visualise features for tensorflow

gistfile1.py

# feature - the tensor you want to visualise in shape [num_examples, dimX, dimY, channels]. If dims are provided, the feature tensor can be flattened. Eg shape: [20, 7, 7, 16]
# grid - the dimensions of a grid on which the features will be displayed. Expects a touple (x,y) where x * y have to equal to channels of the feature. Eg, for 16 channels, grid=[4,4]
# dims (optional) - dimensions of each feature. Eg. for feature [20, 7, 7, 16], dims=[7,7] (although in non-flattened tensor the dims can be ommited)
# max_outputs - number of examples to draw

def visualise_feature(feature, grid, dims=None, max_outputs=10):
    with tf.name_scope('Visualize_filters') as scope:
        original_shape = feature.get_shape().as_list()
        original_len = len(original_shape)

        if (original_len == 4 and dims is None):
            dims = [original_shape[1], original_shape[2]]

        dimA = dims[0]
        dimB = dims[1]

        if (original_len < 4):
            assert dims is not None
            channels = original_shape[2] if original_len == 3 else 1
            feature = tf.reshape(feature, (-1, dimA, dimB, channels))

        V = tf.slice(feature,(0,0,0,0),(-1,-1,-1,-1)) #V[0,...]
        V = tf.reshape(V,(-1, dimA,dimB,1))

        cx=grid[0]
        cy=grid[1]

        V = tf.reshape(V,(-1, dimA,dimB,cy,cx, 1))

        # Reorder so the channels are in the first dimension, x and y follow.
        V = tf.transpose(V, (0, 3, 1, 4, 2, 5))

        V = tf.reshape(V,(-1,cy*dimB,cx*dimA,1))
        
        tf.summary.image("Visualize_kernels", V, max_outputs=max_outputs)