Issues Test Cases on Customized Backend

[administrator2992]

Hello, i have issues when trying test cases on customized backend. actually i copy tflitecpu backend for my customized backend but cause i want to implement in my laptop (using Linux Environtment). so i edited the shell commands and profiler commands as needed.

MyProfiler :

# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
import os
from nn_meter.builder.backends import BaseProfiler

class TFLiteProfiler(BaseProfiler):
    use_gpu = None

    def __init__(self, dst_kernel_path, benchmark_model_path, graph_path='', dst_graph_path='', num_threads=1, num_runs=50, warm_ups=10):
        """
        @params:
        graph_path: graph file. path on host server
        dst_graph_path: graph file. path on device
        kernel_path: dest kernel output file. path on device
        benchmark_model_path: path to benchmark_model
        """
        self._graph_path = graph_path
        self._dst_graph_path = dst_graph_path
        self._dst_kernel_path = dst_kernel_path
        self._benchmark_model_path = benchmark_model_path
        self._num_threads = num_threads
        self._num_runs = num_runs
        self._warm_ups = warm_ups

    def profile(self, graph_path, preserve = False, clean = True, close_xnnpack = False, **kwargs):
        """
        @params:
        preserve: tflite file exists in remote dir. No need to push it again.
        clean: remove tflite file after running.
        """
        model_name = os.path.basename(graph_path)
        remote_graph_path = os.path.join(self._dst_graph_path, model_name)
        kernel_cmd = f'--kernel_path={self._dst_kernel_path}' if self._dst_kernel_path else ''
        close_xnnpack_cmd = f'--use_xnnpack=false' if close_xnnpack else ''

        try:
            kernel_cmd = f'--kernel_path={self._dst_kernel_path}' if self._dst_kernel_path else ''
            close_xnnpack_cmd = f'--use_xnnpack=false' if close_xnnpack else ''
            res = os.system(f' {self._benchmark_model_path} {kernel_cmd} {close_xnnpack_cmd}' \
                               f' --num_threads={self._num_threads}' \
                               f' --num_runs={self._num_runs}' \
                               f' --warmup_runs={self._warm_ups}' \
                               f' --graph={remote_graph_path}' \
                               f' --enable_op_profiling=true' \
                               f' --use_gpu={"true" if self.use_gpu else "false"}')
        except:
            raise
        finally:
            if clean:
                os.system(f"rm {remote_graph_path}")

        return res

MyBackend :

# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
import os
import shutil
import logging
from nn_meter.builder.backends import BaseBackend
from nn_meter.utils.path import get_filename_without_ext
logging = logging.getLogger("nn-Meter")

class TFLiteBackend(BaseBackend):
    parser_class = None
    profiler_class = None

    def update_configs(self):
        """update the config parameters for TFLite platform
        """
        super().update_configs()
        self.profiler_kwargs.update({
            'dst_graph_path': self.configs['REMOTE_MODEL_DIR'],
            'benchmark_model_path': self.configs['BENCHMARK_MODEL_PATH'],
            'dst_kernel_path': self.configs['KERNEL_PATH']
        })

    def convert_model(self, model_path, save_path, input_shape=None):
        """convert the Keras model instance to ``.tflite`` and return model path
        """
        import tensorflow as tf
        model_name = get_filename_without_ext(model_path)
        model = tf.keras.models.load_model(model_path)
        converter = tf.lite.TFLiteConverter.from_keras_model(model)
        tflite_model = converter.convert()
        converted_model = os.path.join(save_path, model_name + '.tflite')
        open(converted_model, 'wb').write(tflite_model)
        shutil.rmtree(model_path)
        return converted_model
    
    def test_connection(self):
        """check the status of backend interface connection
        """
        ...
        logging.keyinfo("hello TFLitex64 backend !")

my backend register config :

builtin_name: TFLitex64
package_location: /home/nn-meter/backends/tflitex64
class_module: cpu
class_name: TFLiteCPUBackend
defaultConfigFile: /home/nn-meter/backends/tflitex64/backend_tflitex64_config.yaml

my backend default config :

REMOTE_MODEL_DIR: /home/nn-meter/models
BENCHMARK_MODEL_PATH: /tensorflow_src/bazel-bin/tensorflow/lite/tools/benchmark/benchmark_model
KERNEL_PATH:

and i used cpu.py in this repo

my custom backend is able to be registered in my laptop but i have issue when i try test fusion rules. i got profile_error.log that error 'int' object has no attribute 'splitlines'. so i try to debug content above line 27 cpu.py because in line 27 call splitlines and the result is 0 (it's integer) so i think that is the issue of error 'int' object has no attribute 'splitlines' but i don't know what is wrong in my configuration or code.

And also i got the issue WARNING:tensorflow:No training configuration found in save file, so the model was *not* compiled. Compile it manually. when i try test fusion rules. which configuration and where should I save the configuration file ?

and i have a question again, does REMOTE_MODEL_DIR must to have any model when running profile_model ?

Thanks

[administrator2992]

i got the solving for my case,

so the issue is from my profiler when running benchmark. i change from os.system to subsprocess,check_output because os.system cannot return shell output. that the code :

res = subprocess.check_output(f' {self._benchmark_model_path} {kernel_cmd} {close_xnnpack_cmd}' \
                               f' --num_threads={self._num_threads}' \
                               f' --num_runs={self._num_runs}' \
                               f' --warmup_runs={self._warm_ups}' \
                               f' --graph={remote_graph_path}' \
                               f' --enable_op_profiling=true' \
                               f' --use_gpu={"true" if self.use_gpu else "false"}', shell=True)

the return must to decode to string as well

return res.decode('utf-8')