Original vggish vs this..

[rohan1561]

Hey doesn't the original tf implementation have only four convolution layers and two fc layers? this one has 6, 3...why the difference? How could the embeddings be identical then?

[KAISER1997]

@rohan1561 Could you find something?. I had a similar doubt.

[eatsleepraverepeat]

@rohan1561 @KAISER1997

this implementation is equal to the one placed In tensorflow/models, and there are, indeed - 6 convolution layers and 3 fully connected ones

here's a list of layers from tf 1.x vggish checkpoint, check the scopes

[
    ('vggish/conv1/biases', [64]),
    ('vggish/conv1/weights', [3, 3, 1, 64]),
    ('vggish/conv2/biases', [128]),
    ('vggish/conv2/weights', [3, 3, 64, 128]),
    ('vggish/conv3/conv3_1/biases', [256]),
    ('vggish/conv3/conv3_1/weights', [3, 3, 128, 256]),
    ('vggish/conv3/conv3_2/biases', [256]),
    ('vggish/conv3/conv3_2/weights', [3, 3, 256, 256]),
    ('vggish/conv4/conv4_1/biases', [512]),
    ('vggish/conv4/conv4_1/weights', [3, 3, 256, 512]),
    ('vggish/conv4/conv4_2/biases', [512]),
    ('vggish/conv4/conv4_2/weights', [3, 3, 512, 512]),
    ('vggish/fc1/fc1_1/biases', [4096]),
    ('vggish/fc1/fc1_1/weights', [12288, 4096]),
    ('vggish/fc1/fc1_2/biases', [4096]),
    ('vggish/fc1/fc1_2/weights', [4096, 4096]),
    ('vggish/fc2/biases', [128]),
    ('vggish/fc2/weights', [4096, 128]),
    ('vggish/logits/biases', [4923]),
    ('vggish/logits/weights', [128, 4923])
]