ubicomplab/rPPG-Toolbox

The issue of the predicted results not matching the actual results

Closed this issue · 3 comments

from main import *

parser = argparse.ArgumentParser()
parser.add_argument('--config_file', required=False,
                    default="./configs/infer_configs/PURE_UBFC-rPPG_TSCAN_BASIC.yaml",
                    type=str, help="The name of the model.")
parser = trainer.BaseTrainer.BaseTrainer.add_trainer_args(parser)
parser = data_loader.BaseLoader.BaseLoader.add_data_loader_args(parser)
args, unknown = parser.parse_known_args()
config = get_config(args)

print(config.TEST.OUTPUT_SAVE_DIR)
test_loader = data_loader.UBFCrPPGLoader.UBFCrPPGLoader
test_data = test_loader(
    name="test",
    data_path=config.TEST.DATA.DATA_PATH,
    config_data=config.TEST.DATA)
data_loader_dict = dict()
data_loader_dict["test"] = DataLoader(
    dataset=test_data,
    num_workers=16,
    batch_size=config.INFERENCE.BATCH_SIZE,
    shuffle=False,
    worker_init_fn=seed_worker,
    generator=general_generator
)
model_trainer = trainer.DeepPhysTrainer.DeepPhysTrainer(config, data_loader_dict)
model_trainer.test(data_loader_dict)

When I use the following code to test the UBFC-rPPG data and load the results saved in */saved_test_outputs/PURE_TSCAN_UBFC-rPPG_outputs.pickle, I use data_out_viz.ipynb for visualization. However, I got the following result: the values predicted by the model seem to be more than 100 times smaller than the actual values.
image

Hi @Amorth,

Is there perhaps some bug in your (I'm assuming) custom code that has snippets of this toolbox in it? For what it's worth, my visualization of the outputs (of a PURE_TSCAN model being tested on UBFC-rPPG) look like this:

image

Quite different from yours. If you're modifying numerous other parts of the toolbox (e.g., pre-processing, post-processing), it's worth noting those differences so that we can try to help you troubleshoot this.

At the very least, I'm guessing your chunk size is bit larger than the default of 180 frames, but it would be good of you to confirm any differences so that anyone trying to help can keep them in mind. Also, the outputs you are trying to visualize are PURE_TSCAN_UBFC-rPPG_outputs.pickle but your code snippet seems to be explicitly related to DeepPhys, why is that the case?

Hi @Amorth,

This issue is getting a bit stale without further discussion, so I'll go ahead and close this. Let us know (by either re-opening this issue or making a new one) if my previous replies aren't consistent with what you observed in an unmodified version of the toolbox and you still need help. We're happy to try and help you troubleshoot this once we get more details about what else you might've modified.