the necessary parameters for training the model using pre-trained weights, adjust the parameters, and run the training process. The presenter also discusses how to check for labeling errors, plot the detected objects' bounding boxes, and save the images in PNG format. Finally, the presenter encourages viewers to explore the predict function's documentation and hints at future videos covering segmentation and post tasks.

How to create GPT powered conversational bot for any website

This video provides a step-by-step process for creating a conversational bot powered by GPT for any website. The presenter explains how to create a web crawler using Scrapy, parse the extracted data, and store output files. They also show how to create an index using Lang chain and use a conversational retrieval chain to implement GPT. Additionally, a Python-based conversational bot using radio is demonstrated with a diffuser-related example. The video offers several code snippets and encourages viewers to provide feedback and support.

Data representations for neural networks

The video covers the basics of data representations for neural networks, specifically tensors, including rank 0, 1, 2, and higher rank tensors. The speaker provides examples of how to represent these tensors using arrays in Python with the NumPy library and highlights the three key attributes that define a tensor, namely rank, shape, and data type. They also demonstrate the manipulation of tensors using slicing operations and negative indexing, as well as the use of matplotlib for displaying images. The video goes on to discuss various types of data representations in neural networks, including vector data, time series data, image data, and video data, with real-world examples such as text documents, stock prices, and trading data. The video ends with a brief summary of the different data tensors and announces that the next video will cover tensor operations.

My First Neural Network using Keras

In this video, the creator goes over the basics of building a simple neural network using Keras for image classification. They use the popular MNIST dataset to classify images of handwritten digits into 10 categories. They start by loading the dataset into Keras and build a sequential model with two dense layers and a 10-way softmax output layer. They then prepare the dataset for fitting by flattening and normalizing it. The creator also explains how to evaluate the model using the training and test loss, and they demonstrate how to make predictions on the test set. The final accuracy obtained is 98.24%, with the creator noting that upcoming videos will cover more technical terms related to deep learning.

Kaggle's 30 Days Of ML (Competition Part 6): Model Stacking

This section of Kaggle's 30 Days of ML competition explores the advanced technique of stacking, which involves creating multiple models and using their predictions as features for a metamodel to generate new predictions. The presenter demonstrates how to implement this process in a Jupyter notebook using five different models and optimizing hyperparameters. They then train a final linear regression model on top of the level one predictions to get the final set of predictions, which they submit to Kaggle. The presenter encourages viewers to experiment with different stacking techniques but cautions against overfitting.

Kaggle's 30 Days Of ML (Competition Part 1): Cross Validation & First Submission on Kaggle

The video covers the first part of Kaggle's 30 Days of ML competition, which involves a regression problem with RMSE as the evaluation metric. The presenter demonstrates how to use K-Fold Cross Validation to create train-validation splits and shows how to prepare and fit both random forest and XGBoost models. They discuss various issues that may arise during the modeling process and provide solutions, including removing unnecessary columns and copying data frames. Additionally, they explain the submission process on Kaggle and caution against submitting more than 10 times per day, while emphasizing the importance of writing code on one's own and trying it out in the competition.