more docstrings

This commit is contained in:
Dhairya Gandhi 2019-10-10 16:48:12 +05:30
parent fe52689cfe
commit f19066ee29

View File

@ -7,23 +7,20 @@ const ϵ = 1e-8
# TODO: should use weak refs
"""
# Descent
Descent(η)
## Description
Classic gradient descent optimiser with learning rate `η`.
For each parameter `p` and its gradient `δp`, this runs `p -= η*δp`
## Constructors
- `Descent()`: Use the default learning rate (η), as described in the parameters section.
- `Descent(η)`: Provide a custom learning rate (η) to the Descent optimiser.
## Parameters
- Learning rate (η): The amount by which the gradients are discounted before updating the weights. Defaults to `0.1`.
- Learning Rate (η): The amount by which the gradients are discounted before updating the weights. Defaults to `0.1`.
## Example
```julia-repl
opt = Descent()
opt = Descent() # uses default η (0.1)
opt = Descent(0.3) # use provided η
ps = params(model)
@ -47,11 +44,18 @@ end
"""
Momentum(η, ρ)
Calls to `Momentum()` default to:
- learning rate (η): 0.01
- decay (ρ): 0.9
Gradient descent with learning rate `η` and momentum `ρ`.
## Parameters
- Learning Rate (`η`): Amount by which gradients are discounted before updating the weights. Defaults to `0.01`.
- Momentum (`ρ`): Parameter that accelerates descent in the relevant direction and dampens oscillations. Defaults to `0.9`.
## Examples
```julia
opt = Momentum() # uses defaults of η = 0.01 and ρ = 0.9
opt = Momentum(0.01, 0.99)
```
"""
mutable struct Momentum
eta::Float64
@ -71,11 +75,18 @@ end
"""
Nesterov(η, ρ)
Calls to `Nesterov()` default to:
- learning rate (η): 0.001
- nesterov momentum (ρ): 0.9
Gradient descent with learning rate `η` and Nesterov momentum `ρ`.
## Parameters
- Learning Rate (η): Amount by which the gradients are dicsounted berfore updating the weights. Defaults to `0.001`.
- Nesterov Momentum (ρ): Paramters controlling the amount of nesterov momentum to be applied. Defaults to `0.9`.
## Examples
```julia
opt = Nesterov() # uses defaults η = 0.001 and ρ = 0.9
opt = Nesterov(0.003, 0.95)
```
"""
mutable struct Nesterov
eta::Float64
@ -96,13 +107,21 @@ end
"""
RMSProp(η, ρ)
Calls to `RMSProp()` default to:
- learning rate (η): 0.001
- rho (ρ): 0.9
Implements the RMSProp algortihm. Often a good choice for recurrent networks. Paramters other than learning rate generally don't need tuning.
## Parameters
- Learning Rate (η): Defaults to `0.001`.
- Rho (ρ): Defaults to `0.9`.
## Examples
```julia
opt = RMSProp() # uses default η = 0.001 and ρ = 0.9
opt = RMSProp(0.002, 0.95)
```
## References
[RMSProp](https://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf)
optimiser. Parameters other than learning rate don't need tuning. Often a good
choice for recurrent networks.
"""
mutable struct RMSProp
eta::Float64
@ -122,10 +141,20 @@ end
"""
ADAM(η, β::Tuple)
Calls to `ADAM()` default to:
- learning rate (η): 0.001
- (beta1, beta2) (β): (0.9, 0.999)
Implements the ADAM optimiser.
## Paramters
- Learning Rate (`η`): Defaults to `0.001`.
- Beta (`β::Tuple`): The first element refers to β1 and the second to β2. Defaults to `(0.9, 0.999)`.
## Examples
```julia
opt = ADAM() # uses the default η = 0.001 and β = (0.9, 0.999)
opt = ADAM(0.001, (0.9, 0.8))
```
## References
[ADAM](https://arxiv.org/abs/1412.6980v8) optimiser.
"""
mutable struct ADAM
@ -149,10 +178,21 @@ end
"""
RADAM(η, β::Tuple)
Calls to `RADAM()` default to:
- learning rate (η): 0.001
- (beta1, beta2) (β): (0.9, 0.999)
Implements the rectified ADAM optimizer.
## Parameters
- Learning Rate (η): Defaults to `0.001`
- Beta (β::Tuple): The first element refers to β1 and the second to β2. Defaults to `(0.9, 0.999)`.
## Examples
```julia
opt = RADAM() # uses the default η = 0.001 and β = (0.9, 0.999)
opt = RADAM(0.001, (0.9, 0.8))
```
## References
[RADAM](https://arxiv.org/pdf/1908.03265v1.pdf) optimiser (Rectified ADAM).
"""
mutable struct RADAM
@ -183,12 +223,20 @@ end
"""
AdaMax(η, β::Tuple)
Calls to `AdaMax()` default to:
- learning rate (η): 0.001
- (beta1, beta2) (β): (0.9, 0.999)
Variant of ADAM based on -norm.
[AdaMax](https://arxiv.org/abs/1412.6980v9) optimiser. Variant of ADAM based on
the -norm.
## Parameters
- Learning Rate (η): Defaults to `0.001`
- Beta (β::Tuple): The first element refers to β1 and the second to β2. Defaults to `(0.9, 0.999)`.
## Examples
```julia
opt = AdaMax() # uses default η and β
opt = AdaMax(0.001, (0.9, 0.995))
```
## References
[AdaMax](https://arxiv.org/abs/1412.6980v9) optimiser.
"""
mutable struct AdaMax
eta::Float64
@ -211,9 +259,19 @@ end
"""
ADAGrad(η)
Calls to `AdaGrad()` default to:
- learning rate (η): 0.1
Implements AdaGrad. It has parameter specific learning rates based on how frequently it is updated.
## Parameters
- Learning Rate (η): Defaults to `0.1`
## Examples
```julia
opt = ADAGrad() # uses default η = 0.1
opt = ADAGrad(0.001)
```
## References
[ADAGrad](http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) optimiser.
Parameters don't need tuning.
"""
@ -234,11 +292,19 @@ end
"""
ADADelta(ρ)
Calls to `ADADelta()` default to:
rho (ρ): 0.9
Version of ADAGrad that adapts learning rate based on a window of past gradient updates. Parameters don't need tuning.
[ADADelta](https://arxiv.org/abs/1212.5701) optimiser. Parameters don't need
tuning.
## Parameters
- Rho (ρ): Factor by which gradient is decayed at each time step. Defaults to `0.9`.
## Examples
```julia
opt = ADADelta() # uses default ρ = 0.9
opt = ADADelta(0.89)
```
## References
[ADADelta](https://arxiv.org/abs/1212.5701) optimiser.
"""
mutable struct ADADelta
rho::Float64
@ -259,12 +325,20 @@ end
"""
AMSGrad(η, β::Tuple)
Calls to `AMSGrad()` default to:
- learning rate (η): 0.001
- (beta1, beta2) (β): (0.9, 0.999)
Implements AMSGrad version of the ADAM optimiser. Parameters don't need tuning.
[AMSGrad](https://openreview.net/forum?id=ryQu7f-RZ) optimiser. Parameters don't need
tuning.
## Parameters
- Learning Rate (η): Defaults to `0.001`.
- Beta (β::Tuple): The first element refers to β1 and the second to β2. Defaults to `(0.9, 0.999)`.
## Examples
```julia
opt = AMSGrad() # uses default η and β
opt = AMSGrad(0.001, (0.89, 0.995))
```
## References
[AMSGrad](https://openreview.net/forum?id=ryQu7f-RZ) optimiser.
"""
mutable struct AMSGrad
eta::Float64
@ -286,12 +360,20 @@ end
"""
NADAM(η, β::Tuple)
Calls to `NADAM()` default to:
- learning rate (η): 0.001
- (beta1, beta2) (β): (0.9, 0.999)
Nesterov variant of ADAM. Parameters don't need tuning.
[NADAM](http://cs229.stanford.edu/proj2015/054_report.pdf) optimiser. Parameters don't need
tuning.
## Parameters
- Learning Rate (η): Defaults to `0.001`.
- Beta (β::Tuple): The first element refers to β1 and the second to β2. Defaults to `(0.9, 0.999)`.
## Examples
```julia
opt = NADAM() # uses default η and β
opt = NADAM(0.002, (0.89, 0.995))
```
## References
[NADAM](http://cs229.stanford.edu/proj2015/054_report.pdf) optimiser.
"""
mutable struct NADAM
eta::Float64
@ -314,11 +396,21 @@ end
"""
ADAMW(η, β::Tuple, decay)
Calls to `ADAMW()` default to:
- learning rate (η) 0.001
- (beta1, beta2) (β): (0.9, 0.999)
Variant of ADAM defined by fixing weight decay regularization.
[ADAMW](https://arxiv.org/abs/1711.05101) fixing weight decay regularization in Adam.
## Parameters
- Learning Rate (η): Defaults to `0.001`.
- Beta (β::Tuple): The first element refers to β1 and the second to β2. Defaults to (0.9, 0.999).
- decay: Decay applied to weights during optimisation. Defaults to 0.
## Examples
```julia
opt = ADAMW() # uses default η, β and decay
opt = ADAMW(0.001, (0.89, 0.995), 0.1)
```
## References
[ADAMW](https://arxiv.org/abs/1711.05101)
"""
ADAMW(η = 0.001, β = (0.9, 0.999), decay = 0) =
Optimiser(ADAM(η, β), WeightDecay(decay))
@ -353,10 +445,12 @@ end
"""
InvDecay(γ)
Calls to `InvDecay()` default to:
- gamma (γ): 0.001
Applies inverse time decay to an optimiser
Apply inverse time decay to an optimiser
## Parameters
- gamma (γ): Defaults to `0.001`
## Example
```julia
Optimiser(InvDecay(..), Opt(..))
```
@ -379,17 +473,20 @@ end
"""
ExpDecay(eta, decay, decay_step, clip)
Calls to `ExpDecay()` default to:
- learning rate (eta): 0.001
- decay: 0.1
- decay_step: 1000
- clip: 1e-4
Discount the learning rate `eta` by `decay` every `decay_step` till a minimum of `clip`.
## Parameters
- Learning Rate (eta): Defaults to `0.001`.
- decay: Factor by which the learning rate is discounted. Defaults to `0.1`.
- decay_step: Schedules decay operations by setting number of steps between two decay operations. Defaults to `1000`.
- clip: Minimum value of learning rate. Defaults to `1e-4`.
## Example
To apply exponential decay to an optimiser:
```julia
Optimiser(ExpDecay(..), Opt(..))
opt = Optimiser(ExpDecay(), ADAM())
```
"""
mutable struct ExpDecay
@ -415,10 +512,10 @@ end
"""
WeightDecay(wd)
Calls to `WeightDecay()` default to:
- weight decay (wd): 0
Decays the weight by `wd`
Decay the weight parameter by `wd`
## Parameters
- weight decay (wd): 0
"""
mutable struct WeightDecay
wd::Real