Add model test CI (#18)

Signed-off-by: Sun, Xuehao <[email protected]>

.azure-pipelines/docker/Dockerfile.devel

@@ -32,6 +32,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends --fix-missing \
     time \
     wget \
     bc \
+    jq \
     vim
 
 RUN ln -sf $(which python3) /usr/bin/python

.azure-pipelines/model-test.yml

@@ -0,0 +1,121 @@
+trigger: none
+
+pr:
+  autoCancel: true
+  drafts: false
+  branches:
+    include:
+      - main
+  paths:
+    include:
+      - onnx_neural_compressor
+      - setup.py
+      - requirements.txt
+      - .azure-pipelines/scripts/models
+      - .azure-pipelines/model-test.yml
+      - .azure-pipelines/template/model-template.yml
+    exclude:
+      - test
+
+variables:
+  OUT_SCRIPT_PATH: $(Build.SourcesDirectory)/.azure-pipelines/scripts/models
+  SCRIPT_PATH: /neural_compressor/.azure-pipelines/scripts
+
+parameters:
+  - name: algorithms
+    type: object
+    default:
+      - SQ
+      - WOQ
+  - name: models
+    type: object
+    default:
+      - bert_base_MRPC
+      - bert_base_MRPC_dynamic
+      - resnet50-v1-12_qdq
+      - resnet50-v1-12
+
+stages:
+  # - stage: ONNX_LLM_Models
+  #   displayName: Run ONNX LLM Model
+  #   pool: ICX-16C
+  #   dependsOn: []
+  #   jobs:
+  #     - ${{ each algorithm in parameters.algorithms }}:
+  #         - job:
+  #           steps:
+  #             - template: template/model-template.yml
+  #               parameters:
+  #                 modelName: "facebook/opt-125m"
+  #                 algorithm: "${{ algorithm }}"
+  #                 script_path: "run_onnxrt_llm_models_trigger.sh"
+
+  - stage: ONNX_Models
+    displayName: Run ONNX Model
+    pool: MODEL_PERF_TEST
+    dependsOn: []
+    jobs:
+      - ${{ each model in parameters.models }}:
+          - job:
+            displayName: ${{ model }}
+            steps:
+              - template: template/model-template.yml
+                parameters:
+                  modelName: "${{ model }}"
+                  algorithm: "Quantize"
+                  script_path: "run_onnxrt_models_trigger.sh"
+
+  - stage: GenerateLogs
+    displayName: Generate Report
+    pool:
+      vmImage: "ubuntu-latest"
+    dependsOn: [ONNX_Models]
+    jobs:
+      - job: GenerateReport
+        steps:
+          - script: |
+              echo ${BUILD_SOURCESDIRECTORY}
+              rm -fr ${BUILD_SOURCESDIRECTORY} || sudo rm -fr ${BUILD_SOURCESDIRECTORY} || true
+              echo y | docker system prune
+            displayName: "Clean workspace"
+          - checkout: self
+            clean: true
+            displayName: "Checkout out Repo"
+          - task: DownloadPipelineArtifact@2
+            inputs:
+              artifact:
+              patterns: "**/result.json"
+              path: $(OUT_SCRIPT_PATH)
+          - task: UsePythonVersion@0
+            displayName: "Use Python 3.10"
+            inputs:
+              versionSpec: "3.10"
+          - script: |
+              cd ${OUT_SCRIPT_PATH}
+              mkdir generated last_generated
+              python -u summarize_results.py --logs_dir $(OUT_SCRIPT_PATH) --output_dir generated
+            displayName: "Summarize all results"
+          - task: DownloadPipelineArtifact@2
+            continueOnError: true
+            inputs:
+              source: "specific"
+              artifact: "FinalReport"
+              patterns: "**.json"
+              path: $(OUT_SCRIPT_PATH)/last_generated
+              project: $(System.TeamProject)
+              pipeline: "onc model test"
+              runVersion: "specific"
+              runId: $(refer_buildId)
+            displayName: "Download last logs"
+          - script: |
+              echo "------ Generating final report.html ------"
+              cd ${OUT_SCRIPT_PATH}
+              pip install jinja2
+              python generate_report.py --json_path generated/summary.json --last_json_path last_generated/summary.json
+            displayName: "Generate report"
+          - task: PublishPipelineArtifact@1
+            inputs:
+              targetPath: $(OUT_SCRIPT_PATH)/generated
+              artifact: FinalReport
+              publishLocation: "pipeline"
+            displayName: "Publish report"

.azure-pipelines/scripts/models/collect_results.py

@@ -0,0 +1,84 @@
+import argparse
+import json
+import os
+import re
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--model", required=True, type=str)
+parser.add_argument("--build_id", required=True, type=str)
+args = parser.parse_args()
+
+URL = (
+    "https://dev.azure.com/lpot-inc/onnx-neural-compressor/_build/results?buildId="
+    + args.build_id
+    + "&view=artifacts&pathAsName=false&type=publishedArtifacts"
+)
+REFER_SUMMARY_PATH = "/neural-compressor/.azure-pipelines/scripts/models/summary.json"
+
+
+def str_to_float(value):
+    try:
+        return round(float(value), 4)
+    except ValueError:
+        return value
+
+
+def get_refer_data():
+    if not os.path.exists(REFER_SUMMARY_PATH):
+        print(f"The file '{REFER_SUMMARY_PATH}' does not exist.")
+        return {}
+
+    with open(REFER_SUMMARY_PATH, "r") as file:
+        refer = json.load(file)
+    return refer
+
+
+def check_status(performance, accuracy):
+    refer = get_refer_data()
+
+    refer_accuracy = refer.get(args.model, {}).get("accuracy", {}).get("value", "N/A")
+    refer_performance = refer.get(args.model, {}).get("performance", {}).get("value", "N/A")
+    print(f"{accuracy=}\n{refer_accuracy=}\n{performance=}\n{refer_performance=}")
+
+    assert accuracy != "N/A" and performance != "N/A"
+    if refer_accuracy != "N/A":
+        assert abs(accuracy - refer_accuracy) <= 0.001
+    if refer_performance != "N/A":
+        assert (refer_performance - performance) / refer_performance <= 0.08
+
+
+def main():
+    result_dict = {
+        args.model: {
+            "performance": {"value": "N/A", "log_path": URL},
+            "accuracy": {"value": "N/A", "log_path": URL},
+        }
+    }
+
+    pattern = {
+        "performance": r"Throughput: ([\d.]+)",
+        "accuracy": r"Accuracy: ([\d.]+)",
+    }
+
+    for mode, _ in result_dict[args.model].items():
+        log_file = f"/neural-compressor/.azure-pipelines/scripts/models/{args.model}/{mode}.log"
+        if not os.path.exists(log_file):
+            print(f"The file '{log_file}' does not exist.")
+            continue
+
+        with open(log_file, "r") as file:
+            log_content = file.read()
+
+        match = re.search(pattern[mode], log_content)
+
+        if match:
+            result_dict[args.model][mode]["value"] = str_to_float(match.group(1))
+
+    with open(f"/neural-compressor/.azure-pipelines/scripts/models/{args.model}/result.json", "w") as json_file:
+        json.dump(result_dict, json_file, indent=4)
+
+    check_status(result_dict[args.model]["performance"]["value"], result_dict[args.model]["accuracy"]["value"])
+
+
+if __name__ == "__main__":
+    main()

.azure-pipelines/scripts/models/env_setup.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+set -eo pipefail
+PATTERN='[-a-zA-Z0-9_]*='
+
+for i in "$@"; do
+    case $i in
+    --model=*)
+        model=${i//${PATTERN}/}
+        ;;
+    *)
+        echo "Parameter $i not recognized."
+        exit 1
+        ;;
+    esac
+done
+
+CONFIG_PATH="/neural-compressor/examples/.config/model_params_onnxrt.json"
+model_src_dir=$(jq -r ".\"onnxrt\".\"$model\".\"model_src_dir\"" "$CONFIG_PATH")
+
+log_dir="/neural-compressor/.azure-pipelines/scripts/models"
+
+$BOLD_YELLOW && echo "======= creat log_dir =========" && $RESET
+if [ -d "${log_dir}/${model}" ]; then
+    $BOLD_GREEN && echo "${log_dir}/${model} already exists, don't need to mkdir." && $RESET
+else
+    $BOLD_GREEN && echo "no log dir ${log_dir}/${model}, create." && $RESET
+    cd "${log_dir}"
+    mkdir "${model}"
+fi
+
+$BOLD_YELLOW && echo "====== install ONC ======" && $RESET
+cd /neural-compressor
+source .azure-pipelines/scripts/change_color.sh
+/bin/bash .azure-pipelines/scripts/install_nc.sh
+
+$BOLD_YELLOW && echo "====== install requirements ======" && $RESET
+cd "/neural-compressor/examples/$model_src_dir"
+pip install -r requirements.txt
+pip list

.azure-pipelines/scripts/models/generate_report.py

@@ -0,0 +1,103 @@
+import argparse
+import json
+import os
+
+from jinja2 import Environment, FileSystemLoader
+
+parser = argparse.ArgumentParser(allow_abbrev=False)
+parser.add_argument("--json_path", type=str, required=True)
+parser.add_argument("--last_json_path", type=str, required=True)
+args = parser.parse_args()
+
+
+def get_data(json_path):
+    """
+    {
+        model: {
+            "performance": {"value": "N/A"|number, "log_path": string},
+            "accuracy": {"value": "N/A"|number, "log_path": string},
+        }
+    }
+    """
+    if os.path.exists(json_path):
+        with open(json_path, "r") as f:
+            return json.load(f)
+    else:
+        return {}
+
+
+def get_ratio(cur, last):
+    if cur == "N/A" or last == "N/A":
+        ratio = "N/A"
+    else:
+        ratio = (float(cur) - float(last)) / float(last) * 100
+        ratio = round(float(ratio), 2)
+    return ratio
+
+
+def get_accuracy_ratio(current_json, last_accuracy_dict):
+    compare_result_dict = []
+    for model, item in current_json.items():
+        current_accuracy = item.get("accuracy", {}).get("value", "N/A")
+        last_accuracy = last_accuracy_dict.get(model, {}).get("accuracy", {}).get("value", "N/A")
+        accuracy_ratio = get_ratio(current_accuracy, last_accuracy)
+
+        current_performance = item.get("performance", {}).get("value", "N/A")
+        last_performance = last_accuracy_dict.get(model, {}).get("performance", {}).get("value", "N/A")
+        performance_ratio = get_ratio(current_performance, last_performance)
+
+        if accuracy_ratio == "N/A" or performance_ratio == "N/A":
+            status = "FAILURE"
+        elif accuracy_ratio != 0:
+            status = "FAILURE"
+        elif performance_ratio > 8 or performance_ratio < -8:
+            status = "FAILURE"
+        else:
+            status = "SUCCESS"
+
+        format_ratio = lambda x: f"{x}%" if x != "N/A" else x
+
+        compare_result_dict.append(
+            {
+                "model": model,
+                "current_accuracy": current_accuracy,
+                "last_accuracy": last_accuracy,
+                "accuracy_ratio": format_ratio(accuracy_ratio),
+                "current_performance": current_performance,
+                "last_performance": last_performance,
+                "performance_ratio": format_ratio(performance_ratio),
+                "status": status,
+            }
+        )
+    return compare_result_dict
+
+
+def generate(rendered_template):
+    with open("generated/report.html", "w") as html_file:
+        html_file.write(rendered_template)
+
+
+def main():
+    path = "{}/templates/".format(os.path.dirname(__file__))
+    BUILD_BUILDID = os.getenv("BUILD_BUILDID")
+
+    loader = FileSystemLoader(path)
+    env = Environment(loader=loader)
+    template = env.get_template("model.jinja2")
+
+    data = get_data(args.json_path)
+    last_data = get_data(args.last_json_path)
+    data = get_accuracy_ratio(data, last_data)
+    info = {
+        "url": f"https://dev.azure.com/lpot-inc/onnx-neural-compressor/_build/results?buildId={BUILD_BUILDID}",
+        "branch": os.getenv("SYSTEM_PULLREQUEST_SOURCEBRANCH"),
+        "commit": os.getenv("BUILD_SOURCEVERSION"),
+        "build_number": BUILD_BUILDID,
+    }
+
+    rendered_template = template.render(data=data, info=info)
+    generate(rendered_template)
+
+
+if __name__ == "__main__":
+    main()