Why does the Nesterov flag not affect the result of calculations? /training/solver.go:

// Update returns the update for a given weight
func (o *SGD) Update(value, gradient float64, iteration, idx int) float64 {
	lr := o.lr / (1 + o.decay*float64(iteration))

	o.moments[idx] = o.momentum*o.moments[idx] - lr*gradient

	if o.nesterov {
		o.moments[idx] = o.momentum*o.moments[idx] - lr*gradient
	}

	return o.moments[idx]
}

Currently there are a few Activation/Loss functions which are provided in go-deep. While internally these functions follow an interface, this is not completely exposed to the clients of this library. For example, If I would like to experiment with a new activation function I would not be able to without modifying the code in go-deep.

If we can expose the ability for clients of this library to provide their own activation functions by providing types that implement the Differentiable and Loss interfaces, I think the flexibility of this library will be greatly increased.

If this is something that sounds reasonable, I would be more than happy to take a crack at implementing this and submitting a merge request.

Hello,

I am trying to save the Neural structure to disk using "encoding / gob" but it seems that it does not work. Am I doing something wrong or should I use something else?

thank you

Would be great to support passing in an interface for the printer that the batch trainer uses, in order to implement custom stopping depending on loss. So loss would need to be part of the interface signature in numeric form.

Is there any way to use the user-defined loss functions so that the model can be trained based on the loss from these defined functions? or am i missing something?

Hi!

There is a code in the activation.go:

func (a ReLU) F(x float64) float64 { return math.Max(x, 0) }

func (a ReLU) Df(y float64) float64 {
	if y > 0 {
		return 1
	}
	return 0
}

This is no correct! The right option:

func (a ReLU) F(x float64) float64 { return math.Max(x, 0) }

func (a ReLU) Df(y float64) float64 {
	if y < 0 {
		return 1
	}
	return 0
}

For example: https://github.com/FluxML/NNlib.jl/blob/master/src/activations.jl

Q

Do you plan to develop the library further? Thanks!

I'm using the library for some project, and I'm finding out how useful a training history feature is useful (as you can find in any soa library in python). Is there any plan for implement it ?

In any case, I'm starting a new one.

Hi, the example in the readme is giving errors:

// params: learning rate, momentum, alpha decay, nesterov optimizer := training.NewSGD(0.05, 0.1, 1e-6, true) // params: optimizer, verbosity (print stats at every 50th iteration) trainer := training.NewTrainer(optimizer, 50)

Cannot use optimizer as (type *SGD) as type Solver

Is connecting a sparse network similar to what is being done in NewNeural() supposed to work? I noticed the batch trainer uses NewNeural() internally.

Mish is a new novel activation function proposed in this paper. It has shown promising results so far and has been adopted in several packages including:

• TensorFlow-Addons
• SpaCy (Tok2Vec Layer)
• Thinc - SpaCy's official NLP based ML library
• Echo AI
• CNTKX - Extension of Microsoft's CNTK
• FastAI-Dev
• Darknet
• Yolov3
• BeeDNN - Library in C++
• Gen-EfficientNet-PyTorch
• dnet

All benchmarks, analysis and links to official package implementations can be found in this repository

It would be nice to have Mish as an option within the activation function group.

This is the comparison of Mish with other conventional activation functions in a SEResNet-50 for CIFAR-10:

I added another activation function, i.e. LeakyReLU. This can be increase numerical efficiency in cases where you have flat cost functions.

Another implementation choice could be add another parameter to ReLU, whose default is 0 and then just use that with other values also for LeakyReLU.