How to Train a Custom Mobile Object Detection Model (with YOLOv4 Tiny and TensorFlow Lite)

Train a custom YOLOv4 tiny object detection model and deploy to TFLite (TensorFlow Lite), so you can deploy your model on device 🚀

Corresponding blog on training custom model and deploy to mobile:

Corresponding YOLOv4 TensorFlow Lite Colab Notebook:

CVAT labeling post:

[kdsv]

11:50 hehe forgot to edit that out :) But thank you so much for the great tutorial!

[alsatwarsravankumar]

I am getting this error!!
When I try to run it!!

CUDA Error: no kernel image is available for execution on the device
CUDA Error: no kernel image is available for execution on the device: File exists
darknet: ./src/utils.c:325: error: Assertion `0' failed.

Unable to debug it!!

[mahadimunna]

Thank you very much for this tutorial, it helped a lot in tflite conversion. I was stuck in it for a while. It was a great help.

[Esudun]

Getting this error
Exception: <unknown>:0: error: is not immutable, try running to prove tensors

[sohailali]

Thanks for great tutorial, it would be great if you can make a video on object detection evaluation metrics.

[sadiyeozdogan]

Hi, I can do all the steps in the colab notebook but it does not save the last step: ': "No such a file or directory" error. Can you help me? Thank you.

[andresdelgado]

CUDA Error: no kernel image is available for execution on the device i cant't solve this error

[jothegamechanger]

Hello roboflow, when i runon android all the got force close

[juandiegoaw]

Could you please teach how to quantize the tflite file?

[izabellaszen]

Getting this error when converting to tflite
error: is not immutable, try running to prove tensors are immutable
Traceback (most recent call last):
File "/usr/local/lib/python3.6/dist-packages/tensorflow/lite/python/convert.py", line 213, in toco_convert_protos
enable_mlir_converter)
File "/usr/local/lib/python3.6/dist-packages/tensorflow/lite/python/wrap_toco.py", line 38, in wrapped_toco_convert
enable_mlir_converter)
Exception: <unknown>:0: error: is not immutable, try running to prove tensors are immutable

[Hammale.Mourad]

When I tried to convert yolov4 tiny weights to Tensorflow Lite I got this .
[[[[Some ops are not supported by the native TFLite runtime, you can enable TF kernels fallback using TF Select.
When I convert model from tf to tflite, a question occur like this, and I add "--target_ops=TFLITE_BUILTINS, SELECT_TF_OPS",

it still can not fix this question, so how do i fix the question.

[robr]

Any suggestions for me? I detected an object but I want to register or align it automatically. Any thoughts on how I can do this?

[Dubb-eegr]

HELP PLEASE:
In the "Configure from Custom Dataset" section. I get the error "FileNotFoundError: [Errno 2] No such file or directory: 'train' "

I've done troubleshooting. No luck. Any Ideas?

[luzanojustinem.]

train custom yolov4 detector

CUDA Error: no kernel image is available for execution on the device
CUDA Error: no kernel image is available for execution on the device: File exists
darknet: ./src/utils.c:325: error: Assertion `0' failed.


help please!!!

[fachriramdhan]

halo sir, i need some help. when i convert tensorflow weight to tf lite i get this problem :
" Variable constant folding is failed. Please consider using enabling flag in the TFLite converter object. For example, = True "

can nobody know how to solve it ?

[educationislah]

Automatic saving failed. This file was updated remotely or in another tab.

[shounakjoshi]

error: is not immutable, try running to prove tensors are immutable
Traceback (most recent call last):

I have been getting this error.

[barbaracardoso]

Great tutorial. I managed to implement on raspberry pi 4. With some modifications and overclocking, the FPS increase from 1.2 to 5.1. I'm using for detect cardboard boxes. It's sort of slow yet, but it's doing the basics that I need.

[abhishekkumarsrivastava]

HI, what's size of yolov4 tiny and converted tflite model for you? For me darknet model size (for yolov4 tiny) is 23MB but the converted TensorFlow model size is 2.7MB only. When I convert it to tflite, model size grew back to 23MB approx (without quantization). Is this ok or am I doing something wrong?

[nafiu]

Is it possible to run the code from colab, locally on a windows pc? It seems not to work if I connect with local runtime on the colab.