Skip to content
Usefull Layers for Knet
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
docs fix rnnoutput, update readme Jan 10, 2019
examples remove cell output Feb 8, 2019
src mnist examples and readme update Feb 6, 2019
test fixes and IndexedDict Jan 25, 2019
.gitignore initial commit Sep 2, 2018
.gitlab-ci.yml Update .gitlab-ci.yml Nov 28, 2018
.travis.yml update travis Dec 19, 2018
LICENSE Create LICENSE Sep 20, 2018
Project.toml training -> enable in dropout; nll with masking Nov 16, 2018 Update Feb 8, 2019
REQUIRE REQUIRE added Sep 17, 2018



KnetLayers provides usefull deep learning layers for Knet, fostering your model development. You are able to use Knet and AutoGrad functionalities without adding them to current workspace.


model = Chain(Dense(input=768, output=128, activation=Sigm()),
	      Dense(input=128, output=10, activation=nothing))

loss(model, x, y) = nll(model(x), y)

Getting Started: Train an MNIST model

using Knet, KnetLayers
import Knet: Data
dtrn,dtst = mnistdata(xsize=(784,:)); # dtrn and dtst = [ (x1,y1), (x2,y2), ... ] where xi,yi are

HIDDEN_SIZES = [100,50]
(m::MLP)(x,y) = nll(m(x),y)
(m::MLP)(d::Data) = mean(m(x,y) for (x,y) in d)
model = MLP(784,HIDDEN_SIZES...,10)


@show 100accuracy(model, dtst)

Example Models



  3. GAN-MLP

  4. ResNet: Residual Networks for Image Recognition

  5. S2S: Sequence to Sequence Reccurent Model

  6. Morse.jl: Morphological Analyzer+Lemmatizer

  7. MAC Network: Memory-Attention-Composition Network for Visual Question Answering

Exported Layers Refence

Example Layers and Usage

using KnetLayers

#Instantiate an MLP model with random parameters
mlp = MLP(100,50,20; activation=Sigm()) # input size=100, hidden=50 and output=20

#Do a prediction with the mlp model
prediction = mlp(randn(Float32,100,1))

#Instantiate a convolutional layer with random parameters
cnn = Conv(height=3, width=3, inout=3=>10, padding=1, stride=1) # A conv layer

#Filter your input with the convolutional layer
output = cnn(randn(Float32,224,224,3,1))

#Instantiate an LSTM model
lstm = LSTM(input=100, hidden=100, embed=50)

#You can use integers to represent one-hot vectors.
#Each integer corresponds to vocabulary index of corresponding element in your data.

#For example a pass over 5-Length sequence
rnnoutput = lstm([3,2,1,4,5];hy=true,cy=true)

#After you get the output, you may acces to hidden states and
#intermediate hidden states produced by the lstm model

#You can also use normal array inputs for low-level control
#One iteration of LSTM with a random input
rnnoutput = lstm(randn(100,1);hy=true,cy=true)

#Pass over a random 10-length sequence:
rnnoutput = lstm(randn(100,1,10);hy=true,cy=true)

#Pass over a mini-batch data which includes unequal length sequences
rnnoutput = lstm([[1,2,3,4],[5,6]];sorted=true,hy=true,cy=true)

#To see and modify rnn params in a structured view


  1. Examples
  2. Special layers such Google's inception
  3. Known embeddings such Gloove
  4. Pretrained Models
You can’t perform that action at this time.