A neural network that creates hashtag recommendations based off of the contents of an image and how similar it is to what the model has seen before.
You will need
- python
- keras
- tensorflow
- cv2
- numpy
To scrape the data from Instagram simply run the scraper.py file. This generates multiple npy files containing lists of the form (id, image_name, hashtag_vector, embedding).
You will need to put the numpy files together such that the id's are contiguous, as they start at 0 for each file. Allocate a portion to be test_csvs.npy and train_csvs.npy. It also creates input.npy files, which map to the same ids as in the csv files. These must be allocated the same way to test_inputs.npy and train_inputs.npy.
Simply run hashtag_generation.py to run the model and the prediction set on test_csvs.npy.
Results are printed in the form
Model bs: 16 / lr: 0.01 / f1: 0.07683573050719152 / pre: 0.07120308663626798 / re: 0.083436087135224
Model bs: 16 / lr: 0.001 / f1: 0.038293216630196934 / pre: 0.05723630417007359 / re: 0.02877106452938759
Model bs: 16 / lr: 0.0001 / f1: 0.035753575357535754 / pre: 0.05403158769742311 / re: 0.026715988491574187
Model bs: 64 / lr: 0.01 / f1: 0.07083782811938152 / pre: 0.06295941195270055 / re: 0.08096999588984792
Model bs: 64 / lr: 0.001 / f1: 0.034636871508379886 / pre: 0.05405405405405406 / re: 0.02548294286888615
Model bs: 64 / lr: 0.0001 / f1: 0.03537338573834924 / pre: 0.05580159433126661 / re: 0.025893958076448828
Model bs: 256 / lr: 0.01 / f1: 0.07895595432300163 / pre: 0.06545847984852583 / re: 0.09946568023016851
Model bs: 256 / lr: 0.001 / f1: 0.035854341736694675 / pre: 0.056288478452066845 / re: 0.026304973284011508
Model bs: 256 / lr: 0.0001 / f1: 0.036211699164345405 / pre: 0.056179775280898875 / re: 0.026715988491574187
Where each line indicates a model and its performance in terms of batch size / learn rate / f1 score / precision / recall.