Update demo and config

Author: lizekai-richard

configs/Demo_Hard_Label.yaml

@@ -3,16 +3,26 @@ real_data_path: "./dataset/"
 ipc: 10
 model_name: "ConvNet-3"
 data_aug_func: "dsa"
-aug_params: {"prob_flip": 0.5, "ratio_rotate": 15.0, "saturation": 2.0, "brightness": 1.0, "contrast": 0.5, "ratio_scale": 1.2, "ratio_crop_pad": 0.125, "ratio_cutout": 0.5}
+aug_params: {
+  "prob_flip": 0.5, 
+  "ratio_rotate": 15.0, 
+  "saturation": 2.0, 
+  "brightness": 1.0, 
+  "contrast": 0.5, 
+  "ratio_scale": 1.2, 
+  "ratio_crop_pad": 0.125, 
+  "ratio_cutout": 0.5
+}
 optimizer: "sgd"
-lr_scheduler: "sgd"
-weight_decay: 0.01
+lr_scheduler: "step"
+weight_decay: 0.0005
 momentum: 0.9
 num_eval: 5
 im_size: (32, 32)
-num_epochs: 300
+num_epochs: 1000
 use_zca: False
 batch_size: 256
 default_lr: 0.01
+num_workers: 4
 save_path: "./results/my_method_hard_label_scores.csv"
 device: "cuda"

configs/Demo_Soft_Label.yaml

@@ -2,20 +2,30 @@ dataset: "CIFAR10"
 real_data_path: "./dataset/"
 ipc: 10
 model_name: "ConvNet-3"
-data_aug_func: "cutmix"
-aug_params: {"cutmix_p": 1.0}
+data_aug_func: "dsa"
+aug_params: {
+    "prob_flip": 0.5,
+    "ratio_rotate": 15.0,
+    "saturation": 2.0,
+    "brightness": 1.0,
+    "contrast": 0.5,
+    "ratio_scale": 1.2,
+    "ratio_crop_pad": 0.125,
+    "ratio_cutout": 0.5
+}
 soft_label_mode: "S"
-soft_label_criterion: "kl"
+soft_label_criterion: "sce"
 optimizer: "sgd"
-lr_scheduler: "cosine"
+lr_scheduler: "step"
 temperature: 20.0
-weight_decay: 0.01
+weight_decay: 0.0005
 momentum: 0.9
 num_eval: 5
 im_size: (32, 32)
-num_epochs: 300
-use_zca: False
+num_epochs: 1000
+use_zca: True
 batch_size: 256
 default_lr: 0.01
+num_workers: 4
 save_path: "./results/my_method_soft_label_scores.csv"
-device: "cuda"
+device: "cuda"

dd_ranking/utils/utils.py

@@ -534,9 +534,6 @@ def validate(
 
     model.eval()
 
-    if aug_func is None:
-        aug_func = default_augmentation
-
     end = time.time()
     last_idx = len(loader) - 1
     with torch.no_grad():
@@ -615,9 +612,6 @@ def validate_dc(
     model = model.to(device)
     model.eval()
 
-    if aug_func is None:
-        aug_func = default_augmentation
-
     for i_batch, datum in enumerate(loader):
         img = datum[0].to(device)
         lab = datum[1].to(device)

demo_hard.py

@@ -8,7 +8,7 @@
 config = Config.from_file("./configs/Demo_Hard_Label.yaml")
 convd3_hard_obj = Hard_Label_Objective_Metrics(config)
 syn_images = torch.load(os.path.join("./DC/CIFAR10/IPC10/", f"images.pt"), map_location='cpu')
-print(convd3_hard_obj.compute_metrics(syn_images))
+print(convd3_hard_obj.compute_metrics(syn_images, syn_lr=0.01))
 
 
 """Use hardcoded parameters"""
@@ -17,6 +17,7 @@
 ipc = 10                              # Specify your IPC
 dataset = "CIFAR10"                   # Specify your dataset name
 data_dir = "./datasets"               # Specify your dataset path
+syn_data_dir = "./DM/CIFAR10/IPC10/"  # Specify your synthetic data path
 model_name = "ConvNet-3"              # Specify your model name
 im_size = (32, 32)                    # Specify your image size
 
@@ -31,13 +32,14 @@
     "ratio_cutout": 0.5,
 }
 
-syn_images = torch.load(os.path.join("./DC/CIFAR10/IPC10/", f"images.pt"), map_location='cpu')
+syn_images = torch.load(os.path.join(syn_data_dir, f"images.pt"), map_location='cpu')
 save_path = f"./results/{dataset}/{model_name}/IPC{ipc}/dm_hard_scores.csv"
 convd3_hard_obj = Hard_Label_Objective_Metrics(
     dataset=dataset, 
     real_data_path=data_dir, 
     ipc=ipc, 
     model_name=model_name,
+    default_lr=0.01,
     optimizer='sgd',             # Use SGD optimizer
     lr_scheduler='step',         # Use StepLR learning rate scheduler
     weight_decay=0.0005,
@@ -47,7 +49,9 @@
     data_aug_func='dsa',         # Use DSA data augmentation
     aug_params=dsa_params,       # Specify DSA parameters
     im_size=im_size,
+    num_epochs=1000,
+    num_workers=4,
     device=device,
     save_path=save_path
 )
-print(convd3_hard_obj.compute_metrics(syn_images))
+print(convd3_hard_obj.compute_metrics(syn_images, syn_lr=0.01))

demo_soft.py

@@ -8,7 +8,8 @@
 config = Config.from_file("./configs/Demo_Soft_Label.yaml")
 convd3_soft_obj = Soft_Label_Objective_Metrics(config)
 syn_images = torch.load(os.path.join(syn_data_dir, f"images.pt"), map_location='cpu')
-print(convd3_soft_obj.compute_metrics(syn_images))
+soft_labels = torch.load(os.path.join(syn_data_dir, f"labels.pt"), map_location='cpu')
+print(convd3_soft_obj.compute_metrics(syn_images, soft_labels, syn_lr=0.01))
 
 
 """Use hardcoded parameters"""
@@ -32,6 +33,7 @@
 }
 
 syn_images = torch.load(os.path.join(syn_data_dir, f"images.pt"), map_location='cpu')
+soft_labels = torch.load(os.path.join(syn_data_dir, f"labels.pt"), map_location='cpu')
 save_path = f"./results/{dataset}/{model_name}/IPC{ipc}/dm_hard_scores.csv"
 convd3_hard_obj = Soft_Label_Objective_Metrics(
     dataset=dataset,
@@ -40,16 +42,19 @@
     model_name=model_name,
     soft_label_criterion='sce',  # Use Soft Cross Entropy Loss
     soft_label_mode='S',         # Use one-to-one image to soft label mapping
+    default_lr=0.01,
     optimizer='sgd',             # Use SGD optimizer
     lr_scheduler='step',         # Use StepLR learning rate scheduler
     weight_decay=0.0005,         
     momentum=0.9,                
-    use_zca=True,                # Use ZCA whitening
+    use_zca=True,                # Use ZCA whitening (please disable it if you didn't use it to distill synthetic data)
     num_eval=5,                  
     data_aug_func='dsa',         # Use DSA data augmentation
     aug_params=dsa_params,       # Specify dsa parameters
     im_size=im_size,
+    num_epochs=1000,
+    num_workers=4,
     device=device,
     save_path=save_path
 )
-print(convd3_hard_obj.compute_metrics(syn_images))
+print(convd3_hard_obj.compute_metrics(syn_images, soft_labels, syn_lr=0.01))