[12:02:26] Device: cuda
[12:02:27] Loaded v10 weights from /home/coder/experiments/2026-04-14-080000-multiscale-flow-v10
[12:02:27] Model parameters: 6,169,586, channels=[56, 112, 224]
[12:02:27] Fine-tune: 10 epochs of 8-step AR with pure SSIM loss
[12:02:31]   43855 sequences
[12:15:01] Epoch 1/10 | loss=0.09345 lr=0.0000098
[12:27:32] Epoch 2/10 | loss=0.09245 lr=0.0000091
[12:29:02]   Val SSIM=0.8883 | {'pong': 0.8811, 'sonic': 0.8354, 'pole_position': 0.9485}
[12:29:02]   New best! SSIM=0.8883
[12:41:38] Epoch 3/10 | loss=0.09166 lr=0.0000081
[12:54:16] Epoch 4/10 | loss=0.09095 lr=0.0000069
[12:55:36]   Val SSIM=0.8887 | {'pong': 0.8824, 'sonic': 0.8352, 'pole_position': 0.9486}
[12:55:36]   New best! SSIM=0.8887
[13:08:12] Epoch 5/10 | loss=0.09033 lr=0.0000055
[13:20:56] Epoch 6/10 | loss=0.08982 lr=0.0000041
[13:22:21]   Val SSIM=0.8885 | {'pong': 0.8824, 'sonic': 0.8347, 'pole_position': 0.9483}
[13:35:14] Epoch 7/10 | loss=0.08941 lr=0.0000029
[13:48:14] Epoch 8/10 | loss=0.08911 lr=0.0000019
[13:49:31]   Val SSIM=0.8883 | {'pong': 0.8822, 'sonic': 0.8344, 'pole_position': 0.9484}
[14:02:17] Epoch 9/10 | loss=0.08888 lr=0.0000012
[14:14:47] Epoch 10/10 | loss=0.08874 lr=0.0000010
[14:16:07]   Val SSIM=0.8881 | {'pong': 0.8815, 'sonic': 0.8343, 'pole_position': 0.9485}
[14:16:07] Experiment dir: 12.4 MB
[14:16:07] Training complete. Best val SSIM: 0.8887