Update audio_diffusion_attacks_forhf/src/test_encoder_attack.py

audio_diffusion_attacks_forhf/src/test_encoder_attack.py

@@ -119,8 +119,9 @@ def poison_audio(waveform, sample_rate, encoders, audio_difference_weights=[1],
             waveform=inputs['input_values'][0]
             #Andy removed: wandb.log({f"unperturbed {audio_name}": wandb.Audio(waveform[0].detach().numpy().flatten(), sample_rate=sample_rate)}, step=0)
             waveform=torch.reshape(waveform, (1, waveform.shape[0], waveform.shape[1]))
-            waveform=waveform.to(device='cuda')
-            inputs["padding_mask"]=inputs["padding_mask"].to(device='cuda')
+            #Andy removed: waveform=waveform.to(device='cuda')
+            #Andy edited: inputs["padding_mask"]=inputs["padding_mask"].to(device='cuda')
+            inputs["padding_mask"]=inputs["padding_mask"]
             
             if method=="encoder":
                 unperturbed_waveform=waveform.clone().detach()
@@ -135,8 +136,10 @@ def poison_audio(waveform, sample_rate, encoders, audio_difference_weights=[1],
                 style_waveform=style_inputs['input_values'][0]
                 #Andy removed: wandb.log({f"transfer style": wandb.Audio(style_waveform[0].detach().numpy().flatten(), sample_rate=sample_rate)}, step=0)
                 style_waveform=torch.reshape(style_waveform, (1, style_waveform.shape[0], style_waveform.shape[1]))
-                style_waveform=style_waveform.to(device='cuda')
-                style_inputs["padding_mask"]=style_inputs["padding_mask"].to(device='cuda')
+                #Andy edited: style_waveform=style_waveform.to(device='cuda')
+                style_waveform=style_waveform
+                #Andy edited: style_inputs["padding_mask"]=style_inputs["padding_mask"].to(device='cuda')
+                style_inputs["padding_mask"]=style_inputs["padding_mask"]
                 # unperturbed_latent=encoder(waveform, inputs["padding_mask"]).audio_values.detach()
                 unperturbed_waveform=style_waveform.clone().detach()
                 unperturbed_latents=[]
@@ -148,7 +151,7 @@ def poison_audio(waveform, sample_rate, encoders, audio_difference_weights=[1],
                 
                             
             noise=torch.normal(torch.zeros(waveform.shape), 0.0)
-            noise=noise.to(device='cuda')
+            #Andy removed: noise=noise.to(device='cuda')
             noise.requires_grad=True
             # waveform=torch.nn.parameter.Parameter(waveform)
             
@@ -169,10 +172,10 @@ def poison_audio(waveform, sample_rate, encoders, audio_difference_weights=[1],
             # )
             
             downsample = torchaudio.transforms.Resample(sample_rate, 22050)
-            downsample=downsample.to(device='cuda')
+            #Andy removed: downsample=downsample.to(device='cuda')
             cos = torch.nn.CosineSimilarity()
             mrstft = auraloss.perceptual.FIRFilter()#auraloss.time.SISDRLoss()#torch.nn.functional.l1_loss
-            mrstft.to(device='cuda')
+            #Andy removed: mrstft.to(device='cuda')
             
             waveform_loss = losses.L1Loss()
             stft_loss = losses.MultiScaleSTFTLoss()