Added Ranked Activation Shift (RAS)

configs/postprocessors/ras.yml

@@ -0,0 +1,5 @@
+postprocessor:
+    name: ras
+    APS_mode: False
+    postprocessor_args: {}
+    postprocessor_sweep: {}

openood/evaluation_api/evaluator.py

@@ -13,6 +13,7 @@
 from openood.networks.ash_net import ASHNet
 from openood.networks.react_net import ReactNet
 from openood.networks.scale_net import ScaleNet
+from openood.networks.ras_net import RASNet
 from openood.networks.adascale_net import AdaScaleANet, AdaScaleLNet
 
 from .datasets import DATA_INFO, data_setup, get_id_ood_dataloader
@@ -90,7 +91,8 @@ def __init__(
         # set up config root
         if config_root is None:
             filepath = os.path.dirname(os.path.abspath(__file__))
-            config_root = os.path.join('/', *filepath.split('/')[:-2], 'configs')
+            config_root = os.path.join('/',
+                                       *filepath.split('/')[:-2], 'configs')
 
         # get postprocessor
         if postprocessor is None:
@@ -121,6 +123,8 @@ def __init__(
             net = AdaScaleANet(net)
         elif postprocessor_name == 'adascale_l':
             net = AdaScaleLNet(net)
+        elif postprocessor_name == 'ras':
+            net = RASNet(net)
 
         # postprocessor setup
         postprocessor.setup(net, dataloader_dict['id'], dataloader_dict['ood'])

openood/evaluation_api/postprocessor.py

@@ -16,8 +16,8 @@
     RMDSPostprocessor, SHEPostprocessor, CIDERPostprocessor, NPOSPostprocessor,
     GENPostprocessor, NNGuidePostprocessor, RelationPostprocessor,
     T2FNormPostprocessor, ReweightOODPostprocessor, fDBDPostprocessor,
-    AdaScalePostprocessor, IODINPostprocessor, NCIPostprocessor,CFOODPostprocessor,
-    VRAPostprocessor, GrOODPostprocessor)
+    AdaScalePostprocessor, IODINPostprocessor, NCIPostprocessor,
+    CFOODPostprocessor, VRAPostprocessor, GrOODPostprocessor, RASPostprocessor)
 from openood.utils.config import Config, merge_configs
 
 postprocessors = {
@@ -73,6 +73,7 @@
     'grood': GrOODPostprocessor,
     'vra': VRAPostprocessor,
     'cfood': CFOODPostprocessor,
+    'ras': RASPostprocessor,
 }
 
 link_prefix = 'https://raw.githubusercontent.com/Jingkang50/OpenOOD/main/configs/postprocessors/'

openood/networks/ras_net.py

@@ -0,0 +1,33 @@
+import torch
+import torch.nn as nn
+
+
+class RASNet(nn.Module):
+    def __init__(self, backbone):
+        super(RASNet, self).__init__()
+        self.backbone = backbone
+        self.mean_curve = None
+
+    def forward(self, x, return_feature=False, return_feature_list=False):
+        try:
+            return self.backbone(x, return_feature, return_feature_list)
+        except TypeError:
+            return self.backbone(x, return_feature)
+
+    def forward_shift(self, x):
+        _, feature = self.backbone(x, return_feature=True)
+        feature = feature.view(feature.size(0), -1)
+
+        sorted_vals, idx = torch.sort(feature, dim=1)
+        mc = self.mean_curve.to(feature.device).expand_as(sorted_vals)
+        shifted = torch.empty_like(feature).scatter_(1, idx, mc)
+
+        logits_cls = self.backbone.get_fc_layer()(shifted)
+        return logits_cls
+
+    def set_mean_curve(self, curve):
+        self.mean_curve = curve
+
+    def get_fc(self):
+        fc = self.backbone.fc
+        return fc.weight.cpu().detach().numpy(), fc.bias.cpu().detach().numpy()

openood/networks/utils.py

@@ -25,6 +25,7 @@
 from .openmax_net import OpenMax
 from .patchcore_net import PatchcoreNet
 from .projection_net import ProjectionNet
+from .ras_net import RASNet
 from .react_net import ReactNet
 from .resnet18_32x32 import ResNet18_32x32
 from .resnet18_64x64 import ResNet18_64x64
@@ -172,6 +173,10 @@ def get_network(network_config):
         backbone = get_network(network_config.backbone)
         net = ReactNet(backbone)
 
+    elif network_config.name == 'ras_net':
+        backbone = get_network(network_config.backbone)
+        net = RASNet(backbone)
+
     elif network_config.name == 'csi_net':
         # don't wrap ddp here cuz we need to modify
         # backbone

openood/postprocessors/__init__.py

@@ -29,6 +29,7 @@
 from .opengan_postprocessor import OpenGanPostprocessor
 from .openmax_postprocessor import OpenMax
 from .patchcore_postprocessor import PatchcorePostprocessor
+from .ras_postprocessor import RASPostprocessor
 from .rd4ad_postprocessor import Rd4adPostprocessor
 from .react_postprocessor import ReactPostprocessor
 from .rmds_postprocessor import RMDSPostprocessor
@@ -50,4 +51,3 @@
 from .grood import GrOODPostprocessor
 from .vra_postprocessor import VRAPostprocessor
 from .cfood_postprocessor import CFOODPostprocessor
-

openood/postprocessors/ras_postprocessor.py

@@ -0,0 +1,45 @@
+from typing import Any
+
+import torch
+import torch.nn as nn
+from tqdm import tqdm
+
+from .base_postprocessor import BasePostprocessor
+
+
+class RASPostprocessor(BasePostprocessor):
+    def __init__(self, config):
+        super(RASPostprocessor, self).__init__(config)
+        self.setup_flag = False
+
+    def setup(self, net: nn.Module, id_loader_dict, ood_loader_dict):
+        if not self.setup_flag:
+            activation_log = []
+            net.eval()
+            with torch.no_grad():
+                loader = id_loader_dict['train']
+                for batch in tqdm(loader,
+                                  desc='Setup: ',
+                                  position=0,
+                                  leave=True):
+                    data = batch['data'].cuda()
+                    data = data.float()
+
+                    _, feature = net(data, return_feature=True)
+                    activation_log.append(feature.data.cpu())
+
+            activation_log = torch.cat(activation_log, dim=0)
+            sorted_vals, _ = torch.sort(activation_log, dim=1)
+            mean_curve = sorted_vals.mean(dim=0).cuda()
+
+            net.set_mean_curve(mean_curve)
+            self.mean_curve = mean_curve
+            self.setup_flag = True
+
+    @torch.no_grad()
+    def postprocess(self, net: nn.Module, data: Any):
+        output = net.forward_shift(data)
+        score = torch.softmax(output, dim=1)
+        _, pred = torch.max(score, dim=1)
+        conf = torch.logsumexp(output.data.cpu(), dim=1)
+        return pred, conf

openood/postprocessors/utils.py

@@ -45,6 +45,7 @@
 from .relation_postprocessor import RelationPostprocessor
 from .grood import GrOODPostprocessor
 from .vra_postprocessor import VRAPostprocessor
+from .ras_postprocessor import RASPostprocessor
 
 
 def get_postprocessor(config: Config):
@@ -94,6 +95,7 @@ def get_postprocessor(config: Config):
         't2fnorm': T2FNormPostprocessor,
         'grood': GrOODPostprocessor,
         'vra': VRAPostprocessor,
+        'ras': RASPostprocessor,
     }
 
     return postprocessors[config.postprocessor.name](config)