run_validation_epoch.py

from pathlib import Path

import torch
from pytorch_lightning import Trainer

from datasets.WIDERFace import WIDERFaceDataModule
from datasets.utils import ReduceBoundingBoxes
from models import ModelMeta
from models.PoolResnet import PoolResnet
from models.Resnet import Resnet
from models.MobilenetV3Backbone import MobilenetV3Backbone

if __name__ == "__main__":
    torch.random.manual_seed(0)

    num_of_patches = nop = 15
    input_shape = (480, 480)
    filters = 128
    lr = 1e-4
    probability_threshold = 0.5
    iou_threshold = 0.01

    name = f"custom_poolresnet_{filters}_{nop}x{nop}_{input_shape[0]}x{input_shape[1]}_sam_adam_all_data"
    log_path = Path(f"logs/out_{name}_single_validate.log")
    log_path.unlink(missing_ok=True)
    model_save_path = f"./saved_models/{name}.pth"

    model = MobilenetV3Backbone(
        filters=filters,
        input_shape=(3, *input_shape),
        num_of_patches=num_of_patches,
        num_of_residual_blocks=10,
        probability_threshold=probability_threshold,
        iou_threshold=iou_threshold,
    ).cuda()

    model.summary()
    model_setup = ModelMeta(
        model=model,
        lr=lr,
        log_path=log_path,
    )

    # checkpoint = torch.load("lightning_logs/custom_poolresnet_128_10x10_480x480_sam_adam_all_data/checkpoints/epoch=69-step=112699.ckpt")
    # checkpoint = torch.load("lightning_logs/custom_resnet_64_15x15_480x480_sam_adam/checkpoints/epoch=52-step=42611.ckpt")
    # checkpoint = torch.load("lightning_logs/custom_poolresnet_64_10x10_480x480_sam_adam/checkpoints/epoch=69-step=56279.ckpt")
    # checkpoint = torch.load("lightning_logs/custom_poolresnet_128_10x10_480x480_sam_adam/checkpoints/epoch=69-step=56279.ckpt")
    checkpoint = torch.load(
        "lightning_logs/pretrained_mobilenetv3backbone_576_15x15_480x480_sam_adam_all_data/checkpoints/epoch=69-step=112699.ckpt"
    )
    model_setup.load_state_dict(checkpoint["state_dict"])
    model_setup.num_of_patches = num_of_patches

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    trainer = Trainer(
        gpus=1 if device == torch.device("cuda") else 0,
        max_epochs=70,
        precision=16,
        progress_bar_refresh_rate=1,
        num_sanity_val_steps=0,
    )
    dm = WIDERFaceDataModule(
        batch_size=8,
        input_shape=input_shape,
        num_of_patches=num_of_patches,
        shuffle=False,
    )
    dm.setup()
    trainer.test(model=model_setup, test_dataloaders=dm.val_dataloader())