[Java] [Python] I don't know why the output is different even though they are the same input(float32) and onnx model.

[shihaim]

Describe the issue

Hi!
It's really important. please help..

[common]
location : shape(1, 1000, 4)
confidence : shape(1, 1000, 2)
landmarks : shape(1, 1000, 10)

limited the system output to five.

[java]

version : 11+
onnxruntime version : 1.14.0

explicit SessionOptions

private void setOnnxOpts(OrtSession.SessionOptions opts) throws OrtException {
	opts.addCPU(true);

	opts.setMemoryPatternOptimization(true);

	opts.disableProfiling();

	opts.setExecutionMode(OrtSession.SessionOptions.ExecutionMode.SEQUENTIAL);

	opts.setOptimizationLevel(OrtSession.SessionOptions.OptLevel.ALL_OPT);

	opts.setInterOpNumThreads(0);

	opts.setIntraOpNumThreads(0);

	opts.setOptimizedModelFilePath("");
	
	opts.addConfigEntry("session.use_ort_model_bytes_directly", "1");
}

[python]

version : 3.7.9
onnxruntime version : 1.14.0

explicit SessionOptions

def set_onnx_opts(self, opts):
	opts.enable_cpu_mem_arena = True

	opts.enable_mem_pattern = True
	
	opts.enable_profiling = False

	opts.execution_mode = onnxruntime.ExecutionMode.ORT_SEQUENTIAL

	opts.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL

	opts.inter_op_num_threads = 0

	opts.intra_op_num_threads = 0

	opts.optimized_model_filepath = ""


	opts.use_deterministic_compute = False

	opts.add_session_config_entry("session.use_ort_model_bytes_directly", "1")

	return opts

To reproduce

n/a

Urgency

No response

Platform

Windows

OS Version

11

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

1.14.0

ONNX Runtime API

Java

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response

[yuslepukhin]

Your version is really old, we do not even build it here. Your best chance is to try the latest. Any fixes that may come would be in the latest version.

[Craigacp]

Ensuring consistent image processing between Java & Python can be tricky, you need to make sure the channels order lines up, that the Java buffer is prepared correctly, and that you're parsing the output right. The output looks like the parsing is ok, so I'd look into the image processing side in Java.

[shihaim]

@yuslepukhin
Of course as you said the old version, but I confirmed that the output (output with the same value) did not change at all when I ran it with the latest version.

[build.gradle.kts]

[Java Output]

[shihaim]

@Craigacp
Thank you..!