Alkimia
Alkimia

Reputation: 33

keras shape mismatch between predict and fit

I have a simple model that flattens sequence embeddings and then sums them. When I run predict I get no errors and the output shape that I expect but when I try to train I get a shape mismatch error

Here is the model:

import numpy as np
from keras import backend as K
from keras.models import Model
from keras.layers.embeddings import Embedding
from keras.layers import Reshape, Lambda

inputs = Input(shape=(20,), name="inputs")
embedding = Embedding(69, 100, name="embeddings")(inputs)
out = Reshape((2000,), name='reshape_embeddings')(embedding)
out = Lambda(lambda x: K.sum(x, axis=1), name='sum_embeddings')(out)
model = Model(inputs, out)
model.compile('adam', 'mean_squared_error')
print(model.summary())

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
inputs (InputLayer)          (None, 20)                0         
_________________________________________________________________
embeddings (Embedding)       (None, 20, 100)           6900      
_________________________________________________________________
reshape_embeddings (Reshape) (None, 2000)              0         
_________________________________________________________________
sum_embeddings (Lambda)      (None,)                   0         
=================================================================
Total params: 6,900
Trainable params: 6,900
Non-trainable params: 0
_________________________________________________________________

Here I build a random x,y sample:

x = np.random.randint(69, size=(500,20))
y = np.random.uniform(0, 1, size=(500,))

When I predict x I get the correct output shape

preds = model.predict(x)
print(preds.shape == y.shape)

When I fit the model I get the following error:

model.fit(x, y, batch_size=50, verbose=1)

ValueError: Error when checking target: expected sum_embeddings to have 1 dimensions, but got array with shape (500, 1)

It feels like I am missing something really simple. Any suggestions would be greatly appreciated

Upvotes: 1

Views: 694

Answers (1)

thushv89
thushv89

Reputation: 11333

Yes, it is a few simple issues with your code. The output of your model needs to have at least rank 2 (in this case (None,1)) (My 2 cents is that optimizer complains when it's not). This is done using keepdims=True. Then you have to add one dimension to y too.

inputs = Input(shape=(20,), name="inputs")
embedding = Embedding(69, 100, name="embeddings")(inputs)
out = Reshape((2000,), name='reshape_embeddings')(embedding)
out = Lambda(lambda x: K.sum(x, axis=1, keepdims=True), name='sum_embeddings')(out)
model = Model(inputs, out)
model.compile('adam', 'mean_squared_error')
print(model.summary())

x = np.random.randint(69, size=(500,20))
y = np.random.uniform(0, 1, size=(500,1))

preds = model.predict(x)
print(preds.shape == y.shape)

model.fit(x, y, batch_size=50, verbose=1)

Upvotes: 1

Related Questions