Some problems when I am using the vela compiler (Input(s) and Output tensors must not be dynamic) about tflite-micro HOT 7 CLOSED

So those warnings pop up when you are trying to convert your keras model to tf lite? Can you please share complete code snippet?

from tflite-micro.

So those warnings pop up when you are trying to convert your keras model to tf lite? Can you please share complete code snippet?

Thanks for your reply!

I don't see this warning when i am trying to convert my keras model to tf lite.

this is my code (ref : https://www.tensorflow.org/lite/performance/post_training_integer_quant):

import tensorflow as tf
import numpy as np
import pathlib

load model

model_path = "E:/tensorflow_lite_example/MNINST_form_tensorflow/save_model/"

model=tf.keras.models.load_model(model_path) 

setting representative_data

(train_images, train_labels),(test_images,test_labels) = mnist.load_data()
train_images = train_images.astype(np.float32)/255.0

def representative_data_gen():
  for input_value in tf.data.Dataset.from_tensor_slices(train_images).batch(1).take(100):
    yield [input_value]

converter to tflite and quantization

converter = tf.lite.TFLiteConverter.from_keras_model(model)

converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.representative_dataset = representative_data_gen
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
converter.inference_input_type = tf.int8
converter.inference_output_type = tf.int8

tflite_model_quant_int8 = converter.convert()

save tflite

tflite_models_dir = pathlib.Path("./tmp/tflie_test/")
tflite_models_dir.mkdir(exist_ok=True, parents=True)

tflite_model_quant_file = tflite_models_dir/"mnist_model_quant_int8.tflite"
tflite_model_quant_file.write_bytes(tflite_model_quant_int8)

from tflite-micro.

Since your goal is to use the ethos-u55, I would recommend starting with one of the TfLite Micro examples, such as the hello_world example.

I'm going to close the current issue, but please feel free to open a new one in the tflite-micro repository for any issues that you run into with the hello_world example.

Also, tagging @freddan80 in case there are any additional ethos-u55 specific examples that might be more relevant here.

from tflite-micro.

Thx for looping me in. We will check it out. Tagging @jenselofsson and @mansnils .

from tflite-micro.

@dexsr415 - As @advaitjain suggested, if you want to get started with Ethos-U55 and Vela in Tflite-micro you could have a look here:
https://github.com/tensorflow/tflite-micro/tree/main/tensorflow/lite/micro/cortex_m_corstone_300
https://github.com/tensorflow/tflite-micro/tree/main/tensorflow/lite/micro/examples/network_tester
https://github.com/tensorflow/tflite-micro/tree/main/tensorflow/lite/micro/kernels/ethos_u

Running the network example will download Vela and convert the person detect int8 model.

In case you run into specific problems with Vela (outside the scope of the network example and benchmarks where Vela is used in Tflite-micro), you could raise an issue in https://community.arm.com/developer/tools-software/oss-platforms/f/machine-learning-forum/.
You can also use support-ml [email protected] for your issues if you are Arm License customer.

from tflite-micro.

@dexsr415 - As @advaitjain suggested, if you want to get started with Ethos-U55 and Vela in Tflite-micro you could have a look here:
https://github.com/tensorflow/tflite-micro/tree/main/tensorflow/lite/micro/cortex_m_corstone_300
https://github.com/tensorflow/tflite-micro/tree/main/tensorflow/lite/micro/examples/network_tester
https://github.com/tensorflow/tflite-micro/tree/main/tensorflow/lite/micro/kernels/ethos_u

Running the network example will download Vela and convert the person detect int8 model.

In case you run into specific problems with Vela (outside the scope of the network example and benchmarks where Vela is used in Tflite-micro), you could raise an issue in https://community.arm.com/developer/tools-software/oss-platforms/f/machine-learning-forum/.
You can also use support-ml [email protected] for your issues if you are Arm License customer.

Thank you for your reply

I will check the information you gave.

Thank you so much

from tflite-micro.

I found that my question is similar to tensorflow/tensorflow#45090 (comment)

This problem seems to be related to dynamic size in keras.

from tflite-micro.