Updated a v1 project to v2. Model trained beautifully but having trouble with the new headers (?) that fastai v2 implements. That’s blocking me from converting the model to Apple’s Core ML format
Notebook (WIP):
https://www.kaggle.com/raymonddashwu/fastai-v2-sugar-vs-tahoe?scriptVersionId=31149139
Repo:
This isn’t a fancy project but a simple blog post to help me start digging into the library and setting up a computer vision project. Hopefully I’ll get started on my larger project towards the middle of the course… it feels a bit to daunting at the moment/not sure where to start but here’s something to help keep me going 
Warning, not for the faint of heart
I hate bug bites, am paranoid about bed bugs, and occasionally get flea bites from my dog (at least that is my assumption) When I do get bit, I sometimes find myself trying to figure out what kind of bite it is. So to save my future self some time, I did some queries on bing image search and started classifying some pretty gross images.
https://colab.research.google.com/drive/15Pm1PcFmVyO04b3mjw0vU9YAMRfvtzOn (WIP)
I’m currently around 58% accuracy and hope to improve that with other data points (location of body and part of the world) and better loss function.
Hmmm probably not going to open that link 
But one thing to consider if you haven’t already is how different people react to the same type of bite, some might be more allergic than others. Also the different reactions for different skin colors will appear differently I guess…super interested to hear how you get on though!
Hi,
I have created a near perfect nuclei segmentation model for microscopic slide images, using just 25 images (and masks), from the dataset and code.
Sample input:
The model was able to predict all the nuclei on an unseen image, close to perfection.
Input:
Prediction:
thanks, @morgan. how allergic one is tough and don’t think I have the data to solve 
the skin color thing is super important and (i think) solvable. I’m glad you mentioned. if I can stomach it, I’ll run the images through OpenCV contours or convert to greyscale and then try retraining.
Does anyone have experience scraping google maps/google street view images and then doing deep learning to identify street signs? Any advice on where to start with this (either the scraping side or the deep learning model selection side) would be welcome.
i’ve not done it and only can offer general thoughts - i googled a bit and found this https://rrwen.github.io/google_streetview/ . there might be better with some more exploration
I have been playing about with training a Danbooru Multi Classifier using V2. So I can use it in colorising and semi super vised projects.
| epoch | train_loss | valid_loss | accuracy_multi | accuracy_multi | time |
|---|---|---|---|---|---|
| 0 | 0.016210 | 0.014924 | 0.995065 | 0.996054 | 6:10:34 |
| 1 | 0.015684 | 0.015714 | 0.995111 | 0.995989 | 6:05:17 |
accuracy mult threshold of 0.2 and 0.4, so I am gonna say my training is going well.
Hi,
I have created a superresolution mobile app based on fastai’s superres notebook.
I didn’t do any interesting deep learning work here (although I’d love to implement this zero-shot learning method in the future!), my main work was on the deployment and demo side: I created both a web-service and a mobile app that enables to take/upload a picture, enhance it, and compare the new vs old. Hopefully, this code can also be reused to demo many of the other cool projects that are shown here (style transfer, semantic segmentation, etc…)!
Unfortunately, the server seems to crash with heavy pictures, so I restricted it to pictures <2MB. If you have any idea how I could solve this issue, please let me know !
Nice results, maybe you could divide the image into 4 parts and process them individually ? combine them up after that
Sounds like a great idea, thanks!
nice Molly! this was somewhere on my todo list.
I never quite got a non-fastai simpson version of ugatit working - using other people’s repo, my training.
I have written a 2 part blog series, part-1 and part-2 about how to build an end to end pipeline for an image segmentation problem.
It builds a solution for TGS Salt Identification Challenge competition in Kaggle.
By following the blog, you would be able to rank in the top 4% of the Kaggle LB. The blogs have step by step explanation of the key concepts like
DataBlocks,
Transformations,
Using unet_learner
Building custom unet architecture based on se_resnext50
Adding hypercolumns.
Adding custom TTA.
Doing K-Fold predictions.
Creating a Kaggle submission file.
At the end of the blog, I have also added a few tricks which you can try to improve the solution further.
One small suggestion, instead of converting the mask values from 255 to 1 in change_image, you can try using IntToFloatTensor(div_mask=255) in batch_tfms.
What will that do exactly? This is the first time I’ve heard of that IE would this help in a case of say 0,1,255?
By default the mask tensor gets divided by 1, but for binary segmentation we may want it to be divided by 255 to make the values in the tensor from 0 to 1.
