HaiyangLiu1997 / Kaggle Kannada Mnist 3rd Solution
Licence: mit
Stars: ✭ 52
Labels
Projects that are alternatives of or similar to Kaggle Kannada Mnist 3rd Solution
Pix2code Template
Build a neural network to code a basic a HTML and CSS website based on a picture of a design mockup.
Stars: ✭ 51 (-1.92%)
Mutual labels: jupyter-notebook
Nn from scratch
Multilayer Neural Network using numpy.
Stars: ✭ 51 (-1.92%)
Mutual labels: jupyter-notebook
Blackbox Attack
Blackbox attacks for deep neural network models
Stars: ✭ 51 (-1.92%)
Mutual labels: jupyter-notebook
Ner blstm Crf
LSTM-CRF for NER with ConLL-2002 dataset
Stars: ✭ 51 (-1.92%)
Mutual labels: jupyter-notebook
Ml securityinformatics
Short Course - Applied Machine Learning for Security Informatics
Stars: ✭ 51 (-1.92%)
Mutual labels: jupyter-notebook
Feature Selection For Machine Learning
Code Repository for the online course Feature Selection for Machine Learning
Stars: ✭ 52 (+0%)
Mutual labels: jupyter-notebook
Reinforcement Learning Introduction
Code from my blog post & online course
Stars: ✭ 51 (-1.92%)
Mutual labels: jupyter-notebook
Coronamaskon
Mask On-Off control with computer vision
Stars: ✭ 52 (+0%)
Mutual labels: jupyter-notebook
Pytorch musicnet
PyTorch DataSet and Jupyter demos for MusicNet
Stars: ✭ 51 (-1.92%)
Mutual labels: jupyter-notebook
Spark Mllib Scala Play
Twitter sentiment analysis based on Apache Spark, MLlib, Scala and Akka.
Stars: ✭ 51 (-1.92%)
Mutual labels: jupyter-notebook
Continuousparetomtl
[ICML 2020] PyTorch Code for "Efficient Continuous Pareto Exploration in Multi-Task Learning"
Stars: ✭ 52 (+0%)
Mutual labels: jupyter-notebook
Tsr Py Faster Rcnn
This repo contains code related to german traffic sign detection and classification using Faster-RCNN
Stars: ✭ 51 (-1.92%)
Mutual labels: jupyter-notebook
Onnx tflite yolov3
A Conversion tool to convert YOLO v3 Darknet weights to TF Lite model (YOLO v3 PyTorch > ONNX > TensorFlow > TF Lite), and to TensorRT (YOLO v3 Pytorch > ONNX > TensorRT).
Stars: ✭ 52 (+0%)
Mutual labels: jupyter-notebook
Ppd599
USC urban data science course series with Python and Jupyter
Stars: ✭ 1,062 (+1942.31%)
Mutual labels: jupyter-notebook
The 3rd solution code for kaggle kannada MNIST playground challenge.
I use it to familiarize myself with the competitions in kaggle.
Only the most basic model and tricks has been used.
Final version settings
- use the 8conv+2linear baseline model in keras.
- use hyper-parameters of optimizer in keras version code.
- TTA not be used in final
- pesudo labels are used in final
- average voting model embedding which contains 5 models are used
- label smoothing, focus loss, etc 10+ tricks are not used (not finish code or useless)
- using 5fold CV to choose model, using all data to train
Why there has pytorch version and keras version
Time line
- I'm more familiar with pytorch, I wrote a framework for this competitions in the beginning.
- I try to reproduce some keras sample baseline's accuracy in pytorch, but failed, still had 0.2%~0.3% gap in final, which is a large difference in this competition.
- The time is limitd. So I try to use keras directly, write a sample version keras code base on some public kernels, the keras version can reproduce the accuracy, so I choose keras version code to continue.
My consideration
- it doesn't mean pytorch can't realize same accuarcy comparing with keras. I already found the their difference in some default settings, but still exist a little gap. I think possible reasons maybe: a. data augmention implementation difference. b. random seed difference.
- keras is not so convenience like pytorch, a. the random seed is hard to fix in keras. b. the keras lib is too high level to rewrite some functions easily.
keras version code 99.420% in private leadboard
- single model, acc around 98.960%(use)
- 5-embedding model, acc around 99.060%(use)
- pesudo label, acc around 99.120%(use)
- 5 * TTA, acc around 99.100%(not use)
- label smoothing, acc around 98.960%(not use)
- seveal tests to choose best augmention parameters(use)
pytorch version code 99.420% in private leadboard
- single model, acc around 98.800%(not use)
- multi-lr, acc decrease(not use)
- choose no weight decay in final, so no bias decay not use
- other tricks in the code, not use because useless in this competitions
Other Notes
- When test, don't train your model again. Save your weight and just read them during test.
- Trust your local cv, and trust youself. I fixed the random seed in the beginning and never change it. And even thouth we fix the random seed, we still need try to identify some method really work or not. Ex. after changing the momentium of batchnorm layer from 0.01 to 0.1, the result change to 98.340 from 98.800. when we swith to another model, the result change to 99.720 from 99.700. So, my conclusion is momentium of batchnorm don't have big influence.
- try better TTA may improve the acc
- using better baseline model may boom up the acc, for this competition, I just want to implement and test tricks in competitions. I have try the MobileNet V3, selfDensenet but results are not so good(I think becasue this task is too sample), I think using NAS to find best model and add tricks is best choice.
Note that the project description data, including the texts, logos, images, and/or trademarks,
for each open source project belongs to its rightful owner.
If you wish to add or remove any projects, please contact us at [email protected].