Model trained with Tensorflow 1.7.1 not working

[aliBUT]

Hi , i was checking your classify example, it works perfectly. But i trained a model using Google code labs and i followed these steps. I downloaded tensorflow 1.7.1 same as tfcsharp plugin . testing works perfectly on python level . but when i load retrained_graph.bytes in your code is dosent load this model . Architecture is mobilenet_0.50_224 and image size is 224 .

this is my retrained model and label

https://www.dropbox.com/s/vw34zpkg2a5klrw/retrained.zip?dl=0

i am not sure what i am doing wrong here

thanks

[Syn-McJ]

Hi, so for your model did you use 0.3 version of Unity plugin that I linked in the readme, or 0.4 version?

[aliBUT]

i used 0.4 version of unity plugin and i am using tensorflow-gpu 1.71 for training

[Syn-McJ]

Ok, I'll check your project later today to see what could be the problem.

[aliBUT]

this is the null exception I am receiving when I try to load my own model. I hope this will help you debug more easily

[Syn-McJ]

I've checked your models and haven't been able to figure out why NRE occurs, unfortunately. If it's critical issue for you then I suggest to use TensorFlow 1.4 for now, otherwise we can wait and see if solution is found, but I'm afraid my TF knowledge is hitting a wall here.

One more thing I would suggest is to check your model with TensorFlowSharp examples and see if it's working there.

[aliBUT]

i tried with tf 1.4 and all the code stopped working and throwing errors. So instead i am going to test my models with tensorflowsharp examples .

[Syn-McJ]

TF 1.4 should work fine, if you provide details to where errors occur I can take a look.

[aliBUT]

I mean errors in python scripts where I was doing my training. Sorry for the confusion.

when I downgraded from 1.7.1 to 1.4 it started giving me this error. Do you think this is because of path variable?

[Syn-McJ]

I see, that's weird. How did you downgrade tensorflow? I usually do pip uninstall tensorflow and then pip install tensorflow==1.4.

In any case, if TesnorflowSharp examples are having same problems with 1.7.1 model, then I think fixing your TF 1.4 installation will be your best bet to make it work in Unity.

[aliBUT]

I did pip install tensoflow==1.4 which caused that error . after doing pip uninstall tensoflow and then pip install tensoflow==1.4 error was no longer there

[aliBUT]

so i did al these steps

1: I downgraded my tensorflow to 1.4 and retrained the model 2: Downgraded the tensorflowsharp plugin from unity to 0.3 in unity project 3: renamed .pb file to bytes 4: Placed the modal and label in resources and gave the references 5: Made the Android build and safe error as above.

this is the model of tf 1.4 here

i think the export of model and the way modal is being used in c# does not match. what python code do you use to export models for image classification?

[Syn-McJ]

Yes, that's might be true. I can see your model is 5smth Mb in size, while my mobilenet models ending up to be around 17Mb.

I use retrain.py script from tensorflow repo: for 1.4: https://github.com/tensorflow/tensorflow/blob/r1.4/tensorflow/examples/image_retraining/retrain.py for 1.7: https://github.com/tensorflow/hub/blob/master/examples/image_retraining/retrain.py

[aliBUT]

i am training it for 5 different categories i think that's why size it that low

https://github.com/tensorflow/tensorflow/blob/r1.4/tensorflow/examples/image_retraining/retrain.py I used this to retain and still the null exception. Anything else I can try?

[aliBUT]

python -m image_retraining.retrain --bottleneck_dir=tf_files/bottlenecks --how_many_training_steps=500 --model_dir=tf_files/models/ --summaries_dir=tf_files/training_summaries/mobilenet_0.50_224 --output_graph=tf_files/retrained_graph.pb --output_labels=tf_files/retrained_labels.txt --architecture=mobilenet_0.50_224 --image_dir=tf_files/DataSet this is my retraining command

[Syn-McJ]

Size doesn't really depend that much on amount of categories, but I think I see why your model is smaller - you use 0.50 mobilenet architecture. Can you try to train full size mobilenet model? --architecture=mobilenet_1.0_224

If it still doesn't work, please send me your trained full size model, I'll check it again.

[aliBUT]

it did not worked . here is the full model

And i also noticed this error inside unity when i run it in editor

[aliBUT]

So just to be sure I prepared a new pc and did everything there, Same null exception issue

[Syn-McJ]

Hi @aliBUT, there is one thing I think you didn't do from the readme: change your INPUT_NAME and OUTPUT_NAME. For mobilenet architecture you have to have this values:

private static string INPUT_NAME = "input";
private static string OUTPUT_NAME = "final_result";

This was the cause of your NullReferenceException. After you fix NRE, you might also experience IndexOutOfRangeException, this one is caused by one extra line in the labels file.

I just checked your last model with correct input/output names and removed empty line in labels, and it works fine. After confirming you can also go back and try your previous models, including 1.7.1 one.

Hope it helps.

[Syn-McJ]

So I just rechecked your original 1.7.1 model with correct INPUT_NAME and OUTPUT_NAME and it works fine with 0.4 plugin. Sorry, totally slipped my mind when I first checked it that I need to change those.

I will close this issue as resolved for now, but feel free to reopen if you still have issues.

[aliBUT]

@Syn-McJ it Works !!!! . how do you usually know what is INPUT_NAME and OUTPUT_NAME ? for all the model architectures?

[Syn-McJ]

@aliBUT usually you can infer that from script or example you used. Like for instance, you can see in this guide that layer names are "Placeholder" and "final_result".

[aliBUT]

So Classify is working fine. I decided to try out object detection. Everything works fine as long as I use your model so I trained my modal by following [this ](https://blogs.msdn.microsoft.com/esmsdn/2018/04/09/post-invitado-part2-step-by-step-how-to-training-your-own-detector-classifier/) tutorial

i am facing same issue as my trained modal doesn't work . I made sure input is image tensor and use SS mobilenet v2 lite for training . i am not sure what is the issue here

this is my model and label map

https://www.dropbox.com/s/sgtolycu2t0hx7b/graph.zip?dl=0

Any help will be appreciated

[Syn-McJ]

Hi @aliBUT, could you open separate issue for that? I actually haven't tried object detention with other model but one in tensorflow repo, so I'll need to investigate this thoroughly when have time.

[aliBUT]

Opened #13