🔢┊digit-recognizer

[steel barn]

I remember the first time I came across this dataset. It was my first taste into computer vision which just blew my mind. Curious as to others' experiences!

[steel barn]

Thanks for sharing! You're also welcome to post in https://discord.com/channels/1101210829807956100/1130784683907612764 for additional visibility.

[tropic magnet]

But it is ok to post here right?
Just to make sure I follow the description of the channel.

[errant ruin]

@tropic magnet to be clear, we'd like to keep all general posts about notebooks/datasets in #🔗┊sharing-projects . You should only share a link to your specific notebook when it's relevant to the discussion.

[tropic magnet]

Ok then thank you for the clarification

[open sequoia]

Hi :> I try to train this model, and I notice that some of the numbers are kind of difficult for the model to learn for practical reasons, for example:

I mean, isn't it a bit tough to teach the model these numbers, without harming it? The number on the left looks more like a 1, so it could be kind of bad for the model's performance to be "correctly" recognised as a 9

[open sequoia]

I just realized that people who scored 100% there most likely just trained the model with the test data as well 😅 I was really at a loss on how it was possible to score 100% considering the concerns I just mentioned at my earlier reply

[ancient plover]

It's crazy what people would do to just achieve the "full-mark"

[ancient plover]

I think they should be considered as outliers. so, by dropping them would make the model generalize even better

[lethal pond]

Guys can you suggest any good video for EDA?

[karmic hedge]

Has anyone seen any Vanilla NN break ~98.5% with roughly the same number of parameters as a CNN?

[fervent echo]

Hello Kaggle community !!! I have released a notebook that using Naive Bayes to join this competition.
But unlike some people using sklearn, I try to do it with only use numpy and pandas.

https://www.kaggle.com/code/jackksoncsie/navie-bayes-from-scratch

[wicked cradle]

My 98% Solution with @magic phoenix and @craggy rampart !
https://www.kaggle.com/code/touhidurrr/recognize-digits-with-cnn

Also, many thanks to @magic phoenix for explaining how CNN's work!

Really helpfull.

[craggy rampart]

This is my notebook about this competition, It achieved pretty high accuracy, I hope you will like it https://www.kaggle.com/code/hoanglongroai/99-accuracy-in-digit-recognization

[sturdy sand]

oh hey i didnt realize this was a channel but i made data augmentation notebook . https://www.kaggle.com/code/elanderos/mnist-data-augmentation-and-cnn-tuning

[fading quartz]

https://www.kaggle.com/code/neupane9sujal/mlp-mnist-from-scratch

Would appreciate feedbacks..

[sacred crest]

trying to finish this one up today/tomorrow. so close to 99%

[sacred crest]

Not live yet but just got this video up Solving Kaggle's MNIST Digit Recognizer with PyTorch: 99% Accuracy
https://youtu.be/2w0pRriQG3A

[thorn plaza]

https://www.kaggle.com/code/pony1013/implementation-of-mlp-no-tf-keras-pytorch
I implemented MLP without using TF/keras/pytorch, would appreciate feedbacks

[inner mauve]

Built a toy NN library in JavaScript (https://npmjs.com/@hycord/neural) and got a 95% on this
Fairly proud of myself for being the first real model I’ve made.

(https://api.masen.dev/neural/digits/predict send a post request with an array of 784 numbers as the body and it will respond with the prediction values)

[green narwhal]

New notebook exploring Digit Recognizer ! Check it out and upvote if you find it helpful ⬆️. Thanks in advance!
https://www.kaggle.com/code/seifwael123/digit-recognizer-top-8

[coral zodiac]

Hello, hope my notebook will be helpful for someone! https://www.kaggle.com/code/kapturovalexander/kapturov-s-solution-of-digit-recognizer

[blazing crescent]

Can anyone guide me with the Deep Learning stuff for this competition, I'm really interested in this image analysis but couldn't find any methods to study it. I'm really confused when to use what model and which activation function. It would be really great!

Thank you

[green narwhal]

https://www.kaggle.com/code/seifwael123/digit-recognizer-99

[true marsh]

anyone interested in doing this project together

[fringe vine]

Hi, I noticed that all labels in sample_submission are equal to 0, is this some kind of error? Or is it just me?

[fallen imp]

for submissions, it tends to just show you the format of how things should be done. the idea is that your submission should be two columns, one being the id, and the second being the digit it is guessing

[mild tusk]

https://youtu.be/tAEVgAa4CZw?si=_1ThlIeIQyJAGpze

[kind crane]

Hi guys, just got a 99%. Did anyone got a 100%? Please LMK.

[kind crane]

Have any of you used data augmentation and got better results? When I used it I just get worst results. Any clue why?

[carmine sand]

https://www.kaggle.com/code/ammanrizwan/digit-recognizer-notebook

96.66 accuracy

[karmic folio]

Guys I have trained a CNN model for digit recognition and got a accuracy of 98.9
But when I submit my prediction in Kaggle I have only gotten a score of 0.52467
this is the model I have used:
Any suggestions??

[rose drum]

Here are some quick suggestions that you can try:
Add more Conv Layer and see how it works.
Use the early stopping callback function. Change the patience parameter to find the best-case scenario.
Try altering filters across the networks.
Apply padding to minimize inconsistent overlapping during the convolve operations.
Try Global Average Pooling or Strided Convolutions since applying Max Pooling on smaller image pixels can result in missing the most prominent feature from each region, causing other important features to be lost and causing information loss.
Use different optimizers such as Adam, RMSprop, Adagrad, Nadam, Adamax
Try lowering batch size (will increase computational processing). In theory, bigger batch size leads to lower variance, therefore, decreasing the learning rate in the process.

[grand chasm]

Hi I’m trying to learn how to choose Neural Network layers. What would the best approach be?

[unreal radish]

hi. I am new to kaggle and I have just completed this challenge. My approach is a 4-layer resnet, along with hyperparameter tuning. Please upvote me if you find my notes useful. https://www.kaggle.com/code/laodriverayu/guide-pytorch-4-layer-resnet-optuna

[upbeat widget]

this community is too quite

[upbeat widget]

if you have had found the way please tell me hhh, i need it too!!😼

[grim cypress]

Well guys, @upbeat widget & @grand chasm I was also looking for a solution to the same problem

But it turns out, layers are also essentially hyperparameters. Therefore, you would probably perform exhaustive search just like you would tune parameters e.g., n_estimators in classification/regression tasks.

However, I did input your query on ChatGPT with Web Search enabled, and here's the response from ChatGPT:

https://chatgpt.com/share/677a261d-6258-8005-916c-eb7a34c8b083

[upbeat widget]

thx so much!

ure so kind

[viscid moth]

i have some tips u might find it helpful:

• double the neurons every next layer (e.g first layer has 8 then 2nd has 16, 3rd has 32,...)
• activation for all layers are relu except the last one is softmax

[grave smelt]

Does it make sense to tune the thresholds as hyperparameters? I would use the negative f1 score as a cost function and try to tune the thresholds for all 10 digits.

The f1 score makes sense, if we want both the precision and the recall to become high.

[grave smelt]

Well, I did it and I did not regret it.

[tiny bluff]

I was working on MNIST dataset and wanted to increase acc using sgd, I got 0.86 acc on kaggle without any transformations on dataset with sgd, any suggestions and also if anyone could explain about model a little bit as I know really less about this model

also if any other metrics I should use?

[frigid haven]

Hey friends!
I just published a Kaggle notebook where I built a Digit Classifier from Scratch with 91.41% accuracy using NumPy and Deep Learning techniques

If you're into ML or starting out with Neural Networks, I’d really appreciate it if you could take a look and leave an upvote if you find it useful 🙏

🔗 https://www.kaggle.com/code/mrmelvin/digit-classifier-from-scratch-with-96-91-accuracy

Thanks so much for your support! 💙

[frigid haven]

Hi everyone! 👋
I’ve created a Digit Classifier from Scratch notebook using only NumPy (no frameworks) and achieved 96.91% accuracy.
It covers full training, forward/backward propagation, and manual optimization — perfect if you're learning how neural networks work under the hood.

📘 Notebook: Digit Classifier from Scratch with 96.91% Accuracy

If you find it useful or learned something new, please consider giving it an upvote ❤️ — it really helps with visibility!

Happy learning! 🙌

If u find something useful,give an upvote

[lament pawn]

https://www.kaggle.com/code/alexandroskanakis/digit-recognizer

[split coral]

Hii

[chrome ingot]

Hi

[limpid flume]

Hiii

[zenith tundra]

https://media.discordapp.net/attachments/1436719817624256534/1436719913518633010/1.JPG?ex=6910a130&is=690f4fb0&hm=6a48397700e40b701b7defba0bc73ccc590e83e58af09eb7035cae318e9fb319&=&format=webp&width=515&height=687
https://media.discordapp.net/attachments/1436719817624256534/1436719914034659408/2.jpg?ex=6910a130&is=690f4fb0&hm=5d3c01e3db0b2fe7135969c69c22cbf49db07bae5ed8cb9a98ac3e18d3c73ce5&=&format=webp&width=515&height=687
https://media.discordapp.net/attachments/1436719817624256534/1436719914512547951/3.jpg?ex=6910a130&is=690f4fb0&hm=59a326eaa4d74733a406431b5c2eb8ee07f6b78d95094102deb1153d2e261407&=&format=webp&width=515&height=687

[sterile ledge]

hiiii

[west solar]

I just did this challenge and it was hard but I'm proud of myself for figuring it out

[shy karma]

Anyone tried Liquid AI stuff on this?

[tidal zinc]

yes i am on the way of it

[alpine sage]

can anyone explain how my model gets 95.22% accuracy so quickly?

epoch 01/20 | loss 0.8047 | test acc 95.22% | 20.5s
epoch 02/20 | loss 0.1108 | test acc 97.16% | 20.1s
epoch 03/20 | loss 0.0664 | test acc 98.08% | 20.0s
epoch 04/20 | loss 0.0496 | test acc 98.56% | 20.1s

It makes me think i am doing something wrong

[alpine sage]

or just lucky, idk