Skip to content

Latest commit

 

History

History
66 lines (42 loc) · 2.9 KB

README.md

File metadata and controls

66 lines (42 loc) · 2.9 KB
Raw

Fine-tune the pixplotml visualization to your image data

PixplotML uses UMAP to cluster the images using a vector representation of each image. To fine-tune PixplotML we create this image vector file by training a classification ML model on your data. The code uses Pytorch-Accelerated to easily and simply train this classification model.

Note: As we're training an ML image model we need to use a machine with a GPU and docker enabled with the NVIDIA extensions.

The inputs:

metadata.csv - one line for each image to visualize. Minimally must contain filename, category columns
Image files - one for each line of the metadata.csv file.

The output:

image_vectors.npy

Process

To train the model on your images and create the image_vectors.npy file, first create the python environment and run the following command:

conda create -n prep_pixplotml python=3.9
conda activate prep_pixplotml
pip install -r ./environment/requirements.txt

Then run the code to create the image_vectors.npy file (using the quickstart coco files as an example - unzip the data/outputs_trained.zip first):

 python main.py --data_fldr ../data/outputs --img_fldr ../data/outputs/images

The image_vectors.npy file will be saved to the location specified in the --data_fldr command line switch.

At this point you can just run the pixplotml server pointing at the --data_fldr folder.

Using Docker

Build the docker image (one-time only):

cd prep_pixplot_files
docker build -t prep_pixplotml:1.0 ./environment

If you have a GPU then we can create image vectors fine-tuned to your data:

docker run --gpus all --ipc=host -v `pwd`/../data/outputs:/data -v `pwd`:/src prep_pixplotml:1.0 python /src/main.py --data_fldr /data --img_fldr /data/images

Then follow the instructions to start Pixplot in the quickstart section of the main README file

If you don't have a GPU then you should just use the imagenet model without finetuning to your data, as it will take a very long time to train a model, but just using the imagenet model will work fine on CPU:

docker run --ipc=host -v `pwd`/../data/outputs:/data -v `pwd`:/src prep_pixplotml:1.0 python /src/main.py --data_fldr /data --img_fldr /data/images --use_imagenet True --device cpu

Optional command line switches

There are some optional command line parameters that you can use if necessary:

1. use_imagenet -- if we wish we can just use the pre-trained imagenet model and not train on your data. This will lead to worse results but is quicker.
2. num_epochs -- defaults to 20
3. min_size -- the size to resize the images to when training the model. The process will also pad the images. This must be <224 but defaults to 60 pixels.
4. device - the torch device to use. If you don't have a GPU then set use_imagenet to True and set the device='cpu'.