Trjtdtk

Why does processed meat contain preservatives, while canned fish needs not?

A ​Note ​on ​N!

How could Tony Stark make this in Endgame?

What are the potential pitfalls when using metals as a currency?

How exactly does Hawking radiation decrease the mass of black holes?

Reducing vertical space in stackrel

How to pronounce 'C++' in Spanish

To say I met a person for the first time

What does it mean to express a gate in Dirac notation?

What does KSP mean?

How can Republicans who favour free markets, consistently express anger when they don't like the outcome of that choice?

French for 'It must be my imagination'?

Why does academia still use scientific journals and not peer-reviewed government funded alternatives?

Will a top journal at least read my introduction?

Examples of subgroups where it's nontrivial to show closure under multiplication?

What is the difference between `command a[bc]d` and `command `ab,cd`

What happened to Captain America in Endgame?

Apply MapThread to all but one variable

Why do games have consumables?

how to find the equation of a circle given points of the circle

What does the "ep" capability mean?

Are Boeing 737-800’s grounded?

Is it possible to determine the symmetric encryption method used by output size?

How to get a plain text file version of a CP/M .BAS (M-BASIC) program?

What encoding to use for my musical vectors?

LSTM neural network for music generationRecurrent neural network multiple types of input KerasRight Way to Input Text Data in Keras Auto EncoderNeural network outputting same result for all inputsHow to optimally train deep learning model using output as new inputPredicting next number in a sequence - data analysisRecommended model for univariate or multivariate multistep ahead time series forecastingArchitecture help for multivariate input and output LSTM modelsHow to reshape data for LSTM training in multivariate sequence predictionAdding context in a sequence to sequence problem

0

$begingroup$

I'm trying to build a music recommendations system using an encoder-decoder sequence-to-sequence architecture using keras. My dataset comprises of playlists containing songs represented as a 13-dimensional feature vector(beat,tempo,key etc). Each playlist acts as a training sample with song vectors for each time step (analogous to words in a sentence). At each time step of the decoder a song from the song vocabulary must be outputed.

The model: Encoder(input layer, single LSTM layer), Decoder(input layer,LSTM layer, softmax layer)

if S[0...N] is a playlist of songs:
encoder inputs = S[0...N-1],
decoder inputs = S[1...N],
decoder targets = decoder inputs shifted by one time step

I am presently using a one hot encoding on the song vocabulary to encode the songs. However this is becoming computationally expensive as the song vocabulary is huge (30000 songs). Furthermore, this limits the network to only learn from context of songs in playlist rather than the feature vectors along with context.

What alternative can I use for the one hot encoding? Is it possible to use the normalized feature vectors as is? If so how would my output layer change and what would be the loss function? Thanks.

machine-learning python keras recurrent-neural-net sequence-to-sequence

share|improve this question

asked Apr 7 at 6:36

FusRhoDa4897FusRhoDa4897

11

$endgroup$

add a comment | 

0

$begingroup$

I'm trying to build a music recommendations system using an encoder-decoder sequence-to-sequence architecture using keras. My dataset comprises of playlists containing songs represented as a 13-dimensional feature vector(beat,tempo,key etc). Each playlist acts as a training sample with song vectors for each time step (analogous to words in a sentence). At each time step of the decoder a song from the song vocabulary must be outputed.

The model: Encoder(input layer, single LSTM layer), Decoder(input layer,LSTM layer, softmax layer)

if S[0...N] is a playlist of songs:
encoder inputs = S[0...N-1],
decoder inputs = S[1...N],
decoder targets = decoder inputs shifted by one time step

I am presently using a one hot encoding on the song vocabulary to encode the songs. However this is becoming computationally expensive as the song vocabulary is huge (30000 songs). Furthermore, this limits the network to only learn from context of songs in playlist rather than the feature vectors along with context.

What alternative can I use for the one hot encoding? Is it possible to use the normalized feature vectors as is? If so how would my output layer change and what would be the loss function? Thanks.

machine-learning python keras recurrent-neural-net sequence-to-sequence

share|improve this question

asked Apr 7 at 6:36

FusRhoDa4897FusRhoDa4897

11

$endgroup$

add a comment | 

$begingroup$

I'm trying to build a music recommendations system using an encoder-decoder sequence-to-sequence architecture using keras. My dataset comprises of playlists containing songs represented as a 13-dimensional feature vector(beat,tempo,key etc). Each playlist acts as a training sample with song vectors for each time step (analogous to words in a sentence). At each time step of the decoder a song from the song vocabulary must be outputed.

The model: Encoder(input layer, single LSTM layer), Decoder(input layer,LSTM layer, softmax layer)

if S[0...N] is a playlist of songs:
encoder inputs = S[0...N-1],
decoder inputs = S[1...N],
decoder targets = decoder inputs shifted by one time step

I am presently using a one hot encoding on the song vocabulary to encode the songs. However this is becoming computationally expensive as the song vocabulary is huge (30000 songs). Furthermore, this limits the network to only learn from context of songs in playlist rather than the feature vectors along with context.

What alternative can I use for the one hot encoding? Is it possible to use the normalized feature vectors as is? If so how would my output layer change and what would be the loss function? Thanks.

machine-learning python keras recurrent-neural-net sequence-to-sequence

share|improve this question

asked Apr 7 at 6:36

FusRhoDa4897FusRhoDa4897

11

$endgroup$

share|improve this question

asked Apr 7 at 6:36

FusRhoDa4897FusRhoDa4897

11

share|improve this question

asked Apr 7 at 6:36

FusRhoDa4897FusRhoDa4897

11

asked Apr 7 at 6:36

FusRhoDa4897FusRhoDa4897

11

asked Apr 7 at 6:36

FusRhoDa4897FusRhoDa4897

11