update ppcls

ppcls/arch/backbone/__init__.py

@@ -69,7 +69,7 @@ from .model_zoo.repvgg import RepVGG_A0, RepVGG_A1, RepVGG_A2, RepVGG_B0, RepVGG
 from .model_zoo.van import VAN_tiny
 from .model_zoo.peleenet import PeleeNet
 from .model_zoo.convnext import ConvNeXt_tiny
-from .model_zoo.cae import cae_base_patch16_224, cae_base_patch16_384, cae_large_patch16_224, cae_large_patch16_384, cae_large_patch16_512, cae_small_patch16_224
+from .model_zoo.cae import cae_base_patch16_224, cae_large_patch16_224
 
 from .variant_models.resnet_variant import ResNet50_last_stage_stride1
 from .variant_models.vgg_variant import VGG19Sigmoid

ppcls/arch/backbone/model_zoo/cae.py

@@ -25,6 +25,17 @@ import paddle
 import paddle.nn as nn
 import paddle.nn.functional as F
 
+from ....utils.download import get_weights_path_from_url
+
+MODEL_URLS = {
+    "cae_base_patch16_224":
+    "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/cae_base_patch16_224_pretrained.pdparams",
+    "cae_large_patch16_224":
+    "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/cae_large_patch16_224_pretrained.pdparams"
+}
+
+__all__ = list(MODEL_URLS.keys())
+
 
 def _ntuple(n):
     def parse(x):
@@ -615,13 +626,22 @@ def _enable_linear_eval(model):
 
 
 def _load_pretrained(pretrained,
+                     pretrained_url,
                      model,
                      model_keys,
                      model_ema_configs,
                      abs_pos_emb,
                      rel_pos_bias,
                      use_ssld=False):
-    checkpoint = paddle.load(pretrained)
+    if pretrained is False:
+        pass
+    elif pretrained is True:
+        local_weight_path = get_weights_path_from_url(pretrained_url).replace(
+            ".pdparams", "")
+        checkpoint = paddle.load(local_weight_path + ".pdparams")
+    elif isinstance(pretrained, str):
+        checkpoint = paddle.load(local_weight_path + ".pdparams")
+
     checkpoint_model = None
     for model_key in model_keys.split('|'):
         if model_key in checkpoint:
@@ -766,48 +786,15 @@ def _load_pretrained(pretrained,
     return
 
 
-def cae_small_patch16_224(**kwargs):
+def cae_base_patch16_224(pretrained=True, use_ssld=False, **kwargs):
     config = kwargs.copy()
     enable_linear_eval = config.pop('enable_linear_eval')
     model_keys = config.pop('model_key')
     model_ema_configs = config.pop('model_ema')
     abs_pos_emb = config.pop('abs_pos_emb')
     rel_pos_bias = config.pop('rel_pos_bias')
-    pretrained = config.pop('pretrained')
-
-    model = VisionTransformer(
-        patch_size=16,
-        embed_dim=384,
-        depth=12,
-        num_heads=12,
-        mlp_ratio=4,
-        qkv_bias=True,
-        norm_layer=partial(
-            nn.LayerNorm, epsilon=1e-6),
-        **kwargs)
-
-    if enable_linear_eval:
-        _enable_linear_eval(model)
-
-    _load_pretrained(
-        pretrained,
-        model,
-        model_keys,
-        model_ema_configs,
-        abs_pos_emb,
-        rel_pos_bias,
-        use_ssld=False)
-    return model
-
-
-def cae_base_patch16_224(**kwargs):
-    config = kwargs.copy()
-    enable_linear_eval = config.pop('enable_linear_eval')
-    model_keys = config.pop('model_key')
-    model_ema_configs = config.pop('model_ema')
-    abs_pos_emb = config.pop('abs_pos_emb')
-    rel_pos_bias = config.pop('rel_pos_bias')
-    pretrained = config.pop('pretrained')
+    if pretrained in config:
+        pretrained = config.pop('pretrained')
 
     model = VisionTransformer(
         patch_size=16,
@@ -825,6 +812,7 @@ def cae_base_patch16_224(**kwargs):
 
     _load_pretrained(
         pretrained,
+        MODEL_URLS["cae_base_patch16_224"],
         model,
         model_keys,
         model_ema_configs,
@@ -835,124 +823,17 @@ def cae_base_patch16_224(**kwargs):
     return model
 
 
-def cae_base_patch16_384(**kwargs):
-    config = kwargs.copy()
-    enable_linear_eval = config.pop('enable_linear_eval')
-    model_keys = config.pop('model_key')
-    model_ema_configs = config.pop('model_ema')
-    abs_pos_emb = config.pop('abs_pos_emb')
-    rel_pos_bias = config.pop('rel_pos_bias')
-    pretrained = config.pop('pretrained')
-
-    model = VisionTransformer(
-        img_size=384,
-        patch_size=16,
-        embed_dim=768,
-        depth=12,
-        num_heads=12,
-        mlp_ratio=4,
-        qkv_bias=True,
-        norm_layer=partial(
-            nn.LayerNorm, epsilon=1e-6),
-        **kwargs)
-
-    if enable_linear_eval:
-        _enable_linear_eval(model)
-
-    _load_pretrained(
-        pretrained,
-        model,
-        model_keys,
-        model_ema_configs,
-        abs_pos_emb,
-        rel_pos_bias,
-        use_ssld=False)
-
-    return model
-
-
-def cae_large_patch16_224(**kwargs):
-    config = kwargs.copy()
-    enable_linear_eval = config.pop('enable_linear_eval')
-    model_keys = config.pop('model_key')
-    model_ema_configs = config.pop('model_ema')
-    abs_pos_emb = config.pop('abs_pos_emb')
-    rel_pos_bias = config.pop('rel_pos_bias')
-    pretrained = config.pop('pretrained')
-
-    model = VisionTransformer(
-        patch_size=16,
-        embed_dim=1024,
-        depth=24,
-        num_heads=16,
-        mlp_ratio=4,
-        qkv_bias=True,
-        norm_layer=partial(
-            nn.LayerNorm, epsilon=1e-6),
-        **kwargs)
-
-    if enable_linear_eval:
-        _enable_linear_eval(model)
-
-    _load_pretrained(
-        pretrained,
-        model,
-        model_keys,
-        model_ema_configs,
-        abs_pos_emb,
-        rel_pos_bias,
-        use_ssld=False)
-
-    return model
-
-
-def cae_large_patch16_384(**kwargs):
-    config = kwargs.copy()
-    enable_linear_eval = config.pop('enable_linear_eval')
-    model_keys = config.pop('model_key')
-    model_ema_configs = config.pop('model_ema')
-    abs_pos_emb = config.pop('abs_pos_emb')
-    rel_pos_bias = config.pop('rel_pos_bias')
-    pretrained = config.pop('pretrained')
-
-    model = VisionTransformer(
-        img_size=384,
-        patch_size=16,
-        embed_dim=1024,
-        depth=24,
-        num_heads=16,
-        mlp_ratio=4,
-        qkv_bias=True,
-        norm_layer=partial(
-            nn.LayerNorm, epsilon=1e-6),
-        **kwargs)
-
-    if enable_linear_eval:
-        _enable_linear_eval(model)
-
-    _load_pretrained(
-        pretrained,
-        model,
-        model_keys,
-        model_ema_configs,
-        abs_pos_emb,
-        rel_pos_bias,
-        use_ssld=False)
-
-    return model
-
-
-def cae_large_patch16_512(**kwargs):
+def cae_large_patch16_224(pretrained=True, use_ssld=False, **kwargs):
     config = kwargs.copy()
     enable_linear_eval = config.pop('enable_linear_eval')
     model_keys = config.pop('model_key')
     model_ema_configs = config.pop('model_ema')
     abs_pos_emb = config.pop('abs_pos_emb')
     rel_pos_bias = config.pop('rel_pos_bias')
-    pretrained = config.pop('pretrained')
+    if pretrained in config:
+        pretrained = config.pop('pretrained')
 
     model = VisionTransformer(
-        img_size=512,
         patch_size=16,
         embed_dim=1024,
         depth=24,
@@ -968,6 +849,7 @@ def cae_large_patch16_512(**kwargs):
 
     _load_pretrained(
         pretrained,
+        MODEL_URLS["cae_large_patch16_224"],
         model,
         model_keys,
         model_ema_configs,

ppcls/configs/CAE/cae_base_patch16_224_finetune.yaml

@@ -4,10 +4,10 @@ Global:
   pretrained_model: null
   output_dir: ./output/
   device: gpu
-  save_interval: 1
+  save_interval: 20
   eval_during_train: True
   eval_interval: 1
-  epochs: 200
+  epochs: 100
   print_batch_step: 10
   use_visualdl: False
   # used for static mode and model export
@@ -17,7 +17,7 @@ Global:
 # model architecture
 Arch:
   name: cae_base_patch16_224
-  class_num: 4
+  class_num: 102
   drop_rate: 0.0
   drop_path_rate: 0.1
   attn_drop_rate: 0.0
@@ -39,7 +39,7 @@ Arch:
     enable_model_ema: False 
     model_ema_decay: 0.9999
     model_ema_force_cpu: False
-  pretrained: ./pretrained/vit_base_cae_pretrained.pdparams
+  pretrained: True
 
 # loss function config for traing/eval process
 Loss:
@@ -60,7 +60,7 @@ Optimizer:
   layerwise_decay: 0.65
   lr:
     name: Cosine
-    learning_rate: 0.004
+    learning_rate: 0.001
     eta_min: 1e-6
     warmup_epoch: 10
     warmup_start_lr: 1e-6
@@ -71,14 +71,14 @@ DataLoader:
   Train:
     dataset:
       name: ImageNetDataset
-      image_root: ./dataset/paddle-job-153869-0/train_eval_data/images
-      cls_label_path: ./dataset/paddle-job-153869-0/train_eval_data/train_data_list.txt
+      image_root: ./dataset/flowers102/
+      cls_label_path: ./dataset/flowers102/train_list.txt
       batch_transform_ops:
         - MixupCutmixHybrid:
             mixup_alpha: 0.8
             cutmix_alpha: 1.0
             switch_prob: 0.5
-            num_classes: 4
+            num_classes: 102
       transform_ops:
         - DecodeImage:
             to_rgb: True
@@ -99,11 +99,10 @@ DataLoader:
             sl: 0.02
             sh: 0.3
             r1: 0.3
-      delimiter: '	'
 
     sampler:
       name: DistributedBatchSampler
-      batch_size: 128
+      batch_size: 2
       drop_last: True
       shuffle: True
     loader:
@@ -113,8 +112,8 @@ DataLoader:
   Eval:
     dataset: 
       name: ImageNetDataset
-      image_root: ./dataset/paddle-job-153869-0/train_eval_data/images
-      cls_label_path: ./dataset/paddle-job-153869-0/train_eval_data/eval_data_list.txt
+      image_root: ./dataset/flowers102/
+      cls_label_path: ./dataset/flowers102/val_list.txt
       transform_ops:
         - DecodeImage:
             to_rgb: True
@@ -128,11 +127,9 @@ DataLoader:
             mean: [0.485, 0.456, 0.406]
             std: [0.229, 0.224, 0.225]
             order: ''
-      delimiter: '	'
-    
     sampler:
       name: DistributedBatchSampler
-      batch_size: 128
+      batch_size: 2
       drop_last: False
       shuffle: False
     loader: