see the blog post for more details
objs
directory from https://github.com/matpalm/procedural_objects/ will get you started./generate_toy_rgb_data.py
./run_random_grasps.py --run 1 --num-cameras 4 --max-frames-to-render 10 &
./run_random_grasps.py --run 2 --num-cameras 4 --max-frames-to-render 10 &
wait
display a 5x5 sample of random frames
./debug_random_frames.py
display a sample of images across frames / cameras columns are a frame in time, rows are specific cameras
./debug_frame_sequence_cameras.py --run 1 --initial-frame 1 --num-frames 4 --cameras 0,1,2,3
show all camera images for a specific frame
./debug_frame.py --run 1 --frame 10
show random (anchor, positive, negative) triples that will be used for training
./debug_random_triples.py
./train.py --help
--img-dir IMG_DIR
--batch-size BATCH_SIZE
note: effective batch is x3 (default: 16)
--embedding-dim EMBEDDING_DIM
image embedding dim (default: 8)
--learning-rate LEARNING_RATE
learning rate for adam (default: 0.001)
--epochs EPOCHS
--steps-per-epoch STEPS_PER_EPOCH
--model-output MODEL_OUTPUT
where to save model (default: model)
calc embeddings for all images
./embed_imgs.py
debug near neighbours
./debug_embedding_near_neighbours.py