Styjun

How to set Output path correctly for a Single Image render?

Is XSS in canonical link possible?

Is it possible to have a strip of cold climate in the middle of a planet?

Difference between -| and |- in TikZ

How will losing mobility of one hand affect my career as a programmer?

Should I install hardwood flooring or cabinets first?

Have I saved too much for retirement so far?

Proving a function is onto where f(x)=|x|.

Can a significant change in incentives void an employment contract?

Can a Necromancer reuse the corpses left behind from slain undead?

Greco-Roman egalitarianism

A social experiment. What is the worst that can happen?

Filling the middle of a torus in Tikz

Folder comparison

How can "mimic phobia" be cured or prevented?

Are sinusoidal travelling waves also normal modes of vibration?

Why has "pence" been used in this sentence, not "pences"?

Why we can't differentiate a polynomial equation as many times as we wish?

Why don’t women say שלא עשני גויה and שפחה, according to Ashkenazim?

How much character growth crosses the line into breaking the character

Do Legal Documents Require Signing In Standard Pen Colors?

Extending the spectral theorem for bounded self adjoint operators to bounded normal operators

What is the gram­mat­i­cal term for “‑ed” words like these?

What major Native American tribes were around Santa Fe during the late 1850s?

ANN when to add new hidden layer

Are there cases where it is better to use sigmoid activation over ReLuUsing relu activation function destroys the modelHow to study the effect of each data on a deep neural network model?TensorFlow: Neural Network accuracy always 100% on train and test setsAccuracy of Model not change, remain at zeroUserWarning: Update your `Dense` call to the Keras 2 API:Something fundamentally wrong in this tensorflow model of two layeris normal behaviour in ANNs for the reduction of the loss function to be dependent on the number of neuron in the hidden layer?How to avoid Softmax activation failure when using ReLU in hidden layers?Neural Network predictions too large

-1

I have dataset based on some stocks.

I am using ANN to train my data. By default have the input and out put layer as follow:

 my_classifi = Sequential()

my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu', input_dim = 11))

my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu'))

my_classifi.add(Dense(output_dim = 1, init = 'uniform', activation = 'sigmoid'))

my_classifi.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])

# Fitting the ANN 
my_classifi.fit(X_train, y_train, batch_size = 50, nb_epoch = 100)

I get Acc of 0.8368

When I add a new hidden layer:

 my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu', input_dim = 11))

my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu'))
my_classifi.add(Dense(output_dim = 4, init = 'uniform', activation = 'relu'))

my_classifi.add(Dense(output_dim = 1, init = 'uniform', activation = 'sigmoid'))

my_classifi.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])

I get Acc of 0.8640

My questions are:

1) when I train the model again (with the extra layer), Accuracy falls back to 0.83. Why is so? I thought it should be around 0.86

2) what is the rule of thumb for adding (or not adding a new hidden layer). if so, what should be the output_dim

Thanks

python-3.x machine-learning neural-network

share|improve this question

asked Mar 21 at 13:41

Fred DowFred Dow

41

• 
  
  
  

  
  1
  

  
  
  
  
  
  

  
  
  Welcome to SO; your question is way too broad, please do take some time to read How to Ask and What topics can I ask about here?.
  
  – desertnaut
  Mar 21 at 13:52
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  
  There are no general rules of thumb; similarly for general reasons regarding variations like that in the accuracy (which are arguably not that great) when adding/removing single layers...
  
  – desertnaut
  Mar 21 at 13:54
  

  
  
  
  

add a comment | 

-1

I have dataset based on some stocks.

I am using ANN to train my data. By default have the input and out put layer as follow:

 my_classifi = Sequential()

my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu', input_dim = 11))

my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu'))

my_classifi.add(Dense(output_dim = 1, init = 'uniform', activation = 'sigmoid'))

my_classifi.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])

# Fitting the ANN 
my_classifi.fit(X_train, y_train, batch_size = 50, nb_epoch = 100)

I get Acc of 0.8368

When I add a new hidden layer:

 my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu', input_dim = 11))

my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu'))
my_classifi.add(Dense(output_dim = 4, init = 'uniform', activation = 'relu'))

my_classifi.add(Dense(output_dim = 1, init = 'uniform', activation = 'sigmoid'))

my_classifi.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])

I get Acc of 0.8640

My questions are:

1) when I train the model again (with the extra layer), Accuracy falls back to 0.83. Why is so? I thought it should be around 0.86

2) what is the rule of thumb for adding (or not adding a new hidden layer). if so, what should be the output_dim

Thanks

python-3.x machine-learning neural-network

share|improve this question

asked Mar 21 at 13:41

Fred DowFred Dow

41

• 
  
  
  

  
  1
  

  
  
  
  
  
  

  
  
  Welcome to SO; your question is way too broad, please do take some time to read How to Ask and What topics can I ask about here?.
  
  – desertnaut
  Mar 21 at 13:52
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  
  There are no general rules of thumb; similarly for general reasons regarding variations like that in the accuracy (which are arguably not that great) when adding/removing single layers...
  
  – desertnaut
  Mar 21 at 13:54
  

  
  
  
  

add a comment | 

I have dataset based on some stocks.

I am using ANN to train my data. By default have the input and out put layer as follow:

 my_classifi = Sequential()

my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu', input_dim = 11))

my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu'))

my_classifi.add(Dense(output_dim = 1, init = 'uniform', activation = 'sigmoid'))

my_classifi.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])

# Fitting the ANN 
my_classifi.fit(X_train, y_train, batch_size = 50, nb_epoch = 100)

I get Acc of 0.8368

When I add a new hidden layer:

 my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu', input_dim = 11))

my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu'))
my_classifi.add(Dense(output_dim = 4, init = 'uniform', activation = 'relu'))

my_classifi.add(Dense(output_dim = 1, init = 'uniform', activation = 'sigmoid'))

my_classifi.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])

I get Acc of 0.8640

My questions are:

1) when I train the model again (with the extra layer), Accuracy falls back to 0.83. Why is so? I thought it should be around 0.86

2) what is the rule of thumb for adding (or not adding a new hidden layer). if so, what should be the output_dim

Thanks

python-3.x machine-learning neural-network

share|improve this question

asked Mar 21 at 13:41

Fred DowFred Dow

41

 my_classifi = Sequential()

my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu', input_dim = 11))

my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu'))

my_classifi.add(Dense(output_dim = 1, init = 'uniform', activation = 'sigmoid'))

my_classifi.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])

# Fitting the ANN 
my_classifi.fit(X_train, y_train, batch_size = 50, nb_epoch = 100)

 my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu', input_dim = 11))

my_classifi.add(Dense(output_dim = 6, init = 'uniform', activation = 'relu'))
my_classifi.add(Dense(output_dim = 4, init = 'uniform', activation = 'relu'))

my_classifi.add(Dense(output_dim = 1, init = 'uniform', activation = 'sigmoid'))

my_classifi.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])

share|improve this question

asked Mar 21 at 13:41

Fred DowFred Dow

41

share|improve this question

asked Mar 21 at 13:41

Fred DowFred Dow

41

asked Mar 21 at 13:41

Fred DowFred Dow

41

asked Mar 21 at 13:41

Fred DowFred Dow

41