A place where "asap" is actually only used when necessary.

Saw this today in the exercise code of a deep learning lecture. These people are doing their PhDs in deep learning 😳

* the snippet has no side effects

Decrease the length of my to-do list

What happens when you feed an LSTM the Godfather movies as input?
It will give you an offer you can't refuse.

Credit: Twitter

The fact that I need to make this shit multimodal is gonna be a whole different level of shitshow. 🤦🤦🤦🤦🤦

Somebody kill me plz.

Today I tried to concatenate a LSTM unit with a FC and was wondering why it was throwing weird shapes at me. 🤦
Yes, I was THE idiot.
Kill me.

Background: I am currently working with a DB that has websocket functionality ("notify a client on insert/etc."). However, you do have to whitelist tables in order to use them with sockets.

I wanted to optimize my code and didn't want to mess with my coworkers dev-data, therefore I duplicated the table. After improving some small things I noticed that the interface does not change with new socket data. I have spent the last hour or so trying to figure out where I broke it.

I just realized that I forgot to whitelist that duplicated table 😐 Most relieving moment today 😅

Bonus side effect: The code is much cleaner now since I refactored a lot of the realtime-logic in order to understand it/fix the bug.

Picking up the work on an existing Wordpress site.

When a non-tech person started a WP page and you as a software engineer should "improve it a little bit, since [the client] already did most of the work"...

By now - unless I really cannot get around it - I refuse to start working extensively on existing wordpress sites.

So I got the LSTM working in keras.
Working from a glorified tutorial.
Why the fuck do people let their github pages go down with no other backup?
Especially if its a link in your blog?
Why would you do that and not post the full script (instead of bits and pieces interspersed with *partial* explanations)?

In any case, its working and training on a test set and examples just to debug my own understanding of the process.

Once thats done I can generate some training data and try training on a small set. If that goes smoothly and the loss looks like it is heading in the right direction, then I'll setup the hardware for the private cloud and start writing the parallel computing component.

I finally got the lstm to a training and validation loss of < 0.05 for predicting the digits of a semiprime's factors.

I used selu activation with lecun normal initialization on a dense decoder, and compiled the model with Adam as the optimizer using mean squared error.

Selu is self-normalizing, meaning it tends to mean 0 and preserves a standard deviation of one, so it eliminates the exploding/vanishing gradient problem. And I can get away with this specifically because selu *only* works on dense layers.

I chose Adam, even though this isn't a spare problem, because Adam excels on noisy problems and non-stationary objectives (definitely this), and because adam typically doesn't require a lot of hyperparameter tuning its ideal here, especially considering because I don't know what the hyperparameters should be to begin with.

I did work out some general guidelines on training quantity vs validation, etc.
The initial set wasn't huge or anything, roughly 110k pairs for training.

It converged pretty quick all things considered, and to the low loss like I mentioned, but even then the system always outputs the same result, regardless of the input, so obviously I'm doing something incorrectly.

The effectiveness of this approach for training and validation makes me question if I haven't got something wildly wrong. Still exploring though and figuring out how to get my answers back out. I'm hoping I just fucked up the output, and not the input as well.

TypeError: can't pickle _thread.lock objects

someone please guide me .....
pick.dump(model.fit(trainX, trainY, epochs=25, batch_size=1, verbose=2), filename)

git rebase is really nice. Can recommend it to clean up a git history :D

Earlier this day, I read that Strapi is dropping its support for MongoDB. I was a bit bummed at first, but their reasoning was good and I moved from MongoDB Atlas to CloudSQL.

From that point on my day got so much better: Now my strapi backend is so much faster than before! I cannot believe, that I just got to migrate to SQL. Should have done this a long time before.

All operations are literally 2-3 times as fast as before. Thank you @strapiDevelopers for forcing me to migrate :D

I am applying CNN-LSTM model to predict the level of interest of a person in a particular image or video. But, I don't have any dataset on EEG sensors
nor do I have any facility to gather such data. Can anyone help me in any way?

Working a week on LSTM based text classifier, getting 89% accuracy only to then get better result with Logistic Regression which was supposed to serve as baseline, lol. Background: 180+ classes of google product categorization taxonomy, 20 million rows of data items (short texts). Had a similar experience once on sentiment classification, where SVMlight outperformed NN models.

Fuck! I frkn hate supabase right now.

Row Level Security seems like a nice feature but it does not fckn work for me.