Keras: How to inspect shapes of tensors

[Falkenjack]

I'm troubleshooting a number of different problems in an experimental network implemented in keras with a theano backend. I'm relatively new to keras and theano (I did some small experiments a few months ago and are returning now, having forgotten most of what I learned then). Lots of my problems would be more transparent if I could inspect the shapes of tensors but try as I might I never find a human readable representation of shapes.

For instance, I have some issues of the "Exception: Input 0 is incompatible with layer c_out: expected ndim=2, found ndim=3" type. While I know that this is a problem with the number of dimensions rather than the size of them, knowing the sizes of the dimensions would help me understand why there is a mismatch and what each dimension contains. Basically, I'm saying that my mental model of tensors and how computations flow through keras/theano is still pretty fuzzy and I need all the help I can get :)

If I print a TensorVariable.shape attribute I get the string "Shape.0", this seems to be an lvector of subtensors but I can't seem to access the values. There is supposed to be a function keras.backend.int_shape(x) which returns the shape as a tuple of integers and/or Nones but when I try it I get "AttributeError: module 'keras.backend' has no attribute 'int_shape'". The input to the c_out layer comes from a merge layer called merge_c, but when I call merge_c.output_shape or merge_c.get_output_shape_at(0) I get "AttributeError: 'TensorVariable' object has no attribute 'get_output_shape_at'".

Can anyone recommend a way to access a human readable representation of tensor shapes?

[bstriner]

@Falkenjack Use keras.backend.shape(x) to get the shape of a tensor or use model.summary() to print the shapes of all of the layers in your model. Please ask usage questions on stackoverflow, slack, or the google group.

Cheers,
Ben

[mrgloom]

Looks like there are several methods? What is the difference?

class_weights_list = [1.0]
class_weights_numpy = np.array(class_weights_list)
class_weights_keras = K.variable(class_weights_numpy)
print('class_weights_keras.get_shape', class_weights_keras.get_shape)
print('K.shape(class_weights_keras)', K.shape(class_weights_keras))

('class_weights_keras.get_shape', <bound method Variable.get_shape of <tf.Variable 'Variable:0' shape=(1,) dtype=float32_ref>>)
('K.shape(class_weights_keras)', <tf.Tensor 'Shape:0' shape=(1,) dtype=int32>)

[Neo-X]

I found int_shape to be much more readable.

keras.backend.int_shape(network)

[chandu8542]

Note the two ways of accessing shape via shape and int_shape. The first one gives the rank and the second one gives the dimension of the tensor.

k=Input(shape=(29,))
import keras.backend as bk
print(bk.shape(k))
Tensor("Shape_7:0", shape=(2,), dtype=int32)

print(bk.int_shape(k))
(None, 29)

[B-Yassine]

You can get your tensor shape as follow:

H, W, n_ch = tensor.shape.as_list()[1:]