Fantastic! Looking forward to it.
Oh those all look quite yummy! Which one is best?
2 Likes
Thanks. I will share too. 
I love Jollof Rice because it is spicy, though I think you will prefer Fried Rice
1 Like
Great work!! What do you mean by “using a better type of spectrogram” ? What were you using before and after?
Finally have my deployment up and running…
You can now classify your guitar model online. Well, if it’s a Gibson or Fender, others are coming up…
It’s based on ResNet-50 and achieves an accuracy of ~ 97.1% for the 11 guitar classes I provided.
I’m currently writing up a blog post and will add the link here.
Tools:
Server: Hetzner VPC instance (1 vCPU, 2GB Ram, 20GB SSD)
Deployment: Dokku
App: Flask, Plot.ly, Bulma.io
The site lives at: http://guitars.cwerner.ai
The GitHub repo for it is:
NOTE: If you downloaded the code before 2018-11-08: there was a nasty bug in my predict function that is now fixed.
19 Likes
That’s a really cool idea! 
1 Like
Is this being ran as a flask web app?
Created the image classification to classify stop, yield and school crossing; Here is the snapshot of my notebook. Need to clean up the date to get more accuracy - currently getting 14% error.
My Input data from google image
Train model/rate
Inference - pretty cool 
1 Like
No this is not a flask app. I am running this project using tool opencv. I found that if we use opencv along with fastai library then we get segmentation fault error. I am not sure why this error occurred. If anyone has encountered this issue and has some resolution, then it will really help me to further built a web app for this current application.
@jeremy @swagman tried the same thing and got similar accuracy ~95%. I think I just lost patience and a few more epochs would take me to ~96-97%.
What model did you use?
(Still working on improving accuracy and sharing this later on)
Made a mushroom classifier (including +/- 100 type) so you can pick some (eatable) mushrooms in the forest
I created the web app in Flask and deployed on Heroku - https://dry-bayou-64303.herokuapp.com/
10 Likes
After a couple data sets that worked well, I tried a really fine grain data set of various similar stringed instruments. The tenor, concert, soprano are all ukuleles for example:
([‘banjo’,
‘banjolele’,
‘bariguitar’,
‘baritone’,
‘bass’,
‘concert’,
‘doublebass’,
‘guitalele’,
‘guitaracoustic’,
‘guitarelectric’,
‘mandolin’,
‘soliduke’,
‘soprano’,
‘tenor’,
‘ubass’,
‘violin’],
16,
1189,
325)
Total time: 01:49
epoch train_loss valid_loss error_rate
1 1.253016 1.463084 0.446154 (00:26)
2 1.170810 1.462633 0.486154 (00:28)
3 1.026609 1.247020 0.415385 (00:26)
4 0.859524 1.187210 0.393846 (00:27)
40% looks bad, but with this data set it did pretty good. the larger numbers in the confusion maxtrix are very similar instruments. also, most of the top losses are effectively miss classified.
have a couple disney princess mugs around the house that have the same colors and design motifs of that princess. repunzel:
i trained a model with images of 5 disney princesses themselves, and then tried to feed in the mug to see if the color or design elements were enough to bucket it as the correct princess. The model to pick between the 5 princesses in the data set was fairly good, about 5 %. it failed however to identify the mugs. The mug above was classified as Tiana, and the Snow White mug was flagged as Repunzel.
i think i did too many epochs here. it looks like it overtrains a little and finds its way again
Total time: 01:37
epoch train_loss valid_loss error_rate
1 0.151588 0.277224 0.106796 (00:09)
2 0.135634 0.228471 0.077670 (00:11)
3 0.104969 0.201389 0.058252 (00:10)
4 0.096066 0.189406 0.048544 (00:09)
5 0.082029 0.162866 0.067961 (00:09)
6 0.071522 0.182156 0.058252 (00:09)
7 0.062653 0.201009 0.058252 (00:09)
8 0.058462 0.186100 0.067961 (00:08)
9 0.052997 0.165251 0.058252 (00:10)
10 0.048404 0.157111 0.048544 (00:08)
resnet50 did a little better but basically the same top losses. here i thought i might of had the learning rate too hight, but it finally settled down:
learn.fit_one_cycle(10, max_lr=slice(8e-5,8e-4))
Total time: 01:53
epoch train_loss valid_loss error_rate
1 0.164247 0.230282 0.067961 (00:11)
2 0.112986 0.227918 0.077670 (00:11)
3 0.096869 0.294595 0.087379 (00:11)
4 0.085803 0.348838 0.126214 (00:11)
5 0.070907 0.180965 0.097087 (00:11)
6 0.081053 0.120536 0.048544 (00:11)
7 0.069127 0.182275 0.048544 (00:11)
8 0.059278 0.175658 0.048544 (00:10)
9 0.051978 0.154336 0.048544 (00:11)
10 0.048301 0.144162 0.038835 (00:11)
1 Like
Session conversion prediction based on clickstream data
Problem statement: Predict probability of conversion in user’s last session(site visit) given clickstream data of users.
Business use cases: Large online businesses(e-commerce, digital media, edtech) selling a product has a notion of conversion(purchase). The product and marketing team wants to know the likelihoods of a user making a purchase based on user’s recent activities and past behaviors. This helps them target users in a promotional/retargeting emails/ad campaigns, and also discover key signals contributing to conversions.
Model definition: p(conversion on last session | events up to last X days, user)
Traditional approach:
- Create RFM(Recency, Frequency, Monetary value) features from clickstream event time series. Excellent read on this topic: https://www.kdd.org/kdd2016/papers/files/adf0160-liuA.pdf
- Create a structured dataset with a target label(user converted or not converted on the last visit) and predictors(features)
- Use GBM or Logistic regression or SVM
Challenges with the traditional approach:
- Clickstream event data is messy, each event contains several dimensions(time, metadata, taxonomy)
- It is extremely hard and time-consuming to manually create features from this large feature space
- Feature engineering requires domain understanding
- Capturing hidden patterns in browsing patterns is hard
- GBM works reasonably well in the presence of good features, not otherwise
Approach with CNN:
- Represent clickstream as a heat map matrix
- Heatmap contains x-axis: time intervals, y-axis: events, cell value as number of event occurrences in a given time interval
- Normalize the heatmap intensity, i.e count values goes from 1 to 50. Mask all cells with no activity
Benefits
- Can use simple aggregation such as sum as the intensity value
- No special treatment for an event. Could easily normalize intensity for each event based on event count distribution across users.
- (Hyperparameter) Use variable time intervals - smaller for recent history, larger for old history
- Automated feature engineering - learns hidden browsing patterns
- Transfer learning: use this approach to learning embeddings of user clickstream, use in other models as features or use it for user segmentation(clustering)
- Less overfitting compared to GBM
Limitations:
- Using different aggregates as cell values
- Not practical to use this approach alone for prediction. Often user attributes need to be included in the model as well, which are numerical/categorical
Exercise
- Training row: User’s activity up to last 7 days since their last session
- Target label: purchase or no purchase in the last session
- Time intervals - 5 mins up to 1 hour, 1 hour up to 1 day, 1 day up to X days
- Dataset - tried this for an e-commerce client, however, this approach is vertical agnostic.
- Dataset size: 15k images, 4% conversion rate. Next step is to test on bigger dataset
- Events - [pageload, product_view, add_to_cart, wishlist, checkout, purchase, email_click, email_open]
Performance:
Update 2: Tried cropping and padding the image at center (0, 0.5) and using the 244x244 image size. This ensure that critical part of the image is not cropped. The accuracy is still the same. At this point, it looks like there is no issue with the learner and we could blame the dataset. Most top losses samples are new users(with no prior history) who converted with few activity. A better dataset and some baseline models to compare the performance would be the next step.
TODO: Benchmark against XGBoost/GBM.
crop_pad_tfms = [crop_pad(size=244, padding_mode='reflection', row_pct = 0.5, col_pct = 0.0)]
tfms = (crop_pad_tfms, crop_pad_tfms)
data = ImageDataBunch.from_name_re(path_img, fnames, pat, bs=bs, ds_tfms=tfms, size=244)
Update 1: Model isn’t performing too well on True labels. 96% accuracy is mainly attributed to TNs. I’m not sizing the image to size=224 as recommended for resnet setting because this takes the image edges out, edges contains the most meaningful information(recent activity). Could this be contributing to the performance? I don’t fully understand the implications of this recommended size setting but I could recreate images such that meaningful information is not lost on resizing to 224. This would verify the assertion. ds_tfms aren’t applied.
TODO: Extend the lookback window(currently 7 days) to borrow more past browsing behaviors
24 Likes
Hi all!
“Inspired by” this great song I decided to make a useless web app that can count the amount of fingers on a hand.
I never build something real with Python, so this was a cool challenge.
I decided to use Django as a framework and deploy it on an $5 / month Ubuntu server, just to see how that would work. Digital Ocean’s tutorials on the server setup were of great help.
The dataset is a combination of Google Images and a Kaggle Sign Language Digits Dataset.
With the Kaggle dataset alone I could get to 0% error rate in a Jupyter Notebook, but this dataset consists of a lot of similar pictures that all have a white background and good lighting.
To make it work with more diverse images proved to be much more difficult. The current accuracy is ± 80%.
Although I would have loved to make something with real value, this was a fun exercise and I learned a lot. Looking forward to next week!
15 Likes
I hope the accuracy is >99%! Could get dangerous!! 
2 Likes
I’ve built whatcar.xyz to guess the Make and Model of photos of cars.
It currently classifies around 400 models of the most popular cars in Australia
It has 88% accuracy on a balanced validation set (although there are some duplicates in my dataset; if any leaked into the validation set this is optimistic). This is pretty impressive since some of the models are really similar
Building an app is a great experience; being able to test my model live with photos from my phone is amazing and helped me understand my model a lot more. My first iteration only had 50 models and I had to hunt for the cars it could classify; now it does most models I see on the street and is generally in the ballpark with classification.
The server code using Starlette is at https://github.com/EdwardJRoss/whatcar. I’ll do a write up and share a training notebook on the weekend.
I’d love to hear feedback if you try it out; unfortunately it won’t work on car models not sold much in Australia.
12 Likes
Pretty cool.
How did you get the data.
Did it manually?