performance and validation opt

micromind/core.py

@@ -5,6 +5,7 @@
 Authors:
     - Francesco Paissan, 2023
 """
+
 from abc import ABC, abstractmethod
 from argparse import Namespace
 from dataclasses import dataclass
@@ -534,12 +535,13 @@ def train(
                     loss_epoch += loss.item()
 
                 self.accelerator.backward(loss)
-                self.accelerator.clip_grad_norm_(self.modules.parameters(), max_norm=max_norm)
+                self.accelerator.clip_grad_norm_(
+                    self.modules.parameters(), max_norm=max_norm
+                )
                 self.opt.step()
 
                 loss_epoch += loss.item()
 
-
                 for m in self.metrics:
                     if (
                         self.current_epoch + 1
@@ -577,10 +579,7 @@ def train(
 
             self.on_train_epoch_end()
 
-            if (
-                self.accelerator.is_local_main_process
-                and self.checkpointer is not None
-            ):
+            if self.accelerator.is_local_main_process and self.checkpointer is not None:
                 self.checkpointer(
                     self,
                     train_metrics,

micromind/networks/yolo.py

@@ -527,6 +527,7 @@ def __init__(
                 skip_tensor_in=False,
             )
 
+
 class Yolov8NeckOpt_gamma2(Yolov8Neck):
     def __init__(
         self, filters=[256, 512, 768], up=[2, 2], heads=[True, True, True], d=1
@@ -536,9 +537,6 @@ def __init__(
         self.up1 = Upsample(up[0], mode="nearest")
         self.up2 = Upsample(up[1], mode="nearest")
 
-        # print(filters, heads)
-        # breakpoint()
-
         self.n1 = XiConv(
             c_in=int(filters[1] + filters[2]),
             c_out=int(filters[1]),
@@ -617,7 +615,7 @@ def __init__(self, nc=80, filters=(), heads=[True, True, True]):
         super().__init__()
         self.reg_max = 16
         self.nc = nc
-        #filters = [f for f, h in zip(filters, heads) if h]
+        # filters = [f for f, h in zip(filters, heads) if h]
         self.nl = len(filters)
         self.no = nc + self.reg_max * 4
         self.stride = torch.tensor([8.0, 16.0, 32.0], dtype=torch.float16)
@@ -700,10 +698,12 @@ def __init__(self, w, r, d, num_classes=80, heads=[True, True, True]):
         super().__init__()
         self.net = Darknet(w, r, d)
         self.fpn = Yolov8Neck(
-            filters=[int(256 * w), int(512 * w), int(512 * w * r)], heads=heads,  d=d
+            filters=[int(256 * w), int(512 * w), int(512 * w * r)], heads=heads, d=d
         )
         self.head = DetectionHead(
-            num_classes, filters=(int(256 * w), int(512 * w), int(512 * w * r)), heads=heads
+            num_classes,
+            filters=(int(256 * w), int(512 * w), int(512 * w * r)),
+            heads=heads,
         )
 
     def forward(self, x):

recipes/object_detection/README.md

@@ -1,5 +1,6 @@
 ## Object Detection using YOLO
 
+**[16 Jan 2024]** Updated training code for better performance. Added ultralytics metrics calculation .<br />
 **[16 Jan 2024]** Added optimized YOLO neck, using XiConv. Fixed compatibility with ultralytics weights.<br />
 **[17 Dec 2023]** Add VOC dataset, selective head option, and instructions for dataset download.<br />
 **[1 Dec 2023]** Fix DDP handling and computational graph.

recipes/object_detection/cfg/data/VOC.yaml

@@ -43,7 +43,7 @@ mixup: 0.0  # (float) image mixup (probability)
 copy_paste: 0.0  # (float) segment copy-paste (probability)
 
 # Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]
-path: /voc
+path: datasets/VOC
 train: # train images (relative to 'path')  16551 images
   - images/train2012
   - images/train2007

recipes/object_detection/cfg/data/coco.yaml

@@ -44,7 +44,7 @@ copy_paste: 0.0  # (float) segment copy-paste (probability)
 
 
 # Dataset location
-path: /work/code/maybe_eccv/recipes/object_detection/data/coco  # dataset root dir
+path: datasets/coco  # dataset root dir
 train: train2017.txt  # train images (relative to 'path') 118287 images
 val: val2017.txt  # val images (relative to 'path') 5000 images
 test: test-dev2017.txt  # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794

recipes/object_detection/cfg/data/coco8.yaml

@@ -44,7 +44,7 @@ copy_paste: 0.0  # (float) segment copy-paste (probability)
 
 
 # Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]
-path: /mnt/data/coco8  # dataset root dir
+path: datasets/coco8  # dataset root dir
 train: images/train  # train images (relative to 'path') 4 images
 val: images/val  # val images (relative to 'path') 4 images
 test:  # test images (optional)