fusion-bench 0.2.30__py3-none-any.whl → 0.2.32__py3-none-any.whl

fusion_bench_config/fabric/auto.yaml

@@ -13,3 +13,4 @@ strategy: auto
 # ``"cpu"``, ``"cuda"``, ``"mps"``, ``"gpu"``, ``"tpu"``, ``"auto"``.
 # for example: fabric.accelerator=cpu
 accelerator: auto
+precision: null

fusion_bench_config/fabric/loggers/mlflow_logger.yaml

@@ -1,2 +1,6 @@
 # https://mlflow.org/
 _target_: lightning.pytorch.loggers.MLFlowLogger
+tracking_uri: "sqlite:///${path.output_dir}/mlflow.db"
+experiment_name: ${hydra:job.config_name}
+run_name: ${now:%Y-%m-%d_%H-%M-%S}
+save_dir: ${path.log_dir}

fusion_bench_config/hydra/default.yaml

@@ -4,7 +4,9 @@ defaults:
 run:
   dir: ${path.log_dir}
 sweep:
-  dir: ${path.log_dir}
+  # the directory where all multirun outputs are stored
+  # can not refer to ${path.log_dir} because this is evaluated before constructing the separate run configs
+  dir: ${oc.env:FUSION_BENCH_PROJECT_ROOT,"."}/outputs/multirun/${hydra.job.config_name}/${now:%Y-%m-%d_%H-%M-%S}
   subdir: ${hydra.job.num}
 job:
   env_set:

fusion_bench_config/method/concrete_subspace/clip_concrete_tsvm.yaml

@@ -0,0 +1,38 @@
+_target_: fusion_bench.method.concrete_subspace.clip_concrete_tsvm.ConcreteTSVMForOpenCLIP
+# === Concrete Subspace parameters ===
+# batch size per gpu
+# if you have multiple gpus, the total batch size will be `batch_size * num_gpus`
+dataloader_kwargs:
+  batch_size: 16
+  num_workers: 8
+optimizer:
+  _target_: torch.optim.AdamW
+  lr: 1e-3
+  weight_decay: 0.01
+  fused: null
+lr_scheduler: null
+merge_dtype: null
+max_steps: 2000
+save_interval: 500
+initial_logits: 0
+temperature: 0.5
+# "discrete" or "continuous", this is the mask applied for evaluation, not during training
+# the performance of final model are expected to be similar
+eval_mask_type: continuous
+mask_checkpoint: null
+# if `clamp_weights` is true, the weights will be clamped to [0, 1]
+clamp_weights: false
+# arguments of `functional_call`
+tie_weights: true
+strict: false
+# directory to cache zero-shot classification heads
+cache_dir: outputs
+skip_training: false
+# === TSVM parameters ===
+exclude_keys: null
+# alpha (also known as scaling factor) is a float or a list of floats
+# example:
+# alpha: 1
+# alpha: [1, 0.5, 0.25]
+alpha: 1
+return_single_task_models: false

fusion_bench_config/method/dop/dop_general.yaml

@@ -0,0 +1,33 @@
+_target_: fusion_bench.method.dop.dop_general.DOPMerging
+
+# the random seed to use
+seed: null
+# shuffle the order of the models
+shuffle_order: true
+# save the merged model on every step
+save_on_every_step: false
+# evaluate the merged model on every step
+evaluate_on_every_step: true
+
+# optimizer (learning rate)
+lr: 1e-4
+# optimizer (num_steps)
+num_steps: 200
+
+# weighted loss
+# if mgda is true, use mgda to optimize the loss weights
+mgda: true
+# if mgda is false, this is the weight for the loss of the first task
+alpha: 0.8
+# if mgda is true and ema is ture, using exponential moving average (ema), alpha is the initial value
+ema: true
+# if mgda is true and ema is ture, using exponential moving average (ema), beta is the decay rate
+ema_beta: 0.999
+
+# epsilon for svd (the proportion of energy retained)
+svd_epsilon: 0.99999
+# the space to project the delta w (left singular vectors, right singular vectors, or both)
+svd_proj_space: uv # u or v or uv
+
+# the number of ray actors to use for parallel merging (0 means no ray)
+num_ray_actors: 0

fusion_bench_config/method/emr_merging/emr_merging.yaml

@@ -0,0 +1 @@
+_target_: fusion_bench.method.emr_merging.EMRMerging

fusion_bench_config/method/opcm/opcm_general.yaml

@@ -0,0 +1,18 @@
+# =============================================================================
+# FusionBench Method Configuration: OPCM
+# =============================================================================
+# Incrementally merges models via SVD projection and evaluation per step.
+# =============================================================================
+_target_: fusion_bench.method.opcm.opcm_general.OPCM
+# shuffle the order of the models
+shuffle_order: true
+# the scaling factor for the SVD projection
+alpha: 0.5
+# the random seed to use
+seed: null
+# save the merged model on every step
+save_on_every_step: true
+# evaluate the merged model on every step
+evaluate_on_every_step: true
+# the number of ray actors to use for distributed merging
+num_ray_actors: 0

fusion_bench_config/modelpool/ConvNextForImageClassification/convnext-base-224_8-tasks.yaml

@@ -0,0 +1,15 @@
+_target_: fusion_bench.modelpool.ConvNextForImageClassificationPool
+_recursive_: False
+models:
+  _pretrained_: facebook/convnext-base-224
+  sun397: tanganke/convnext-base-224_sun397_sgd_batch-size-64_lr-0.01_steps-4000
+  stanford-cars: tanganke/convnext-base-224_stanford-cars_sgd_batch-size-64_lr-0.01_steps-4000
+  resisc45: tanganke/convnext-base-224_resisc45_sgd_batch-size-64_lr-0.01_steps-4000
+  eurosat: tanganke/convnext-base-224_eurosat_sgd_batch-size-64_lr-0.01_steps-4000
+  svhn: tanganke/convnext-base-224_svhn_sgd_batch-size-64_lr-0.01_steps-4000
+  gtsrb: tanganke/convnext-base-224_gtsrb_sgd_batch-size-64_lr-0.01_steps-4000
+  mnist: tanganke/convnext-base-224_mnist_sgd_batch-size-64_lr-0.01_steps-4000
+  dtd: tanganke/convnext-base-224_dtd_sgd_batch-size-64_lr-0.01_steps-4000
+train_datasets: null
+val_datasets: null
+test_datasets: null

fusion_bench_config/taskpool/ImageClassificationTaskPool/convnext-base-224_8-tasks.yaml

@@ -0,0 +1,17 @@
+defaults:
+  - /dataset/image_classification/test@test_datasets:
+      - sun397
+      - stanford-cars
+      - resisc45
+      - eurosat
+      - svhn
+      - gtsrb
+      - mnist
+      - dtd
+  - _self_
+_target_: fusion_bench.taskpool.image_classification.ImageClassificationTaskPool
+_recursive_: False
+processor: facebook/convnext-base-224
+dataloader_kwargs:
+  batch_size: 32
+  num_workers: 0