Merge pull request #194 from Nota-NetsPresso/177-detection-head-re-impl

Detection head re-implementation for PyNetsPresso compatibility

CHANGELOG.md

@@ -14,7 +14,7 @@ No changes to highlight.
 
 ## Other Changes:
 
-- Fix Faster R-CNN detection head by `@illian01` in [PR 184](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/184)
+- Fix Faster R-CNN detection head by `@illian01` in [PR 184](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/184), [PR 194](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/194)
 - Refactoring models/op module by `@illian01` in [PR 189](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/189), [PR 190](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/190)
 - Release NetsPresso Trainer colab tutorial by `@illian01` in [PR 191](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/191)
 - Parameterize activation function of BasicBlock and Bottleneck by `@illian01` in [PR193](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/193)

src/netspresso_trainer/dataloaders/detection/transforms.py

@@ -35,9 +35,6 @@ def val_transforms_efficientformer(conf_augmentation):
     return val_transforms_composed
 
 def create_transform_detection(model_name: str, is_training=False):
-
-    if model_name == 'efficientformer':
-        if is_training:
-            return train_transforms_efficientformer
-        return val_transforms_efficientformer
-    raise ValueError(f"No such model named: {model_name} !!!")
+    if is_training:
+        return train_transforms_efficientformer
+    return val_transforms_efficientformer

src/netspresso_trainer/metrics/detection/metric.py

@@ -171,39 +171,39 @@ class DetectionMetric(BaseMetric):
     def __init__(self, **kwargs):
         super().__init__()
 
-    def calibrate(self, pred, target, **kwargs):
+    def calibrate(self, predictions, targets, **kwargs):
         result_dict = {k: 0. for k in self.metric_names}
 
         iou_thresholds = np.linspace(0.5, 0.95, 10)
         stats = []
 
         # Gather matching stats for predictions and targets
-
-        predicted_objs_bbox, predicted_objs_class, predicted_objs_confidence = pred
-        true_objs_bbox, true_objs_class = target
-
-        true_objs = np.concatenate((true_objs_bbox, true_objs_class[..., np.newaxis]), axis=-1)
-        predicted_objs = np.concatenate((predicted_objs_bbox, predicted_objs_class[..., np.newaxis], predicted_objs_confidence[..., np.newaxis]), axis=-1)
-
-        if predicted_objs.shape[0] == 0 and true_objs.shape[0]:
-            stats.append(
-                (
-                    np.zeros((0, iou_thresholds.size), dtype=bool),
-                    *np.zeros((2, 0)),
-                    true_objs[:, 4],
+        for pred, target in zip(predictions, targets):
+            predicted_objs_bbox, predicted_objs_class, predicted_objs_confidence = pred['post_boxes'], pred['post_labels'], pred['post_scores']
+            true_objs_bbox, true_objs_class = target['boxes'], target['labels']
+
+            true_objs = np.concatenate((true_objs_bbox, true_objs_class[..., np.newaxis]), axis=-1)
+            predicted_objs = np.concatenate((predicted_objs_bbox, predicted_objs_class[..., np.newaxis], predicted_objs_confidence[..., np.newaxis]), axis=-1)
+
+            if predicted_objs.shape[0] == 0 and true_objs.shape[0]:
+                stats.append(
+                    (
+                        np.zeros((0, iou_thresholds.size), dtype=bool),
+                        *np.zeros((2, 0)),
+                        true_objs[:, 4],
+                    )
                 )
-            )
 
-        if true_objs.shape[0]:
-            matches = match_detection_batch(predicted_objs, true_objs, iou_thresholds)
-            stats.append(
-                (
-                    matches,
-                    predicted_objs[:, 5],
-                    predicted_objs[:, 4],
-                    true_objs[:, 4],
+            if true_objs.shape[0]:
+                matches = match_detection_batch(predicted_objs, true_objs, iou_thresholds)
+                stats.append(
+                    (
+                        matches,
+                        predicted_objs[:, 5],
+                        predicted_objs[:, 4],
+                        true_objs[:, 4],
+                    )
                 )
-            )
 
         # Compute average precisions if any matches exist
         if stats:

src/netspresso_trainer/models/heads/detection/experimental/detection/roi_heads.py

@@ -211,48 +211,8 @@ def postprocess_detections(
         all_labels = []
 
         # Apply Non-maximum suppression
-        # Now, it only implemented on batch size 1
-        boxes, scores, image_shape = pred_boxes_list[0], pred_scores_list[0], image_shapes
-
-        boxes = det_utils.clip_boxes_to_image(boxes, image_shape)
-
-        # create labels for each prediction
-        labels = (torch.ones_like(class_logits[0, :], dtype=torch.int64).cumsum(0) - 1).to(device)
-        labels = labels.view(1, -1).expand_as(scores)
-
-        # remove predictions with the background label
-        boxes = boxes[:, 1:]
-        scores = scores[:, 1:]
-        labels = labels[:, 1:]
-
-        # batch everything, by making every class prediction be a separate instance
-        boxes = boxes.reshape(-1, 4)
-        scores = scores.reshape(-1)
-        labels = labels.reshape(-1)
-
-        # remove low scoring boxes
-        inds = torch.where(scores > self.score_thresh)[0]
-        boxes, scores, labels = boxes[inds], scores[inds], labels[inds]
-
-        # remove empty boxes
-        keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
-        boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
-
-        # non-maximum suppression, independently done per class
-        keep = det_utils._batched_nms_vanilla(boxes, scores, labels, self.nms_thresh, self.class_ids)
-        # keep only topk scoring predictions
-        keep = keep[: self.detections_per_img]
-        boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
-
-        all_boxes.append(boxes)
-        all_scores.append(scores)
-        all_labels.append(labels)
-
-        # TODO
-        # Apply NMS on various batch
-        '''
-        for boxes, scores, image_shape in zip(pred_boxes_list, pred_scores_list, image_shapes):
-            boxes = det_utils.clip_boxes_to_image(boxes, image_shape)
+        for boxes, scores in zip(pred_boxes_list, pred_scores_list):
+            boxes = det_utils.clip_boxes_to_image(boxes, image_shapes)
 
             # create labels for each prediction
             labels = torch.arange(num_classes, device=device)
@@ -285,7 +245,6 @@ def postprocess_detections(
             all_boxes.append(boxes)
             all_scores.append(scores)
             all_labels.append(labels)
-        '''
 
         return all_boxes, all_scores, all_labels
 

src/netspresso_trainer/models/heads/detection/experimental/detection/rpn.py

@@ -235,36 +235,11 @@ def filter_proposals(
 
         final_boxes = []
         final_scores = []
-
-        # Apply Non-maximum suppression
-        # Now, it only implemented on batch size 1
-        boxes, scores, lvl, img_shape = proposals[0], objectness_prob[0], levels[0], image_shapes
-        boxes = det_utils.clip_boxes_to_image(boxes, img_shape)
-
-        # remove small boxes
-        keep = box_ops.remove_small_boxes(boxes, self.min_size)
-        boxes, scores, lvl = boxes[keep], scores[keep], lvl[keep]
-
-        # remove low scoring boxes
-        # use >= for Backwards compatibility
-        keep = torch.where(scores >= self.score_thresh)[0]
-        boxes, scores, lvl = boxes[keep], scores[keep], lvl[keep]
 
-        # non-maximum suppression, independently done per level
-        keep = det_utils._batched_nms_vanilla(boxes, scores, lvl, self.nms_thresh, list(range(len(num_anchors_per_level))))
-
-        # keep only topk scoring predictions
-        keep = keep[: self.post_nms_top_n()]
-        boxes, scores = boxes[keep], scores[keep]
-
-        final_boxes.append(boxes)
-        final_scores.append(scores)
+        # Apply Non-maximum suppression
 
-        # TODO
-        # Apply NMS on various batch
-        '''
-        for boxes, scores, lvl, img_shape in zip(proposals, objectness_prob, levels, image_shapes):
-            boxes = box_ops.clip_boxes_to_image(boxes, img_shape)
+        for boxes, scores, lvl in zip(proposals, objectness_prob, levels,):
+            boxes = box_ops.clip_boxes_to_image(boxes, image_shapes)
 
             # remove small boxes
             keep = box_ops.remove_small_boxes(boxes, self.min_size)
@@ -284,7 +259,7 @@ def filter_proposals(
 
             final_boxes.append(boxes)
             final_scores.append(scores)
-        '''
+
         return final_boxes, final_scores
 
     def forward(

src/netspresso_trainer/pipelines/detection.py

@@ -1,11 +1,16 @@
+import copy
 import logging
 import os
+from pathlib import Path
 
 import numpy as np
 import torch
 from omegaconf import OmegaConf
 
-from ..models.utils import DetectionModelOutput
+from ..models import build_model
+from ..models.utils import DetectionModelOutput, load_from_checkpoint
+from ..utils.fx import save_graphmodule
+from ..utils.onnx import save_onnx
 from .base import BasePipeline
 
 logger = logging.getLogger("netspresso_trainer")
@@ -17,6 +22,15 @@ def __init__(self, conf, task, model_name, model, devices, train_dataloader, eva
                                                 train_dataloader, eval_dataloader, class_map, **kwargs)
         self.num_classes = train_dataloader.dataset.num_classes
 
+        # Re-compose torch.fx backbone and nn.Module head
+        # To load head weights, config should have head_checkpoint value.
+        if kwargs['is_graphmodule_training']:
+            model = build_model(conf.model, task, self.num_classes, None, conf.augmentation.img_size)
+            model.backbone = self.model
+            model.head = load_from_checkpoint(model.head, conf.model.head_checkpoint)
+            model = model.to(device=devices)
+            self.model = model
+
     def train_step(self, batch):
         self.model.train()
         images, labels, bboxes = batch['pixel_values'], batch['label'], batch['bbox']
@@ -39,7 +53,7 @@ def train_step(self, batch):
 
         # generate proposals for training
         proposals = rpn_features['boxes']
-        proposals, matched_idxs, labels, regression_targets = head.roi_heads.select_training_samples(proposals, targets)
+        proposals, matched_idxs, roi_head_labels, regression_targets = head.roi_heads.select_training_samples(proposals, targets)
 
         # forward to roi head
         roi_features = head.roi_heads(features, proposals, head.image_size)
@@ -48,33 +62,34 @@ def train_step(self, batch):
         out = DetectionModelOutput()
         out.update(rpn_features)
         out.update(roi_features)
-        out.update({'labels': labels, 'regression_targets': regression_targets})
+        out.update({'labels': roi_head_labels, 'regression_targets': regression_targets})
 
+        # Compute loss
         self.loss_factory.calc(out, target=targets, phase='train')
 
         self.loss_factory.backward()
         self.optimizer.step()
 
-        # TODO: metric update
-        # out = {k: v.detach() for k, v in out.items()}
-        # self.metric_factory(out['pred'], target=targets, mode='train')
-
         if self.conf.distributed:
             torch.distributed.barrier()
 
     def valid_step(self, batch):
         self.model.eval()
         images, labels, bboxes = batch['pixel_values'], batch['label'], batch['bbox']
+        bboxes = [bbox.to(self.devices) for bbox in bboxes]
+        labels = [label.to(self.devices) for label in labels]
         images = images.to(self.devices)
-        #targets = [{"boxes": box.to(self.devices), "labels": label.to(self.devices)}
-        #           for box, label in zip(bboxes, labels)]
+        targets = [{"boxes": box, "labels": label} for box, label in zip(bboxes, labels)]
 
         out = self.model(images)
-        # TODO: compute loss for validation
-        #self.loss_factory.calc(out, target=targets, phase='valid')
 
-        # TODO: metric update
-        # self.metric_factory(out['pred'], (labels, bboxes), mode='valid')
+        # Compute loss
+        head = self.model.head
+        matched_idxs, roi_head_labels = head.roi_heads.assign_targets_to_proposals(out['boxes'], bboxes, labels)
+        matched_gt_boxes = [bbox[idx] for idx, bbox in zip(matched_idxs, bboxes)]
+        regression_targets = head.roi_heads.box_coder.encode(matched_gt_boxes, out['boxes'])
+        out.update({'labels': roi_head_labels, 'regression_targets': regression_targets})
+        self.loss_factory.calc(out, target=targets, phase='valid')
 
         if self.conf.distributed:
             torch.distributed.barrier()
@@ -102,18 +117,77 @@ def test_step(self, batch):
         return results
 
     def get_metric_with_all_outputs(self, outputs):
-        targets = np.empty((0, 4))
-        preds = np.empty((0, 5))  # with confidence score
-        targets_indices = np.empty(0)
-        preds_indices = np.empty(0)
+        pred = []
+        targets = []
         for output_batch in outputs:
             for detection, class_idx in output_batch['target']:
-                targets = np.vstack([targets, detection])
-                targets_indices = np.append(targets_indices, class_idx)
+                target_on_image = {}
+                target_on_image['boxes'] = detection
+                target_on_image['labels'] = class_idx
+                targets.append(target_on_image)
 
             for detection, class_idx in output_batch['pred']:
-                preds = np.vstack([preds, detection])
-                preds_indices = np.append(preds_indices, class_idx)
-
-        pred_bbox, pred_confidence = preds[..., :4], preds[..., -1]  # (N x 4), (N,)
-        self.metric_factory.calc((pred_bbox, preds_indices, pred_confidence), (targets, targets_indices), phase='valid')
+                pred_on_image = {}
+                pred_on_image['post_boxes'] = detection[..., :4]
+                pred_on_image['post_scores'] = detection[..., -1]
+                pred_on_image['post_labels'] = class_idx
+                pred.append(pred_on_image)
+        self.metric_factory.calc(pred, target=targets, phase='valid')
+
+    def save_checkpoint(self, epoch: int):
+
+        # Check whether the valid loss is minimum at this epoch
+        valid_losses = {epoch: record['valid_losses'].get('total') for epoch, record in self.training_history.items()
+                        if 'valid_losses' in record}
+        best_epoch = min(valid_losses, key=valid_losses.get)
+        save_best_model = best_epoch == epoch
+
+        model = self.model.module if hasattr(self.model, 'module') else self.model
+        if self.save_dtype == torch.float16:
+            model = copy.deepcopy(model).type(self.save_dtype)
+        result_dir = self.train_logger.result_dir
+        model_path = Path(result_dir) / f"{self.task}_{self.model_name}_epoch_{epoch}.ext"
+        best_model_path = Path(result_dir) / f"{self.task}_{self.model_name}_best.ext"
+        optimizer_path = Path(result_dir) / f"{self.task}_{self.model_name}_epoch_{epoch}_optimzer.pth"
+
+        if self.save_optimizer_state:
+            optimizer = self.optimizer.module if hasattr(self.optimizer, 'module') else self.optimizer
+            save_dict = {'optimizer': optimizer.state_dict(), 'start_epoch_at_one': self.start_epoch_at_one, 'last_epoch': epoch}
+            torch.save(save_dict, optimizer_path)
+            logger.debug(f"Optimizer state saved at {str(optimizer_path)}")
+
+        if self.is_graphmodule_training:
+            # Just save graphmodule checkpoint
+            torch.save(model, (model_path.parent / f"{model_path.stem}_backbone").with_suffix(".pth"))
+            logger.debug(f"PyTorch FX model saved at {(model_path.parent / f'{model_path.stem}_backbone').with_suffix('.pth')}")
+            torch.save(model.head.state_dict(), (model_path.parent / f"{model_path.stem}_head").with_suffix(".pth"))
+            logger.info(f"Detection head saved at {(model_path.parent / f'{model_path.stem}_head').with_suffix('.pth')}")
+            if save_best_model:
+                save_onnx(model, best_model_path.with_suffix(".onnx"), sample_input=self.sample_input.type(self.save_dtype))
+                logger.info(f"ONNX model converting and saved at {str(best_model_path.with_suffix('.onnx'))}")
+
+                torch.save(model.backbone, (model_path.parent / f"{best_model_path.stem}_backbone").with_suffix(".pt"))
+                logger.info(f"Best model saved at {(model_path.parent / f'{best_model_path.stem}_backbone').with_suffix('.pt')}")
+                # save head separately
+                torch.save(model.head.state_dict(), (model_path.parent / f"{best_model_path.stem}_head").with_suffix(".pth"))
+                logger.info(f"Detection head saved at {(model_path.parent / f'{best_model_path.stem}_head').with_suffix('.pth')}")
+            return
+        torch.save(model.state_dict(), model_path.with_suffix(".pth"))
+        logger.debug(f"PyTorch model saved at {str(model_path.with_suffix('.pth'))}")
+        if save_best_model:
+            torch.save(model.state_dict(), best_model_path.with_suffix(".pth"))
+            logger.info(f"Best model saved at {str(best_model_path.with_suffix('.pth'))}")
+
+            try:
+                save_onnx(model, best_model_path.with_suffix(".onnx"), sample_input=self.sample_input.type(self.save_dtype))
+                logger.info(f"ONNX model converting and saved at {str(best_model_path.with_suffix('.onnx'))}")
+
+                # fx backbone
+                save_graphmodule(model.backbone, (model_path.parent / f"{best_model_path.stem}_backbone_fx").with_suffix(".pt"))
+                logger.info(f"PyTorch FX model tracing and saved at {(model_path.parent / f'{best_model_path.stem}_backbone_fx').with_suffix('.pt')}")
+                # save head separately
+                torch.save(model.head.state_dict(), (model_path.parent / f"{best_model_path.stem}_head").with_suffix(".pth"))
+                logger.info(f"Detection head saved at {(model_path.parent / f'{best_model_path.stem}_head').with_suffix('.pth')}")
+            except Exception as e:
+                logger.error(e)
+                pass