How to run the streamlit app and train/add your model¶

This manual shows you how to run the streamlit app. It also explains how to train your model based on the csv files and the images using the scripts code/labelCNN/combine_labels.py and code/labelCNN/training.py. For details please look at the source code. Also, have a look at /net/projects/scratch/winter/valid_until_31_July_2020/asparagus/katha/ to see what a possible setup looks like or how to use the grid (train.sge) to train your model.

Clone the GitHub repository¶

git clone https://github.com/CogSciUOS/asparagus.git

Activate the asparagus_env¶

In the repository’s root directory, you can find the asparagus_env.yml file to create an environment . It is most probably not platform independent, so please use a Linux system (e.g. the university computers). Use the command conda env create -f asparagus_env.yml to set up the environment. Then activate the environment by using conda activate asparagus_env

Regarding the requirements¶

You do not need to use the provided environment. Everything should work if you install the following Python packages: tensorflow h5py numpy pandas scikit-image streamlit.

Preparation¶

The image paths are assumed to have this structure: ${IMAGE_DIR}/0/0/0000_a.jpg, this way we can more or less reliably read them from the csv files.
Look at the easy alternative below step 2!
Prepare the labels, so that they are in one file:python combine_labels.py ${LABELS_FOLDER} labels.csv This merges all csv files inside the directory labels properly into labels.csv You can find the script in the folder code. Also have a look at the bash script merge_labels.sh to see which folder I used and with which command line arguments I called the script. You find it in /net/projects/scratch/winter/valid_until_31_July_2020/asparagus/katha/.

Easy alternative to step 2.: You can also use the labels.csv file I already created at /net/projects/scratch/winter/valid_until_31_July_2020/asparagus/katha/.
Skip this part if you want to use a trained model (it is as easy as selecting it in the app):
To train, run training.py: python training.py labels.csv /net/projects/scratch/winter/valid_until_31_July_2020/asparagus/preprocessed_images/without_background_pngs/ model_name Where images is the ${IMAGE_DIR} from step 1 and labels.csv the file from step 2 and a name for the model. Make sure you save your trained model to the folder code/labelCNN/models. Also have a look at the bash script train.sge to see with which command line arguments I called the script. You find it in /net/projects/scratch/winter/valid_until_31_July_2020/asparagus/katha/.
To inspect the training process, run tensorboard --logdir logs and open a browser at http://localhost:6006

To inspect the results, you can use streamlit: streamlit run app_from_img_to_categories.py(for details follow the next steps).

Navigate to the streamlit app¶

You can find the app that supplies an end-to-end solution from a csv file containing manually labeled features and paths to images of asparagus pieces to the prediction of the quality class of the asparagus piece. You can find it in the code folder. It is named app_from_img_to_categories.py. In order to run the app, type streamlit run app_from_img_to_categories.py in the terminal.

Open browser at the provided URL¶

In the terminal you will find output similar to this:

  You can now view your Streamlit app in your browser.

  Network URL: http://192.168.0.8:8501
  External URL: http://31.17.252.51:8501

Copy the Network URL into the address field of the browser of your choice.

Select the path to image folder¶

Now you should see that the app is running and greets you with “Asparagus label prediction”. The next step is to select the path of the image folder e.g. /net/projects/scratch/winter/valid_until_31_July_2020/asparagus/preprocessed_images/without_background_pngs/. If the one you want to specify does not appear in the selection bar, please select None and the app allows you to specify the path manually.

Select csv file¶

Select a corresponding csv file e.g. /net/projects/scratch/winter/valid_until_31_July_2020/asparagus/katha/labels.csv or the one you created in step 2.

Looking at the app¶

Go through the app step by step. The app automatically updates after you made a selection.

If something is not working¶

Please let me know when you encounter problems or have some suggestions on how to improve the app. I would also be very happy about pull requests that add new features.

Things you can try on your own when running into problems:

refresh
check all the check boxes from top to bottom
check the source code

Cuda removal¶

I had to remove cudnn and cudatoolkit to run everything without a problem. That is why they are not included in the asparagus_env. If that leads to problems, install the environment and remove the two packages selectively without removing their dependencies. See: conda documentation.

Finally¶

I hope we will add many cool models and explore our data further :) Have fun playing around with the app!

Port forwarding¶

Port forwarding is still under active investigation. I will keep the manual updated.