configs/yolov5/yolov5x.yaml

@@ -3,13 +3,13 @@ __BASE__: [
   './hyp.scratch-high.yaml',
 ]
 
-per_batch_size: 32  # 32 * 8 = 256
+per_batch_size: 16  # 16 * 8 = 128
 img_size: 640
-sync_bn: False
-
-# backbone/head calculate using fp16, loss fp32
-ms_amp_level: O3
-keep_loss_fp32: True
+sync_bn: True
+recompute: False
+recompute_layers: 0
+ms_amp_level: O0
+#keep_loss_fp32: True
 
 
 network:

ma-pre-start.sh

@@ -0,0 +1,4 @@
+pip install albumentations==1.0.3 -i https://pypi.tuna.tsinghua.edu.cn/simple
+pip install opencv-python==4.7.0.68 -i https://pypi.tuna.tsinghua.edu.cn/simple
+pip install opencv-python-headless==4.7.0.68 -i https://pypi.tuna.tsinghua.edu.cn/simple
+echo "ModelArts: pip install albumentations/opencv-python finish."

mindyolo/engine/enginer.py

@@ -278,7 +278,7 @@ class Enginer:
             if self.ema:
                 self.ema.update()
             self.scaler.adjust(grads_finite)
-            if not grads_finite:
+            if not grads_finite and (cur_step % self.cfg.log_interval == 0):
                 if self.cfg.overflow_still_update:
                     logger.warning(f"overflow, still update, loss scale adjust to {self.scaler.scale_value.asnumpy()}")
                 else:

mindyolo/engine/utils.py

@@ -1,7 +1,6 @@
-import random, os, yaml, glob, re
+import random, os, yaml
 import numpy as np
 from datetime import datetime
-from pathlib import Path
 
 import mindspore as ms
 from mindspore import context
@@ -84,5 +83,5 @@ def set_default(cfg):
         cfg.data.train_set = os.path.join(cfg.data_dir, cfg.data.train_set)
         cfg.data.val_set = os.path.join(cfg.data_dir, cfg.data.val_set)
         cfg.data.test_set = os.path.join(cfg.data_dir, cfg.data.test_set)
-        cfg.weight = cfg.ckpt_dir if cfg.ckpt_dir else ''  # os.path.join(cfg.ckpt_dir, cfg.weight) if cfg.weight else ''
+        cfg.weight = cfg.ckpt_dir if cfg.ckpt_dir else ''
         cfg.ema_weight = os.path.join(cfg.ckpt_dir, cfg.ema_weight) if cfg.ema_weight else ''

mindyolo/models/layers/__init__.py

@@ -11,7 +11,7 @@ from .upsample import *
 __all__ = ['Swish',
            'Shortcut', 'Concat', 'ReOrg', 'Identity',
            'ConvNormAct', 'RepConv', 'DownC',
-           'Bottleneck', 'C3',
+           'Bottleneck', 'C3', 'C2f',
            'ImplicitA', 'ImplicitM',
            'MP', 'SP', 'MaxPool2d',
            'SPPCSPC', 'SPPF',

mindyolo/models/layers/bottleneck.py

@@ -5,20 +5,18 @@ from .conv import ConvNormAct
 
 class Bottleneck(nn.Cell):
     # Standard bottleneck
-    # ch_in, ch_out, shortcut, groups, expansion
-    def __init__(self, c1, c2, shortcut=True, e=0.5, momentum=0.97, eps=1e-3, sync_bn=False):
-        super(Bottleneck, self).__init__()
+    def __init__(self, c1, c2, shortcut=True, k=(1, 3), g=(1, 1), e=0.5, momentum=0.97, eps=1e-3, sync_bn=False):  # ch_in, ch_out, shortcut, groups, kernels, expand
+        super().__init__()
         c_ = int(c2 * e)  # hidden channels
-        self.conv1 = ConvNormAct(c1, c_, 1, 1, momentum=momentum, eps=eps, sync_bn=sync_bn)
-        self.conv2 = ConvNormAct(c_, c2, 3, 1, momentum=momentum, eps=eps, sync_bn=sync_bn)
+        self.conv1 = ConvNormAct(c1, c_, k[0], 1, g=g[0], momentum=momentum, eps=eps, sync_bn=sync_bn)
+        self.conv2 = ConvNormAct(c_, c2, k[1], 1, g=g[1], momentum=momentum, eps=eps, sync_bn=sync_bn)
         self.add = shortcut and c1 == c2
 
     def construct(self, x):
-        c1 = self.conv1(x)
-        c2 = self.conv2(c1)
-        out = c2
         if self.add:
-            out = x + out
+            out = x + self.conv2(self.conv1(x))
+        else:
+            out = self.conv2(self.conv1(x))
         return out
 
 
@@ -31,7 +29,7 @@ class C3(nn.Cell):
         self.conv2 = ConvNormAct(c1, c_, 1, 1, momentum=momentum, eps=eps, sync_bn=sync_bn)
         self.conv3 = ConvNormAct(2 * c_, c2, 1, momentum=momentum, eps=eps, sync_bn=sync_bn)  # act=FReLU(c2)
         self.m = nn.SequentialCell(
-            [Bottleneck(c_, c_, shortcut, e=1.0, momentum=momentum, eps=eps, sync_bn=sync_bn) for _ in range(n)])
+            [Bottleneck(c_, c_, shortcut, k=(1, 3), e=1.0, momentum=momentum, eps=eps, sync_bn=sync_bn) for _ in range(n)])
         self.concat = ops.Concat(axis=1)
 
     def construct(self, x):
@@ -42,3 +40,27 @@ class C3(nn.Cell):
         c5 = self.conv3(c4)
 
         return c5
+
+
+class C2f(nn.Cell):
+    # CSP Bottleneck with 2 convolutions
+    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5, momentum=0.97, eps=1e-3, sync_bn=False):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        _c = int(c2 * e)  # hidden channels
+        self.cv1 = ConvNormAct(c1, 2 * _c, 1, 1, momentum=momentum, eps=eps, sync_bn=sync_bn)
+        self.cv2 = ConvNormAct((2 + n) * _c, c2, 1, momentum=momentum, eps=eps, sync_bn=sync_bn)  # optional act=FReLU(c2)
+        self.m = nn.CellList([
+            Bottleneck(_c, _c, shortcut, k=(3, 3), g=(1, g), e=1.0, momentum=momentum, eps=eps, sync_bn=sync_bn) for _ in range(n)
+        ])
+
+    def construct(self, x):
+        y = ()
+        x = self.cv1(x)
+        x_tuple = ops.split(x, axis=1, output_num=2)
+        y += x_tuple
+        for i in range(len(self.m)):
+            m = self.m[i]
+            out = m(y[-1])
+            y += (out,)
+
+        return self.cv2(ops.concat(y, axis=1))

mindyolo/models/model_factory.py

@@ -106,15 +106,15 @@ def parse_model(d, ch, nc, sync_bn=False):  # model_dict, input_channels(3)
                 pass
 
         n = max(round(n * gd), 1) if n > 1 else n  # depth gain
-        if m in (nn.Conv2d, ConvNormAct, RepConv, DownC, SPPCSPC, SPPF, C3, Bottleneck):
+        if m in (nn.Conv2d, ConvNormAct, RepConv, DownC, SPPCSPC, SPPF, C3, C2f, Bottleneck):
             c1, c2 = ch[f], args[0]
             if c2 != no:  # if not output
                 c2 = math.ceil(c2 * gw / 8) * 8
 
             args = [c1, c2, *args[1:]]
-            if m in (ConvNormAct, RepConv, DownC, SPPCSPC, SPPF, C3, Bottleneck):
+            if m in (ConvNormAct, RepConv, DownC, SPPCSPC, SPPF, C3, C2f, Bottleneck):
                 kwargs["sync_bn"] = sync_bn
-            if m in (DownC, SPPCSPC, C3):
+            if m in (DownC, SPPCSPC, C3, C2f):
                 args.insert(2, n)  # number of repeats
                 n = 1
         elif m in (nn.BatchNorm2d, nn.SyncBatchNorm):

mindyolo/utils/config.py

@@ -86,7 +86,7 @@ def get_args_train(parents=None):
     parser.add_argument('--accumulate', type=int, default=1,
                         help='grad accumulate step, recommended when batch-size is less than 64')
     parser.add_argument('--auto_accumulate', type=ast.literal_eval, default=False, help='auto accumulate')
-    parser.add_argument('--log_interval', type=int, default=1, help='log interval')
+    parser.add_argument('--log_interval', type=int, default=100, help='log interval')
     parser.add_argument('--multi_scale', type=ast.literal_eval, default=False, help='vary img-size +/- 50%')
     parser.add_argument('--single_cls', type=ast.literal_eval, default=False,
                         help='train multi-class data as single-class')
@@ -127,7 +127,7 @@ def get_args_test(parents=None):
     parser.add_argument('--single_cls', type=ast.literal_eval, default=False,
                         help='train multi-class data as single-class')
     parser.add_argument('--rect', type=ast.literal_eval, default=False, help='rectangular training')
-    parser.add_argument('--nms_time_limit', type=float, default=20.0, help='time limit for NMS')
+    parser.add_argument('--nms_time_limit', type=float, default=60.0, help='time limit for NMS')
     parser.add_argument('--recompute', type=ast.literal_eval, default=False, help='Recompute')
     parser.add_argument('--recompute_layers', type=int, default=0)
     parser.add_argument('--conf_thres', type=float, default=0.001, help='object confidence threshold')
@@ -151,7 +151,7 @@ def get_args_test(parents=None):
                         help='Whether to run eval during training')
     parser.add_argument('--multi_scale', type=ast.literal_eval, default=False, help='vary img-size +/- 50%')
     parser.add_argument('--overflow_still_update', type=ast.literal_eval, default=False, help='overflow still update')
-    parser.add_argument('--log_interval', type=int, default=1, help='log interval')
+    parser.add_argument('--log_interval', type=int, default=100, help='log interval')
     parser.add_argument('--freeze', type=list, default=[],
                         help='Freeze layers: backbone of yolov7=50, first3=0 1 2')
 
@@ -199,7 +199,7 @@ def get_args_infer(parents=None):
                         help='Whether to run eval during training')
     parser.add_argument('--multi_scale', type=ast.literal_eval, default=False, help='vary img-size +/- 50%')
     parser.add_argument('--overflow_still_update', type=ast.literal_eval, default=False, help='overflow still update')
-    parser.add_argument('--log_interval', type=int, default=1, help='log interval')
+    parser.add_argument('--log_interval', type=int, default=100, help='log interval')
     parser.add_argument('--freeze', type=list, default=[],
                         help='Freeze layers: backbone of yolov7=50, first3=0 1 2')
     parser.add_argument('--log_level', type=str, default='INFO', help='save dir')
@@ -207,7 +207,7 @@ def get_args_infer(parents=None):
     parser.add_argument('--conf_thres', type=float, default=0.001, help='object confidence threshold')
     parser.add_argument('--iou_thres', type=float, default=0.65, help='IOU threshold for NMS')
     parser.add_argument('--rect', type=ast.literal_eval, default=False, help='rectangular training')
-    parser.add_argument('--nms_time_limit', type=float, default=20.0, help='time limit for NMS')
+    parser.add_argument('--nms_time_limit', type=float, default=60.0, help='time limit for NMS')
     parser.add_argument('--single_cls', type=ast.literal_eval, default=False,
                         help='train multi-class data as single-class')
 

requirements.txt

@@ -0,0 +1,9 @@
+# MindYOLO requirements
+# Usage: pip install -r requirements.txt
+
+# Base
+opencv-python-headless>=4.7.0.68
+
+# Extras
+albumentations>=1.0.3
+pycocotools>=2.0.2