权重改为fp32, 其他部分保持bf16

目前参数--encoder_mode有三种选择: 1. fp32: 全部使用fp32, 适合显存充足的情况 2. autocast: 使用torch.cuda.amp.autocast自动混合精度, 稍微快一些, psnr下降较多 3. bf16_full: 全部使用bf16, 精度较高
2026-01-18 20:24:37 +08:00
parent 44379f3e31
commit e1b029201e
4 changed files with 75 additions and 16 deletions
--- a/unitree_g1_pack_camera/case1/output/tensorboard/events.out.tfevents.1768734530.node-0.417718.0
+++ b/unitree_g1_pack_camera/case1/output/tensorboard/events.out.tfevents.1768734530.node-0.417718.0
--- a/unitree_g1_pack_camera/case1/run_world_model_interaction.sh
+++ b/unitree_g1_pack_camera/case1/run_world_model_interaction.sh
@@ -23,5 +23,5 @@ dataset="unitree_g1_pack_camera"
        --perframe_ae \
        --diffusion_dtype bf16 \
        --projector_dtype bf16 \
-        --encoder_dtype bf16
+        --encoder_mode autocast #fp32/autocast/bf16_full
 } 2>&1 | tee "${res_dir}/output.log"