update the accuracy

Author: chenjun2hao

README.md

@@ -39,7 +39,7 @@ download the pretrained model on imagenet or the segmentation model from the [of
 
 ## VAL
 
-use the [official pretrained model](https://github.com/ydhongHIT/DDRNet) and our `eval.py` code. so the result may different from official.
+use the [official pretrained model](https://github.com/ydhongHIT/DDRNet) and our `eval.py` code. with [ydhongHIT's](https://github.com/ydhongHIT) advice now can reach the same accuracy in the paper. Thanks.
 
 ```python
 cd ${PROJECT}
@@ -48,14 +48,14 @@ python tools/eval.py --cfg experiments/cityscapes/ddrnet23_slim.yaml
 
 | model | Train Set | Test Set | OHEM | Multi-scale| Flip | mIoU | Link |
 | :--: | :--: | :--: | :--: | :--: | :--: | :--: | :--: |
-| DDRNet23_slim | unknown | eval | Yes | No | No | 76.83 | [official](https://github.com/ydhongHIT/DDRNet) |
-| DDRNet23_slim | unknown | eval | Yes | No | Yes| 77.40 | [official](https://github.com/ydhongHIT/DDRNet) |
-| DDRNet23      | unknown | eval | Yes | No | No | 78.41 | [official](https://github.com/ydhongHIT/DDRNet) |
-| DDRNet23      | unknown | eval | Yes | No | Yes| 78.85 | [official](https://github.com/ydhongHIT/DDRNet) |
+| DDRNet23_slim | unknown | eval | Yes | No | No | 77.83 | [official](https://github.com/ydhongHIT/DDRNet) |
+| DDRNet23_slim | unknown | eval | Yes | No | Yes| 78.42 | [official](https://github.com/ydhongHIT/DDRNet) |
+| DDRNet23      | unknown | eval | Yes | No | No | 79.51 | [official](https://github.com/ydhongHIT/DDRNet) |
+| DDRNet23      | unknown | eval | Yes | No | Yes| 79.98 | [official](https://github.com/ydhongHIT/DDRNet) |
 
 
 **Note**
-- [the official repository](https://github.com/ydhongHIT/DDRNet) on DDRNet23_slim without Flip can reach 77.40 on cityscapes dataset
+- with the `ALIGN_CORNERS: false` in `***.yaml` will reach higher accuracy.
 
 
 ## TRAIN
@@ -78,6 +78,7 @@ python -m torch.distributed.launch --nproc_per_node=2 tools/train.py --cfg exper
 | DDRNet39      | train | eval | Yes | No | Yes | ~ | None |
 
 **Note**
+- set the `ALIGN_CORNERS: true` in `***.yaml`, because i use the default setting in [HRNet-Semantic-Segmentation OCR](https://github.com/HRNet/HRNet-Semantic-Segmentation/tree/HRNet-OCR).
 - Multi-scale with scales: 0.5,0.75,1.0,1.25,1.5,1.75. it runs too slow.
 - from [ydhongHIT](https://github.com/ydhongHIT), can change the `align_corners=True` with better performance, the default option is `False`
 

experiments/cityscapes/ddrnet23.yaml

@@ -18,6 +18,7 @@ MODEL:
   NAME: ddrnet_23
   NUM_OUTPUTS: 2
   PRETRAINED: "pretrained_models/DDRNet23_imagenet.pth"
+  ALIGN_CORNERS: false
 LOSS:
   USE_OHEM: true
   OHEMTHRES: 0.9
@@ -49,7 +50,7 @@ TEST:
   - 1024
   BASE_SIZE: 2048
   BATCH_SIZE_PER_GPU: 4
-  FLIP_TEST: false
+  FLIP_TEST: true
   MULTI_SCALE: false
   MODEL_FILE: "pretrained_models/best_val.pth"
   OUTPUT_INDEX: 0

experiments/cityscapes/ddrnet23_slim.yaml

@@ -18,6 +18,7 @@ MODEL:
   NAME: ddrnet_23_slim
   NUM_OUTPUTS: 2
   PRETRAINED: "pretrained_models/DDRNet23s_imagenet.pth"
+  ALIGN_CORNERS: false
 LOSS:
   USE_OHEM: true
   OHEMTHRES: 0.9
@@ -49,9 +50,9 @@ TEST:
   - 1024
   BASE_SIZE: 2048
   BATCH_SIZE_PER_GPU: 4
-  FLIP_TEST: true
+  FLIP_TEST: false
   MULTI_SCALE: false
   SCALE_LIST: [1]
   #0.5,0.75,1.0,1.25,1.5,1.75
-  MODEL_FILE: "pretrained_models/best_val_smaller.pth"
+  # MODEL_FILE: "pretrained_models/best_val_smaller.pth"
   OUTPUT_INDEX: 0

tools/to_onnx.py

@@ -27,7 +27,7 @@ def parse_args():
 
     parser.add_argument('--cfg',
                         help='experiment configure file name',
-                        default="experiments/map/map_hrnet_ocr_w18_small_v2_512x1024_sgd_lr1e-2_wd5e-4_bs_12_epoch484.yaml",
+                        default="experiments/cityscapes/ddrnet23_slim.yaml",
                         type=str)
     parser.add_argument('opts',
                         help="Modify config options using the command-line",
@@ -46,7 +46,7 @@ def __init__(self, model):
 
     def forward(self, x):
         x1, x2 = self.backone(x)
-        y = F.interpolate(x2, size=(480,640), mode='bilinear')
+        y = F.interpolate(x1, size=(480,640), mode='bilinear')
         # y = F.softmax(y, dim=1)
         y = torch.argmax(y, dim=1)
 
@@ -94,18 +94,19 @@ def main():
     model.load_state_dict(model_dict)
 
     net = onnx_net(model)
+    net = net.eval()
 
     # x = torch.randn((1, 3, 512, 384))
     x = torch.randn((1,3,480,640))
     torch_out = net(x)
 
     # output_path = "output/tensorrt/resnet50/resnet50_bilinear.onnx"
-    output_path = "output/hrnetv2_w18_480_640.onnx"
+    output_path = "output/ddrnet23_slim.onnx"
     torch.onnx.export(net,               # model being run
                     x,                         # model input (or a tuple for multiple inputs)
                     output_path,   # where to save the model (can be a file or file-like object)
                     export_params=True,        # store the trained parameter weights inside the model file
-                    opset_version=9,          # the ONNX version to export the model to
+                    opset_version=11,          # the ONNX version to export the model to
                     do_constant_folding=True,  # whether to execute constant folding for optimization
                     input_names = ['inputx'],   # the model's input names
                     output_names = ['outputy'], # the model's output names