CODEWITHSUNDEEP
pythonneural-networktensorflowconvolutionkeras
KFox
KFox

Reputation: 1266

Keras convolution filter size issue

I'm very new to Keras, and in my early experiments I've been encountering an error when I create convolutional layers. When I run this code (found here; "Shared vision model" section)

from keras.models import *
from keras.layers import *
from keras.layers.convolutional import *
from keras.layers.pooling import *

digit_input = Input(shape=(1, 27, 27))
x = Convolution2D(64, 3, 3)(digit_input)
x = MaxPooling2D((2, 2))(x)

mdl = Model(digit_input, x)

I get the following error (traceback points to the Convolution2D line):

ValueError: Filter must not be larger than the input: Filter: (3, 3) Input: (1, 27)

I'd assume I was misusing the interface--especially since no one else seems to have this particular problem--but I pulled the code directly from the Keras docs. The problem persists when I use the model.add() syntax. Interestingly though, when I shrink the filter size to 1x1 or increase the number of input channels to 3, the error goes away. I have tried manually setting dim_ordering='tf' for the convolution and my keras.json file reads:

{
"image_dim_ordering": "tf",
"epsilon": 1e-07,
"floatx": "float32",
"backend": "tensorflow"
}

Keras looks like a great tool, but I'm completely unable to use it so far.

P.S. I am using Keras 1.1.0 and Tensorflow 0.10.

Upvotes: 1

Views: 1519

Answers (3)

Naomi Fridman
Naomi Fridman

Reputation: 2144

Its a channel first-last problem. The default is channel last, and if your image will be in the shape (27,27,1) it will work.

digit_input = Input(shape=(27, 27, 1))
x = Convolution2D(64, 3, 3)(digit_input)  
x = MaxPooling2D((2, 2))(x)
mdl = Model(digit_input, x)
mdl.summary()

You will get:

Model: "model_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_3 (InputLayer)         [(None, 27, 27, 1)]       0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 9, 9, 64)          640       
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 4, 4, 64)          0         
=================================================================
Total params: 640
Trainable params: 640
Non-trainable params: 0

Upvotes: 0

clab
clab

Reputation: 76

Try this

digit_input = Input(shape=(27, 27, 1))
x = Convolution2D(64, 3, 3, dim_ordering='tf')(digit_input)

Upvotes: 0

Dwight Crow
Dwight Crow

Reputation: 368

'image_dim_ordering' : 'tf' is setting the input's third dimension as the channel input to the 2D Convolution: ie you've currently put a 1x27 input with 27 channels in to be convolved with a 3x3 kernel and no padding.

You should be able to switch 'image_dim_ordering' to 'th' and have this code immediately work. I think 'th' used to be the default and you've unfortunately hit an obsolete tutorial. Here's the Keras documentation from Convolution2D - it's there but not particularly prominent:

dim_ordering: 'th' or 'tf'. In 'th' mode, the channels dimension (the depth) is at index 1, in 'tf' mode is it at index 3. It defaults to the image_dim_ordering value found in your Keras config file at ~/.keras/keras.json. If you never set it, then it will be "tf".

Hope that helps!

Upvotes: 2

Related Questions