Hi everyone. I've built a classification model using the Larq framework. After training, I obtained an h5 format model file and converted it to tflite format using the Larq Compute Engine. I observed the model using Netron, with the h5 model on the left and the tflite model on the right.
I noticed that after the reshape operation in the tflite model, there are consecutive LceQuantize and LceDequantize operators. As marked by the red box, I think the consecutive use of these two operators seems meaningless. Why does this happen during the conversion process?
I am using Larq version 0.13.1, Larq Compute Engine version 0.11.1, TensorFlow version 2.3.0, and Python 3.8.
My model code is as follows:
kwargs = dict(input_quantizer="ste_sign",
kernel_quantizer="ste_sign",
kernel_constraint="weight_clip",
use_bias=False)
vww_model = tf.keras.models.Sequential([
# In the first layer we only quantize the weights and not the input
lq.layers.QuantConv2D(32, 3,
kernel_quantizer="ste_sign",
kernel_constraint="weight_clip",
use_bias=False,
input_shape=(96, 96, 3),
strides=(2, 2),
padding="same",
pad_values=1),
tf.keras.layers.BatchNormalization(momentum=0.999, scale=False),
lq.layers.QuantConv2D(32, 3, padding="same", pad_values=1, strides=(2, 2), **kwargs),
tf.keras.layers.BatchNormalization(momentum=0.999, scale=False),
lq.layers.QuantConv2D(32, 3, padding="same", pad_values=1, **kwargs),
tf.keras.layers.BatchNormalization(momentum=0.999, scale=False),
lq.layers.QuantConv2D(64, 3, padding="same", pad_values=1, strides=(2, 2), **kwargs),
tf.keras.layers.BatchNormalization(momentum=0.999, scale=False),
lq.layers.QuantConv2D(64, 3, padding="same", pad_values=1, **kwargs),
tf.keras.layers.BatchNormalization(momentum=0.999, scale=False),
lq.layers.QuantConv2D(128, 3, padding="same", pad_values=1, strides=(2, 2), **kwargs),
tf.keras.layers.BatchNormalization(momentum=0.999, scale=False),
lq.layers.QuantConv2D(128, 3, padding="same", pad_values=1, **kwargs),
tf.keras.layers.BatchNormalization(momentum=0.999, scale=False),
lq.layers.QuantConv2D(256, 3, padding="same", pad_values=1, strides=(2, 2), **kwargs),
tf.keras.layers.BatchNormalization(momentum=0.999, scale=False),
lq.layers.QuantConv2D(256, 3, padding="same", pad_values=1, **kwargs),
tf.keras.layers.BatchNormalization(momentum=0.999, scale=False),
tf.keras.layers.MaxPool2D(pool_size=(3, 3), strides=(1, 1)),
tf.keras.layers.Flatten(),
lq.layers.QuantDense(128, **kwargs),
tf.keras.layers.BatchNormalization(momentum=0.999, scale=False),
lq.layers.QuantDense(64, **kwargs),
tf.keras.layers.BatchNormalization(momentum=0.999, scale=False),
lq.layers.QuantDense(2, **kwargs),
tf.keras.layers.BatchNormalization(momentum=0.999, scale=False),
tf.keras.layers.Activation("softmax"),
])Conversion code is as follows:
model_name_in = "vww1_full.h5"
model_name_out = "vww1_binary.tflite"
custom_objects = {
"QuantConv2D": layers.QuantConv2D,
}
model = load_model(model_name_in, custom_objects=custom_objects)
lq.models.summary(model)
# Convert our Keras model to a TFLite flatbuffer file
with open(model_name_out, "wb") as flatbuffer_file:
flatbuffer_bytes = lce.convert_keras_model(model)
flatbuffer_file.write(flatbuffer_bytes)I would greatly appreciate it if anyone could offer a solution.
Hi everyone. I've built a classification model using the Larq framework. After training, I obtained an h5 format model file and converted it to tflite format using the Larq Compute Engine. I observed the model using Netron, with the h5 model on the left and the tflite model on the right.
I noticed that after the reshape operation in the tflite model, there are consecutive LceQuantize and LceDequantize operators. As marked by the red box, I think the consecutive use of these two operators seems meaningless. Why does this happen during the conversion process?
I am using Larq version 0.13.1, Larq Compute Engine version 0.11.1, TensorFlow version 2.3.0, and Python 3.8.
My model code is as follows:
Conversion code is as follows:
I would greatly appreciate it if anyone could offer a solution.