CODEWITHSUNDEEP
keras
VincentDM
VincentDM

Reputation: 489

Unknown value in Incompatible shapes error from Keras/TF

I'm getting started with Keras and I think I'm missing something regarding how layers interact together.

I have this shape of data :

x shape : (696, 5, 6)
y shape : (696, 5, 2)

And a really simple model that I'm just trying to make work to keep exploring

inputShape = (xtrain.shape[1], xtrain.shape[2])    
batchSize = xtrain.shape[0] / 6
outputDim = ytrain.shape[2]

model = Sequential()
model.add(Dense(500, activation='relu', input_shape=inputShape, batch_size=batchSize))
model.add(Dense(500, activation='relu'))
model.add(Dense(outputDim, activation='softmax'))
model.compile(optimizer='rmsprop', loss='mse')

And I can't figure out where is the 32 coming from in this error

tensorflow.python.framework.errors_impl.InvalidArgumentError:  Incompatible shapes: [116,5] vs. [32,5]

Upvotes: 1

Views: 605

Answers (1)

rvinas
rvinas

Reputation: 11895

You almost have it, Vincent. Let me explain you what is going on:

  • When passing batch_size=batchSize to the first Dense layer, the model is expecting all the inputs to have batch size batchSize.

  • However, when you fit your model via:

    model.fit(xtrain, ytrain)

    the error is raised, because the default value for the argument batch_size of model.fit is 32, which is what model.fit uses to create the batches from xtrain and ytrain.

There are two possible solutions to fix this issue:

  1. You can set the argument batch_size of model.fit to batchSize.
  2. A more flexible solution is to remove batch_size (or set batch_size=None) of the first Dense layer. This allows to have dynamic batch sizes - the network works for all batch sizes (e.g. you can set batch_size of model.fit or model.predict to an arbitrary value).

Upvotes: 2

Related Questions