TFLite file larger than expected

[andynader]

Hi, When I train the following architecture:

kwargs = {
    "input_quantizer": "ste_sign",
    "kernel_quantizer": "ste_sign",
    "kernel_constraint": "weight_clip",
}



binary_net = tf.keras.Sequential(
    [
        tf.keras.layers.Input(shape=(x_train.shape[1], x_train.shape[2])),
        lq.layers.QuantConv1D(32, (3,), padding="same", use_bias=False, **kwargs),
        lq.layers.QuantConv1D(32, (3,), padding="same",  use_bias=False, **kwargs),
        tf.keras.layers.MaxPooling1D(pool_size=2),
        lq.layers.QuantConv1D(64, (3,), padding="same",  use_bias=False, **kwargs),
        lq.layers.QuantConv1D(64, (3,), padding="same", use_bias=False, **kwargs),
        tf.keras.layers.MaxPooling1D(pool_size=4),
        lq.layers.QuantConv1D(128, (3,), padding="same",  use_bias=False, **kwargs),
        lq.layers.QuantConv1D(128, (3,), padding="same",  use_bias=False, **kwargs),
        tf.keras.layers.MaxPooling1D(pool_size=6),
        lq.layers.QuantConv1D(256, (3,), padding="same",  use_bias=False, **kwargs),
        lq.layers.QuantConv1D(256, (3,), padding="same", use_bias=False, **kwargs),
        tf.keras.layers.MaxPooling1D(pool_size=6),
        lq.layers.QuantConv1D(256, (3,), padding="same",  use_bias=False, **kwargs),
        lq.layers.QuantConv1D(256, (3,), padding="same", use_bias=False, **kwargs),
        tf.keras.layers.UpSampling1D(size=6),
        lq.layers.QuantConv1D(128, (3,), padding="same",  use_bias=False, **kwargs),
        lq.layers.QuantConv1D(128, (3,), padding="same",  use_bias=False, **kwargs),
        tf.keras.layers.UpSampling1D(size=6),
        lq.layers.QuantConv1D(64, (3,), padding="same",  use_bias=False, **kwargs),
        lq.layers.QuantConv1D(64, (3,), padding="same",  use_bias=False, **kwargs),
        tf.keras.layers.UpSampling1D(size=4),
        lq.layers.QuantConv1D(32, (3,), padding="same", use_bias=False, **kwargs),
        lq.layers.QuantConv1D(32, (3,), padding="same", use_bias=False, **kwargs),
        tf.keras.layers.UpSampling1D(size=2),
        lq.layers.QuantConv1D(3, (3,), padding="same",  use_bias=False, **kwargs),
        lq.layers.QuantConv1D(3, (3,), padding="same",  use_bias=False, **kwargs),
        tf.keras.layers.Flatten(),
        lq.layers.QuantDense(288, use_bias=False, **kwargs)
    ]
)

lq.models.summary() suggests that the model should be around 150KB, yet when I convert it to a TFLite file using lce.convert_keras_model(), the file ends up being 5MB, which is the size of the floating point equivalent of the network. In addition, when I benchmark this model on a Pi 4, it ends up being equivalent in inference time to a full precision network. Any idea why this is happening?

[Tombana]

Hi @andynader,

Our tflite support currently includes 2-dimensional binary convolutions but not 1-dimensional ones. It should work if you use a 2D convolution but with height (or width) set to 1 and filter size (3,1) (or (1,3)). The UpSampling1D layers become UpSampling2D with size=(1,x).

[andynader]

Yup, that worked, thank you so much!