6cb84f4b742861e823015cac18535efa468b0d81,autokeras/layer_transformer.py,,conv_dense_to_wider_layer,#Any#Any#Any#,128

Before Change

    new_pre_layer.set_weights((student_w1, student_b1))
    new_next_layer = Dense(n_units)
    new_next_layer.build((None, n_units))
    new_next_layer.set_weights((student_w2.flatten(), teacher_b2))
    return new_pre_layer, new_next_layer

After Change

    pre_filter_shape = pre_layer.kernel_size
    conv_func = get_conv_layer_func(len(pre_filter_shape))
    n_pre_filters = pre_layer.filters
    n_units = next_layer.units

    teacher_w1 = pre_layer.get_weights()[0]
    teacher_w2 = next_layer.get_weights()[0]
    n_total_weights = int(reduce(mul, teacher_w2.shape))
    teacher_w2 = teacher_w2.reshape(int(n_total_weights / n_pre_filters / n_units), n_pre_filters, n_units)
    teacher_b1 = pre_layer.get_weights()[1]
    teacher_b2 = next_layer.get_weights()[1]
    rand = np.random.randint(n_pre_filters, size=n_add_filters)
    replication_factor = np.bincount(rand)
    student_w1 = teacher_w1.copy()
    student_w2 = teacher_w2.copy()
    student_b1 = teacher_b1.copy()
    // target layer update (i)
    for i in range(len(rand)):
        teacher_index = rand[i]
        new_weight = teacher_w1[:, :, :, teacher_index]
        new_weight = new_weight[:, :, :, np.newaxis]
        student_w1 = np.concatenate((student_w1, new_weight), axis=3)
        student_b1 = np.append(student_b1, teacher_b1[teacher_index])
    // next layer update (i+1)
    for i in range(len(rand)):
        teacher_index = rand[i]
        factor = replication_factor[teacher_index] + 1
        new_weight = teacher_w2[:, teacher_index, :] * (1. / factor)
        new_weight_re = new_weight[:, np.newaxis, :]
        student_w2 = np.concatenate((student_w2, new_weight_re), axis=1)
        student_w2[:, teacher_index, :] = new_weight

    new_pre_layer = conv_func(n_pre_filters + n_add_filters,
                              kernel_size=pre_filter_shape,
                              activation="relu",
                              padding="same",
                              input_shape=pre_layer.input_shape[1:])
    new_pre_layer.build((None,) * (len(pre_filter_shape) + 1) + (pre_layer.input_shape[-1],))
    new_pre_layer.set_weights((student_w1, student_b1))
    new_next_layer = Dense(n_units, activation="relu")
    n_new_total_weights = int(reduce(mul, student_w2.shape))
    input_dim = int(n_new_total_weights / n_units)
    new_next_layer.build((None, input_dim))
    new_next_layer.set_weights((student_w2.reshape(input_dim, n_units), teacher_b2))
    return new_pre_layer, new_next_layer

Instances