python - 如何确定更快的 RCNN (PyTorch) 的验证损失？

Question

我按照本教程进行对象检测: https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html

及其包含以下 train_one_epoch 和 evaluate 函数的 GitHub 存储库:

https://github.com/pytorch/vision/blob/main/references/detection/engine.py

但是，我想计算验证期间的损失。我为评估损失实现了这个，基本上为了获得损失，model.train() 需要打开:

@torch.no_grad()
def evaluate_loss(model, data_loader, device):
    val_loss = 0
    model.train()
    for images, targets in data_loader:
        images = list(image.to(device) for image in images)
        targets = [{k: v.to(device) for k, v in t.items()} for t in targets]

        loss_dict = model(images, targets)

        losses = sum(loss for loss in loss_dict.values())

        # reduce losses over all GPUs for logging purposes
        loss_dict_reduced = utils.reduce_dict(loss_dict)
        losses_reduced = sum(loss for loss in loss_dict_reduced.values())
        val_loss += losses_reduced
  
  validation_loss = val_loss/ len(data_loader)    
  return validation_loss

然后我将它放在我的 for 循环中的学习率调度程序步骤之后:

 for epoch in range(args.num_epochs):
        # train for one epoch, printing every 10 iterations
        train_one_epoch(model, optimizer, train_data_loader, device, epoch, print_freq=10)
    
        # update the learning rate
        lr_scheduler.step()

        validation_loss = evaluate_loss(model, valid_data_loader, device=device)

        # evaluate on the test dataset
        evaluate(model, valid_data_loader, device=device)

这看起来正确吗？它会干扰训练或产生不准确的验证损失吗？

如果可以，通过使用它，是否有一种简单的方法可以为验证损失应用提前停止？

我正在考虑在上面显示的评估模型函数之后添加类似这样的东西:

torch.save({
            'epoch': epoch,
            'model_state_dict': net.state_dict(),
            'optimizer_state_dict': optimizer.state_dict(),
            'validation loss': valid_loss,
            }, PATH)

我还打算在每个时期保存模型以用于检查点目的。但是，我需要确定保存“最佳”模型的验证“损失”。

Answer 1

因此，当 model.eval() 设置时，pytorch fasterrcnn 返回损失没有任何阶段。但是，您可以手动使用 forward 代码在评估模式下生成损失:

from typing import Tuple, List, Dict, Optional
import torch
from torch import Tensor
from collections import OrderedDict
from torchvision.models.detection.roi_heads import fastrcnn_loss
from torchvision.models.detection.rpn import concat_box_prediction_layers
def eval_forward(model, images, targets):
    # type: (List[Tensor], Optional[List[Dict[str, Tensor]]]) -> Tuple[Dict[str, Tensor], List[Dict[str, Tensor]]]
    """
    Args:
        images (list[Tensor]): images to be processed
        targets (list[Dict[str, Tensor]]): ground-truth boxes present in the image (optional)
    Returns:
        result (list[BoxList] or dict[Tensor]): the output from the model.
            It returns list[BoxList] contains additional fields
            like `scores`, `labels` and `mask` (for Mask R-CNN models).
    """
    model.eval()

    original_image_sizes: List[Tuple[int, int]] = []
    for img in images:
        val = img.shape[-2:]
        assert len(val) == 2
        original_image_sizes.append((val[0], val[1]))

    images, targets = model.transform(images, targets)

    # Check for degenerate boxes
    # TODO: Move this to a function
    if targets is not None:
        for target_idx, target in enumerate(targets):
            boxes = target["boxes"]
            degenerate_boxes = boxes[:, 2:] <= boxes[:, :2]
            if degenerate_boxes.any():
                # print the first degenerate box
                bb_idx = torch.where(degenerate_boxes.any(dim=1))[0][0]
                degen_bb: List[float] = boxes[bb_idx].tolist()
                raise ValueError(
                    "All bounding boxes should have positive height and width."
                    f" Found invalid box {degen_bb} for target at index {target_idx}."
                )

    features = model.backbone(images.tensors)
    if isinstance(features, torch.Tensor):
        features = OrderedDict([("0", features)])
    model.rpn.training=True
    #model.roi_heads.training=True


    #####proposals, proposal_losses = model.rpn(images, features, targets)
    features_rpn = list(features.values())
    objectness, pred_bbox_deltas = model.rpn.head(features_rpn)
    anchors = model.rpn.anchor_generator(images, features_rpn)

    num_images = len(anchors)
    num_anchors_per_level_shape_tensors = [o[0].shape for o in objectness]
    num_anchors_per_level = [s[0] * s[1] * s[2] for s in num_anchors_per_level_shape_tensors]
    objectness, pred_bbox_deltas = concat_box_prediction_layers(objectness, pred_bbox_deltas)
    # apply pred_bbox_deltas to anchors to obtain the decoded proposals
    # note that we detach the deltas because Faster R-CNN do not backprop through
    # the proposals
    proposals = model.rpn.box_coder.decode(pred_bbox_deltas.detach(), anchors)
    proposals = proposals.view(num_images, -1, 4)
    proposals, scores = model.rpn.filter_proposals(proposals, objectness, images.image_sizes, num_anchors_per_level)

    proposal_losses = {}
    assert targets is not None
    labels, matched_gt_boxes = model.rpn.assign_targets_to_anchors(anchors, targets)
    regression_targets = model.rpn.box_coder.encode(matched_gt_boxes, anchors)
    loss_objectness, loss_rpn_box_reg = model.rpn.compute_loss(
        objectness, pred_bbox_deltas, labels, regression_targets
    )
    proposal_losses = {
        "loss_objectness": loss_objectness,
        "loss_rpn_box_reg": loss_rpn_box_reg,
    }

    #####detections, detector_losses = model.roi_heads(features, proposals, images.image_sizes, targets)
    image_shapes = images.image_sizes
    proposals, matched_idxs, labels, regression_targets = model.roi_heads.select_training_samples(proposals, targets)
    box_features = model.roi_heads.box_roi_pool(features, proposals, image_shapes)
    box_features = model.roi_heads.box_head(box_features)
    class_logits, box_regression = model.roi_heads.box_predictor(box_features)

    result: List[Dict[str, torch.Tensor]] = []
    detector_losses = {}
    loss_classifier, loss_box_reg = fastrcnn_loss(class_logits, box_regression, labels, regression_targets)
    detector_losses = {"loss_classifier": loss_classifier, "loss_box_reg": loss_box_reg}
    boxes, scores, labels = model.roi_heads.postprocess_detections(class_logits, box_regression, proposals, image_shapes)
    num_images = len(boxes)
    for i in range(num_images):
        result.append(
            {
                "boxes": boxes[i],
                "labels": labels[i],
                "scores": scores[i],
            }
        )
    detections = result
    detections = model.transform.postprocess(detections, images.image_sizes, original_image_sizes)  # type: ignore[operator]
    model.rpn.training=False
    model.roi_heads.training=False
    losses = {}
    losses.update(detector_losses)
    losses.update(proposal_losses)
    return losses, detections

测试这段代码给我:

import torchvision
from torchvision.models.detection.faster_rcnn import FastRCNNPredictor

# load a model pre-trained on COCO
model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)

# replace the classifier with a new one, that has
# num_classes which is user-defined
num_classes = 2  # 1 class (person) + background
# get number of input features for the classifier
in_features = model.roi_heads.box_predictor.cls_score.in_features
# replace the pre-trained head with a new one
model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)
losses, detections = eval_forward(model,torch.randn([1,3,300,300]),[{'boxes':torch.tensor([[100,100,200,200]]),'labels':torch.tensor([0])}])

{'loss_classifier': tensor(0.6594, grad_fn=<NllLossBackward0>),
'loss_box_reg': tensor(0., grad_fn=<DivBackward0>),
 'loss_objectness': tensor(0.5108, grad_fn=<BinaryCrossEntropyWithLogitsBackward0>),
 'loss_rpn_box_reg': tensor(0.0160, grad_fn=<DivBackward0>)}