Devlog #7: Hello! I added a bunch of visual fields in the blocks, such as a dimension size thing - that shows the sizes of the dimensions (they all need to somehow match! using flatten is easiest). There is stuff underneath the run button - which includes the inference, etc.
Devlog #6: I added set seed as a extension. This took forever, because for some reason Tensorflow.js does not have the set_seed feature built in (unlike Python). Other than that, I fixed a couple bugs.
Devlog #5: The website is really really slow when trying to run the models, so I swapped to Web Workers to handle the actual code execution (in the browser still) so it wouldn’t lag down the frontend of the website. I also added two more blocks: Separable Convolutional Network and Permute Layer. Seperable has higher inference speeds (and is used in stuff like MobileNet - a lightweight, mobile computer vision models). Permute modifies the axes of a dataset without actually changing the values of the data.
Devlog 4: I added support for XOR datasets, and fixed a lot of bugs, along with rewriting some of the code to be easier to use and quicker to use. I also cleaned up the toolbox.
Devlog #3: I built up a second page for the saved projects and deletion, and modified the README file for the definitions of some words.
Devlog #2: Hello! I’ve added a ton of new features into the software. Right now, I’ve added Conv2d Layers, Activation Layers, Dropout Layers, LSTM Layers,Max Pooling Layers, Multi-head attention (still a W.I.P), Embedding Layers for text, Batch/Layer Normalization, ability to choose custom loss functions, more activation types ,math expressions for math dataset, JSON/JSONL-dataset support, RNNs, a vocabulary list of all the fancy words, custom inference options for raw-inference, and a default template (seen in the image below).
Devlog 1: Hello! I’ve built out what I was thinking for the design of this project. It uses blockly for the scratch style engine, and a custom compiler in the background to convert it to actual code. I’m planning to add more types of data, like the ability to import custom data for the model training, and in the future, more types of layers.
Devlog 2: I cleaned up the graph by making the lines smaller, plus I added a output layer + and a label for each of the layers. I also reorganized the UI so it can fit on one screen.
Devlog 1: I’ve built up the website, but now I’m going to continue working on the activation chart. Its supposed to look like the “brain visualizers” that show the connections between neurons. Currently, each node is 48 neurons (in the AI model) since a 1:1 ratio would need over 14,000 nodes (which will cause a lot of lag).