Merge branch 'develop' into ConvNeXt

ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_base_224.yaml

@@ -42,11 +42,12 @@ Optimizer:
   no_weight_decay_name: pos_embed cls_token .bias norm 
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 1.25e-4
+    learning_rate: 2.5e-4
-    eta_min: 1.25e-6
+    eta_min: 2.5e-6
     warmup_epoch: 20
-    warmup_start_lr: 1.25e-7
+    warmup_start_lr: 2.5e-7
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_base_384.yaml

@@ -42,11 +42,12 @@ Optimizer:
   no_weight_decay_name: pos_embed cls_token .bias norm 
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 6.25e-5
+    learning_rate: 1.25e-4
-    eta_min: 6.25e-7
+    eta_min: 1.25e-6
     warmup_epoch: 20
-    warmup_start_lr: 6.25e-8
+    warmup_start_lr: 1.25e-7
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_large_224.yaml

@@ -42,11 +42,12 @@ Optimizer:
   no_weight_decay_name: pos_embed cls_token .bias norm 
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 1.25e-4
+    learning_rate: 2.5e-4
-    eta_min: 1.25e-6
+    eta_min: 2.5e-6
     warmup_epoch: 20
-    warmup_start_lr: 1.25e-7
+    warmup_start_lr: 2.5e-7
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_large_384.yaml

@@ -42,11 +42,12 @@ Optimizer:
   no_weight_decay_name: pos_embed cls_token .bias norm 
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 3.125e-5
+    learning_rate: 6.25e-5
-    eta_min: 3.125e-7
+    eta_min: 6.25e-7
     warmup_epoch: 20
-    warmup_start_lr: 3.125e-8
+    warmup_start_lr: 6.25e-8
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_small_224.yaml

@@ -42,11 +42,12 @@ Optimizer:
   no_weight_decay_name: pos_embed cls_token .bias norm 
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 2.5e-4
+    learning_rate: 5e-4
-    eta_min: 2.5e-6
+    eta_min: 5e-6
     warmup_epoch: 20
-    warmup_start_lr: 2.5e-7
+    warmup_start_lr: 5e-7
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_tiny_224.yaml

@@ -42,11 +42,12 @@ Optimizer:
   no_weight_decay_name: pos_embed cls_token .bias norm 
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 5e-6
+    eta_min: 1e-5
     warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/DeiT/DeiT_base_distilled_patch16_224.yaml

@@ -40,11 +40,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token pos_embed dist_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 1e-3
+    learning_rate: 2e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 # data loader for train and eval
 DataLoader:

ppcls/configs/ImageNet/DeiT/DeiT_base_distilled_patch16_384.yaml

@@ -40,11 +40,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token pos_embed dist_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 1e-3
+    learning_rate: 2e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 # data loader for train and eval
 DataLoader:

ppcls/configs/ImageNet/DeiT/DeiT_base_patch16_224.yaml

@@ -40,11 +40,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token pos_embed dist_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 1e-3
+    learning_rate: 2e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 # data loader for train and eval
 DataLoader:

ppcls/configs/ImageNet/DeiT/DeiT_base_patch16_384.yaml

@@ -40,11 +40,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token pos_embed dist_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 1e-3
+    learning_rate: 2e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 # data loader for train and eval
 DataLoader:

ppcls/configs/ImageNet/DeiT/DeiT_small_distilled_patch16_224.yaml

@@ -41,10 +41,10 @@ Optimizer:
   one_dim_param_no_weight_decay: True
   lr:
     name: Cosine
-    learning_rate: 1e-3
+    learning_rate: 2e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 # data loader for train and eval
 DataLoader:

ppcls/configs/ImageNet/DeiT/DeiT_small_patch16_224.yaml

@@ -40,11 +40,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token pos_embed dist_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 1e-3
+    learning_rate: 2e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 # data loader for train and eval
 DataLoader:

ppcls/configs/ImageNet/DeiT/DeiT_tiny_distilled_patch16_224.yaml

@@ -40,11 +40,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token pos_embed dist_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 1e-3
+    learning_rate: 2e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 # data loader for train and eval
 DataLoader:

ppcls/configs/ImageNet/DeiT/DeiT_tiny_patch16_224.yaml

@@ -40,11 +40,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token pos_embed dist_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 1e-3
+    learning_rate: 2e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 # data loader for train and eval
 DataLoader:

ppcls/configs/ImageNet/Distillation/mv3_large_x1_0_distill_mv3_small_x1_0.yaml

@@ -49,9 +49,8 @@ Loss:
         model_name_pairs:
         - ["Student", "Teacher"]
   Eval:
-    - DistillationGTCELoss:
+    - CELoss:
         weight: 1.0
-        model_names: ["Student"]
         
 
 Optimizer:

ppcls/configs/ImageNet/Distillation/resnet34_distill_resnet18_afd.yaml

@@ -88,10 +88,8 @@ Loss:
         s_shapes: *s_shapes
         t_shapes: *t_shapes
   Eval:
-    - DistillationGTCELoss:
+    - CELoss:
         weight: 1.0
-        model_names: ["Student"]
-        
 
 Optimizer:
   name: Momentum

ppcls/configs/ImageNet/PVTV2/PVT_V2_B0.yaml

@@ -43,11 +43,12 @@ Optimizer:
   no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 5e-6
+    eta_min: 1e-5
     warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/PVTV2/PVT_V2_B1.yaml

@@ -43,11 +43,12 @@ Optimizer:
   no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 5e-6
+    eta_min: 1e-5
     warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/PVTV2/PVT_V2_B2.yaml

@@ -43,11 +43,12 @@ Optimizer:
   no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 5e-6
+    eta_min: 1e-5
     warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/PVTV2/PVT_V2_B2_Linear.yaml

@@ -43,11 +43,12 @@ Optimizer:
   no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 5e-6
+    eta_min: 1e-5
     warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/PVTV2/PVT_V2_B3.yaml

@@ -44,11 +44,12 @@ Optimizer:
   no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 5e-6
+    eta_min: 1e-5
     warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/PVTV2/PVT_V2_B4.yaml

@@ -44,11 +44,12 @@ Optimizer:
   no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 5e-6
+    eta_min: 1e-5
     warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/PVTV2/PVT_V2_B5.yaml

@@ -44,11 +44,12 @@ Optimizer:
   no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 5e-6
+    eta_min: 1e-5
     warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_base_patch4_window12_384.yaml

@@ -41,11 +41,12 @@ Optimizer:
   no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm 
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 20
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_base_patch4_window7_224.yaml

@@ -41,11 +41,12 @@ Optimizer:
   no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm 
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 20
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_large_patch4_window12_384.yaml

@@ -41,11 +41,12 @@ Optimizer:
   no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm 
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 20
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_large_patch4_window7_224.yaml

@@ -41,11 +41,12 @@ Optimizer:
   no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm 
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 20
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_small_patch4_window7_224.yaml

@@ -41,11 +41,12 @@ Optimizer:
   no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm 
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 20
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_tiny_patch4_window7_224.yaml

@@ -41,11 +41,12 @@ Optimizer:
   no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm 
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 20
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/Twins/alt_gvt_base.yaml

@@ -43,11 +43,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token proj.0.weight proj.1.weight proj.2.weight proj.3.weight pos_block
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/Twins/alt_gvt_large.yaml

@@ -43,11 +43,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token proj.0.weight proj.1.weight proj.2.weight proj.3.weight pos_block
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/Twins/alt_gvt_small.yaml

@@ -43,11 +43,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token proj.0.weight proj.1.weight proj.2.weight proj.3.weight pos_block
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/Twins/pcpvt_base.yaml

@@ -43,11 +43,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token proj.0.weight proj.1.weight proj.2.weight proj.3.weight pos_block
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/Twins/pcpvt_large.yaml

@@ -43,11 +43,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token proj.0.weight proj.1.weight proj.2.weight proj.3.weight pos_block
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 
 # data loader for train and eval

ppcls/configs/ImageNet/Twins/pcpvt_small.yaml

@@ -43,11 +43,12 @@ Optimizer:
   no_weight_decay_name: norm cls_token proj.0.weight proj.1.weight proj.2.weight proj.3.weight pos_block
   one_dim_param_no_weight_decay: True
   lr:
+    # for 8 cards
     name: Cosine
-    learning_rate: 5e-4
+    learning_rate: 1e-3
-    eta_min: 1e-5
+    eta_min: 2e-5
     warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6
 
 
 # data loader for train and eval

ppcls/engine/engine.py

@@ -262,12 +262,17 @@ class Engine(object):
             self.model_ema = ExponentialMovingAverage(
                 self.model, self.config['EMA'].get("decay", 0.9999))
 
-        # for distributed
+        # check the gpu num
         world_size = dist.get_world_size()
         self.config["Global"]["distributed"] = world_size != 1
-        if world_size != 4 and self.mode == "train":
+        if self.mode == "train":
-            msg = f"The training strategy in config files provided by PaddleClas is based on 4 gpus. But the number of gpus is {world_size} in current training. Please modify the stategy (learning rate, batch size and so on) if use config files in PaddleClas to train."
+            std_gpu_num = 8 if self.config["Optimizer"][
-            logger.warning(msg)
+                "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)
+
+        # for distributed
         if self.config["Global"]["distributed"]:
             dist.init_parallel_env()
             self.model = paddle.DataParallel(self.model)

ppcls/engine/evaluation/classification.py

@@ -80,22 +80,17 @@ def classification_eval(engine, epoch_id=0):
         current_samples = batch_size * paddle.distributed.get_world_size()
         accum_samples += current_samples
 
+        if isinstance(out, dict) and "Student" in out:
+            out = out["Student"]
+        if isinstance(out, dict) and "logits" in out:
+            out = out["logits"]
+
         # gather Tensor when distributed
         if paddle.distributed.get_world_size() > 1:
             label_list = []
             paddle.distributed.all_gather(label_list, batch[1])
             labels = paddle.concat(label_list, 0)
 
-            if isinstance(out, dict):
-                if "Student" in out:
-                    out = out["Student"]
-                    if isinstance(out, dict):
-                        out = out["logits"]
-                elif "logits" in out:
-                    out = out["logits"]
-                else:
-                    msg = "Error: Wrong key in out!"
-                    raise Exception(msg)
             if isinstance(out, list):
                 preds = []
                 for x in out:

ppcls/loss/deephashloss.py

@@ -20,6 +20,7 @@ class DSHSDLoss(nn.Layer):
     """
     # DSHSD(IEEE ACCESS 2019)
     # paper [Deep Supervised Hashing Based on Stable Distribution](https://ieeexplore.ieee.org/document/8648432/)
+    # code reference: https://github.com/swuxyj/DeepHash-pytorch/blob/master/DSHSD.py
     """
 
     def __init__(self, alpha, multi_label=False):
@@ -62,6 +63,7 @@ class DSHSDLoss(nn.Layer):
 class LCDSHLoss(nn.Layer):
     """
     # paper [Locality-Constrained Deep Supervised Hashing for Image Retrieval](https://www.ijcai.org/Proceedings/2017/0499.pdf)
+    # code reference: https://github.com/swuxyj/DeepHash-pytorch/blob/master/LCDSH.py
     """
 
     def __init__(self, n_class, _lambda):
@@ -100,6 +102,7 @@ class DCHLoss(paddle.nn.Layer):
     """
     # paper [Deep Cauchy Hashing for Hamming Space Retrieval]
     URL:(http://ise.thss.tsinghua.edu.cn/~mlong/doc/deep-cauchy-hashing-cvpr18.pdf)
+    # code reference: https://github.com/swuxyj/DeepHash-pytorch/blob/master/DCH.py
     """
 
     def __init__(self, gamma, _lambda, n_class):

ppcls/loss/emlloss.py

@@ -23,6 +23,11 @@ from .comfunc import rerange_index
 
 
 class EmlLoss(paddle.nn.Layer):
+    """Ensemble Metric Learning Loss
+    paper: [Large Scale Strongly Supervised Ensemble Metric Learning, with Applications to Face Verification and Retrieval](https://arxiv.org/pdf/1212.6094.pdf)
+    code reference: https://github.com/PaddlePaddle/models/blob/develop/PaddleCV/metric_learning/losses/emlloss.py
+    """
+
     def __init__(self, batch_size=40, samples_each_class=2):
         super(EmlLoss, self).__init__()
         assert (batch_size % samples_each_class == 0)

ppcls/loss/googlenetloss.py

@@ -18,11 +18,13 @@ import paddle.nn.functional as F
 class GoogLeNetLoss(nn.Layer):
     """
     Cross entropy loss used after googlenet
+    reference paper: [https://arxiv.org/pdf/1409.4842v1.pdf](Going Deeper with Convolutions)
     """
+
     def __init__(self, epsilon=None):
         super().__init__()
-        assert (epsilon is None or epsilon <= 0 or epsilon >= 1), "googlenet is not support label_smooth"
+        assert (epsilon is None or epsilon <= 0 or
-            
+                epsilon >= 1), "googlenet is not support label_smooth"
 
     def forward(self, inputs, label):
         input0, input1, input2 = inputs

ppcls/loss/msmloss.py

@@ -21,7 +21,9 @@ from .comfunc import rerange_index
 
 class MSMLoss(paddle.nn.Layer):
     """
-    MSMLoss Loss, based on triplet loss. USE P * K samples.
+    paper : [Margin Sample Mining Loss: A Deep Learning Based Method for Person Re-identification](https://arxiv.org/pdf/1710.00478.pdf)
+    code reference: https://github.com/michuanhaohao/keras_reid/blob/master/reid_tripletcls.py
+    Margin Sample Mining Loss, based on triplet loss. USE P * K samples.
     the batch size is fixed. Batch_size = P * K;  but the K may vary between batches.
     same label gather together
 

ppcls/loss/npairsloss.py

@@ -5,6 +5,11 @@ import paddle
 
 
 class NpairsLoss(paddle.nn.Layer):
+    """Npair_loss_
+    paper [Improved deep metric learning with multi-class N-pair loss objective](https://dl.acm.org/doi/10.5555/3157096.3157304)
+    code reference: https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/contrib/losses/metric_learning/npairs_loss
+    """
+
     def __init__(self, reg_lambda=0.01):
         super(NpairsLoss, self).__init__()
         self.reg_lambda = reg_lambda

ppcls/loss/pairwisecosface.py

@@ -23,6 +23,11 @@ import paddle.nn.functional as F
 
 
 class PairwiseCosface(nn.Layer):
+    """
+    paper: Circle Loss: A Unified Perspective of Pair Similarity Optimization
+    code reference: https://github.com/leoluopy/circle-loss-demonstration/blob/main/circle_loss.py
+    """
+
     def __init__(self, margin, gamma):
         super(PairwiseCosface, self).__init__()
         self.margin = margin
@@ -36,8 +41,10 @@ class PairwiseCosface(nn.Layer):
         dist_mat = paddle.matmul(embedding, embedding, transpose_y=True)
 
         N = dist_mat.shape[0]
-        is_pos = targets.reshape([N,1]).expand([N,N]).equal(paddle.t(targets.reshape([N,1]).expand([N,N]))).astype('float')
+        is_pos = targets.reshape([N, 1]).expand([N, N]).equal(
-        is_neg = targets.reshape([N,1]).expand([N,N]).not_equal(paddle.t(targets.reshape([N,1]).expand([N,N]))).astype('float')
+            paddle.t(targets.reshape([N, 1]).expand([N, N]))).astype('float')
+        is_neg = targets.reshape([N, 1]).expand([N, N]).not_equal(
+            paddle.t(targets.reshape([N, 1]).expand([N, N]))).astype('float')
 
         # Mask scores related to itself
         is_pos = is_pos - paddle.eye(N, N)
@@ -46,10 +53,12 @@ class PairwiseCosface(nn.Layer):
         s_n = dist_mat * is_neg
 
         logit_p = -self.gamma * s_p + (-99999999.) * (1 - is_pos)
-        logit_n = self.gamma * (s_n + self.margin) + (-99999999.) * (1 - is_neg)
+        logit_n = self.gamma * (s_n + self.margin) + (-99999999.) * (1 - is_neg
+                                                                     )
 
-        loss = F.softplus(paddle.logsumexp(logit_p, axis=1) + paddle.logsumexp(logit_n, axis=1)).mean()
+        loss = F.softplus(
+            paddle.logsumexp(
+                logit_p, axis=1) + paddle.logsumexp(
+                    logit_n, axis=1)).mean()
 
         return {"PairwiseCosface": loss}
-
-

ppcls/loss/rkdloss.py

@@ -29,6 +29,7 @@ def pdist(e, squared=False, eps=1e-12):
 
 
 class RKdAngle(nn.Layer):
+    # paper : [Relational Knowledge Distillation](https://arxiv.org/abs/1904.05068?context=cs.LG)
     # reference: https://github.com/lenscloth/RKD/blob/master/metric/loss.py
     def __init__(self, target_size=None):
         super().__init__()
@@ -64,6 +65,7 @@ class RKdAngle(nn.Layer):
 
 
 class RkdDistance(nn.Layer):
+    # paper : [Relational Knowledge Distillation](https://arxiv.org/abs/1904.05068?context=cs.LG)
     # reference: https://github.com/lenscloth/RKD/blob/master/metric/loss.py
     def __init__(self, eps=1e-12, target_size=1):
         super().__init__()

ppcls/loss/supconloss.py

@@ -4,6 +4,7 @@ from paddle import nn
 
 class SupConLoss(nn.Layer):
     """Supervised Contrastive Learning: https://arxiv.org/pdf/2004.11362.pdf.
+    code reference: https://github.com/HobbitLong/SupContrast/blob/master/losses.py
     It also supports the unsupervised contrastive loss in SimCLR"""
 
     def __init__(self,

ppcls/loss/trihardloss.py

@@ -22,6 +22,8 @@ from .comfunc import rerange_index
 
 class TriHardLoss(paddle.nn.Layer):
     """
+    paper: In Defense of the Triplet Loss for Person Re-Identification
+    code reference: https://github.com/VisualComputingInstitute/triplet-reid/blob/master/loss.py
     TriHard Loss, based on triplet loss. USE P * K samples.
     the batch size is fixed. Batch_size = P * K;  but the K may vary between batches.
     same label gather together

ppcls/loss/triplet.py

@@ -1,3 +1,17 @@
+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
@@ -8,6 +22,8 @@ import paddle.nn as nn
 
 class TripletLossV2(nn.Layer):
     """Triplet loss with hard positive/negative mining.
+    paper : [Facenet: A unified embedding for face recognition and clustering](https://arxiv.org/pdf/1503.03832.pdf)
+    code reference: https://github.com/okzhili/Cartoon-face-recognition/blob/master/loss/triplet_loss.py
     Args:
         margin (float): margin for triplet.
     """

ppcls/optimizer/__init__.py

@@ -118,8 +118,6 @@ def build_optimizer(config, epochs, step_each_epoch, model_list=None):
                     if hasattr(model_list[i], optim_scope):
                         optim_model.append(getattr(model_list[i], optim_scope))
 
-        assert len(optim_model) == 1, \
-            "Invalid optim model for optim scope({}), number of optim_model={}".format(optim_scope, len(optim_model))
         optim = getattr(optimizer, optim_name)(
             learning_rate=lr, grad_clip=grad_clip,
             **optim_cfg)(model_list=optim_model)

test_tipc/config/CSWinTransformer/CSWinTransformer_tiny_224_train_infer_python.txt

@@ -13,7 +13,7 @@ train_infer_img_dir:./dataset/ILSVRC2012/val
 null:null
 ##
 trainer:norm_train
-norm_train:tools/train.py -c ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_tiny_224.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False
+norm_train:tools/train.py -c ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_tiny_224.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False -o Global.print_batch_step=1
 pact_train:null
 fpgm_train:null
 distill_train:null

test_tipc/config/MobileViT/MobileViT_S_train_infer_python.txt

@@ -13,7 +13,7 @@ train_infer_img_dir:./dataset/ILSVRC2012/val
 null:null
 ##
 trainer:norm_train
-norm_train:tools/train.py -c ppcls/configs/ImageNet/MobileViT/MobileViT_S.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False
+norm_train:tools/train.py -c ppcls/configs/ImageNet/MobileViT/MobileViT_S.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False -o Global.print_batch_step=1
 pact_train:null
 fpgm_train:null
 distill_train:null

test_tipc/config/PVTV2/PVT_V2_B2_Linear_train_infer_python.txt

@@ -13,7 +13,7 @@ train_infer_img_dir:./dataset/ILSVRC2012/val
 null:null
 ##
 trainer:norm_train
-norm_train:tools/train.py -c ppcls/configs/ImageNet/PVTV2/PVT_V2_B2_Linear.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False
+norm_train:tools/train.py -c ppcls/configs/ImageNet/PVTV2/PVT_V2_B2_Linear.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False -o Global.print_batch_step=1
 pact_train:null
 fpgm_train:null
 distill_train:null