nohup: ignoring input ----------------- Options --------------- arch: CLIP:ViT-L/14 batch_size: 32 [default: 10] beta1: 0.9 blur_sig: 0 checkpoints_dir: ./checkpoints class_bal: False data_label: train epoch: 50 [default: 100] fake_list_path: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake fine_tune: False fix_backbone: False fix_encoder: False gpu_ids: 1 isTrain: True [default: None] jpg_method: cv2 jpg_qual: 75 loss_freq: 100 lr: 2e-09 name: lipfd_train [default: experiment_name] num_threads: 0 optim: adam pretrained_model: ./checkpoints/experiment_name/model_epoch_29.pth real_list_path: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/0_real rz_interp: bilinear save_epoch_freq: 1 serial_batches: False suffix: train_split: train val_split: val weight_decay: 0.0001 ----------------- End ------------------- Length of data loader: 1483 Length of val loader: 189 epoch: 0 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 109.3497085571289 step: 100 Train loss: 115.77810668945312 step: 200 Train loss: 87.44400787353516 step: 300 Train loss: 121.6205062866211 step: 400 Train loss: 98.32791137695312 step: 500 Train loss: 105.16156005859375 step: 600 Train loss: 126.1203384399414 step: 700 Train loss: 97.93379211425781 step: 800 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 115.77715301513672 step: 900 Train loss: 102.43711853027344 step: 1000 Train loss: 98.23110961914062 step: 1100 Train loss: 107.89041137695312 step: 1200 Train loss: 121.88636779785156 step: 1300 Train loss: 93.9386978149414 step: 1400 saving the model at the end of epoch 0 validating... (Val @ epoch 0) acc: 0.5619205298013245 ap: 0.5358050929646152 fpr: 0.5423841059602649 fnr: 0.3337748344370861 epoch: 1 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 80.58499145507812 step: 1500 Train loss: 93.76890563964844 step: 1600 Train loss: 99.49382019042969 step: 1700 Train loss: 114.85481262207031 step: 1800 Train loss: 111.15044403076172 step: 1900 Train loss: 98.20150756835938 step: 2000 Train loss: 108.94288635253906 step: 2100 Train loss: 97.31046295166016 step: 2200 Train loss: 111.14601135253906 step: 2300 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 99.89578247070312 step: 2400 Train loss: 104.52080535888672 step: 2500 Train loss: 102.98640441894531 step: 2600 Train loss: 110.73353576660156 step: 2700 Train loss: 110.9141616821289 step: 2800 Train loss: 102.97096252441406 step: 2900 saving the model at the end of epoch 1 validating... (Val @ epoch 1) acc: 0.5668874172185431 ap: 0.5403939443490584 fpr: 0.6112582781456953 fnr: 0.25496688741721857 epoch: 2 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 103.4789810180664 step: 3000 Train loss: 103.80142211914062 step: 3100 Train loss: 104.60528564453125 step: 3200 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 104.47315979003906 step: 3300 Train loss: 117.82144165039062 step: 3400 Train loss: 107.26661682128906 step: 3500 Train loss: 107.14081573486328 step: 3600 Train loss: 120.1827163696289 step: 3700 Train loss: 106.52513122558594 step: 3800 Train loss: 113.37628173828125 step: 3900 Train loss: 95.05484008789062 step: 4000 Train loss: 103.19268035888672 step: 4100 Train loss: 98.6756591796875 step: 4200 Train loss: 95.58908081054688 step: 4300 Train loss: 109.18098449707031 step: 4400 saving the model at the end of epoch 2 validating... (Val @ epoch 2) acc: 0.7384105960264901 ap: 0.6719250858323706 fpr: 0.22251655629139072 fnr: 0.30066225165562915 epoch: 3 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 95.00332641601562 step: 4500 Train loss: 99.6748046875 step: 4600 Train loss: 92.31637573242188 step: 4700 Train loss: 116.52984619140625 step: 4800 Train loss: 127.14207458496094 step: 4900 Train loss: 109.15467834472656 step: 5000 Train loss: 116.30558776855469 step: 5100 Train loss: 109.86700439453125 step: 5200 Train loss: 109.25109100341797 step: 5300 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 104.51777648925781 step: 5400 Train loss: 113.69713592529297 step: 5500 Train loss: 111.95301818847656 step: 5600 Train loss: 103.20728302001953 step: 5700 Train loss: 94.30152130126953 step: 5800 Train loss: 98.17637634277344 step: 5900 saving the model at the end of epoch 3 validating... (Val @ epoch 3) acc: 0.7683774834437086 ap: 0.7047433801685381 fpr: 0.22119205298013245 fnr: 0.24205298013245033 epoch: 4 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 105.40957641601562 step: 6000 Train loss: 90.19412231445312 step: 6100 Train loss: 119.46781158447266 step: 6200 Train loss: 111.60482788085938 step: 6300 Train loss: 91.49681854248047 step: 6400 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 105.05416870117188 step: 6500 Train loss: 93.8360595703125 step: 6600 Train loss: 113.2322998046875 step: 6700 Train loss: 94.4849853515625 step: 6800 Train loss: 104.68202209472656 step: 6900 Train loss: 109.71572875976562 step: 7000 Train loss: 101.18453216552734 step: 7100 Train loss: 110.65788269042969 step: 7200 Train loss: 94.93778991699219 step: 7300 Train loss: 98.08305358886719 step: 7400 saving the model at the end of epoch 4 validating... (Val @ epoch 4) acc: 0.8084437086092715 ap: 0.7422743564712114 fpr: 0.15132450331125827 fnr: 0.23178807947019867 epoch: 5 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 119.11969757080078 step: 7500 Train loss: 101.54550170898438 step: 7600 Train loss: 125.54096221923828 step: 7700 Train loss: 107.59408569335938 step: 7800 Train loss: 90.32762145996094 step: 7900 Train loss: 105.04289245605469 step: 8000 Train loss: 118.83558654785156 step: 8100 Train loss: 108.56706237792969 step: 8200 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 79.27407836914062 step: 8300 Train loss: 78.02943420410156 step: 8400 Train loss: 100.42111206054688 step: 8500 Train loss: 111.81028747558594 step: 8600 Train loss: 87.120849609375 step: 8700 Train loss: 95.80047607421875 step: 8800 saving the model at the end of epoch 5 validating... (Val @ epoch 5) acc: 0.8337748344370861 ap: 0.7723560796362805 fpr: 0.1380794701986755 fnr: 0.1943708609271523 epoch: 6 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 101.03681945800781 step: 8900 Train loss: 100.72366333007812 step: 9000 Train loss: 100.6364517211914 step: 9100 Train loss: 92.65306854248047 step: 9200 Train loss: 90.16053771972656 step: 9300 Train loss: 96.73390197753906 step: 9400 Train loss: 106.46004486083984 step: 9500 Train loss: 103.21466064453125 step: 9600 Train loss: 116.55767822265625 step: 9700 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 85.55316162109375 step: 9800 Train loss: 96.09256744384766 step: 9900 Train loss: 102.47693634033203 step: 10000 Train loss: 81.52528381347656 step: 10100 Train loss: 108.37000274658203 step: 10200 Train loss: 103.06450653076172 step: 10300 saving the model at the end of epoch 6 validating... (Val @ epoch 6) acc: 0.8528145695364239 ap: 0.7921117857965466 fpr: 0.10927152317880795 fnr: 0.18509933774834436 epoch: 7 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 98.67803955078125 step: 10400 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 76.78703308105469 step: 10500 Train loss: 85.51482391357422 step: 10600 Train loss: 92.47024536132812 step: 10700 Train loss: 114.7383804321289 step: 10800 Train loss: 102.96634674072266 step: 10900 Train loss: 87.48692321777344 step: 11000 Train loss: 90.52207946777344 step: 11100 Train loss: 111.1252212524414 step: 11200 Train loss: 87.54425811767578 step: 11300 Train loss: 105.98014831542969 step: 11400 Train loss: 90.05438232421875 step: 11500 Train loss: 113.97802734375 step: 11600 Train loss: 102.34657287597656 step: 11700 Train loss: 84.95005798339844 step: 11800 saving the model at the end of epoch 7 validating... (Val @ epoch 7) acc: 0.8683774834437086 ap: 0.8040541849307241 fpr: 0.06556291390728476 fnr: 0.19768211920529802 epoch: 8 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 85.20083618164062 step: 11900 Train loss: 87.54823303222656 step: 12000 Train loss: 113.52896881103516 step: 12100 Train loss: 129.43238830566406 step: 12200 Train loss: 114.55952453613281 step: 12300 Train loss: 103.117431640625 step: 12400 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 91.63424682617188 step: 12500 Train loss: 95.65892028808594 step: 12600 Train loss: 90.61808776855469 step: 12700 Train loss: 96.25401306152344 step: 12800 Train loss: 97.88920593261719 step: 12900 Train loss: 86.9149398803711 step: 13000 Train loss: 102.21635437011719 step: 13100 Train loss: 109.04608917236328 step: 13200 Train loss: 125.03392028808594 step: 13300 saving the model at the end of epoch 8 validating... (Val @ epoch 8) acc: 0.8783112582781457 ap: 0.8186793368354413 fpr: 0.06920529801324503 fnr: 0.1741721854304636 epoch: 9 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 82.39558410644531 step: 13400 Train loss: 95.42220306396484 step: 13500 Train loss: 86.77865600585938 step: 13600 Train loss: 85.80287170410156 step: 13700 Train loss: 90.50550842285156 step: 13800 Train loss: 117.10944366455078 step: 13900 Train loss: 101.46859741210938 step: 14000 Train loss: 129.0438995361328 step: 14100 Train loss: 91.16072845458984 step: 14200 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 101.34473419189453 step: 14300 Train loss: 111.5604019165039 step: 14400 Train loss: 113.9402847290039 step: 14500 Train loss: 102.88592529296875 step: 14600 Train loss: 91.57630920410156 step: 14700 Train loss: 105.73886108398438 step: 14800 saving the model at the end of epoch 9 validating... (Val @ epoch 9) acc: 0.8816225165562914 ap: 0.8196018733092221 fpr: 0.052980132450331126 fnr: 0.1837748344370861 epoch: 10 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 97.87602233886719 step: 14900 Train loss: 95.77507781982422 step: 15000 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 95.00900268554688 step: 15100 Train loss: 107.40650939941406 step: 15200 Train loss: 96.98809814453125 step: 15300 Train loss: 112.39302062988281 step: 15400 Train loss: 88.38487243652344 step: 15500 Train loss: 101.34196472167969 step: 15600 Train loss: 102.19132995605469 step: 15700 Train loss: 113.40335845947266 step: 15800 Train loss: 89.87989044189453 step: 15900 Train loss: 104.63227081298828 step: 16000 Train loss: 98.57897186279297 step: 16100 Train loss: 86.28732299804688 step: 16200 Train loss: 101.2440185546875 step: 16300 saving the model at the end of epoch 10 validating... (Val @ epoch 10) acc: 0.8892384105960265 ap: 0.8295498309705669 fpr: 0.04933774834437086 fnr: 0.17218543046357615 epoch: 11 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 108.46096801757812 step: 16400 Train loss: 105.14181518554688 step: 16500 Train loss: 93.53913116455078 step: 16600 Train loss: 116.83553314208984 step: 16700 Train loss: 90.81753540039062 step: 16800 Train loss: 125.16203308105469 step: 16900 Train loss: 93.55419921875 step: 17000 Train loss: 86.73138427734375 step: 17100 Train loss: 94.1943588256836 step: 17200 Train loss: 82.45521545410156 step: 17300 Train loss: 96.87867736816406 step: 17400 Train loss: 79.26760864257812 step: 17500 Train loss: 99.71473693847656 step: 17600 Train loss: 101.4261474609375 step: 17700 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png saving the model at the end of epoch 11 validating... (Val @ epoch 11) acc: 0.895364238410596 ap: 0.8384289589972633 fpr: 0.04933774834437086 fnr: 0.15993377483443708 epoch: 12 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 99.41751861572266 step: 17800 Train loss: 100.92225646972656 step: 17900 Train loss: 100.92166900634766 step: 18000 Train loss: 80.60731506347656 step: 18100 Train loss: 84.21905517578125 step: 18200 Train loss: 99.12002563476562 step: 18300 Train loss: 97.47544860839844 step: 18400 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 95.96418762207031 step: 18500 Train loss: 92.93983459472656 step: 18600 Train loss: 102.73402404785156 step: 18700 Train loss: 115.28011322021484 step: 18800 Train loss: 78.9912109375 step: 18900 Train loss: 96.33280944824219 step: 19000 Train loss: 86.17536926269531 step: 19100 Train loss: 99.39048767089844 step: 19200 saving the model at the end of epoch 12 validating... (Val @ epoch 12) acc: 0.8948675496688742 ap: 0.8358056681580142 fpr: 0.041721854304635764 fnr: 0.16854304635761588 epoch: 13 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 97.01101684570312 step: 19300 Train loss: 99.92607116699219 step: 19400 Train loss: 94.14410400390625 step: 19500 Train loss: 108.56205749511719 step: 19600 Train loss: 100.33192443847656 step: 19700 Train loss: 119.96015167236328 step: 19800 Train loss: 100.62596130371094 step: 19900 Train loss: 89.90122985839844 step: 20000 Train loss: 104.45906066894531 step: 20100 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 109.33349609375 step: 20200 Train loss: 104.80198669433594 step: 20300 Train loss: 101.63945770263672 step: 20400 Train loss: 93.26719665527344 step: 20500 Train loss: 89.44151306152344 step: 20600 Train loss: 88.83653259277344 step: 20700 saving the model at the end of epoch 13 validating... (Val @ epoch 13) acc: 0.9021523178807948 ap: 0.8468278718816762 fpr: 0.04304635761589404 fnr: 0.15264900662251657 epoch: 14 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 92.56515502929688 step: 20800 Train loss: 112.10537719726562 step: 20900 Train loss: 92.1612548828125 step: 21000 Train loss: 100.8740234375 step: 21100 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 89.64739227294922 step: 21200 Train loss: 105.49078369140625 step: 21300 Train loss: 81.24507904052734 step: 21400 Train loss: 93.61893463134766 step: 21500 Train loss: 105.0328369140625 step: 21600 Train loss: 81.52472686767578 step: 21700 Train loss: 82.63867950439453 step: 21800 Train loss: 93.63219451904297 step: 21900 Train loss: 88.57166290283203 step: 22000 Train loss: 95.90657806396484 step: 22100 Train loss: 103.16031646728516 step: 22200 saving the model at the end of epoch 14 validating... (Val @ epoch 14) acc: 0.9026490066225166 ap: 0.8462686098988782 fpr: 0.0380794701986755 fnr: 0.1566225165562914 epoch: 15 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 114.83440399169922 step: 22300 Train loss: 108.83474731445312 step: 22400 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 88.62757873535156 step: 22500 Train loss: 96.61287689208984 step: 22600 Train loss: 115.21736145019531 step: 22700 Train loss: 89.04222106933594 step: 22800 Train loss: 88.95414733886719 step: 22900 Train loss: 85.84208679199219 step: 23000 Train loss: 114.43083190917969 step: 23100 Train loss: 111.42693328857422 step: 23200 Train loss: 101.97541809082031 step: 23300 Train loss: 103.68124389648438 step: 23400 Train loss: 88.98529052734375 step: 23500 Train loss: 88.13697814941406 step: 23600 Train loss: 96.48775482177734 step: 23700 saving the model at the end of epoch 15 validating... (Val @ epoch 15) acc: 0.904635761589404 ap: 0.8482638373065826 fpr: 0.03443708609271523 fnr: 0.1562913907284768 epoch: 16 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 82.79197692871094 step: 23800 Train loss: 92.29851531982422 step: 23900 Train loss: 101.32656860351562 step: 24000 Train loss: 82.86312866210938 step: 24100 Train loss: 99.44190979003906 step: 24200 Train loss: 114.21587371826172 step: 24300 Train loss: 93.10935974121094 step: 24400 Train loss: 110.8064193725586 step: 24500 Train loss: 107.86590576171875 step: 24600 Train loss: 104.00509643554688 step: 24700 Train loss: 105.76403045654297 step: 24800 Train loss: 107.95489501953125 step: 24900 Train loss: 103.10502624511719 step: 25000 Train loss: 96.59196472167969 step: 25100 Train loss: 99.68357849121094 step: 25200 saving the model at the end of epoch 16 validating... (Val @ epoch 16) acc: 0.9076158940397351 ap: 0.852086660247679 fpr: 0.032119205298013244 fnr: 0.15264900662251657 epoch: 17 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 92.31273651123047 step: 25300 Train loss: 84.96509552001953 step: 25400 Train loss: 86.664306640625 step: 25500 Train loss: 108.47801208496094 step: 25600 Train loss: 84.15589904785156 step: 25700 Train loss: 106.80313873291016 step: 25800 Train loss: 100.79236602783203 step: 25900 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 103.33578491210938 step: 26000 Train loss: 88.80908966064453 step: 26100 Train loss: 99.35258483886719 step: 26200 Train loss: 92.63272094726562 step: 26300 Train loss: 99.22555541992188 step: 26400 Train loss: 81.28981018066406 step: 26500 Train loss: 92.94877624511719 step: 26600 saving the model at the end of epoch 17 validating... (Val @ epoch 17) acc: 0.9135761589403973 ap: 0.8610761985014281 fpr: 0.032119205298013244 fnr: 0.14072847682119205 epoch: 18 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 104.60865783691406 step: 26700 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 88.01316833496094 step: 26800 Train loss: 95.64061737060547 step: 26900 Train loss: 96.63502502441406 step: 27000 Train loss: 97.56912231445312 step: 27100 Train loss: 96.4901123046875 step: 27200 Train loss: 96.18511962890625 step: 27300 Train loss: 85.32862854003906 step: 27400 Train loss: 102.43637084960938 step: 27500 Train loss: 101.28318786621094 step: 27600 Train loss: 110.92155456542969 step: 27700 Train loss: 92.72991943359375 step: 27800 Train loss: 100.44973754882812 step: 27900 Train loss: 99.80447387695312 step: 28000 Train loss: 98.5571517944336 step: 28100 saving the model at the end of epoch 18 validating... (Val @ epoch 18) acc: 0.9124172185430464 ap: 0.8580459269467713 fpr: 0.02748344370860927 fnr: 0.147682119205298 epoch: 19 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 89.30645751953125 step: 28200 Train loss: 95.78688049316406 step: 28300 Train loss: 100.19153594970703 step: 28400 Train loss: 89.90309143066406 step: 28500 Train loss: 103.15199279785156 step: 28600 Train loss: 95.88514709472656 step: 28700 Train loss: 77.65068817138672 step: 28800 Train loss: 95.78532409667969 step: 28900 Train loss: 91.47555541992188 step: 29000 Train loss: 102.67447662353516 step: 29100 Train loss: 85.68608093261719 step: 29200 Train loss: 81.28839111328125 step: 29300 Train loss: 87.3929672241211 step: 29400 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 127.06197357177734 step: 29500 Train loss: 88.83319854736328 step: 29600 saving the model at the end of epoch 19 validating... (Val @ epoch 19) acc: 0.9149006622516557 ap: 0.8619897276184927 fpr: 0.028145695364238412 fnr: 0.14205298013245032 epoch: 20 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 103.29493713378906 step: 29700 Train loss: 95.25360107421875 step: 29800 Train loss: 95.465576171875 step: 29900 Train loss: 87.75604248046875 step: 30000 Train loss: 92.74017333984375 step: 30100 Train loss: 86.07817077636719 step: 30200 Train loss: 98.4383544921875 step: 30300 Train loss: 95.09693908691406 step: 30400 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 93.51901245117188 step: 30500 Train loss: 93.16593933105469 step: 30600 Train loss: 94.72856140136719 step: 30700 Train loss: 96.36589050292969 step: 30800 Train loss: 81.53694152832031 step: 30900 Train loss: 111.12959289550781 step: 31000 Train loss: 91.42665100097656 step: 31100 saving the model at the end of epoch 20 validating... (Val @ epoch 20) acc: 0.9147350993377483 ap: 0.8610072160162654 fpr: 0.025496688741721854 fnr: 0.14503311258278145 epoch: 21 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 98.46931457519531 step: 31200 Train loss: 88.39454650878906 step: 31300 Train loss: 86.25625610351562 step: 31400 Train loss: 91.50813293457031 step: 31500 Train loss: 103.72244262695312 step: 31600 Train loss: 96.97940063476562 step: 31700 Train loss: 129.3763427734375 step: 31800 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 87.54872131347656 step: 31900 Train loss: 110.87643432617188 step: 32000 Train loss: 94.82489013671875 step: 32100 Train loss: 91.09426879882812 step: 32200 Train loss: 141.98573303222656 step: 32300 Train loss: 91.30592346191406 step: 32400 Train loss: 91.48355102539062 step: 32500 Train loss: 84.7899169921875 step: 32600 saving the model at the end of epoch 21 validating... (Val @ epoch 21) acc: 0.9175496688741722 ap: 0.8655735635765534 fpr: 0.026490066225165563 fnr: 0.13841059602649006 epoch: 22 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 81.1429443359375 step: 32700 Train loss: 80.49852752685547 step: 32800 Train loss: 103.57378387451172 step: 32900 Train loss: 115.09785461425781 step: 33000 Train loss: 100.0188980102539 step: 33100 Train loss: 117.90495300292969 step: 33200 Train loss: 115.60507202148438 step: 33300 Train loss: 95.4669189453125 step: 33400 Train loss: 92.07056427001953 step: 33500 Train loss: 94.75532531738281 step: 33600 Train loss: 104.36040496826172 step: 33700 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 102.83029174804688 step: 33800 Train loss: 107.6717529296875 step: 33900 Train loss: 95.68386840820312 step: 34000 Train loss: 99.3794174194336 step: 34100 saving the model at the end of epoch 22 validating... (Val @ epoch 22) acc: 0.9193708609271524 ap: 0.8676197734205308 fpr: 0.02384105960264901 fnr: 0.13741721854304637 epoch: 23 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 107.59786987304688 step: 34200 Train loss: 114.41547393798828 step: 34300 Train loss: 98.87919616699219 step: 34400 Train loss: 103.96318054199219 step: 34500 Train loss: 100.78030395507812 step: 34600 Train loss: 102.7208480834961 step: 34700 Train loss: 107.3170166015625 step: 34800 Train loss: 70.77960205078125 step: 34900 Train loss: 107.55149841308594 step: 35000 Train loss: 95.44448852539062 step: 35100 Train loss: 94.88804626464844 step: 35200 Train loss: 86.25000762939453 step: 35300 Train loss: 102.06913757324219 step: 35400 Train loss: 92.03421020507812 step: 35500 saving the model at the end of epoch 23 validating... (Val @ epoch 23) acc: 0.9218543046357616 ap: 0.871453551393735 fpr: 0.02384105960264901 fnr: 0.13245033112582782 epoch: 24 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 101.97744750976562 step: 35600 Train loss: 81.46246337890625 step: 35700 Train loss: 94.9013671875 step: 35800 Train loss: 91.16714477539062 step: 35900 Train loss: 95.36466979980469 step: 36000 Train loss: 106.69194030761719 step: 36100 Train loss: 95.71537780761719 step: 36200 Train loss: 95.00849914550781 step: 36300 Train loss: 104.0283203125 step: 36400 Train loss: 98.74684143066406 step: 36500 Train loss: 102.49458312988281 step: 36600 Train loss: 77.29928588867188 step: 36700 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 98.31216430664062 step: 36800 Train loss: 99.293212890625 step: 36900 Train loss: 83.95875549316406 step: 37000 saving the model at the end of epoch 24 validating... (Val @ epoch 24) acc: 0.919205298013245 ap: 0.8667115617611424 fpr: 0.02152317880794702 fnr: 0.1400662251655629 epoch: 25 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 106.17198181152344 step: 37100 Train loss: 95.02005004882812 step: 37200 Train loss: 95.54281616210938 step: 37300 Train loss: 85.25819396972656 step: 37400 Train loss: 95.11949157714844 step: 37500 Train loss: 99.20923614501953 step: 37600 Train loss: 107.23602294921875 step: 37700 Train loss: 87.11701965332031 step: 37800 Train loss: 94.58992004394531 step: 37900 Train loss: 121.3162841796875 step: 38000 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 91.7796401977539 step: 38100 Train loss: 107.27289581298828 step: 38200 Train loss: 94.27934265136719 step: 38300 Train loss: 95.79428100585938 step: 38400 Train loss: 103.22969055175781 step: 38500 saving the model at the end of epoch 25 validating... (Val @ epoch 25) acc: 0.9216887417218543 ap: 0.8704344488944431 fpr: 0.02119205298013245 fnr: 0.13543046357615893 epoch: 26 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 96.3729476928711 step: 38600 Train loss: 98.09357452392578 step: 38700 Train loss: 95.33633422851562 step: 38800 Train loss: 108.59393310546875 step: 38900 Train loss: 90.77816772460938 step: 39000 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 106.17906188964844 step: 39100 Train loss: 98.71460723876953 step: 39200 Train loss: 107.44701385498047 step: 39300 Train loss: 102.05140686035156 step: 39400 Train loss: 90.86222839355469 step: 39500 Train loss: 83.28683471679688 step: 39600 Train loss: 95.26985168457031 step: 39700 Train loss: 92.15057373046875 step: 39800 Train loss: 99.08290100097656 step: 39900 Train loss: 88.71820068359375 step: 40000 saving the model at the end of epoch 26 validating... (Val @ epoch 26) acc: 0.9230132450331126 ap: 0.8722919306182526 fpr: 0.02052980132450331 fnr: 0.13344370860927152 epoch: 27 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 95.718505859375 step: 40100 Train loss: 83.5436019897461 step: 40200 Train loss: 90.60633850097656 step: 40300 Train loss: 95.76900482177734 step: 40400 Train loss: 77.22334289550781 step: 40500 Train loss: 106.1387939453125 step: 40600 Train loss: 105.90230560302734 step: 40700 Train loss: 98.82454681396484 step: 40800 Train loss: 80.53895568847656 step: 40900 Train loss: 96.0670166015625 step: 41000 Train loss: 96.09965515136719 step: 41100 Train loss: 102.98234558105469 step: 41200 Train loss: 94.43991088867188 step: 41300 Train loss: 82.52233123779297 step: 41400 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 99.63735961914062 step: 41500 saving the model at the end of epoch 27 validating... (Val @ epoch 27) acc: 0.9231788079470199 ap: 0.8720708628922907 fpr: 0.018874172185430464 fnr: 0.1347682119205298 epoch: 28 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 94.73419189453125 step: 41600 Train loss: 86.86043548583984 step: 41700 Train loss: 94.25665283203125 step: 41800 Train loss: 87.87420654296875 step: 41900 Train loss: 94.3466796875 step: 42000 Train loss: 87.02870178222656 step: 42100 Train loss: 94.72970581054688 step: 42200 Train loss: 73.56180572509766 step: 42300 Train loss: 107.19182586669922 step: 42400 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 86.6338882446289 step: 42500 Train loss: 90.05741882324219 step: 42600 Train loss: 96.47938537597656 step: 42700 Train loss: 87.6587142944336 step: 42800 Train loss: 99.31858825683594 step: 42900 Train loss: 88.81858825683594 step: 43000 saving the model at the end of epoch 28 validating... (Val @ epoch 28) acc: 0.9263245033112583 ap: 0.8769619526443624 fpr: 0.018874172185430464 fnr: 0.12847682119205298 epoch: 29 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 109.44593811035156 step: 43100 Train loss: 84.59053039550781 step: 43200 Train loss: 84.24394226074219 step: 43300 Train loss: 119.02490234375 step: 43400 Train loss: 101.92208862304688 step: 43500 Train loss: 105.8259048461914 step: 43600 Train loss: 98.34077453613281 step: 43700 Train loss: 105.81497192382812 step: 43800 Train loss: 91.91256713867188 step: 43900 Train loss: 99.22040557861328 step: 44000 Train loss: 94.55714416503906 step: 44100 Train loss: 97.93988037109375 step: 44200 Train loss: 84.58506774902344 step: 44300 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 80.71188354492188 step: 44400 saving the model at the end of epoch 29 validating... (Val @ epoch 29) acc: 0.9253311258278145 ap: 0.875024502327229 fpr: 0.017549668874172187 fnr: 0.1317880794701987 epoch: 30 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 107.75877380371094 step: 44500 Train loss: 99.62721252441406 step: 44600 Train loss: 98.58932495117188 step: 44700 Train loss: 80.76273345947266 step: 44800 Train loss: 95.53893280029297 step: 44900 Train loss: 94.35447692871094 step: 45000 Train loss: 98.68751525878906 step: 45100 Train loss: 91.52035522460938 step: 45200 Train loss: 85.08636474609375 step: 45300 Train loss: 99.96929931640625 step: 45400 Train loss: 110.87091064453125 step: 45500 Train loss: 102.12557983398438 step: 45600 Train loss: 80.78931427001953 step: 45700 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 99.07274627685547 step: 45800 Train loss: 88.63421630859375 step: 45900 saving the model at the end of epoch 30 validating... (Val @ epoch 30) acc: 0.9245033112582781 ap: 0.873643483930774 fpr: 0.017218543046357615 fnr: 0.1337748344370861 epoch: 31 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 95.41173553466797 step: 46000 Train loss: 91.78175354003906 step: 46100 Train loss: 99.63282775878906 step: 46200 Train loss: 126.46421813964844 step: 46300 Train loss: 91.29084777832031 step: 46400 Train loss: 94.41521453857422 step: 46500 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 87.62366485595703 step: 46600 Train loss: 83.82557678222656 step: 46700 Train loss: 119.25001525878906 step: 46800 Train loss: 80.16643524169922 step: 46900 Train loss: 96.1780014038086 step: 47000 Train loss: 94.11321258544922 step: 47100 Train loss: 69.96588134765625 step: 47200 Train loss: 97.95004272460938 step: 47300 Train loss: 91.87395477294922 step: 47400 saving the model at the end of epoch 31 validating... (Val @ epoch 31) acc: 0.9253311258278145 ap: 0.8743518106575674 fpr: 0.015231788079470199 fnr: 0.13410596026490065 epoch: 32 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 94.93876647949219 step: 47500 Train loss: 74.20389556884766 step: 47600 Train loss: 93.996337890625 step: 47700 Train loss: 91.61723327636719 step: 47800 Train loss: 80.5855484008789 step: 47900 Train loss: 105.62805938720703 step: 48000 Train loss: 77.38821411132812 step: 48100 Train loss: 90.43487548828125 step: 48200 Train loss: 104.10159301757812 step: 48300 Train loss: 96.5937271118164 step: 48400 Train loss: 115.57551574707031 step: 48500 Train loss: 83.31884002685547 step: 48600 Train loss: 94.00515747070312 step: 48700 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 91.11808013916016 step: 48800 Train loss: 95.9767074584961 step: 48900 saving the model at the end of epoch 32 validating... (Val @ epoch 32) acc: 0.9326158940397351 ap: 0.8868095052590573 fpr: 0.018543046357615896 fnr: 0.1162251655629139 epoch: 33 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 77.57847595214844 step: 49000 Train loss: 114.95893859863281 step: 49100 Train loss: 89.53807830810547 step: 49200 Train loss: 102.10774993896484 step: 49300 Train loss: 91.46820068359375 step: 49400 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 107.26177978515625 step: 49500 Train loss: 79.87150573730469 step: 49600 Train loss: 80.25978088378906 step: 49700 Train loss: 86.90876007080078 step: 49800 Train loss: 108.16836547851562 step: 49900 Train loss: 84.42555236816406 step: 50000 Train loss: 98.3057632446289 step: 50100 Train loss: 94.54742431640625 step: 50200 Train loss: 105.42430114746094 step: 50300 Train loss: 76.19158935546875 step: 50400 saving the model at the end of epoch 33 validating... (Val @ epoch 33) acc: 0.9317880794701987 ap: 0.8850802343352012 fpr: 0.017218543046357615 fnr: 0.11920529801324503 epoch: 34 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 125.81848907470703 step: 50500 Train loss: 92.07084655761719 step: 50600 Train loss: 73.8358383178711 step: 50700 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 125.55552673339844 step: 50800 Train loss: 83.88774108886719 step: 50900 Train loss: 90.35460662841797 step: 51000 Train loss: 93.7743148803711 step: 51100 Train loss: 102.3692626953125 step: 51200 Train loss: 87.55474090576172 step: 51300 Train loss: 105.97702026367188 step: 51400 Train loss: 76.19060516357422 step: 51500 Train loss: 94.39215087890625 step: 51600 Train loss: 108.25830078125 step: 51700 Train loss: 86.64149475097656 step: 51800 Train loss: 94.81710815429688 step: 51900 saving the model at the end of epoch 34 validating... (Val @ epoch 34) acc: 0.9312913907284768 ap: 0.8838867964195967 fpr: 0.015894039735099338 fnr: 0.12152317880794702 epoch: 35 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 76.47958374023438 step: 52000 Train loss: 102.25346374511719 step: 52100 Train loss: 100.02899169921875 step: 52200 Train loss: 76.23573303222656 step: 52300 Train loss: 90.76263427734375 step: 52400 Train loss: 96.04088592529297 step: 52500 Train loss: 95.24079895019531 step: 52600 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 97.91497802734375 step: 52700 Train loss: 97.41355895996094 step: 52800 Train loss: 97.85961151123047 step: 52900 Train loss: 114.14883422851562 step: 53000 Train loss: 94.67399597167969 step: 53100 Train loss: 102.08966064453125 step: 53200 Train loss: 91.06591796875 step: 53300 saving the model at the end of epoch 35 validating... (Val @ epoch 35) acc: 0.9319536423841059 ap: 0.8851402935697372 fpr: 0.016556291390728478 fnr: 0.1195364238410596 epoch: 36 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 105.64775848388672 step: 53400 Train loss: 109.65997314453125 step: 53500 Train loss: 105.07560729980469 step: 53600 Train loss: 107.36741638183594 step: 53700 Train loss: 83.0212631225586 step: 53800 Train loss: 90.62057495117188 step: 53900 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 106.80844116210938 step: 54000 Train loss: 94.00015258789062 step: 54100 Train loss: 97.37210083007812 step: 54200 Train loss: 100.07282257080078 step: 54300 Train loss: 86.8153076171875 step: 54400 Train loss: 98.42520141601562 step: 54500 Train loss: 84.40351867675781 step: 54600 Train loss: 94.61422729492188 step: 54700 Train loss: 94.72000122070312 step: 54800 saving the model at the end of epoch 36 validating... (Val @ epoch 36) acc: 0.9354304635761589 ap: 0.890912769477931 fpr: 0.017218543046357615 fnr: 0.1119205298013245 epoch: 37 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 91.30091857910156 step: 54900 Train loss: 91.02377319335938 step: 55000 Train loss: 105.36109924316406 step: 55100 Train loss: 100.18766784667969 step: 55200 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 87.48077392578125 step: 55300 Train loss: 90.96659851074219 step: 55400 Train loss: 102.36908721923828 step: 55500 Train loss: 93.92560577392578 step: 55600 Train loss: 98.06395721435547 step: 55700 Train loss: 87.72573852539062 step: 55800 Train loss: 106.69154357910156 step: 55900 Train loss: 104.80895233154297 step: 56000 Train loss: 113.35851287841797 step: 56100 Train loss: 94.14604187011719 step: 56200 Train loss: 86.715576171875 step: 56300 saving the model at the end of epoch 37 validating... (Val @ epoch 37) acc: 0.9336092715231789 ap: 0.8871650741233064 fpr: 0.01456953642384106 fnr: 0.11821192052980133 epoch: 38 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 94.313232421875 step: 56400 Train loss: 94.04849243164062 step: 56500 Train loss: 109.79649353027344 step: 56600 Train loss: 98.18072509765625 step: 56700 Train loss: 89.97587585449219 step: 56800 Train loss: 109.39126586914062 step: 56900 Train loss: 83.82635498046875 step: 57000 Train loss: 109.32949829101562 step: 57100 Train loss: 97.58815002441406 step: 57200 Train loss: 99.30084228515625 step: 57300 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 82.94303894042969 step: 57400 Train loss: 99.66937255859375 step: 57500 Train loss: 82.54581451416016 step: 57600 Train loss: 90.79081726074219 step: 57700 Train loss: 93.85063171386719 step: 57800 saving the model at the end of epoch 38 validating... (Val @ epoch 38) acc: 0.93658940397351 ap: 0.8921524878940557 fpr: 0.015231788079470199 fnr: 0.11158940397350993 epoch: 39 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 101.46546936035156 step: 57900 Train loss: 86.85536193847656 step: 58000 Train loss: 86.88607788085938 step: 58100 Train loss: 94.00170135498047 step: 58200 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 101.704833984375 step: 58300 Train loss: 86.93431854248047 step: 58400 Train loss: 106.63441467285156 step: 58500 Train loss: 105.39498138427734 step: 58600 Train loss: 102.37200927734375 step: 58700 Train loss: 98.33232879638672 step: 58800 Train loss: 94.26693725585938 step: 58900 Train loss: 106.35066223144531 step: 59000 Train loss: 98.30419921875 step: 59100 Train loss: 79.83206176757812 step: 59200 Train loss: 90.36638641357422 step: 59300 saving the model at the end of epoch 39 validating... (Val @ epoch 39) acc: 0.9347682119205298 ap: 0.8886050604448648 fpr: 0.013245033112582781 fnr: 0.11721854304635762 epoch: 40 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 83.80421447753906 step: 59400 Train loss: 97.71980285644531 step: 59500 Train loss: 86.69020080566406 step: 59600 Train loss: 114.35474395751953 step: 59700 Train loss: 97.22096252441406 step: 59800 Train loss: 88.50094604492188 step: 59900 Train loss: 97.62434387207031 step: 60000 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 97.61849212646484 step: 60100 Train loss: 118.01669311523438 step: 60200 Train loss: 83.99238586425781 step: 60300 Train loss: 86.89488983154297 step: 60400 Train loss: 98.79199981689453 step: 60500 Train loss: 78.77233123779297 step: 60600 Train loss: 83.23431396484375 step: 60700 Train loss: 86.27637481689453 step: 60800 saving the model at the end of epoch 40 validating... (Val @ epoch 40) acc: 0.9346026490066225 ap: 0.8877273707600164 fpr: 0.011258278145695364 fnr: 0.1195364238410596 epoch: 41 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 91.28060913085938 step: 60900 Train loss: 105.48225402832031 step: 61000 Train loss: 97.78602600097656 step: 61100 Train loss: 105.30380249023438 step: 61200 Train loss: 105.22386169433594 step: 61300 Train loss: 94.03921508789062 step: 61400 Train loss: 75.66316223144531 step: 61500 Train loss: 83.37248992919922 step: 61600 Train loss: 82.9482192993164 step: 61700 Train loss: 105.40867614746094 step: 61800 Train loss: 117.3250732421875 step: 61900 Train loss: 87.09813690185547 step: 62000 Train loss: 116.06901550292969 step: 62100 Train loss: 94.27781677246094 step: 62200 saving the model at the end of epoch 41 validating... (Val @ epoch 41) acc: 0.9352649006622517 ap: 0.8892963302952812 fpr: 0.012913907284768211 fnr: 0.11655629139072848 epoch: 42 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 109.22545623779297 step: 62300 Train loss: 90.06454467773438 step: 62400 Train loss: 97.81288146972656 step: 62500 Train loss: 94.01913452148438 step: 62600 Train loss: 89.94760131835938 step: 62700 Train loss: 113.94859313964844 step: 62800 Train loss: 109.82142639160156 step: 62900 Train loss: 94.44390869140625 step: 63000 Train loss: 121.51100158691406 step: 63100 Train loss: 90.37667846679688 step: 63200 Train loss: 90.16975402832031 step: 63300 Train loss: 101.73966979980469 step: 63400 Train loss: 77.35667419433594 step: 63500 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 90.05270385742188 step: 63600 Train loss: 102.1438980102539 step: 63700 saving the model at the end of epoch 42 validating... (Val @ epoch 42) acc: 0.9407284768211921 ap: 0.8985567058009383 fpr: 0.01423841059602649 fnr: 0.10430463576158941 epoch: 43 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 94.28565979003906 step: 63800 Train loss: 88.5025405883789 step: 63900 Train loss: 84.7755126953125 step: 64000 Train loss: 95.95376586914062 step: 64100 Train loss: 83.29885864257812 step: 64200 Train loss: 94.64238739013672 step: 64300 Train loss: 97.70860290527344 step: 64400 Train loss: 96.45343780517578 step: 64500 Train loss: 94.48623657226562 step: 64600 Train loss: 117.8350830078125 step: 64700 Train loss: 96.92566680908203 step: 64800 Train loss: 77.26060485839844 step: 64900 Train loss: 102.19673156738281 step: 65000 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 94.15009307861328 step: 65100 Train loss: 97.46650695800781 step: 65200 saving the model at the end of epoch 43 validating... (Val @ epoch 43) acc: 0.9400662251655629 ap: 0.8972597832630945 fpr: 0.013576158940397352 fnr: 0.10629139072847682 epoch: 44 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 105.65857696533203 step: 65300 Train loss: 85.96650695800781 step: 65400 Train loss: 95.0820541381836 step: 65500 Train loss: 94.70610809326172 step: 65600 Train loss: 93.39395141601562 step: 65700 Train loss: 100.08433532714844 step: 65800 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 101.28034973144531 step: 65900 Train loss: 89.05029296875 step: 66000 Train loss: 90.05953216552734 step: 66100 Train loss: 87.03773498535156 step: 66200 Train loss: 109.26770782470703 step: 66300 Train loss: 82.67086791992188 step: 66400 Train loss: 76.64437103271484 step: 66500 Train loss: 86.54629516601562 step: 66600 Train loss: 76.48202514648438 step: 66700 saving the model at the end of epoch 44 validating... (Val @ epoch 44) acc: 0.938907284768212 ap: 0.8951623726650679 fpr: 0.012913907284768211 fnr: 0.10927152317880795 epoch: 45 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 102.70532989501953 step: 66800 Train loss: 98.54911804199219 step: 66900 Train loss: 80.60196685791016 step: 67000 Train loss: 89.60649108886719 step: 67100 Train loss: 101.90727233886719 step: 67200 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 110.23324584960938 step: 67300 Train loss: 90.51313781738281 step: 67400 Train loss: 82.82902526855469 step: 67500 Train loss: 114.34927368164062 step: 67600 Train loss: 104.7222671508789 step: 67700 Train loss: 104.95767211914062 step: 67800 Train loss: 93.2955322265625 step: 67900 Train loss: 103.30267333984375 step: 68000 Train loss: 101.23458862304688 step: 68100 Train loss: 90.28176879882812 step: 68200 saving the model at the end of epoch 45 validating... (Val @ epoch 45) acc: 0.9380794701986755 ap: 0.8932968424573797 fpr: 0.011258278145695364 fnr: 0.11258278145695365 epoch: 46 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 97.50723266601562 step: 68300 Train loss: 97.13092803955078 step: 68400 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 89.87747192382812 step: 68500 Train loss: 90.36524963378906 step: 68600 Train loss: 101.256591796875 step: 68700 Train loss: 99.16970825195312 step: 68800 Train loss: 109.6037826538086 step: 68900 Train loss: 105.37516784667969 step: 69000 Train loss: 94.57360076904297 step: 69100 Train loss: 97.98590850830078 step: 69200 Train loss: 97.60783386230469 step: 69300 Train loss: 85.90118408203125 step: 69400 Train loss: 97.27726745605469 step: 69500 Train loss: 116.9710922241211 step: 69600 Train loss: 97.3941650390625 step: 69700 saving the model at the end of epoch 46 validating... (Val @ epoch 46) acc: 0.9397350993377483 ap: 0.895867613538835 fpr: 0.010927152317880795 fnr: 0.10960264900662252 epoch: 47 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 97.43257141113281 step: 69800 Train loss: 94.33767700195312 step: 69900 Train loss: 101.50326538085938 step: 70000 Train loss: 76.67514038085938 step: 70100 Train loss: 90.25025939941406 step: 70200 Train loss: 82.4213638305664 step: 70300 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 105.03654479980469 step: 70400 Train loss: 92.296630859375 step: 70500 Train loss: 91.99394989013672 step: 70600 Train loss: 76.64805603027344 step: 70700 Train loss: 95.92701721191406 step: 70800 Train loss: 90.02147674560547 step: 70900 Train loss: 90.73638153076172 step: 71000 Train loss: 89.86149597167969 step: 71100 saving the model at the end of epoch 47 validating... (Val @ epoch 47) acc: 0.9405629139072847 ap: 0.897211416794766 fpr: 0.010927152317880795 fnr: 0.10794701986754966 epoch: 48 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 97.4541015625 step: 71200 Train loss: 102.45660400390625 step: 71300 Train loss: 69.94125366210938 step: 71400 Train loss: 109.10926818847656 step: 71500 Train loss: 97.28998565673828 step: 71600 Train loss: 100.94027709960938 step: 71700 Train loss: 94.3294677734375 step: 71800 Train loss: 79.23173522949219 step: 71900 Train loss: 107.92404174804688 step: 72000 Train loss: 96.29600524902344 step: 72100 Train loss: 93.08766174316406 step: 72200 Train loss: 98.92930603027344 step: 72300 Train loss: 102.63845825195312 step: 72400 Train loss: 93.48680114746094 step: 72500 Train loss: 105.561767578125 step: 72600 saving the model at the end of epoch 48 validating... (Val @ epoch 48) acc: 0.9395695364238411 ap: 0.8954934582458408 fpr: 0.010596026490066225 fnr: 0.11026490066225166 epoch: 49 /opt/conda/envs/LipFD/lib/python3.10/site-packages/torchvision/transforms/functional.py:1603: UserWarning: The default value of the antialias parameter of all the resizing transforms (Resize(), RandomResizedCrop(), etc.) will change from None to True in v0.17, in order to be consistent across the PIL and Tensor backends. To suppress this warning, directly pass antialias=True (recommended, future default), antialias=None (current default, which means False for Tensors and True for PIL), or antialias=False (only works on Tensors - PIL will still use antialiasing). This also applies if you are using the inference transforms from the models weights: update the call to weights.transforms(antialias=True). warnings.warn( Train loss: 89.19844818115234 step: 72700 Train loss: 105.79002380371094 step: 72800 Train loss: 94.37416076660156 step: 72900 Train loss: 97.48074340820312 step: 73000 Train loss: 76.3938980102539 step: 73100 Train loss: 93.292236328125 step: 73200 Train loss: 93.73245239257812 step: 73300 Train loss: 111.39205932617188 step: 73400 Train loss: 79.65298461914062 step: 73500 Train loss: 83.60038757324219 step: 73600 Train loss: 90.1827392578125 step: 73700 Train loss: 86.8857421875 step: 73800 WARNING: Failed to read image, skipping: /apdcephfs_gy4/share_303628665/joywu/research/LipFD/datasets/FairTalking-Bench/train/1_fake/Sonic_1135_Fake_Sonic_7.png Train loss: 89.89358520507812 step: 73900 Train loss: 97.02238464355469 step: 74000 Train loss: 101.51692962646484 step: 74100 saving the model at the end of epoch 49 validating... (Val @ epoch 49) acc: 0.9405629139072847 ap: 0.896998055371985 fpr: 0.010264900662251655 fnr: 0.10860927152317881