next

pix2pixHD_attack/models/pix2pixHD_model.py

@@ -242,7 +242,7 @@ class Pix2PixHDModel(BaseModel):
             input_concat = input_label  
 
         input_adv = torch.clamp(input_concat + perturb, min=-1, max=1)   
-           
+        
         with torch.no_grad():
             fake_image = self.netG.forward(input_adv)
         

pix2pixHD_attack/test.py

@@ -34,6 +34,10 @@ if not opt.engine and not opt.onnx:
         print(model)
 else:
     from run_engine import run_trt_engine, run_onnx
+
+# Initialize Metrics
+l1_error, l2_error, min_dist, l0_error = 0.0
+n_samples = 0
     
 for i, data in enumerate(dataset):
     if i >= opt.how_many:
@@ -59,14 +63,26 @@ for i, data in enumerate(dataset):
     elif opt.onnx:
         generated = run_onnx(opt.onnx, opt.data_type, minibatch, [data['label'], data['inst']])
     else:        
-        # generated = model.inference(data['label'], data['inst'], data['image'])
+        generated_noattack = model.inference(data['label'], data['inst'], data['image'])
         adv_image, perturb = model.attack(data['label'], data['inst'], data['image'])
         generated, adv_img = model.inference_attack(data['label'], data['inst'], data['image'], perturb)
         
     visuals = OrderedDict([('input_label', util.tensor2label(adv_img.data[0], opt.label_nc)),
-                           ('synthesized_image', util.tensor2im(generated.data[0]))])
+                           ('attacked_image', util.tensor2im(generated.data[0])),
+                           ('noattack', util.tensor2im(generated_noattack.data[0]))])
     img_path = data['path']
     print('process image... %s' % img_path)
     visualizer.save_images(webpage, visuals, img_path)
 
+    # Compute metrics
+    l1_error += F.l1_loss(generated, generated_noattack)
+    l2_error += F.mse_loss(generated, generated_noattack)
+    l0_error += (generated - generated_noattack).norm(0)
+    min_dist += (generated - generated_noattack).norm(float('-inf'))
+    n_samples += 1
+
+# Print metrics
+print('{} images. L1 error: {}. L2 error: {}. L0 error: {}. L_-inf error: {}. Perceptual error: {}.'.format(n_samples, 
+l1_error / n_samples, l2_error / n_samples, l0_error / n_samples, min_dist / n_samples, perceptual_error / n_samples))
+
 webpage.save()