Hello,
my super simple modification of the lesson 1 experiment.
Instead of only detecting birds it can detect if:
Is it a bird? Is it a plane? No, it’s Superman!
1 Like
Hello everyone,
I’ve just learned lesson 1 and 2. After seeing your wonderful projects above, I couldn’t wait to have a try and build one.
Now I just made some changes on the example code to tell apple from peach. (I think they look like sometimes.) which fruit it is: apple or peach?
Some notes:
At first I use “apple” and “peach” as the search keyword, the trained model mistook the “peaches on the tree photo” as “apple”, and the probability of apple is 0.52.
My Solution:
I figured maybe it had seen many photos of only one apple or peach, but not many photos of apples or peaches. So I changed the search keyword to be “apples” and “peaches”. I also increased the number of “apple/peach tree photos” from 30 to 50. This time it can tell it’s peach tree, and the probability of apple is 0.0076.
1 Like
Hello, iluu, good job! Your log of experiment is great. Did you take it using code or by hand? After looking at your note I also wondered why data augmentation didn’t improve the result.
I just trained a model to tell whether the fruit is apple or peach: which fruit it is: apple or peach. Next I’d like to put the project on huggingface like you.
1 Like
Hello, Yifan123, cute project, cucumber v.s zucchini classifier. Likewise, I think the fruit apple and peach are alike, so I trained a model to tell apples from peaches.
Hi! I am really excited to share my work after the completion of lesson 2. I created a galaxy classifier. Galaxies are classified into three general types based on their shape:
spiral, elliptical and irregular. The model is completely based on lesson 2 of this course. It is quite accurate at classifying the galaxies.
Link to the demo:
https://huggingface.co/spaces/fierce74/Galaxy_classifier
4 Likes
Hello,
Excited as everyone, after Lesson 1 I wanted to see if the model can recognize a hawk & a falcon from a group of hawks and falcons.
But to my surprise, it can recognize a falcon with 99% probability but a hawk with only 1% probability.
Here it is: hawk vs falcon
Am I doing something wrong?
Hey Leonard,
Your solution is correct, its just you are not using the result of the predict method correctly. I made exactly the same mistake 
The predict method returns a label, index and an array of probabilities (probs).
In order to read the correct probability of the predicted label you have to use the retuned index on the probs array.
Because falcon is the label associated with index 0, you get the correct result, but in order to get the correct hawk probability you have to update the method as follows:
is_hawk,idx,probs = learn.predict(PILImage.create('hawk.jpg'))
print(f"This is a: {is_hawk}.")
print(f"Probability it's a hawk: {probs[idx]:.4f}")
Hope this helps 
1 Like
This is my work for lesson 1. It runs on 7 different image categories including dolphins and whales. I thought maybe dolphins and whales would be too similar, but it worked well.
I used good programming practices to make the code less repetitive. And a check to not download images if they are already there, which makes it run much faster when rerunning it.I also used the current API for ddg (not the deprecated one in the lesson), so I had to look up how generators work.
I had some headaches with the learner, but got some good help in the forum! 
1 Like
Hi, I’m 13, I created a biome detector that can recognize 3 different biomes. Check it out on: Biome Detector
2 Likes
Decoding thoughts with fast.ai. Predicting if the user is thinking about a digit or not using EEG data recorded from a Muse headset.
1 Like
this is so cool
Had a lot of fun building a Fish or Face classifier yesterday evening. Thanks to the webcam input in Gradio, we posed and the model classfied. Sometimes, the model getting it wrong is fun too.
1 Like
Hello Everyone!
Just completed the first lecture today and it was awesome!
Here’s my first attempt at training NNs:
Helmet detector
https://www.kaggle.com/code/yashjindal/helmetdetector
Currently it doesn’t perform up to the mark, this could be due to the small dataset.
I’d really appreciate if you go through it and provide your valuable suggestions and feedback for improvement!
Hi, I just finished lesson 2, and have made a slightly inaccurate(but fun-to-play-with) paintings classifier.
And here’s a blog post where I discuss the reasons for it being a little inaccurate.
I didn’t properly utilize Quarto’s “freeze render” option so all of my code blocks are just screen shots. I’ll attempt to fix that in my next blog post.
Thank you all and thanks fast.ai for such an awesome course.
1 Like
Hello everyone, this is my first post.
I tinkered around with the first lesson and wanted to try my hand at a few things. First I just modified the code to discriminate between koalas and sheep because why not. After realizing I just did not load in enough photos I then moved on to trying other categories. I realized that using search is a messy process and my data was fairly dirty, but I was able to create a discriminator that can distinguish between which console a game comes from; limited to N64, PS1, and Sega Genesis. I hoped this would be a good test as they were relatively contemporary but used different core graphic architectures. This being the case there should be enough information in game screenshots to tell which console it come from.
After some tinkering this is my final result. I could test it more, but I think moving on to the next lesson will be more effective and building up from this:
This is the notebook, but it’s pretty rough, and you will see the bones of previous attempts.
2 Likes
Hey everyone,
Posted a blog just now that uses object detection and image classification to count cards in Blackjack live. Check it out if interested:
3 Likes
Created a Youtube thumbnail (video cover photo) click-thru rate predictor!
Try this model on 
Hugging Face Space
This is a Resnet model that was fine-tunned by a dataset from my wife’s Youtube channel
I have deployed the model to a custom website which is more accessible for my wife. I used GitHub page to host it and the hugging face embedding method to add the model because the API method mentioned in the course is not valid now.
I have added two same modules to the website so my wife can upload and compare two candidate cover photos for her next Youtube video. The model works well based on the data from her recent videos!
Thanks Jeremy for this wonderful course! 
3 Likes
Hi guys!
First post here. Just restarted my DL journey a few days ago. (I’ve tried several times in the past but would always hit a roadblock. Tried a few books but could never get past initial few chapters.) Liking fastai’s approach so far and I’m excited to hopefully the stay the course this time around.
I really love pizza and thought it’d be cool to build out a pizza classifier. I trained it on images of NY and Detroit style pizza. I used the resnet34 as my pretrained model. I noticed some interesting cases where I’d feed it a picture of a whole detroit style pizza, it would predict its a ny style pizza. The opposite did not occur at least from the sample images I tried. It would do better on cases where the image of a Detroit style piza was being lifted up and you could see the crust of the pizza. I downloaded the images via fastai download_images helper function.
Who else loves pizza?
1 Like
Hi, I want to make a face recognition system that can recognize faces of around 200 people with good speed and accuracy. When I googled around I found that ArcFace is a good option. But I don’t know how to use ArcFace and what are all the preprocessing steps involved in preparing the dataset. Also I saw that MTCNN is the in built face detection model in ArcFace. Should I try using another face detection model or sould I just use the default MTCNN. Kindly help me. I did google around but did not know what to do. If you could just give me an outline of ArcFace and how do I use it for real-time application it would be wonderful. Thank you in advance.