Merge branch 'main' of github.com:sajal2692/sajal2692.github.io

Author: sajal2692

astro.config.ts

@@ -5,6 +5,7 @@ import remarkToc from "remark-toc";
 import remarkCollapse from "remark-collapse";
 import sitemap from "@astrojs/sitemap";
 import { SITE } from "./src/config";
+import partytown from "@astrojs/partytown";
 
 // https://astro.build/config
 export default defineConfig({
@@ -15,6 +16,11 @@ export default defineConfig({
     }),
     react(),
     sitemap(),
+    partytown({
+      config: {
+        forward: ["dataLayer.push"],
+      },
+    }),
   ],
   markdown: {
     remarkPlugins: [

package-lock.json

@@ -16,6 +16,7 @@
   },
   "dependencies": {
     "@astrojs/check": "^0.4.1",
+    "@astrojs/partytown": "^2.0.4",
     "@astrojs/rss": "^4.0.1",
     "@resvg/resvg-js": "^2.6.0",
     "astro": "^4.1.1",

package.json

@@ -1,13 +1,13 @@
 import type { Site, SocialObjects } from "./types";
 
 export const SITE: Site = {
-  website: "https://sajal2692.github.io", // "https://sajalsharma.com/", // replace this with your deployed domain
+  website: "https://sajalsharma.com/", // replace this with your deployed domain
   author: "Sajal Sharma",
   desc: "Internet home of Sajal Sharma, a software engineer specializing in AI and ML.",
   title: "Sajal Sharma",
   ogImage: "astropaper-og.jpg",
   lightAndDarkMode: true,
-  postPerPage: 3,
+  postPerPage: 5,
   scheduledPostMargin: 15 * 60 * 1000, // 15 minutes
 };
 
@@ -47,7 +47,7 @@ export const SOCIALS: SocialObjects = [
     name: "Twitter",
     href: "https://twitter.com/sajal2692",
     linkTitle: `${SITE.title} on Twitter`,
-    active: true,
+    active: false,
   },
   {
     name: "Twitch",

src/assets/images/blog/corrective-rag-langgraph/basic-rag-flow.png

@@ -1,6 +1,6 @@
 ---
 title: "Building an Image Classifier Really Fast Using Fastai"
-author: "Your Name"  # Replace with the actual author's name
+author: "Sajal Sharma" # Replace with the actual author's name
 pubDatetime: 2022-10-28T00:00:00Z
 slug: building-image-classifier-fastai
 featured: false
@@ -10,7 +10,7 @@ tags:
   - Computer Vision
   - fastai
 description: "In this post, I demonstrate how to quickly build an image classifier using the fastai library, a powerful tool for practical deep learning. The project involves classifying images of fruit as either rotten or fresh."
-canonicalURL: ""  # Add if the article is published elsewhere
+canonicalURL: "" # Add if the article is published elsewhere
 ---
 
 ## Table of contents
@@ -19,15 +19,14 @@ canonicalURL: ""  # Add if the article is published elsewhere
 
 I recently started the [fast.ai](https://course.fast.ai/Lessons/lesson1.html) course to build up my practical deep learning skills. In order to better retain what I learn, I'm going to be writing a series of posts/notebooks, implementing my own models based on the course content. This notebook is written based on what I learned from the first week of the course.
 
-In this notebook we'll build an image classifier using the [fastai](https://docs.fast.ai), a deep learning library built on top of Pytorch that provides both high-level and low-level components to quickly build state-of-the-art models for common deep learning domains. 
+In this notebook we'll build an image classifier using the [fastai](https://docs.fast.ai), a deep learning library built on top of Pytorch that provides both high-level and low-level components to quickly build state-of-the-art models for common deep learning domains.
 
 We'll build a model that can classify images of fruit into a binary category: rotten or not. You can imagine such a model being used inside refrigerators to detect if produce kept inside it has gone bad.
 
 When I started learning ML in 2016, building such models was a non-trivial task. Libraries to build deep neural networks were still in their infancy (Pytorch was introduced in late 2016), and building accurate image classification models required a certain degree of specialized knowledge. All that has changed and, as you'll notice in the notebook, we can build an image classifier using just a few lines of code.
 
 Let's get started!
 
-
 ```python
 import os
 # !pip install -Uqq fastai duckduckgo_search
@@ -39,7 +38,6 @@ We'll be needing the `duckduckgo_search` package to quickly search for, and down
 
 <br />
 
-
 ```python
 from duckduckgo_search import ddg_images
 from fastcore.all import *
@@ -52,16 +50,14 @@ def search_images(term, max_images=40):
 urls = search_images('rotten fruit', max_images=1)
 urls[0]
 ```
+
 ```output
 Searching for 'rotten fruit'
 'https://i.pinimg.com/originals/13/2e/48/132e481c0ef6f1516de2b5b80a553b6a.jpg'
 ```
 
-
-
 Let's download this image and open it.
 
-
 ```python
 from fastdownload import download_url
 dest = 'rotten.jpg'
@@ -72,33 +68,23 @@ im = Image.open(dest)
 im.to_thumb(256,256)
 ```
 
-
-
-
 ![rotten_or_not](@assets/images/blog/building-image-classifier-fastai/is-the-fruit-rotten-or-not_8_0.png)
-    
-
-
 
 Doing something similar for fresh fruit.
 
-
 ```python
 download_url(search_images('fresh fruit', max_images=1)[0], 'fresh.jpg', show_progress=False)
 Image.open('fresh.jpg').to_thumb(256,256)
 ```
+
 ```output
 Searching for 'fresh fruit'
 ```
-    
-![rotten_or_not](@assets/images/blog/building-image-classifier-fastai/is-the-fruit-rotten-or-not_10_1.png)
-    
-
 
+![rotten_or_not](@assets/images/blog/building-image-classifier-fastai/is-the-fruit-rotten-or-not_10_1.png)
 
 Now that we know what duckduckgo image search is working fine, we can download images for both rotten and fresh fruit and store them in their respective directories. We use time.sleep to avoid spamming the search API.
 
-
 ```python
 searches = 'rotten', 'fresh'
 path=Path('rotten_or_fresh')
@@ -116,6 +102,7 @@ for o in searches:
     download_images(dest, urls=search_images(f'{o} vegetables'))
     resize_images(path/o, max_size=400, dest=path/o)
 ```
+
 ```output
 Searching for 'rotten fruit'
 Searching for 'rotten apple'
@@ -127,14 +114,12 @@ Searching for 'fresh banana'
 Searching for 'fresh vegetables'
 ```
 
-
 ## Training our model
 
 We have our images and the next step is to train a model. Again, it blows my mind how simple this is using fastai. I'll briefly explain what the below blocks of code are doing.
 
 First, we check if all image files can be opened correctly using a fastai vision library utility verify_images. If it can't be opened, then we unlink it from our path so that is is not used in model training.
 
-
 ```python
 # validate images
 failed=verify_images(get_image_files(path))
@@ -146,21 +131,18 @@ len(failed)
 0
 ```
 
-
-
 Next, we'll use another building block from the fastai library, the `DataBlock` class, which we can use to represent our training data, the labels, data splitting criteria, and any data transformations.
 
 `blocks=(ImageBlock, CategoryBlock)` is used to specify what kind of data is in the DataBlock. We have images, and categories - hence a tuple of ImageBlock and CategoryBlock classes.
 
-`get_items` takes the function `get_image_files` as its parameter. `get_image_files` is used to find the paths of our input images. 
+`get_items` takes the function `get_image_files` as its parameter. `get_image_files` is used to find the paths of our input images.
 
-`splitter=RandomSplitter(valid_pct=0.2, seed=42)` specifies that we want to randomly split our input data into training and validation sets, using 20% data for validation. 
+`splitter=RandomSplitter(valid_pct=0.2, seed=42)` specifies that we want to randomly split our input data into training and validation sets, using 20% data for validation.
 
 `get_y=parent_label` specifies that the labels for an image file is its parent (the directory that the file belongs to).
 
 `item_tfms=[Resize(192, method='squish')]` specifies the transformation performed on each file. Here we are resizing each image to 192x192 pixels by squishing it. Another option could be to `crop` the image.
 
-
 ```python
 dls = DataBlock(
     blocks=(ImageBlock, CategoryBlock),
@@ -173,18 +155,12 @@ dls = DataBlock(
 dls.show_batch(max_n=6)
 ```
 
-
-    
 ![rotten_or_not](@assets/images/blog/building-image-classifier-fastai/is-the-fruit-rotten-or-not_16_0.png)
 
-    
-
-
 Above you can see a batch of images from our DataBlock, along with their labels. This is a nice way of quickly knowing if a sample from our data is correct (images/labels).
 
 To train our model we will fine-tune the resnet18, which is one of the most widely used computer vision models, on our dataset.
 
-
 ```python
 clf = vision_learner(dls, resnet18, metrics=error_rate)
 clf.fine_tune(5)
@@ -211,7 +187,6 @@ clf.fine_tune(5)
   </tbody>
 </table>
 
-
 <table border="1" class="dataframe">
   <thead>
     <tr style="text-align: left;">
@@ -261,17 +236,15 @@ clf.fine_tune(5)
   </tbody>
 </table>
 
-
 ## Using the model
 
 It's finally time to use our model and see how it does predicting if a fruit is rotten or not.
 
-
 ```python
 Image.open('rotten.jpg').to_thumb(256,256)
 ```
-![rotten_or_not](@assets/images/blog/building-image-classifier-fastai/is-the-fruit-rotten-or-not_20_0.png)
 
+![rotten_or_not](@assets/images/blog/building-image-classifier-fastai/is-the-fruit-rotten-or-not_20_0.png)
 
 <br />
 
@@ -292,21 +265,14 @@ Probability it's rotten: 1.0000
 Image.open('fresh.jpg').to_thumb(256,256)
 ```
 
-
-
-
-    
 ![rotten_or_not](@assets/images/blog/building-image-classifier-fastai/is-the-fruit-rotten-or-not_22_0.png)
-    
-
-
-
 
 ```python
 is_rotten,_,probs = clf.predict(PILImage.create('fresh.jpg'))
 print(f"This fruit/vegetable is: {is_rotten}.")
 print(f"Probability it's fresh: {probs[0]:.4f}")
 ```
+
 ```output
 This fruit/vegetable is: fresh.
 Probability it's fresh: 1.0000
@@ -316,23 +282,19 @@ Probability it's fresh: 1.0000
 
 Let's see if our model can predict if a given image is of a rotten orange or a fresh orange. We haven't explicitly downloaded images of fresh/rotten oranges for our training set, so it would be a good generalization on "unseen data".
 
-
 ```python
 download_url(search_images('fresh orange', max_images=1)[0], 'fresh orange.jpg', show_progress=False)
 Image.open('fresh orange.jpg').to_thumb(256,256)
 ```
-    
-![rotten_or_not](@assets/images/blog/building-image-classifier-fastai/is-the-fruit-rotten-or-not_25_1.png)
-    
-
-
 
+![rotten_or_not](@assets/images/blog/building-image-classifier-fastai/is-the-fruit-rotten-or-not_25_1.png)
 
 ```python
 is_rotten,_,probs = clf.predict(PILImage.create('fresh orange.jpg'))
 print(f"This fruit/vegetable is: {is_rotten}.")
 print(f"Probability it's fresh: {probs[0]:.4f}")
 ```
+
 ```output
 This fruit/vegetable is: fresh.
 Probability it's fresh: 0.9748
@@ -345,31 +307,31 @@ download_url(search_images('rotten orange', max_images=1)[0], 'rotten orange.jpg
 Image.open('rotten orange.jpg').to_thumb(256,256)
 ```
 
-    
 ![rotten_or_not](@assets/images/blog/building-image-classifier-fastai/is-the-fruit-rotten-or-not_27_1.png)
 
-
 ```python
 is_rotten,_,probs = clf.predict(PILImage.create('rotten orange.jpg'))
 print(f"This fruit/vegetable is: {is_rotten}.")
 print(f"Probability it's rotten: {probs[1]:.4f}")
 ```
+
 ```output
     This fruit/vegetable is: rotten.
 Probability it's rotten: 0.9899
 ```
 
-Not bad at all. The model seems to generalize fine. Though, a more accurate measure of generalizability would involve creating a separate test set and calculating performance metrics. 
+Not bad at all. The model seems to generalize fine. Though, a more accurate measure of generalizability would involve creating a separate test set and calculating performance metrics.
 
 ## Summary
 
 There you have it! With a few lines of code we have created our own image classification model by fine-tuning off the shelf models with fastai. The high level apis that the library provides makes the process of building an initial model a breeze.
-If you want to run the notebook for yourself, you can check it out on Kaggle [here](https://www.kaggle.com/code/sajalsharma26/is-the-fruit-rotten-or-not). I urge you to try building your own 
+If you want to run the notebook for yourself, you can check it out on Kaggle [here](https://www.kaggle.com/code/sajalsharma26/is-the-fruit-rotten-or-not). I urge you to try building your own
 classification model on images from duckduckgo search.
 
 I'll be going over the rest of the fastai course in the coming weeks. Even though I have only done the first two weeks till now, I highly recommend it for anyone
 interested in Machine Learning, more so for people with a coding background.
 
 ## Resources
+
 - fastai Course: https://course.fast.ai/
-- Notebook on Kaggle: https://www.kaggle.com/code/sajalsharma26/is-the-fruit-rotten-or-not
+- Notebook on Kaggle: https://www.kaggle.com/code/sajalsharma26/is-the-fruit-rotten-or-not