[onnxruntime.training.artifacts] convert the'. h5 'file to onnx for training

[laborchen]

Use 'generate_artifacts' to convert the'. h5 'file to onnx for training
But the training effect in Qt is completely inferior to using Python's'. h5 '
I suspect it's a code conversion issue with 'onnxrruntime. training. artifacts' that caused the fusion of operators in training_madel.onnx, which can only predict but cannot be trained

import os
import onnx
import tf2onnx
import numpy as np
import tensorflow as tf
from tensorflow.keras.models import load_model, save_model
from tensorflow.keras.layers import Activation
from tensorflow.keras.models import clone_model
from tensorflow.keras.activations import relu
import onnxruntime.training.artifacts as artifacts

def export_initial_models(keras_model_path, output_dir):
"""
Args:
keras_model_path: Keras模型文件路径 (.h5)
output_dir: 输出目录
"""
# 创建输出目录
baseline_dir = os.path.join(output_dir, 'baseline')
current_dir = os.path.join(output_dir, 'current')
os.makedirs(baseline_dir, exist_ok=True)
os.makedirs(current_dir, exist_ok=True)
# 加载Keras模型
print("Loading Keras model...")
keras_model = load_model(keras_model_path)
# ==========================================================
# 导出 ONNX + ORT Training Artifacts
# ==========================================================
try:
print("开始导出 ONNX...")
# ONNX输出路径
onnx_path = os.path.join(baseline_dir, 'inference_model.onnx')
spec = (
tf.TensorSpec(
(1, 1, 5, 1000),
tf.float32,
name="input"
),
)
model_proto, _ = tf2onnx.convert.from_keras(
keras_model,
input_signature=spec,
opset=17
)
with open(onnx_path, "wb") as f:
f.write(model_proto.SerializeToString())
print("✓ inference_model.onnx 导出成功")
# ======================================================
# 生成 ORT Training Artifacts
# ======================================================
print("开始生成 ORT Training Artifacts...")
onnx_model = onnx.load(onnx_path)

    # 自动打印参数名
    print("=== 模型参数 ===")
    trainable_params = []
    for init in onnx_model.graph.initializer:
        name = init.name
        print(name)
        # 只训练真正权重
        if any(k in name.lower() for k in [
            "weight",
            "bias",
            "kernel",
            "gamma",
            "beta",
            "matmul",
            "conv"
        ]):
            trainable_params.append(name)
    print("\n=== 可训练参数 ===")
    for p in trainable_params:
        print(p)

    # 生成训练图
    artifacts.generate_artifacts(
        onnx_model,
        requires_grad=trainable_params,
        frozen_params=[],
        loss=artifacts.LossType.CrossEntropyLoss,
        optimizer=artifacts.OptimType.AdamW,
        artifact_directory=baseline_dir
    )
    print("✓ ORT Training Artifacts 生成成功")
except Exception as e:
    print("ORT Artifacts 导出失败:")
    print(e)

# 5. 复制到current目录
import shutil
for item in os.listdir(baseline_dir):
    src = os.path.join(baseline_dir, item)
    dst = os.path.join(current_dir, item)
    if os.path.isfile(src):
        shutil.copy2(src, dst)
    elif os.path.isdir(src):
        if os.path.exists(dst):
            shutil.rmtree(dst)
        shutil.copytree(src, dst)
print(f"\n✅ Initial models exported to {output_dir}")
print("\nGenerated files:")
for item in os.listdir(baseline_dir):
    item_path = os.path.join(baseline_dir, item)
    if os.path.isfile(item_path):
        size = os.path.getsize(item_path) / 1024
        print(f"  - {item} ({size:.2f} KB)")
    else:
        print(f"  - {item}/ (directory)")

if name == "main":
keras_model_path = "./Model/CrossSubject/Net.h5"
export_initial_models(keras_model_path, "./models")

[Sagi Amangeldi (capCafu)]

Probably not fusion. Nothing is trainable unless you name it.

generate_artifacts only gives gradients to parameters listed in requires_grad, and that argument defaults to None. Leave it out and the model will run a forward pass and compute a loss but never update a weight, which is what you're seeing.

The names have to match the initializers tf2onnx produced, not your Keras layer names, so easiest to just list them:

model = onnx.load("model.onnx")
names = [i.name for i in model.graph.initializer]
artifacts.generate_artifacts(model, requires_grad=names, loss=..., optimizer=...)

Anything you don't want trained goes in frozen_params.

[jinwoo (dev-jinwoohong)]

Your code already passes requires_grad=trainable_params, so the missing argument is not the explanation here. The name filter still needs checking. You can inspect what was actually saved as trainable in the checkpoint:

from onnxruntime.training.api import CheckpointState

state = CheckpointState.load_checkpoint("models/baseline/checkpoint")
for name, parameter in state.parameters:
    if parameter.requires_grad:
        print(name, parameter.data.shape, parameter.data.dtype)

Those checkpoint properties let you compare the result with the Keras weights you intend to update. Selecting every ONNX initializer can also select constants used for shapes or indexing.

Another concrete difference to check is the loss: artifacts.LossType.CrossEntropyLoss adds SoftmaxCrossEntropyLoss and creates INT64 labels. For ordinary classification, it expects raw class scores and class-index labels. If your exported Keras model already ends in Softmax, attaching this loss applies Softmax again and changes the training objective.

The model and Qt loop would be needed to identify the cause. Comparing the same batch's predictions and loss before the first update would help separate an export/loss mismatch from an update-loop problem.