Keep multiview Berzelius training script

train_dememwm_curric_event_trigger_8a100_berzelius.sh

@@ -1,95 +0,0 @@
-#!/usr/bin/env bash
-#SBATCH --job-name=dememwm-curric-event-anchorclean-causal-8a100
-#SBATCH --nodes=1
-#SBATCH --ntasks-per-node=8
-#SBATCH --cpus-per-task=8
-#SBATCH --time=72:00:00
-#SBATCH --output=slurm_logs/dememwm-curric-event-anchorclean-causal-8a100-%j.out
-#SBATCH --error=slurm_logs/dememwm-curric-event-anchorclean-causal-8a100-%j.out
-#SBATCH --account=berzelius-2025-436
-#SBATCH --gres=gpu:A100-SXM4-80GB:8
-#SBATCH --chdir=/proj/cvl/users/x_fahkh2/DeMemWM
-
-set -eo pipefail
-
-module load buildenv-gcccuda/12.1.1-gcc12.3.0
-source $(conda info --base)/etc/profile.d/conda.sh
-
-export PYTHONPATH="./:${PYTHONPATH:-}"
-export HF_HOME=/proj/cvl/users/x_fahkh2/caches
-export TORCH_HOME=/proj/cvl/users/x_fahkh2/caches
-export PIP_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export TMPDIR=/proj/cvl/users/x_fahkh2/caches
-export TRITON_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export CUDA_HOME=$CUDA_ROOT
-export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export HYDRA_FULL_ERROR=1
-
-git checkout main
-
-RUN_NAME=train_dememwm_curric_event_trigger_anchorclean_causal_a2d4r2_route_ntok8_8a100_bs8_global64_400k
-OUTPUT_DIR=/proj/cvl/users/x_fahkh2/DeMemWM/results/${RUN_NAME}
-CURRICULUM='{enabled:true,stages:[{name:no_updown,until_step:120000,dataset:{wo_updown:true}},{name:with_updown,until_step:400000,dataset:{wo_updown:false}}]}'
-
-srun python -m main \
-    +name=${RUN_NAME} \
-    +output_dir=$OUTPUT_DIR/ \
-    auto_resume=true \
-    wandb.mode=offline \
-    "experiment.tasks=[training]" \
-    algorithm=dememwm_base \
-    +customized_load=true \
-    +seperate_load=true \
-    +diffusion_model_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/oasis500m.safetensors \
-    +vae_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/vit-l-20.safetensors \
-    dataset=video_minecraft_dememwm_latent \
-    dataset.save_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft \
-    dataset.precomputed_feature_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft/vae_features \
-    dataset.n_frames=8 \
-    dataset.n_frames_valid=1100 \
-    dataset.context_length=100 \
-    dataset.single_eval_clip=true \
-    dataset.memory_selection.enabled=true \
-    dataset.memory_selection.causal=true \
-    dataset.memory_selection.max_anchor_frames=2 \
-    dataset.memory_selection.max_dynamic_frames=4 \
-    dataset.memory_selection.max_revisit_frames=2 \
-    dataset.memory_selection.pose_similarity_threshold=0.6 \
-    dataset.memory_selection.training_use_plucker=true \
-    dataset.memory_selection.training_plucker_weight=1.0 \
-    dataset.memory_selection.fov_overlap_threshold=0.6 \
-    dataset.memory_selection.min_total_selected_coverage=0.1 \
-    dataset.memory_selection.local_context_exclusion_frames=8 \
-    dataset.memory_selection.anchor_diverse_selection=true \
-    +dataset.memory_selection.pose_preselect_topk=64 \
-    +dataset.memory_selection.candidate_chunk_size=64 \
-    dataset.memory_selection.dynamic.selection_policy=event_triggered \
-    algorithm.n_tokens=8 \
-    algorithm.context_frames=600 \
-    algorithm.metric_report_segment=100 \
-    algorithm.chunk_size=1 \
-    algorithm.warmup_steps=10000 \
-    algorithm.trainability.train_full_dit=true \
-    algorithm.trainability.full_dit_start_step=40000 \
-    algorithm.trainability.lr.memory_modules=4.0e-5 \
-    algorithm.trainability.lr.base_dit=1.0e-5 \
-    algorithm.log_video=true \
-    algorithm.save_local=true \
-    algorithm.diffusion.sampling_timesteps=20 \
-    algorithm.noise_route.anchor=high \
-    algorithm.noise_route.dynamic=low \
-    algorithm.noise_route.revisit=all \
-    algorithm.memory_noise.enabled=true \
-    algorithm.memory_noise.mode=random_cleaner_fraction \
-    algorithm.memory_noise.anchor_max_fraction=0.0 \
-    algorithm.memory_noise.dynamic_max_fraction=0.25 \
-    algorithm.memory_noise.revisit_max_fraction=0.25 \
-    "experiment.training.curriculum=${CURRICULUM}" \
-    experiment.training.batch_size=8 \
-    experiment.training.optim.accumulate_grad_batches=1 \
-    experiment.validation.batch_size=1 \
-    experiment.validation.limit_batch=1 \
-    experiment.training.checkpointing.every_n_train_steps=2000 \
-    experiment.training.checkpointing.save_last_k=10 \
-    experiment.validation.val_every_n_step=2000 \
-    experiment.training.max_steps=400000

train_dememwm_curric_event_trigger_8h200_berzelius.sh

@@ -1,95 +0,0 @@
-#!/usr/bin/env bash
-#SBATCH --job-name=dememwm-curric-event-anchorclean-causal-8h200
-#SBATCH --nodes=1
-#SBATCH --ntasks-per-node=8
-#SBATCH --cpus-per-task=8
-#SBATCH --time=72:00:00
-#SBATCH --output=slurm_logs/dememwm-curric-event-anchorclean-causal-8h200-%j.out
-#SBATCH --error=slurm_logs/dememwm-curric-event-anchorclean-causal-8h200-%j.out
-#SBATCH --account=berzelius-2025-436
-#SBATCH --gres=gpu:8
-#SBATCH --chdir=/proj/cvl/users/x_fahkh2/DeMemWM
-
-set -eo pipefail
-
-module load buildenv-gcccuda/12.9.1-gcc11
-source $(conda info --base)/etc/profile.d/conda.sh
-
-export PYTHONPATH="./:${PYTHONPATH:-}"
-export HF_HOME=/proj/cvl/users/x_fahkh2/caches
-export TORCH_HOME=/proj/cvl/users/x_fahkh2/caches
-export PIP_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export TMPDIR=/proj/cvl/users/x_fahkh2/caches
-export TRITON_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export CUDA_HOME=$CUDA_ROOT
-export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export HYDRA_FULL_ERROR=1
-
-git checkout main
-
-RUN_NAME=train_dememwm_curric_event_trigger_anchorclean_causal_a2d4r2_route_ntok8_8h200_bs16_global128_150k
-OUTPUT_DIR=/proj/cvl/users/x_fahkh2/DeMemWM/results/${RUN_NAME}
-CURRICULUM='{enabled:true,stages:[{name:no_updown,until_step:30000,dataset:{wo_updown:true}},{name:with_updown,until_step:150000,dataset:{wo_updown:false}}]}'
-
-srun python -m main \
-    +name=${RUN_NAME} \
-    +output_dir=$OUTPUT_DIR/ \
-    auto_resume=true \
-    wandb.mode=offline \
-    "experiment.tasks=[training]" \
-    algorithm=dememwm_base \
-    +customized_load=true \
-    +seperate_load=true \
-    +diffusion_model_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/oasis500m.safetensors \
-    +vae_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/vit-l-20.safetensors \
-    dataset=video_minecraft_dememwm_latent \
-    dataset.save_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft \
-    dataset.precomputed_feature_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft/vae_features \
-    dataset.n_frames=8 \
-    dataset.n_frames_valid=1100 \
-    dataset.context_length=100 \
-    dataset.single_eval_clip=true \
-    dataset.memory_selection.enabled=true \
-    dataset.memory_selection.causal=true \
-    dataset.memory_selection.max_anchor_frames=2 \
-    dataset.memory_selection.max_dynamic_frames=4 \
-    dataset.memory_selection.max_revisit_frames=2 \
-    dataset.memory_selection.pose_similarity_threshold=0.6 \
-    dataset.memory_selection.training_use_plucker=true \
-    dataset.memory_selection.training_plucker_weight=1.0 \
-    dataset.memory_selection.fov_overlap_threshold=0.6 \
-    dataset.memory_selection.min_total_selected_coverage=0.1 \
-    dataset.memory_selection.local_context_exclusion_frames=8 \
-    dataset.memory_selection.anchor_diverse_selection=true \
-    +dataset.memory_selection.pose_preselect_topk=64 \
-    +dataset.memory_selection.candidate_chunk_size=64 \
-    dataset.memory_selection.dynamic.selection_policy=event_triggered \
-    algorithm.n_tokens=8 \
-    algorithm.context_frames=600 \
-    algorithm.metric_report_segment=100 \
-    algorithm.chunk_size=1 \
-    algorithm.warmup_steps=1000 \
-    algorithm.trainability.train_full_dit=true \
-    algorithm.trainability.full_dit_start_step=10000 \
-    algorithm.trainability.lr.memory_modules=8.0e-5 \
-    algorithm.trainability.lr.base_dit=2.0e-5 \
-    algorithm.log_video=true \
-    algorithm.save_local=true \
-    algorithm.diffusion.sampling_timesteps=20 \
-    algorithm.noise_route.anchor=high \
-    algorithm.noise_route.dynamic=low \
-    algorithm.noise_route.revisit=all \
-    algorithm.memory_noise.enabled=true \
-    algorithm.memory_noise.mode=random_cleaner_fraction \
-    algorithm.memory_noise.anchor_max_fraction=0.0 \
-    algorithm.memory_noise.dynamic_max_fraction=0.25 \
-    algorithm.memory_noise.revisit_max_fraction=0.25 \
-    "experiment.training.curriculum=${CURRICULUM}" \
-    experiment.training.batch_size=16 \
-    experiment.training.optim.accumulate_grad_batches=1 \
-    experiment.validation.batch_size=1 \
-    experiment.validation.limit_batch=1 \
-    experiment.training.checkpointing.every_n_train_steps=2000 \
-    experiment.training.checkpointing.save_last_k=10 \
-    experiment.validation.val_every_n_step=2000 \
-    experiment.training.max_steps=150000

train_dememwm_curric_event_trigger_anchorclean_causal_route_safe_8h200_berzelius.sh

@@ -1,102 +0,0 @@
-#!/usr/bin/env bash
-#SBATCH --job-name=dememwm-curric-event-anchorclean-causal-route-safe-8h200
-#SBATCH --nodes=1
-#SBATCH --ntasks-per-node=8
-#SBATCH --cpus-per-task=8
-#SBATCH --time=72:00:00
-#SBATCH --output=slurm_logs/dememwm-curric-event-anchorclean-causal-route-safe-8h200-%j.out
-#SBATCH --error=slurm_logs/dememwm-curric-event-anchorclean-causal-route-safe-8h200-%j.out
-#SBATCH --account=berzelius-2025-436
-#SBATCH --gres=gpu:8
-#SBATCH --chdir=/proj/cvl/users/x_fahkh2/DeMemWM
-
-set -eo pipefail
-
-module load buildenv-gcccuda/12.9.1-gcc11
-source $(conda info --base)/etc/profile.d/conda.sh
-
-export PYTHONPATH="./:${PYTHONPATH:-}"
-export HF_HOME=/proj/cvl/users/x_fahkh2/caches
-export TORCH_HOME=/proj/cvl/users/x_fahkh2/caches
-export PIP_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export TMPDIR=/proj/cvl/users/x_fahkh2/caches
-export TRITON_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export CUDA_HOME=$CUDA_ROOT
-export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export HYDRA_FULL_ERROR=1
-
-git checkout main
-
-RUN_NAME=train_dememwm_curric_event_trigger_anchorclean_causal_route_safe_a2d4r2_ntok8_8h200_bs16_global128_120k
-OUTPUT_DIR=/proj/cvl/users/x_fahkh2/DeMemWM/results/${RUN_NAME}
-CURRICULUM='{enabled:true,stages:[{name:no_updown,until_step:60000,dataset:{wo_updown:true}},{name:with_updown,until_step:120000,dataset:{wo_updown:false}}]}'
-
-srun python -m main \
-    +name=${RUN_NAME} \
-    +output_dir=$OUTPUT_DIR/ \
-    auto_resume=true \
-    wandb.mode=offline \
-    "experiment.tasks=[training]" \
-    algorithm=dememwm_base \
-    +customized_load=true \
-    +seperate_load=true \
-    +diffusion_model_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/oasis500m.safetensors \
-    +vae_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/vit-l-20.safetensors \
-    dataset=video_minecraft_dememwm_latent \
-    dataset.save_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft \
-    dataset.precomputed_feature_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft/vae_features \
-    dataset.n_frames=8 \
-    dataset.n_frames_valid=1100 \
-    dataset.context_length=100 \
-    dataset.single_eval_clip=true \
-    dataset.memory_selection.enabled=true \
-    dataset.memory_selection.causal=true \
-    dataset.memory_selection.max_anchor_frames=2 \
-    dataset.memory_selection.max_dynamic_frames=4 \
-    dataset.memory_selection.max_revisit_frames=2 \
-    dataset.memory_selection.pose_similarity_threshold=0.6 \
-    dataset.memory_selection.training_use_plucker=true \
-    dataset.memory_selection.training_plucker_weight=1.0 \
-    dataset.memory_selection.fov_overlap_threshold=0.6 \
-    dataset.memory_selection.min_total_selected_coverage=0.1 \
-    dataset.memory_selection.local_context_exclusion_frames=8 \
-    dataset.memory_selection.anchor_diverse_selection=true \
-    +dataset.memory_selection.pose_preselect_topk=64 \
-    +dataset.memory_selection.candidate_chunk_size=64 \
-    dataset.memory_selection.dynamic.selection_policy=event_triggered \
-    algorithm.causal=true \
-    algorithm.n_tokens=8 \
-    algorithm.context_frames=600 \
-    algorithm.metric_report_segment=100 \
-    +algorithm.log_per_frame_metrics=true \
-    algorithm.chunk_size=1 \
-    algorithm.warmup_steps=5000 \
-    algorithm.trainability.train_full_dit=true \
-    algorithm.trainability.full_dit_start_step=40000 \
-    algorithm.trainability.lr.memory_modules=4.0e-5 \
-    algorithm.trainability.lr.base_dit=1.0e-5 \
-    algorithm.log_video=true \
-    algorithm.save_local=true \
-    algorithm.diffusion.sampling_timesteps=20 \
-    algorithm.noise_route.anchor=high \
-    algorithm.noise_route.dynamic=low \
-    algorithm.noise_route.revisit=all \
-    algorithm.memory_noise.enabled=true \
-    algorithm.memory_noise.mode=random_cleaner_fraction \
-    algorithm.memory_noise.anchor_max_fraction=0.0 \
-    algorithm.memory_noise.dynamic_max_fraction=0.50 \
-    algorithm.memory_noise.revisit_max_fraction=0.50 \
-    +algorithm.memory_noise.validation_noisy_memory=false \
-    "experiment.training.curriculum=${CURRICULUM}" \
-    experiment.training.batch_size=16 \
-    experiment.training.precision=16-mixed \
-    experiment.training.optim.accumulate_grad_batches=1 \
-    experiment.training.optim.gradient_clip_val=1.0 \
-    experiment.validation.batch_size=1 \
-    experiment.validation.limit_batch=1 \
-    experiment.validation.val_every_n_step=2000 \
-    experiment.training.checkpointing.every_n_train_steps=2000 \
-    experiment.training.checkpointing.save_top_k=5 \
-    experiment.training.checkpointing.monitor=lpips \
-    experiment.training.checkpointing.mode=min \
-    experiment.training.max_steps=120000

train_dememwm_curric_event_trigger_anchorclean_causal_route_safe_nodynamic_8h200_berzelius.sh

@@ -1,102 +0,0 @@
-#!/usr/bin/env bash
-#SBATCH --job-name=dememwm-curric-event-anchorclean-causal-route-safe-nodynamic-8h200
-#SBATCH --nodes=1
-#SBATCH --ntasks-per-node=8
-#SBATCH --cpus-per-task=8
-#SBATCH --time=72:00:00
-#SBATCH --output=slurm_logs/dememwm-curric-event-anchorclean-causal-route-safe-nodynamic-8h200-%j.out
-#SBATCH --error=slurm_logs/dememwm-curric-event-anchorclean-causal-route-safe-nodynamic-8h200-%j.out
-#SBATCH --account=berzelius-2025-436
-#SBATCH --gres=gpu:8
-#SBATCH --chdir=/proj/cvl/users/x_fahkh2/DeMemWM
-
-set -eo pipefail
-
-module load buildenv-gcccuda/12.9.1-gcc11
-source $(conda info --base)/etc/profile.d/conda.sh
-
-export PYTHONPATH="./:${PYTHONPATH:-}"
-export HF_HOME=/proj/cvl/users/x_fahkh2/caches
-export TORCH_HOME=/proj/cvl/users/x_fahkh2/caches
-export PIP_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export TMPDIR=/proj/cvl/users/x_fahkh2/caches
-export TRITON_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export CUDA_HOME=$CUDA_ROOT
-export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export HYDRA_FULL_ERROR=1
-
-git checkout main
-
-RUN_NAME=train_dememwm_curric_event_trigger_anchorclean_causal_route_safe_a2d0r2_ntok8_8h200_bs16_global128_120k
-OUTPUT_DIR=/proj/cvl/users/x_fahkh2/DeMemWM/results/${RUN_NAME}
-CURRICULUM='{enabled:true,stages:[{name:no_updown,until_step:60000,dataset:{wo_updown:true}},{name:with_updown,until_step:120000,dataset:{wo_updown:false}}]}'
-
-srun python -m main \
-    +name=${RUN_NAME} \
-    +output_dir=$OUTPUT_DIR/ \
-    auto_resume=true \
-    wandb.mode=offline \
-    "experiment.tasks=[training]" \
-    algorithm=dememwm_base \
-    +customized_load=true \
-    +seperate_load=true \
-    +diffusion_model_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/oasis500m.safetensors \
-    +vae_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/vit-l-20.safetensors \
-    dataset=video_minecraft_dememwm_latent \
-    dataset.save_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft \
-    dataset.precomputed_feature_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft/vae_features \
-    dataset.n_frames=8 \
-    dataset.n_frames_valid=1100 \
-    dataset.context_length=100 \
-    dataset.single_eval_clip=true \
-    dataset.memory_selection.enabled=true \
-    dataset.memory_selection.causal=true \
-    dataset.memory_selection.max_anchor_frames=2 \
-    dataset.memory_selection.max_dynamic_frames=0 \
-    dataset.memory_selection.max_revisit_frames=2 \
-    dataset.memory_selection.pose_similarity_threshold=0.6 \
-    dataset.memory_selection.training_use_plucker=true \
-    dataset.memory_selection.training_plucker_weight=1.0 \
-    dataset.memory_selection.fov_overlap_threshold=0.6 \
-    dataset.memory_selection.min_total_selected_coverage=0.1 \
-    dataset.memory_selection.local_context_exclusion_frames=8 \
-    dataset.memory_selection.anchor_diverse_selection=true \
-    +dataset.memory_selection.pose_preselect_topk=64 \
-    +dataset.memory_selection.candidate_chunk_size=64 \
-    dataset.memory_selection.dynamic.selection_policy=event_triggered \
-    algorithm.causal=true \
-    algorithm.n_tokens=8 \
-    algorithm.context_frames=600 \
-    algorithm.metric_report_segment=100 \
-    +algorithm.log_per_frame_metrics=true \
-    algorithm.chunk_size=1 \
-    algorithm.warmup_steps=5000 \
-    algorithm.trainability.train_full_dit=true \
-    algorithm.trainability.full_dit_start_step=40000 \
-    algorithm.trainability.lr.memory_modules=4.0e-5 \
-    algorithm.trainability.lr.base_dit=1.0e-5 \
-    algorithm.log_video=true \
-    algorithm.save_local=true \
-    algorithm.diffusion.sampling_timesteps=20 \
-    algorithm.noise_route.anchor=high \
-    algorithm.noise_route.dynamic=low \
-    algorithm.noise_route.revisit=all \
-    algorithm.memory_noise.enabled=true \
-    algorithm.memory_noise.mode=random_cleaner_fraction \
-    algorithm.memory_noise.anchor_max_fraction=0.0 \
-    algorithm.memory_noise.dynamic_max_fraction=0.0 \
-    algorithm.memory_noise.revisit_max_fraction=0.50 \
-    +algorithm.memory_noise.validation_noisy_memory=false \
-    "experiment.training.curriculum=${CURRICULUM}" \
-    experiment.training.batch_size=16 \
-    experiment.training.precision=16-mixed \
-    experiment.training.optim.accumulate_grad_batches=1 \
-    experiment.training.optim.gradient_clip_val=1.0 \
-    experiment.validation.batch_size=1 \
-    experiment.validation.limit_batch=1 \
-    experiment.validation.val_every_n_step=2000 \
-    experiment.training.checkpointing.every_n_train_steps=2000 \
-    experiment.training.checkpointing.save_top_k=5 \
-    experiment.training.checkpointing.monitor=lpips \
-    experiment.training.checkpointing.mode=min \
-    experiment.training.max_steps=120000

train_dememwm_curric_event_trigger_anchorclean_noncausal_8a100_berzelius.sh

@@ -1,95 +0,0 @@
-#!/usr/bin/env bash
-#SBATCH --job-name=dememwm-curric-event-anchorclean-noncausal-8a100
-#SBATCH --nodes=1
-#SBATCH --ntasks-per-node=8
-#SBATCH --cpus-per-task=8
-#SBATCH --time=72:00:00
-#SBATCH --output=slurm_logs/dememwm-curric-event-anchorclean-noncausal-8a100-%j.out
-#SBATCH --error=slurm_logs/dememwm-curric-event-anchorclean-noncausal-8a100-%j.out
-#SBATCH --account=berzelius-2025-436
-#SBATCH --gres=gpu:A100-SXM4-80GB:8
-#SBATCH --chdir=/proj/cvl/users/x_fahkh2/DeMemWM
-
-set -eo pipefail
-
-module load buildenv-gcccuda/12.1.1-gcc12.3.0
-source $(conda info --base)/etc/profile.d/conda.sh
-
-export PYTHONPATH="./:${PYTHONPATH:-}"
-export HF_HOME=/proj/cvl/users/x_fahkh2/caches
-export TORCH_HOME=/proj/cvl/users/x_fahkh2/caches
-export PIP_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export TMPDIR=/proj/cvl/users/x_fahkh2/caches
-export TRITON_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export CUDA_HOME=$CUDA_ROOT
-export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export HYDRA_FULL_ERROR=1
-
-git checkout main
-
-RUN_NAME=train_dememwm_curric_event_trigger_anchorclean_noncausal_a2d4r2_route_ntok8_8a100_bs8_global64_400k
-OUTPUT_DIR=/proj/cvl/users/x_fahkh2/DeMemWM/results/${RUN_NAME}
-CURRICULUM='{enabled:true,stages:[{name:no_updown,until_step:120000,dataset:{wo_updown:true}},{name:with_updown,until_step:400000,dataset:{wo_updown:false}}]}'
-
-srun python -m main \
-    +name=${RUN_NAME} \
-    +output_dir=$OUTPUT_DIR/ \
-    auto_resume=true \
-    wandb.mode=offline \
-    "experiment.tasks=[training]" \
-    algorithm=dememwm_base \
-    +customized_load=true \
-    +seperate_load=true \
-    +diffusion_model_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/oasis500m.safetensors \
-    +vae_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/vit-l-20.safetensors \
-    dataset=video_minecraft_dememwm_latent \
-    dataset.save_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft \
-    dataset.precomputed_feature_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft/vae_features \
-    dataset.n_frames=8 \
-    dataset.n_frames_valid=1100 \
-    dataset.context_length=100 \
-    dataset.single_eval_clip=true \
-    dataset.memory_selection.enabled=true \
-    dataset.memory_selection.causal=false \
-    dataset.memory_selection.max_anchor_frames=2 \
-    dataset.memory_selection.max_dynamic_frames=4 \
-    dataset.memory_selection.max_revisit_frames=2 \
-    dataset.memory_selection.pose_similarity_threshold=0.6 \
-    dataset.memory_selection.training_use_plucker=true \
-    dataset.memory_selection.training_plucker_weight=1.0 \
-    dataset.memory_selection.fov_overlap_threshold=0.6 \
-    dataset.memory_selection.min_total_selected_coverage=0.1 \
-    dataset.memory_selection.local_context_exclusion_frames=8 \
-    dataset.memory_selection.anchor_diverse_selection=true \
-    +dataset.memory_selection.pose_preselect_topk=64 \
-    +dataset.memory_selection.candidate_chunk_size=64 \
-    dataset.memory_selection.dynamic.selection_policy=event_triggered \
-    algorithm.n_tokens=8 \
-    algorithm.context_frames=600 \
-    algorithm.metric_report_segment=100 \
-    algorithm.chunk_size=1 \
-    algorithm.warmup_steps=10000 \
-    algorithm.trainability.train_full_dit=true \
-    algorithm.trainability.full_dit_start_step=40000 \
-    algorithm.trainability.lr.memory_modules=4.0e-5 \
-    algorithm.trainability.lr.base_dit=1.0e-5 \
-    algorithm.log_video=true \
-    algorithm.save_local=true \
-    algorithm.diffusion.sampling_timesteps=20 \
-    algorithm.noise_route.anchor=high \
-    algorithm.noise_route.dynamic=low \
-    algorithm.noise_route.revisit=all \
-    algorithm.memory_noise.enabled=true \
-    algorithm.memory_noise.mode=random_cleaner_fraction \
-    algorithm.memory_noise.anchor_max_fraction=0.0 \
-    algorithm.memory_noise.dynamic_max_fraction=0.25 \
-    algorithm.memory_noise.revisit_max_fraction=0.25 \
-    "experiment.training.curriculum=${CURRICULUM}" \
-    experiment.training.batch_size=8 \
-    experiment.training.optim.accumulate_grad_batches=1 \
-    experiment.validation.batch_size=1 \
-    experiment.validation.limit_batch=1 \
-    experiment.training.checkpointing.every_n_train_steps=2000 \
-    experiment.training.checkpointing.save_last_k=10 \
-    experiment.validation.val_every_n_step=2000 \
-    experiment.training.max_steps=400000

train_dememwm_curric_event_trigger_anchorclean_noncausal_8h200_berzelius.sh

@@ -1,95 +0,0 @@
-#!/usr/bin/env bash
-#SBATCH --job-name=dememwm-curric-event-anchorclean-noncausal-8h200
-#SBATCH --nodes=1
-#SBATCH --ntasks-per-node=8
-#SBATCH --cpus-per-task=8
-#SBATCH --time=72:00:00
-#SBATCH --output=slurm_logs/dememwm-curric-event-anchorclean-noncausal-8h200-%j.out
-#SBATCH --error=slurm_logs/dememwm-curric-event-anchorclean-noncausal-8h200-%j.out
-#SBATCH --account=berzelius-2025-436
-#SBATCH --gres=gpu:8
-#SBATCH --chdir=/proj/cvl/users/x_fahkh2/DeMemWM
-
-set -eo pipefail
-
-module load buildenv-gcccuda/12.9.1-gcc11
-source $(conda info --base)/etc/profile.d/conda.sh
-
-export PYTHONPATH="./:${PYTHONPATH:-}"
-export HF_HOME=/proj/cvl/users/x_fahkh2/caches
-export TORCH_HOME=/proj/cvl/users/x_fahkh2/caches
-export PIP_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export TMPDIR=/proj/cvl/users/x_fahkh2/caches
-export TRITON_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export CUDA_HOME=$CUDA_ROOT
-export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export HYDRA_FULL_ERROR=1
-
-git checkout main
-
-RUN_NAME=train_dememwm_curric_event_trigger_anchorclean_noncausal_a2d4r2_route_ntok8_8h200_bs16_global128_150k
-OUTPUT_DIR=/proj/cvl/users/x_fahkh2/DeMemWM/results/${RUN_NAME}
-CURRICULUM='{enabled:true,stages:[{name:no_updown,until_step:30000,dataset:{wo_updown:true}},{name:with_updown,until_step:150000,dataset:{wo_updown:false}}]}'
-
-srun python -m main \
-    +name=${RUN_NAME} \
-    +output_dir=$OUTPUT_DIR/ \
-    auto_resume=true \
-    wandb.mode=offline \
-    "experiment.tasks=[training]" \
-    algorithm=dememwm_base \
-    +customized_load=true \
-    +seperate_load=true \
-    +diffusion_model_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/oasis500m.safetensors \
-    +vae_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/vit-l-20.safetensors \
-    dataset=video_minecraft_dememwm_latent \
-    dataset.save_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft \
-    dataset.precomputed_feature_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft/vae_features \
-    dataset.n_frames=8 \
-    dataset.n_frames_valid=1100 \
-    dataset.context_length=100 \
-    dataset.single_eval_clip=true \
-    dataset.memory_selection.enabled=true \
-    dataset.memory_selection.causal=false \
-    dataset.memory_selection.max_anchor_frames=2 \
-    dataset.memory_selection.max_dynamic_frames=4 \
-    dataset.memory_selection.max_revisit_frames=2 \
-    dataset.memory_selection.pose_similarity_threshold=0.6 \
-    dataset.memory_selection.training_use_plucker=true \
-    dataset.memory_selection.training_plucker_weight=1.0 \
-    dataset.memory_selection.fov_overlap_threshold=0.6 \
-    dataset.memory_selection.min_total_selected_coverage=0.1 \
-    dataset.memory_selection.local_context_exclusion_frames=8 \
-    dataset.memory_selection.anchor_diverse_selection=true \
-    +dataset.memory_selection.pose_preselect_topk=64 \
-    +dataset.memory_selection.candidate_chunk_size=64 \
-    dataset.memory_selection.dynamic.selection_policy=event_triggered \
-    algorithm.n_tokens=8 \
-    algorithm.context_frames=600 \
-    algorithm.metric_report_segment=100 \
-    algorithm.chunk_size=1 \
-    algorithm.warmup_steps=1000 \
-    algorithm.trainability.train_full_dit=true \
-    algorithm.trainability.full_dit_start_step=10000 \
-    algorithm.trainability.lr.memory_modules=8.0e-5 \
-    algorithm.trainability.lr.base_dit=2.0e-5 \
-    algorithm.log_video=true \
-    algorithm.save_local=true \
-    algorithm.diffusion.sampling_timesteps=20 \
-    algorithm.noise_route.anchor=high \
-    algorithm.noise_route.dynamic=low \
-    algorithm.noise_route.revisit=all \
-    algorithm.memory_noise.enabled=true \
-    algorithm.memory_noise.mode=random_cleaner_fraction \
-    algorithm.memory_noise.anchor_max_fraction=0.0 \
-    algorithm.memory_noise.dynamic_max_fraction=0.25 \
-    algorithm.memory_noise.revisit_max_fraction=0.25 \
-    "experiment.training.curriculum=${CURRICULUM}" \
-    experiment.training.batch_size=16 \
-    experiment.training.optim.accumulate_grad_batches=1 \
-    experiment.validation.batch_size=1 \
-    experiment.validation.limit_batch=1 \
-    experiment.training.checkpointing.every_n_train_steps=2000 \
-    experiment.training.checkpointing.save_last_k=10 \
-    experiment.validation.val_every_n_step=2000 \
-    experiment.training.max_steps=150000

train_dememwm_curric_event_trigger_no_route_8a100_berzelius.sh

@@ -1,95 +0,0 @@
-#!/usr/bin/env bash
-#SBATCH --job-name=dememwm-curric-event-noroute-anchorclean-8a100
-#SBATCH --nodes=1
-#SBATCH --ntasks-per-node=8
-#SBATCH --cpus-per-task=8
-#SBATCH --time=72:00:00
-#SBATCH --output=slurm_logs/dememwm-curric-event-noroute-anchorclean-8a100-%j.out
-#SBATCH --error=slurm_logs/dememwm-curric-event-noroute-anchorclean-8a100-%j.out
-#SBATCH --account=berzelius-2025-436
-#SBATCH --gres=gpu:A100-SXM4-80GB:8
-#SBATCH --chdir=/proj/cvl/users/x_fahkh2/DeMemWM
-
-set -eo pipefail
-
-module load buildenv-gcccuda/12.1.1-gcc12.3.0
-source $(conda info --base)/etc/profile.d/conda.sh
-
-export PYTHONPATH="./:${PYTHONPATH:-}"
-export HF_HOME=/proj/cvl/users/x_fahkh2/caches
-export TORCH_HOME=/proj/cvl/users/x_fahkh2/caches
-export PIP_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export TMPDIR=/proj/cvl/users/x_fahkh2/caches
-export TRITON_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export CUDA_HOME=$CUDA_ROOT
-export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export HYDRA_FULL_ERROR=1
-
-git checkout main
-
-RUN_NAME=train_dememwm_curric_event_trigger_no_route_anchorclean_a2d4r2_ntok8_8a100_bs8_global64_400k
-OUTPUT_DIR=/proj/cvl/users/x_fahkh2/DeMemWM/results/${RUN_NAME}
-CURRICULUM='{enabled:true,stages:[{name:no_updown,until_step:120000,dataset:{wo_updown:true}},{name:with_updown,until_step:400000,dataset:{wo_updown:false}}]}'
-
-srun python -m main \
-    +name=${RUN_NAME} \
-    +output_dir=$OUTPUT_DIR/ \
-    auto_resume=true \
-    wandb.mode=offline \
-    "experiment.tasks=[training]" \
-    algorithm=dememwm_base \
-    +customized_load=true \
-    +seperate_load=true \
-    +diffusion_model_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/oasis500m.safetensors \
-    +vae_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/vit-l-20.safetensors \
-    dataset=video_minecraft_dememwm_latent \
-    dataset.save_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft \
-    dataset.precomputed_feature_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft/vae_features \
-    dataset.n_frames=8 \
-    dataset.n_frames_valid=1100 \
-    dataset.context_length=100 \
-    dataset.single_eval_clip=true \
-    dataset.memory_selection.enabled=true \
-    dataset.memory_selection.causal=true \
-    dataset.memory_selection.max_anchor_frames=2 \
-    dataset.memory_selection.max_dynamic_frames=4 \
-    dataset.memory_selection.max_revisit_frames=2 \
-    dataset.memory_selection.pose_similarity_threshold=0.6 \
-    dataset.memory_selection.training_use_plucker=true \
-    dataset.memory_selection.training_plucker_weight=1.0 \
-    dataset.memory_selection.fov_overlap_threshold=0.6 \
-    dataset.memory_selection.min_total_selected_coverage=0.1 \
-    dataset.memory_selection.local_context_exclusion_frames=8 \
-    dataset.memory_selection.anchor_diverse_selection=true \
-    +dataset.memory_selection.pose_preselect_topk=64 \
-    +dataset.memory_selection.candidate_chunk_size=64 \
-    dataset.memory_selection.dynamic.selection_policy=event_triggered \
-    algorithm.n_tokens=8 \
-    algorithm.context_frames=600 \
-    algorithm.metric_report_segment=100 \
-    algorithm.chunk_size=1 \
-    algorithm.warmup_steps=10000 \
-    algorithm.trainability.train_full_dit=true \
-    algorithm.trainability.full_dit_start_step=40000 \
-    algorithm.trainability.lr.memory_modules=4.0e-5 \
-    algorithm.trainability.lr.base_dit=1.0e-5 \
-    algorithm.log_video=true \
-    algorithm.save_local=true \
-    algorithm.diffusion.sampling_timesteps=20 \
-    algorithm.noise_route.anchor=all \
-    algorithm.noise_route.dynamic=all \
-    algorithm.noise_route.revisit=all \
-    algorithm.memory_noise.enabled=true \
-    algorithm.memory_noise.mode=random_cleaner_fraction \
-    algorithm.memory_noise.anchor_max_fraction=0.0 \
-    algorithm.memory_noise.dynamic_max_fraction=0.25 \
-    algorithm.memory_noise.revisit_max_fraction=0.25 \
-    "experiment.training.curriculum=${CURRICULUM}" \
-    experiment.training.batch_size=8 \
-    experiment.training.optim.accumulate_grad_batches=1 \
-    experiment.validation.batch_size=1 \
-    experiment.validation.limit_batch=1 \
-    experiment.training.checkpointing.every_n_train_steps=2000 \
-    experiment.training.checkpointing.save_last_k=10 \
-    experiment.validation.val_every_n_step=2000 \
-    experiment.training.max_steps=400000

train_dememwm_curric_event_trigger_no_route_8h200_berzelius.sh

@@ -1,95 +0,0 @@
-#!/usr/bin/env bash
-#SBATCH --job-name=dememwm-curric-event-noroute-anchorclean-8h200
-#SBATCH --nodes=1
-#SBATCH --ntasks-per-node=8
-#SBATCH --cpus-per-task=8
-#SBATCH --time=72:00:00
-#SBATCH --output=slurm_logs/dememwm-curric-event-noroute-anchorclean-8h200-%j.out
-#SBATCH --error=slurm_logs/dememwm-curric-event-noroute-anchorclean-8h200-%j.out
-#SBATCH --account=berzelius-2025-436
-#SBATCH --gres=gpu:8
-#SBATCH --chdir=/proj/cvl/users/x_fahkh2/DeMemWM
-
-set -eo pipefail
-
-module load buildenv-gcccuda/12.9.1-gcc11
-source $(conda info --base)/etc/profile.d/conda.sh
-
-export PYTHONPATH="./:${PYTHONPATH:-}"
-export HF_HOME=/proj/cvl/users/x_fahkh2/caches
-export TORCH_HOME=/proj/cvl/users/x_fahkh2/caches
-export PIP_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export TMPDIR=/proj/cvl/users/x_fahkh2/caches
-export TRITON_CACHE_DIR=/proj/cvl/users/x_fahkh2/caches
-export CUDA_HOME=$CUDA_ROOT
-export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export HYDRA_FULL_ERROR=1
-
-git checkout main
-
-RUN_NAME=train_dememwm_curric_event_trigger_no_route_anchorclean_a2d4r2_ntok8_8h200_bs16_global128_150k
-OUTPUT_DIR=/proj/cvl/users/x_fahkh2/DeMemWM/results/${RUN_NAME}
-CURRICULUM='{enabled:true,stages:[{name:no_updown,until_step:30000,dataset:{wo_updown:true}},{name:with_updown,until_step:150000,dataset:{wo_updown:false}}]}'
-
-srun python -m main \
-    +name=${RUN_NAME} \
-    +output_dir=$OUTPUT_DIR/ \
-    auto_resume=true \
-    wandb.mode=offline \
-    "experiment.tasks=[training]" \
-    algorithm=dememwm_base \
-    +customized_load=true \
-    +seperate_load=true \
-    +diffusion_model_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/oasis500m.safetensors \
-    +vae_path=/proj/cvl/users/x_fahkh2/WorldMem_Repro/checkpoints/vit-l-20.safetensors \
-    dataset=video_minecraft_dememwm_latent \
-    dataset.save_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft \
-    dataset.precomputed_feature_dir=/proj/cvl/users/x_fahkh2/WorldMem_Repro/datasets/minecraft/vae_features \
-    dataset.n_frames=8 \
-    dataset.n_frames_valid=1100 \
-    dataset.context_length=100 \
-    dataset.single_eval_clip=true \
-    dataset.memory_selection.enabled=true \
-    dataset.memory_selection.causal=true \
-    dataset.memory_selection.max_anchor_frames=2 \
-    dataset.memory_selection.max_dynamic_frames=4 \
-    dataset.memory_selection.max_revisit_frames=2 \
-    dataset.memory_selection.pose_similarity_threshold=0.6 \
-    dataset.memory_selection.training_use_plucker=true \
-    dataset.memory_selection.training_plucker_weight=1.0 \
-    dataset.memory_selection.fov_overlap_threshold=0.6 \
-    dataset.memory_selection.min_total_selected_coverage=0.1 \
-    dataset.memory_selection.local_context_exclusion_frames=8 \
-    dataset.memory_selection.anchor_diverse_selection=true \
-    +dataset.memory_selection.pose_preselect_topk=64 \
-    +dataset.memory_selection.candidate_chunk_size=64 \
-    dataset.memory_selection.dynamic.selection_policy=event_triggered \
-    algorithm.n_tokens=8 \
-    algorithm.context_frames=600 \
-    algorithm.metric_report_segment=100 \
-    algorithm.chunk_size=1 \
-    algorithm.warmup_steps=1000 \
-    algorithm.trainability.train_full_dit=true \
-    algorithm.trainability.full_dit_start_step=10000 \
-    algorithm.trainability.lr.memory_modules=8.0e-5 \
-    algorithm.trainability.lr.base_dit=2.0e-5 \
-    algorithm.log_video=true \
-    algorithm.save_local=true \
-    algorithm.diffusion.sampling_timesteps=20 \
-    algorithm.noise_route.anchor=all \
-    algorithm.noise_route.dynamic=all \
-    algorithm.noise_route.revisit=all \
-    algorithm.memory_noise.enabled=true \
-    algorithm.memory_noise.mode=random_cleaner_fraction \
-    algorithm.memory_noise.anchor_max_fraction=0.0 \
-    algorithm.memory_noise.dynamic_max_fraction=0.25 \
-    algorithm.memory_noise.revisit_max_fraction=0.25 \
-    "experiment.training.curriculum=${CURRICULUM}" \
-    experiment.training.batch_size=16 \
-    experiment.training.optim.accumulate_grad_batches=1 \
-    experiment.validation.batch_size=1 \
-    experiment.validation.limit_batch=1 \
-    experiment.training.checkpointing.every_n_train_steps=2000 \
-    experiment.training.checkpointing.save_last_k=10 \
-    experiment.validation.val_every_n_step=2000 \
-    experiment.training.max_steps=150000

train_dememwm_curric_event_trigger_anchorclean_causal_nodynamic_8h200_berzelius.sh → train_dememwm_curric_multiview_8h200_berzelius.sh

@@ -1,12 +1,12 @@
 #!/usr/bin/env bash
-#SBATCH --job-name=dememwm-curric-event-anchorclean-causal-nodynamic-8h200
+#SBATCH --job-name=dememwm-curric-multiview-anchorclean-causal-8h200
 #SBATCH --nodes=1
 #SBATCH --ntasks-per-node=8
 #SBATCH --cpus-per-task=8
 #SBATCH --time=72:00:00
-#SBATCH --output=slurm_logs/dememwm-curric-event-anchorclean-causal-nodynamic-8h200-%j.out
-#SBATCH --error=slurm_logs/dememwm-curric-event-anchorclean-causal-nodynamic-8h200-%j.out
-#SBATCH --account=berzelius-2025-436
+#SBATCH --output=slurm_logs/dememwm-curric-multiview-anchorclean-causal-8h200-%j.out
+#SBATCH --error=slurm_logs/dememwm-curric-multiview-anchorclean-causal-8h200-%j.out
+#SBATCH --account=berzelius-2026-196
 #SBATCH --gres=gpu:8
 #SBATCH --chdir=/proj/cvl/users/x_fahkh2/DeMemWM
 
@@ -27,9 +27,9 @@ export HYDRA_FULL_ERROR=1
 
 git checkout main
 
-RUN_NAME=train_dememwm_curric_event_trigger_anchorclean_causal_a2d0r2_route_ntok8_8h200_bs16_global128_120k
+RUN_NAME=train_dememwm_curric_multiview_anchorclean_causal_a2d4r2_route_ntok8_8h200_bs16_global128_150k
 OUTPUT_DIR=/proj/cvl/users/x_fahkh2/DeMemWM/results/${RUN_NAME}
-CURRICULUM='{enabled:true,stages:[{name:no_updown,until_step:30000,dataset:{wo_updown:true}},{name:with_updown,until_step:120000,dataset:{wo_updown:false}}]}'
+CURRICULUM='{enabled:true,stages:[{name:no_updown,until_step:40000,dataset:{wo_updown:true}},{name:with_updown,until_step:120000,dataset:{wo_updown:false}}]}'
 
 srun python -m main \
     +name=${RUN_NAME} \
@@ -52,41 +52,34 @@ srun python -m main \
     dataset.memory_selection.enabled=true \
     dataset.memory_selection.causal=true \
     dataset.memory_selection.max_anchor_frames=2 \
-    dataset.memory_selection.max_dynamic_frames=0 \
+    dataset.memory_selection.max_dynamic_frames=4 \
     dataset.memory_selection.max_revisit_frames=2 \
-    dataset.memory_selection.pose_similarity_threshold=0.6 \
-    dataset.memory_selection.training_use_plucker=true \
-    dataset.memory_selection.training_plucker_weight=1.0 \
-    dataset.memory_selection.fov_overlap_threshold=0.6 \
-    dataset.memory_selection.min_total_selected_coverage=0.1 \
-    dataset.memory_selection.local_context_exclusion_frames=8 \
-    dataset.memory_selection.anchor_diverse_selection=true \
-    +dataset.memory_selection.pose_preselect_topk=64 \
-    +dataset.memory_selection.candidate_chunk_size=64 \
-    dataset.memory_selection.dynamic.selection_policy=event_triggered \
+    dataset.memory_selection.dynamic.selection_policy=multiview \
+    dataset.memory_selection.dynamic.multiview_selector=fov_greedy \
     algorithm.n_tokens=8 \
     algorithm.context_frames=600 \
     algorithm.metric_report_segment=100 \
+    algorithm.memory_attention_key_only_geometry=true \
+    algorithm.add_timestamp_embedding=false \
     algorithm.chunk_size=1 \
-    algorithm.warmup_steps=1000 \
+    algorithm.warmup_steps=500 \
     algorithm.trainability.train_full_dit=true \
     algorithm.trainability.full_dit_start_step=10000 \
     algorithm.trainability.lr.memory_modules=8.0e-5 \
     algorithm.trainability.lr.base_dit=2.0e-5 \
     algorithm.log_video=true \
-    algorithm.save_local=true \
+    algorithm.save_local=false \
     algorithm.diffusion.sampling_timesteps=20 \
     algorithm.noise_route.anchor=high \
     algorithm.noise_route.dynamic=low \
     algorithm.noise_route.revisit=all \
     algorithm.memory_noise.enabled=true \
     algorithm.memory_noise.mode=random_cleaner_fraction \
-    algorithm.memory_noise.anchor_max_fraction=0.0 \
-    algorithm.memory_noise.dynamic_max_fraction=0.0 \
-    algorithm.memory_noise.revisit_max_fraction=0.25 \
+    algorithm.memory_noise.anchor_max_fraction=0.25 \
+    algorithm.memory_noise.dynamic_max_fraction=0.50 \
+    algorithm.memory_noise.revisit_max_fraction=0.50 \
     "experiment.training.curriculum=${CURRICULUM}" \
     experiment.training.batch_size=16 \
-    experiment.training.optim.accumulate_grad_batches=1 \
     experiment.validation.batch_size=1 \
     experiment.validation.limit_batch=1 \
     experiment.training.checkpointing.every_n_train_steps=2000 \