ALl the classifiers supported by our code are trained on the 1000 class ImageNet dataset by default. If you want to see the labels available, please refer to Classify-read-me. If you would like to train the classifiers on other data sets, here you can find a list of some existing options:
The most straightforward method for creating models is to use the platform supervisely. This platform provides a way to annotate data, prepare a synthetic data set, train models, and download them without ever having to bring up an ipython notebook.
(nb: supervisely is a web platform, but needs cloud configuration or CUDA hardware. TODO) (Q: should I be explaining how to use supervisely in more detail? I have been reading about it but not sure)
Another way to train models is to use the process documented in tensorflow for poets. This is a great way to train a basic image classifier with various categories, which doesn't require any bounding box annotation (you label the training set by putting them in appropriate folders).
The third and perhaps most flexible way to train models is using keras, and generalising the methodology from this excellent 11-part series on keras and python.
In the tensorflow directory there are scripts to setup a ready-made environment in Docker for training models using tensorflow and keras. You can also refer to the README] for a brief explanation of how to install and use Docker.
cd tensorflow [Q: is this directory supposed to be one of the folders that are included in the github download? I don't see any directory called tensorflow at the moment; alternatively, are they supposed to download tensorflow manually?]
sh setup.sh # downloads tensorflow models and builds Docker image locally
sh run.sh # starts Docker container with appropriate volume/port mappingVisit http://localhost:8080 and use the token displayed in the console after running the last command. Tensorboard is also available at http://localhost:6006
the excellent CLI google_images_download is part of the Pipfile. to retrieve more than 100 images at a time, you need to download chromedriver, unzip, and pass the appropriate path to the binary. the suggestion is to put the binary in /usr/local/bin, and then you can copy and paste the following command to download images for a search:
googleimagesdownload --keywords "tanks" --limit 1000 --chromedriver /usr/local/bin/chromedriveranother handy tool is findimagedupes, especially if you are creating datasets by interweaving google searches (which will inevitably have overlapping images returned).