CLIP+MLP Aesthetic Score Predictor, using ViT-H, based on improved-aesthetic-predictor
project by Miao Ju and hlky
Datasets used:
I had some issues with clip-retrieval, so this just computes the image embeds for every image in input-dir, and checks if the path already exists in output-dir (in case you need to resume)
usage: imageembeds.py [-h] --input-dir INPUT_DIR
--output-dir OUTPUT_DIR
[--batch-size BATCH_SIZE]
[--num_preprocess_threads NUM_PREPROCESS_THREADS]
[--gpu-id GPU_ID] [--cpu]
imageembeds.py: error: the following arguments are required: --input-dir, --output-dir
precomputed ViT-H embeds for sac+ava+logos
prepares the data for training
dataset ratings files (to use with precomputed embeds)
usage: sacavalogo.py [-h] --laion-logo-parquet LAION_LOGO_PARQUET --laion-logo-embeddings-dir LAION_LOGO_EMBEDDINGS_DIR --ava-txt AVA_TXT --ava-embeddings-dir AVA_EMBEDDINGS_DIR --sac-sqlite
SAC_SQLITE --sac-embeddings-dir SAC_EMBEDDINGS_DIR --output-dir OUTPUT_DIR [--x-only] [--y-only]
sacavalogo.py: error: the following arguments are required: --laion-logo-parquet, --laion-logo-embeddings-dir, --ava-txt, --ava-embeddings-dir, --sac-sqlite, --sac-embeddings-dir, --output-dir
use --x-only or --y-only to only prepare one part of the data, in case you want to try something different for the ratings only
usage: train_predictor.py [-h] --x-npy X_NPY --y-npy Y_NPY [--output-dir OUTPUT_DIR] [--save-name SAVE_NAME] [--batch-size BATCH_SIZE] [--lr LR] [--optimizer {adam,adamw}]
[--val-percentage VAL_PERCENTAGE] [--epochs EPOCHS] [--val-count VAL_COUNT]
train_predictor.py: error: the following arguments are required: --x-npy, --y-npy
gradio demo
usage: app.py [-h] --model-path MODEL_PATH [--device {cuda,cpu}] [--port PORT]
app.py: error: the following arguments are required: --model-path
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent