I watched the Part 1 video last year, with fastai v0.7, and just i was amazed to see how much better fastai performs in comparison to other deep learning libraries. I then wondered how would it perform against itself, and needless to say, the library did not let me down. I found a paper written by one of my college senior in early 2019, using a thermal image dataset. At the time, they got a best case accuracy of 97.08% and a validation loss of 11% using resnet101 and fastai v0.7.1, which was achieved after multiple parameter modifications and model tuning.
In July, 2019, I present, fastai v1.0, resnet50 and 10 minutes of coding:
Model Accuracy: 99.38%
Training Loss: 1.4%
Validation Loss: 1.7%
After taking the first two class of part 2 of 2019, I was able to get how autograd works and how it was used in pytorch. And to be sure I grasp the concept, I created a simple autograd in javascript and also create a pytorch like implementation in javascript.
It was implemented in obsevablehq js interactive notebookhere
and also you can help me check this medium post draft explaining the basic concept here
Spending some time applying the approach really helped me understand the concepts more deeply, although I’m still not sure I fully understand how the time series aspect works with regards to trend. I noticed that in the Rossmann notebook there are no ‘recent trend’ type variables i.e. something that describes the days and weeks leading up to each observation. Of course, this information is still there in the other observations, so I’m guessing that the ‘date’ (day, month, year) embeddings encompass this in the network e.g. October’s embedding encodes in some way that September is ‘nearby’, meaning that the model can account for recent trend (assuming it is predictive).
Loved the first part of the course so going onto part 2 now!
Created a web app and deployed to Heroku. Despite an okay confusion matrix (see the notebook), the model often gets confused (try images from the homepage, which are in the training set, or Google). Need a bigger dataset?
Hi everyone! I just started ths course and I’m super pumped!
Following the suggested practice for lesson 1, I built a classifier of paintings into their art period. I used 300 google image search results for each of the 18 classes. My model has ~50% accuracy. I am not entirely sure how good this is so I would appreciate any hints or insights into the work.
Hi @champs.jaideep,
I took a look at your kernel and studied the paper. From what I understand, Arcface loss might be applicable to problems other than computer vision.
Does replacing Linear layers from other models (a text classifier for exmaple) with ArcMarginProduct make sense?
What do you think?
well i havent tried on the models other than the classification model. It produced fantastic results for kaggle whale identification form hump back which had 5000 classes plus. You are welcome to try it out for other models.Arc face is also useful for face detection by learning deep features.
As a homework for lesson 1, I have created a dataset of images of text of four languages: English, Punjabi, Hindi and Urdu.
The images contain text of either of the four languages (some of the images have text of two languages). The effort is to detect the langauge of the text in the images.
With whatever I learned from lesson 1, I was able to achieve an error rate of 0.303125 using resnet34.
Believe me, data is very noisy. I will try to improve it with whatever I will learn from next lessons.
Update: 12/08/19
I applied data cleaning on the data as taught by Jeremy in lesson 2.
With it, I was able to reduce the error rate to 0.240385.
I have experimented tonight with changing the batchsize and nothing else on my a network trained on anime character faces and the results are not what I expected. I have placed a notebook at https://github.com/Dakini/BatchSize that show my results.
Training for 5 epochs with fit_one_cycle with a batchsize of 728,got an accuracy of ~58%. While a batchsize of 32 got 79%. There seemed to be no difference between the training times for them either.
Trained a CNN with the amazing techniques mentioned throughout the course and deployed it as a web app. Currently at ~95% accuracy but with room for improvement.
The stats:
Training data: 26 classes, approx. 400 samples each, batch size of 16 (see current list of classes)
CNN learner with resnet50: cnn_learner(data, arch, metrics=accuracy, wd=0.1, ps=0.01)