A deep learning library built using javascript to build and train machine learning models in the browsers.
npm install aksquare.js --save
via Script tag Add the following code to an HTML file:
<html>
<head>
<script>https://cdn.jsdelivr.net/npm/aksquare.js@1.0.1/aksquare/aksquare.min.js</script>
<script>
let text = "The king is a man who rules over a nation, he always have a woman beside him called the queen.
text_lower = text.toLocaleLowerCase()
text_list = text_lower.split("\n")
var stopwords = ["a","in","when","the","of","is","who"]
let [word_list, all_text] = gen_word(5,text_list);
let unique_dict = unique_word(all_text)
let n_words = obj_len(unique_dict);
console.log(n_words);
let [data, label] = create_data(word_list)
let embed_dim = 50;
let model = new aksquare.Sequential([
new aksquare.Linear(n_words,embed_dim),
new aksquare.Linear(embed_dim,n_words),
new aksquare.Softmax()
]);
let optim = new aksquare.OptimSGD(model,lr=0.001);
epoch = 50
for(let i=0; i< epoch; i++){
let total_loss = 0;
for(let j=0; j < data.length; j++){
let x_data = data[j]
let y_data = label[j]
let x = new aksquare.Tensor(1,n_words, false);
x.setFrom(x_data)
model.forward(x)
// console.log(-Math.log(model.out.out[y_data-1]))
let loss = new aksquare.Loss(y_data-1,model)
// console.log(loss.out);
total_loss += loss.out
loss.backward()
optim.step();
optim.grad_zero()
}
console.log(`for epoch ${i} Loss is ${total_loss/data.length}`)
}
//get embedding weight
let embed_weight = get_weight(model.models[0].W)
console.log(embed_weight[0].length)
</script>
</head>
<body>
</body>
</html>
Via NPM
Install aksquare.js in your project using yarn or npm
import * as aksquare from 'aksquare.js'
let embed_dim = 50;
let model = new aksquare.Sequential([
new aksquare.Linear(n_words,embed_dim),
new aksquare.Linear(embed_dim,n_words),
new aksquare.Softmax()
]);
let optim = new aksquare.OptimSGD(model,lr=0.001);
let x_data = data[j]
let y_data = label[j]
let x = new aksquare.Tensor(1,n_words, false);
x.setFrom(x_data)
model.forward(x)
let loss = new aksquare.Loss(y_data-1,model);
total_loss += loss.out
loss.backward()
optim.step();
optim.grad_zero()
}
let embed_weight = get_weight(model.models[0].W)
console.log(embed_weight[0].length)