mehrdadh commented on code in PR #13978: URL: https://github.com/apache/tvm/pull/13978#discussion_r1107466692
########## tests/micro/common/test_mlperftiny.py: ########## @@ -0,0 +1,301 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +import os +from urllib.parse import urlparse + +import pytest +import tensorflow as tf +import numpy as np +import tarfile +import tempfile +import pathlib + +import tvm +import tvm.testing +from tvm import relay +from tvm.relay.backend import Executor, Runtime +from tvm.micro.project_api import server +from tvm.contrib.download import download_testdata +from tvm.micro import export_model_library_format +from tvm.micro.model_library_format import generate_c_interface_header +from tvm.micro.testing import create_aot_session, predict_labels_aot +from tvm.micro.testing.utils import ( + create_header_file, + mlf_extract_workspace_size_bytes, +) + +MLPERF_TINY_MODELS = { + "kws": { + "name": "Keyword Spotting", + "index": 1, + "url": "https://github.com/mlcommons/tiny/raw/bceb91c5ad2e2deb295547d81505721d3a87d578/benchmark/training/keyword_spotting/trained_models/kws_ref_model.tflite";, + "sample": "https://github.com/tlc-pack/web-data/raw/main/testdata/microTVM/data/keyword_spotting_int8_6.pyc.npy";, + }, + "vww": { + "name": "Visual Wake Words", + "index": 2, + "url": "https://github.com/mlcommons/tiny/raw/bceb91c5ad2e2deb295547d81505721d3a87d578/benchmark/training/visual_wake_words/trained_models/vww_96_int8.tflite";, + "sample": "https://github.com/tlc-pack/web-data/raw/main/testdata/microTVM/data/visual_wake_word_int8_1.npy";, + }, + # Note: The reason we use quantized model with float32 I/O is + # that TVM does not handle the int8 I/O correctly and accuracy + # would drop significantly. + "ad": { + "name": "Anomaly Detection", + "index": 3, + "url": "https://github.com/mlcommons/tiny/raw/bceb91c5ad2e2deb295547d81505721d3a87d578/benchmark/training/anomaly_detection/trained_models/ToyCar/baseline_tf23/model/model_ToyCar_quant_fullint_micro.tflite";, + "sample": "https://github.com/tlc-pack/web-data/raw/main/testdata/microTVM/data/anomaly_detection_normal_id_01.npy";, + # This model takes in a (1, 640) vector, so it must be called 40 times - + # once for each time slice. + }, + "ic": { + "name": "Image Classification", + "index": 4, + "url": "https://github.com/mlcommons/tiny/raw/bceb91c5ad2e2deb295547d81505721d3a87d578/benchmark/training/image_classification/trained_models/pretrainedResnet_quant.tflite";, + "sample": "https://github.com/tlc-pack/web-data/raw/main/testdata/microTVM/data/image_classification_int8_0.npy";, + }, +} + + +def mlperftiny_get_module(model_name: str): + model_url = MLPERF_TINY_MODELS[model_name]["url"] + url = urlparse(model_url) + model_path = download_testdata(model_url, os.path.basename(url.path), module="model") + + tflite_model_buf = open(model_path, "rb").read() + try: + import tflite + + tflite_model = tflite.Model.GetRootAsModel(tflite_model_buf, 0) + except AttributeError: + import tflite.Model + + tflite_model = tflite.Model.Model.GetRootAsModel(tflite_model_buf, 0) + + interpreter = tf.lite.Interpreter(model_path=str(model_path)) + interpreter.allocate_tensors() + input_details = interpreter.get_input_details() + output_details = interpreter.get_output_details() + + model_info = { + "input_name": input_details[0]["name"], + "input_shape": tuple(input_details[0]["shape"]), + "input_dtype": np.dtype(input_details[0]["dtype"]).name, + "output_name": output_details[0]["name"], + "output_shape": tuple(output_details[0]["shape"]), + "output_dtype": np.dtype(output_details[0]["dtype"]).name, + } + + if model_name != "ad": + model_info["quant_output_scale"] = output_details[0]["quantization_parameters"]["scales"][0] + model_info["quant_output_zero_point"] = output_details[0]["quantization_parameters"][ + "zero_points" + ][0] + + relay_mod, params = relay.frontend.from_tflite( + tflite_model, + shape_dict={model_info["input_name"]: model_info["input_shape"]}, + dtype_dict={model_info["input_name"]: model_info["input_dtype"]}, + ) + return relay_mod, params, model_info + + +def get_test_data(model_name: str) -> list: + sample_url = MLPERF_TINY_MODELS[model_name]["sample"] + url = urlparse(sample_url) + sample_path = download_testdata(sample_url, os.path.basename(url.path), module="data") + return [np.load(sample_path)] + + +def predict_ad_labels_aot(session, aot_executor, input_data, runs_per_sample=1): + """A special version of tvm/micro/testing/evaluation.py's predict_labels_aot. + The runtime returned for each sample is the median of the runtimes for all slices + in that sample.""" + + assert runs_per_sample > 0 + assert aot_executor.get_num_inputs() == 1 + assert aot_executor.get_num_outputs() == 1 + + sample_counter = 0 + for sample in input_data: + output_fp32 = np.empty_like(sample) + slice_runtimes = [] + + for i, time_slice in enumerate(sample): + aot_executor.get_input(0).copyfrom(time_slice.reshape((1, 640))) + result = aot_executor.module.time_evaluator( + "run", session.device, number=runs_per_sample + )() + slice_runtimes.append(result.mean) + output_fp32[i, :] = aot_executor.get_output(0).numpy() + + sample_counter += 1 + errors = np.mean(np.square(sample - output_fp32), axis=1) + yield np.mean(errors), np.median(slice_runtimes) + + [email protected]("model_name", ["kws", "vww", "ad", "ic"]) [email protected]("project_type", ["host_driven", "mlperftiny"]) [email protected]_micro [email protected]_boards( + ["mps2_an521", "mps3_an547", "stm32f746g_disco", "nucleo_f746zg", "nrf5340dk_nrf5340_cpuapp"] +) +def test_mlperftiny_models(platform, board, workspace_dir, serial_number, model_name, project_type): + """MLPerfTiny models test. + Testing MLPerfTiny models using host_driven project. In this case one input sample is used + to verify the end to end execution. Accuracy is not checked in this test. + + Also, this test builds each model in standalone mode that can be used with EEMBC runner. + """ + if platform != "zephyr": + pytest.skip(reason="Other platforms are not supported yet.") + + use_cmsis_nn = False + relay_mod, params, model_info = mlperftiny_get_module(model_name) + target = tvm.micro.testing.get_target(platform, board) + project_options = {"config_main_stack_size": 4000, "serial_number": serial_number} + + if use_cmsis_nn: + project_options["cmsis_path"] = os.getenv("CMSIS_PATH") + + if model_name == "ad": + predictor = predict_ad_labels_aot + else: + predictor = predict_labels_aot + + if project_type == "host_driven": + with create_aot_session( + platform, + board, + target, + relay_mod, + params, + build_dir=workspace_dir, + # The longest models take ~5 seconds to infer, but running them + # ten times (with NUM_TESTING_RUNS_PER_SAMPLE) makes that 50 + timeout_override=server.TransportTimeouts( + session_start_retry_timeout_sec=300, + session_start_timeout_sec=150, + session_established_timeout_sec=150, + ), + project_options=project_options, + use_cmsis_nn=use_cmsis_nn, + ) as session: + aot_executor = tvm.runtime.executor.aot_executor.AotModule( + session.create_aot_executor() + ) + args = { + "session": session, + "aot_executor": aot_executor, + "input_data": get_test_data(model_name), + "runs_per_sample": 10, + } + predicted_labels, runtimes = zip(*predictor(**args)) + + avg_runtime = float(np.mean(runtimes) * 1000) + print(f"Model {model_name} average runtime: {avg_runtime}") + + elif project_type == "mlperftiny": + runtime = Runtime("crt") + executor = Executor( + "aot", {"unpacked-api": True, "interface-api": "c", "workspace-byte-alignment": 8} + ) + + config = {"tir.disable_vectorize": True} + if use_cmsis_nn: + from tvm.relay.op.contrib import cmsisnn + + config["relay.ext.cmsisnn.options"] = {"mcpu": target.mcpu} + relay_mod = cmsisnn.partition_for_cmsisnn(relay_mod, params, mcpu=target.mcpu) + + with tvm.transform.PassContext(opt_level=3, config=config): + module = tvm.relay.build( + relay_mod, target=target, params=params, runtime=runtime, executor=executor + ) + + temp_dir = tvm.contrib.utils.tempdir() + model_tar_path = temp_dir / "model.tar" + export_model_library_format(module, model_tar_path) + workspace_size = mlf_extract_workspace_size_bytes(model_tar_path) + + extra_tar_dir = tvm.contrib.utils.tempdir() + extra_tar_file = extra_tar_dir / "extra.tar" + with tarfile.open(extra_tar_file, "w:gz") as tf: + with tempfile.TemporaryDirectory() as tar_temp_dir: + model_files_path = os.path.join(tar_temp_dir, "include") + os.mkdir(model_files_path) + header_path = generate_c_interface_header( + module.libmod_name, + [model_info["input_name"]], + [model_info["output_name"]], + [], + {}, + [], + 0, + model_files_path, + {}, + {}, + ) + tf.add(header_path, arcname=os.path.relpath(header_path, tar_temp_dir)) + + create_header_file( + "output_data", + np.zeros( + shape=model_info["output_shape"], + dtype=model_info["output_dtype"], + ), + "include/tvm", + tf, + ) + + input_total_size = 1 + input_shape = model_info["input_shape"] + for i in range(len(input_shape)): + input_total_size *= input_shape[i] + + template_project_path = pathlib.Path(tvm.micro.get_microtvm_template_projects(platform)) + project_options.update( + { + "extra_files_tar": str(extra_tar_file), + "project_type": project_type, + "board": board, + "compile_definitions": [ + f"-DWORKSPACE_SIZE={workspace_size + 512}", # Memory workspace size, 512 is a temporary offset Review Comment: yeah because we are not building with USMP yet. I should rename it to `TVM_WORKSPACE_SIZE_BYTES` though -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
