Skip to content

Latest commit

 

History

History
84 lines (59 loc) · 3.31 KB

3d_compat.md

File metadata and controls

84 lines (59 loc) · 3.31 KB
Raw

Instructions for running SIRI/PLAD on 3D Compat challenge dataset

The Visual Shape Inference Challenge is being organized as a part of the ...

Loading the Dataset

Please download the data from the Challenge website. Save the data in the following format:

data_dir/
  | -- 3d_csg/...
  | -- 3DCoMPaT/
  | ----- | -- voxel_train.h5
  | ----- | -- voxel_val.h5
  | ----- | -- voxel_test.h5

Installation

Please follow the instruction in README.md for installation.

Training

  1. Pretrain the model on synthetic data. This will run the pretrain, and the best model will be saved at project_dir/models/pretrain/best_model.pt.
python scripts/train.py --config-file configs/pretrain.py --name pretrain
  1. Finetune with SIRI: Let $model_path denote the path of the best pretrained model from step 1.
python scripts/train.py --config-file configs/siri.py --name siri --cfg.plad.starting_weights $model_path --cfg.ws_config.starting_weights $model_path --target 3DCoMPaT

Inference

Inference (with rewriting) depends on two pickle files which are generated during the training. $model_path which is the "best" model on the validation set, and $subexpr_cache_path which are all the subexpressions discovered during the code grafting process (The previous command would save it as project_dir/models/siri/all_subexpr.pkl). With these two files, you can run inference on the test set using:

python scripts/eval_sequential.py --config-file configs/eval.py --name eval --cfg.siri.rewriters.CGRewriter.cache_config.subexpr_load_path $subexpr_cache_path --cfg.trainer.load_weights $model_path --target 3DCoMPaT

Note that we run this process sequentially (no batched beam search or multiple processes for PO Rewriter) so that we can measure the inference time of the system. The average per-shape inference time will be printed at the end of the evaluation. This will save all the expressions in a pickle file project_dir/logs/eval/final_programs.pkl. This will contain the expressions containing no parameter primitive expressions such as NoParamCuboid3D and NoParamSphere3D. You can map them to the challenge language as follows:

import geolipi.symbolic as gls 
import _pickle as cPickle
mapper = {
    gls.NoParamCuboid3D: (gls.Cuboid3D, (0.5, 0.5, 0.5)),
    gls.NoParamSphere3D: (gls.Sphere3D, (0.5,)),
}

def remap(expr):
    if isinstance(expr, gls.GLFunction):
        args = []
        for arg in expr.args:
            args.append(remap(arg))
        expr_cls = expr.__class__
        if expr_cls in mapper:
            cur_expr, param = mapper[expr_cls]
            args.append(param)
        else:
            cur_expr = expr.__class__
        return cur_expr(*args)
    else:
        return expr
  
    

expr_file = "final_programs.pkl"
save_file = "converted.pkl"

expressions = cPickle.load(open(expr_file, "rb"))
expressions = [remap(x) for x in expressions]

cPickle.dump(expressions, open(save_file, "wb"))

Submit the file to the evaluation server here. Remeber to report the measured per shape inference time as well.

Note: You can run the process on the validation set to generate the programs for the val set, and use the evaluation scripts provided in the VSIC repository to measure the evaluation metrics on the val set.