refine code

ppcls/configs/Pedestrian/strong_baseline_baseline.yaml

@@ -71,6 +71,7 @@ DataLoader:
           - ResizeImage:
               size: [128, 256]
               return_numpy: False
+              backend: "pil"
           - RandFlipImage:
               flip_code: 1
           - Pad:
@@ -101,6 +102,7 @@ DataLoader:
           - ResizeImage:
               size: [128, 256]
               return_numpy: False
+              backend: "pil"
           - ToTensor:
           - Normalize:
               mean: [0.485, 0.456, 0.406]
@@ -124,6 +126,7 @@ DataLoader:
           - ResizeImage:
               size: [128, 256]
               return_numpy: False
+              backend: "pil"
           - ToTensor:
           - Normalize:
               mean: [0.485, 0.456, 0.406]

ppcls/configs/Pedestrian/strong_baseline_m1.yaml

@@ -90,6 +90,7 @@ DataLoader:
           - ResizeImage:
               size: [128, 256]
               return_numpy: False
+              backend: "pil"
           - RandFlipImage:
               flip_code: 1
           - Pad:
@@ -126,6 +127,7 @@ DataLoader:
           - ResizeImage:
               size: [128, 256]
               return_numpy: False
+              backend: "pil"
           - ToTensor:
           - Normalize:
               mean: [0.485, 0.456, 0.406]
@@ -149,6 +151,7 @@ DataLoader:
           - ResizeImage:
               size: [128, 256]
               return_numpy: False
+              backend: "pil"
           - ToTensor:
           - Normalize:
               mean: [0.485, 0.456, 0.406]

ppcls/configs/Pedestrian/strong_baseline_m1_centerloss.yaml

@@ -64,7 +64,7 @@ Loss:
         weight: 0.0005
         num_classes: *class_num
         feat_dim: *feat_dim
-        feat_from: "backbone"
+        feature_from: "backbone"
   Eval:
     - CELoss:
         weight: 1.0
@@ -101,6 +101,7 @@ DataLoader:
           - ResizeImage:
               size: [128, 256]
               return_numpy: False
+              backend: "pil"
           - RandFlipImage:
               flip_code: 1
           - Pad:
@@ -137,6 +138,7 @@ DataLoader:
           - ResizeImage:
               size: [128, 256]
               return_numpy: False
+              backend: "pil"
           - ToTensor:
           - Normalize:
               mean: [0.485, 0.456, 0.406]
@@ -160,6 +162,7 @@ DataLoader:
           - ResizeImage:
               size: [128, 256]
               return_numpy: False
+              backend: "pil"
           - ToTensor:
           - Normalize:
               mean: [0.485, 0.456, 0.406]

ppcls/data/preprocess/__init__.py

@@ -25,6 +25,7 @@ from ppcls.data.preprocess.ops.operators import DecodeImage
 from ppcls.data.preprocess.ops.operators import ResizeImage
 from ppcls.data.preprocess.ops.operators import CropImage
 from ppcls.data.preprocess.ops.operators import RandCropImage
+from ppcls.data.preprocess.ops.operators import RandCropImageV2
 from ppcls.data.preprocess.ops.operators import RandFlipImage
 from ppcls.data.preprocess.ops.operators import NormalizeImage
 from ppcls.data.preprocess.ops.operators import ToCHWImage

ppcls/data/preprocess/ops/operators.py

@@ -63,6 +63,8 @@ class UnifiedResize(object):
                 resample = random.choice(resample)
             if isinstance(src, np.ndarray):
                 pil_img = Image.fromarray(src)
+            else:
+                pil_img = src
             pil_img = pil_img.resize(size, resample)
             if return_numpy:
                 return np.asarray(pil_img)

ppcls/engine/engine.py

@@ -254,8 +254,9 @@ class Engine(object):
         world_size = dist.get_world_size()
         self.config["Global"]["distributed"] = world_size != 1
         if self.mode == "train":
-            std_gpu_num = 8 if self.config["Optimizer"][
+            std_gpu_num = 8 if isinstance(
-                "name"] == "AdamW" else 4
+                self.config["Optimizer"],
+                dict) and self.config["Optimizer"]["name"] == "AdamW" else 4
             if world_size != std_gpu_num:
                 msg = f"The training strategy provided by PaddleClas is based on {std_gpu_num} gpus. But the number of gpu is {world_size} in current training. Please modify the stategy (learning rate, batch size and so on) if use this config to train."
                 logger.warning(msg)