java.lang.IllegalArgumentException: Cannot copy to a TensorFlowLite tensor (serving_default_input_1:0) with 4915200 bytes from a Java Buffer with 1228800 bytes

[hanspres1999]

I am running my custom trained YOLO5 model on the example app and end up with the error:

E/AndroidRuntime(17777): java.lang.IllegalArgumentException: Cannot copy to a TensorFlowLite tensor (serving_default_input_1:0) with 4915200 bytes from a Java Buffer with 1228800 bytes.
E/AndroidRuntime(17777): 	at org.tensorflow.lite.TensorImpl.throwIfSrcShapeIsIncompatible(TensorImpl.java:416)
E/AndroidRuntime(17777): 	at org.tensorflow.lite.TensorImpl.setTo(TensorImpl.java:140)
E/AndroidRuntime(17777): 	at org.tensorflow.lite.NativeInterpreterWrapper.run(NativeInterpreterWrapper.java:251)
E/AndroidRuntime(17777): 	at org.tensorflow.lite.InterpreterImpl.runForMultipleInputsOutputs(InterpreterImpl.java:101)
E/AndroidRuntime(17777): 	at org.tensorflow.lite.Interpreter.runForMultipleInputsOutputs(Interpreter.java:95)
E/AndroidRuntime(17777): 	at com.ultralytics.ultralytics_yolo.predict.detect.TfliteDetector.runInference(TfliteDetector.java:250)
E/AndroidRuntime(17777): 	at com.ultralytics.ultralytics_yolo.predict.detect.TfliteDetector.lambda$predict$0$com-ultralytics-ultralytics_yolo-predict-detect-TfliteDetector(TfliteDetector.java:196)
E/AndroidRuntime(17777): 	at com.ultralytics.ultralytics_yolo.predict.detect.TfliteDetector$$ExternalSyntheticLambda0.run(Unknown Source:2)
E/AndroidRuntime(17777): 	at android.os.Handler.handleCallback(Handler.java:958)
E/AndroidRuntime(17777): 	at android.os.Handler.dispatchMessage(Handler.java:99)
E/AndroidRuntime(17777): 	at android.os.Looper.loopOnce(Looper.java:230)
E/AndroidRuntime(17777): 	at android.os.Looper.loop(Looper.java:319)
E/AndroidRuntime(17777): 	at android.app.ActivityThread.main(ActivityThread.java:8906)
E/AndroidRuntime(17777): 	at java.lang.reflect.Method.invoke(Native Method)
E/AndroidRuntime(17777): 	at com.android.internal.os.RuntimeInit$MethodAndArgsCaller.run(RuntimeInit.java:608)
E/AndroidRuntime(17777): 	at com.android.internal.os.ZygoteInit.main(ZygoteInit.java:1103)
I/Process (17777): Sending signal. PID: 17777 SIG: 9
Lost connection to device.

Exited.

I've changed the imgsz parameter on the metadata.yaml file to 640, 640 (and tried keeping it at 320,320) and yet no difference.

[glenn-jocher]

@hanspres1999 it seems like there's a mismatch between the input tensor size and the buffer size. Ensure your model's input dimensions match the expected input size in the app. Double-check the model's input shape and adjust the imgsz parameter accordingly. If the issue persists, try updating to the latest version of the app and model.

[dingniaoji]

exactly same error with yolov8 f16 imgsz=640 tflite model.
I have tried f16 with imgsz=320, it still outputs same error.

do I must use int8 model with imgsz=320?

[glenn-jocher]

@dingniaoji it seems like a tensor size mismatch. Ensure your input dimensions match the model's expected size. Try using int8 with imgsz=320 if the issue persists, and verify with the latest package version.

[ice6]

after a source code debugging, I finally solve the problem.

just to make sure that the imgsz in your metadata.yaml is same with your model.

I tried 320 with fp16 and 320 with fp32, both are ok.

[ice6]

@hanspres1999 in your case, it seems that in the metadata.yaml, imgsz should be 320 x 320, because 320 * 320 * 3 * 4 = 1228800.
and the imgsz in your model is 640x640, because640 * 640 * 3 * 4 = 4915200.
3 is the channel count(r/g/b) , 4 is the float32 size in Byte.

make the the size same may solve your problem.

[hanspres1999]

@ice6 I've kept my imgsz in metadata as 640 x 640. It still throws the same error. If it helps, the model was developed with the YOLOv5s backend.

[dingniaoji]

I have never tried yolov5, I think it should be fine, because the underlying is tensorflow-lite.

use netron(https://netron.app/) to inspect your model. and click the inputs_0, and paste the INPUTS and OUTPUTS in the right panel here.

paste the content of metadata.yaml here will also help.

[glenn-jocher]

Please inspect your model using Netron to verify the input dimensions. Share the INPUTS and OUTPUTS details, along with the metadata.yaml content, so we can assist you further.

[johnebere58]

Search for this and rename it to the size of the model you used

public static final int INPUT_SIZE = 320;

[pderrenger]

Ensure the INPUT_SIZE in your code matches the model's input dimensions. Adjust it to the size specified in your metadata.yaml.