Custom transfer learning tensorflowlite model gives error in android code for image classification

General Discussion
,
#1

Hello, I am trying to create chess pieces image classification program running on my phone. I trained my model using transfer learning technique. Made 6 classes for the six figures and converted my tf model to tflite model. When I use the model with a labelmap in the android project I get an error stating: Caused by: java.lang.IllegalArgumentException: Label number 6 mismatch the shape on axis 1. The whole error log is here: https://pastebin.com/bxdq9x1r.
This is my tflite model: https://www.pastefile.com/vpg57x
This is my label map: https://www.pastefile.com/9rg9v7

1 Like
#2

The label map URL seems unreachable. Can you check it?

2 Likes
#3

This is my label map: https://www.pastefile.com/ncfyht .
The link was giving error and I pasted it again.
If needed I can show the code for training the model and converting it to tflite as well as the java code for the android app.

1 Like
#4

I am using this command for the conversion of tf model to tflite model: command = "tflite_convert \ --saved_model_dir={} \ --output_file={} \ --input_shapes=1,300,300,3 \ --input_arrays=normalized_input_image_tensor \ --output_arrays='TFLite_Detection_PostProcess','TFLite_Detection_PostProcess:1','TFLite_Detection_PostProcess:2','TFLite_Detection_PostProcess:3' \ --inference_type=FLOAT \ --allow_custom_ops".format(FROZEN_TFLITE_PATH, TFLITE_MODEL, ).
If you look at input_shapes, I think this is the image shape that the model will be used to, so maybe when i make a picture with my phone I get different shape for example 2000x3000px and then the error occurs. Can it be that one?

2 Likes
#5

In the interpreter API you have an utility to resize your input.

Check

3 Likes
#6

Hello again, so I checked it and the error was coming from the size of the input image. So I changed the size to match the input size of the model or in my case 640x640. But now I am facing two new errors one is: Cannot copy from a TensorFlowLite tensor with shape [1, 10, 4] to a Java object with shape [1, 6], which means that my shape(which consists of 6 chess figures) does not match the java code which expects 10, 4 so I find this line in the class: imgData = ByteBuffer.allocateDirect( 4 * DIM_IMG_SIZE_X * DIM_IMG_SIZE_Y * DIM_PIXEL_SIZE); and I delete the 4 so this error dissapears, but new one appears: java.nio.BufferOverflowException. I apply my interpreter class now, please help me I really am in a pickle: https://pastebin.com/Hti2zeJm. This exception appears at line 189. Thanks in advance.

1 Like
#7

Hi Ivan,

I do not think your assumption is correct. Check an error that I forced to appear in my AS project. E/EXECUTOR: something went wrong: Cannot copy to a TensorFlowLite tensor (input) with 1769472 bytes from a Java Buffer with 2211840 bytes. So my model expects 1769472 input and gets 2211840. In your case it expects input [1, 6] and gets [1, 10, 4].
Please check your .tflite file with netron.app to get a visualization of your .tflite file. You are always welcome to give us a link of your project to build and help you.

Best

2 Likes
#8

Hello, thank you a lot for the answer! I am adding my github repo link so you can download the project and look at the error yourself. My github link: https://github.com/ivanpetrov95/chesshelper/tree/master.
Now for the error, I simulated such error by replacing my 640 px size to something bigger or lower, but again I am newbie in the tensorflow area so I might be wrong. My model is in the assets as well as the labelmap and everything needed. I checked my tflite file in netron and I saw that way my needed input must be 640x640x3(3 I think is the RGB, but again I am a newbie), but I could not “read” the output.
Thank you again for looking at my problem and answering me!

1 Like
#9

Hi Ivan,

I will look into your project and I will get back to you when I figure out what is going on with the inputs to the Interpreter. But you have to definetely check the conversion of your model to tflite because I have also seen that netron.app does not show results for the output of the model. Something is going on there!

Best

1 Like
#10

Hi Ivan,

I built your project and took a look into it. The result is that your model outputs an array of [1][10][4] and you are givving it an output array of [1][6]. So the problem is at your output of the model. I think you have to check again the conversion to tflite (if pure tensorflow code works correctly) and check it with netron.app. I really do not know what the output should be, so if it has to be an array of [1][6] after the correct convertion you will see this at netron.app.
Besides that there are some other coding issues at your project that I managed to overcome…but they are minor and firstly you have to check the conversion.
Happy to help again. Ping me anytime.

Best

4 Likes
#11

Hello, I looked at my model in netron and indeed I sawo 10 nodes at the output I suppose you talk about them when you say 1, 10, 4. Attaching an image:
https://pasteboard.co/K3mBbgg.png
I have followed a tutoril and I am pretty new to this area, so sorry if I do not have good knowledge about that. I will try building another model and look at it and get in touch with you again. THANK you a lot for the help! :slight_smile:

2 Likes
#12

Hello again, I have a problem with converting my .pb file to tflite. I am using tensorflow 1.14 now and I am trying to create the tflite file with tflite_converter. The command is: tflite_convert \ --graph_def_file=tmp/frozen.pb \ --output_file=tmp/model.tflite \ --input_format=TENSORFLOW_GRAPHDEF \ --output_format=TFLITE \ --input_arrays=input_tensor \ --output_arrays=output_pred \ --input_shapes=1,224,224,3 \. The error I get is: ValueError: The shape of tensor 'input_tensor' cannot be changed from (?, ?) to [1, 224, 224, 3]. Shapes must be equal rank, but are 2 and 4

1 Like
#13

Also this is the frozen graph file I am trying to convert to tflite with the command: https://www.pastefile.com/ahahwt.
Thank you for the help.

1 Like
#14

Let’s tag @Sayak_Paul and ask him if he has a colab notebook demonstrating the appropriate way of converting TensorFlow 1.14 .pb files to tflite.

1 Like
#15

Here’s one:

3 Likes
#16

@thea is it possible to install plugins for better code renditions? I would suggest installing at least two: regular Python code and Jupyter Notebook code.

2 Likes