DepthwiseConv2D Layers cannot be clustered or sparsely pruned

Inizia eseguendo la riproduzione MNIST fornita con le API cluster_weights e sparse pruning di TensorFlow Model Optimization. Traccia il modo in cui vengono gestiti i kernel di DepthwiseConv2D, quindi verifica che il clustering produca tre pesi unici e che il pruning m-by-n produca il pattern di sparsity previsto.

When attempting to sparsely prune or cluster DepthwiseConv2D Layers it appears that no clustering or sparse pruning actually occurs.

System information

TensorFlow version (installed from source or binary): 2.16.1/2.15.1

TensorFlow Model Optimization version (installed from source or binary): 0.8.0/0.7.5

Python version: 3.9.18

Describe the expected behavior

I should see that the kernel for the Depthwise layer should have 3 unique weights, not 9.

Describe the current behavior

The Depthwise layer has 9 unique weights

The same issue occurs for sparse m by n pruning, the weights are not pruned correctly in an m by n manner

Code to reproduce the issue

import tensorflow as tf
import tensorflow_model_optimization as tfmot
from tensorflow import keras

import numpy as np

(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.mnist.load_data()
train_images = train_images / 255.0
test_images = test_images / 255.0

model = keras.Sequential([
keras.layers.InputLayer(input_shape=(28, 28)),
keras.layers.Reshape(target_shape=(28, 28, 1)),
keras.layers.DepthwiseConv2D(kernel_size=(3, 3), activation=tf.nn.relu),
keras.layers.MaxPooling2D(pool_size=(2, 2)),
keras.layers.Flatten(),
keras.layers.Dense(10)
])

model.compile(optimizer='adam',
loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])

model.fit(
train_images,
train_labels,
validation_split=0.1,
epochs=10
)

cluster_weights = tfmot.clustering.keras.cluster_weights
CentroidInitialization = tfmot.clustering.keras.CentroidInitialization

clustering_params = {
'number_of_clusters': 3,
'cluster_centroids_init': CentroidInitialization.LINEAR
}

clustered_model = cluster_weights(model, **clustering_params)

opt = keras.optimizers.Adam(learning_rate=1e-5)

clustered_model.compile(
loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer=opt,
metrics=['accuracy'])
clustered_model.fit(
train_images,
train_labels,
batch_size=500,
epochs=1,
validation_split=0.1)

for layer in model.layers:
for weight in layer.weights:
print(weight.name)
print(len(np.unique(weight)))