File size: 1,521 Bytes
59630ba | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 |
dataset="realestate10k_rotate" # "realestate10k"
algorithm="dfot_geometry_forcing"
exp_name="eval_360_rotation"
result_dir="$output_dir/train/$exp_name"
checkpoint_path="checkpoints/geometry_forcing_state_dict.ckpt"
eval_result_dir="$output_dir/geometry_forcing_rotation"
echo "Result directory: $result_dir"
echo "Checkpoint path: $checkpoint_path"
# ## construct evaluation args
checkpoints_dir=${result_dir}/checkpoints
python -m main +name=rotation dataset=$dataset \
algorithm=$algorithm \
experiment=video_generation @diffusion/continuous \
algorithm.alignment.apply_unnormalize_recon=True \
algorithm.alignment.latents_info=2 \
load=$checkpoint_path \
dataset.num_eval_videos=1000 \
'experiment.tasks=[validation]' experiment.validation.data.shuffle=False experiment.test.data.shuffle=False dataset.frame_skip=1 \
dataset.n_frames=16 \
algorithm.logging.max_num_videos=1000 \
dataset.context_length=1 \
algorithm.tasks.prediction.history_guidance.name=stabilized_vanilla \
+algorithm.tasks.prediction.history_guidance.guidance_scale=4.0 \
+algorithm.tasks.prediction.history_guidance.stabilization_level=0.02 \
experiment.validation.batch_size=1 \
algorithm.tasks.interpolation.history_guidance.name=vanilla \
+algorithm.tasks.interpolation.history_guidance.guidance_scale=1.5 \
'algorithm.logging.metrics=[fvd,fid,psnr,lpips,ssim]'
hydra.run.dir=$eval_result_dir |