Replies: 5 comments 12 replies
-
|
For instance, a custom loss function would greatly improve training times when using reinforcement learning. |
Beta Was this translation helpful? Give feedback.
-
|
OMG 🤣 Soooo... A lot more digging, and I found out that this was already solved 🥰 |
Beta Was this translation helpful? Give feedback.
-
|
While Is there perhaps a way to set a custom loss/cost/error function for training? If not, I'd be willing to implement the feature into For the past few years, I've been using Brain.js way more than tf.js; honestly, this would make it possible to switch over to using Brain.js exclusively. In the meantime, I'll run some tests to determine how fit |
Beta Was this translation helpful? Give feedback.
-
|
Yea, let's support it. Hell is greatly appreciated. |
Beta Was this translation helpful? Give feedback.
-
This line in this.backwardPropagate[this.outputLayer] = this.gpu.createKernelMap(For CPU neural nets, I'm thinking about this line here in this.calculateDeltas(value.output);As for the API, I'm thinking about just adding another function, I have much bigger ideas for the custom loss function than a simple HELM, but yeah. I'll update you once I have something nice for you, Robert 💞 |
Beta Was this translation helpful? Give feedback.
-
|
Side note: I suggest playing around with that ELM implementation in your spare time. The training pattern is quite interesting!! For the standard XOR training set, it instantly reaches an error of ~17% by iteration 1000, but then struggles for any further iterations to drop below 17%. Once this threshold is reached, it any further progress seems extremely slow. I'm thinking if we leverage this in our HELM implementation by rushing the ELM's peak optimization and then switch over to standard backpropagation for all further training, we can take advantage of the extremely fast rush to the ELM's threshold while totally avoiding the ELM's threshold weakness (the 17% I mentioned in the case of the standard XOR training set). |
Beta Was this translation helpful? Give feedback.
-
|
btw the work is being done in this branch. Already have a working implementation for function loss(
this: IKernelFunctionThis,
actual: KernelOutput, // The actual output data of the neural network during this run.
expected: KernelOutput, // The output data expected to be produced by the neural network given the provided input data.
inputs: InputType, // The input data provided to entire neural network (the data passed into the `NeuralNetwork.run(input)` method).
state: KernelOutput[][], // The current training state of the neural network; see below for more information.
) {
return expected[this.thread.x] - actual[this.thread.x];
}Where:
The aspect of the An example designed to optimize training for XOR might look like: function loss(
this: IKernelFunctionThis,
actual: KernelOutput, // The actual output data of the neural network during this run.
expected: KernelOutput, // The output data expected to be produced by the neural network given the provided input data.
inputs: InputType, // The input data provided to entire neural network (the data passed into the `NeuralNetwork.run(input)` method).
state: KernelOutput[][], // The current training state of the neural network; see above for more information.
) {
// Perform a standard loss function as our foundation to build upon.
let neuronLoss = expected[this.thread.x] - actual[this.thread.x];
// if ( o == i0 ^ i1 ) then return 10% of the loss value.
// Otherwise, return the full loss value.
if (Math.round(actual[0]) === Math.round(inputs[0]) ^ Math.round(inputs[1])) neuronLoss *= 0.1;
return neuronLoss;
}which allows for you to train a neural network to predict XOR calculations without ever even having been fed a data set. Just stream random data as training input, allow it to produce random output at first, and train on the error from the calculation. Instead of telling the network, "Hey! There's a bounty reward for whoever can solve for Y given X," you can instead define a rule to determine what is or is not good behavior, or even create a hybrid loss function. The above kernel function is a simple example of a hybrid loss function that I kind of just tossed together to train a standard feed-forward network to solve for XOR by using the standard XOR training set. In my experience, this greatly improves training times when done correctly. Sending tons of positivity and good vibes your way friend 💕 |
Beta Was this translation helpful? Give feedback.
-
|
I just realized that a |
Beta Was this translation helpful? Give feedback.
-
|
Here's some links in case you're interested 😊
|
Beta Was this translation helpful? Give feedback.
-
|
Darn... I ran into a major roadblock. For some odd reason, it's refusing to let me add any parameters past "target" in the subkernel, and I can't possibly imagine why. It gives this error: any clues as to why gpu.js won't let me add more than 2 parameters to the subkernel? |
Beta Was this translation helpful? Give feedback.
-
|
Simplifying the subkernel didn't resolve the error either, so it's likely not parsing related... |
Beta Was this translation helpful? Give feedback.
-
|
IT WORKS!!! AHHHHH OMG!!! Okay, okay, so it's working on both CPU and GPU now!!! The catch is that I need to inform |
Beta Was this translation helpful? Give feedback.
-
Happy little benchmarks
|
Beta Was this translation helpful? Give feedback.
-
|
Okay, this is honestly the coolest feature I've ever added to anything. This has me so excited!!! Like, to give you an idea of why I'm so excited: A custom loss function made training an autoencoder much faster, and within a single |
Beta Was this translation helpful? Give feedback.
-
Is it possible to define a custom loss function for a neural network in Brain.js like in Tensorflow?
If not, where might I begin to implement it? I took a look in the codebase and searched for terms like "loss", "cost", etc. without much luck. I found the error calculation functions, but I'm still lost tbh. Any help would be greatly appreciated 🥰
Beta Was this translation helpful? Give feedback.
All reactions