fusion-bench 0.2.6__py3-none-any.whl → 0.2.8__py3-none-any.whl

fusion_bench_config/clip-vit-base-patch32_robustness_corrupted.yaml

@@ -11,7 +11,7 @@ _target_: fusion_bench.programs.FabricModelFusionProgram
 _recursive_: false
 fast_dev_run: false # Run a single batch of data to test the model or method
 # Run the script without actually running the experiment, use with `print_config=true`. 
-# You can also use `--cfg` or `-c` to  show the configuration instead of runing.
+# You can also use `--cfg` or `-c` to show the configuration instead of running.
 dry_run: false
 print_config: true # Print the configuration to the console
 merged_model_save_path: null # path to save the merged model, use "{log_dir}" to refer to the logger directory, for example `merged_model_save_path=\{log_dir\}/merged_model`

fusion_bench_config/dataset/image_classification/README.md

@@ -0,0 +1,6 @@
+# Image Classification Dataset Configurations
+
+This folder contains the dataset configuration for image classification tasks.
+
+- Each dataset should have 'image' and 'label' columns.
+- If a dataset has no test split, we will use the validation split as the test split and create the validation set from the training set.

fusion_bench_config/dataset/image_classification/test/TALL14.yaml

@@ -0,0 +1,20 @@
+# The 14 task used in the paper: 
+# Wang et al. Localizing Task Information for Improved Model Merging and Compression
+# http://arxiv.org/abs/2405.07813
+defaults:
+  # eight tasks in the task arithmetic paper
+  - sun397
+  - stanford-cars
+  - resisc45
+  - eurosat
+  - svhn
+  - gtsrb
+  - mnist
+  - dtd
+  # additional 6 tasks in the TALL mask paper
+  - oxford_flowers102
+  - pcam
+  - fer2013
+  - oxford-iiit-pet
+  - stl10
+  - cifar100

fusion_bench_config/dataset/image_classification/test/TALL20.yaml

@@ -0,0 +1,28 @@
+# The 20 task used in the paper: 
+# Wang et al. Localizing Task Information for Improved Model Merging and Compression
+# http://arxiv.org/abs/2405.07813
+defaults:
+  # eight tasks in the task arithmetic paper
+  - sun397
+  - stanford-cars
+  - resisc45
+  - eurosat
+  - svhn
+  - gtsrb
+  - mnist
+  - dtd
+  # additional 6 tasks in the TALL mask paper (TALL 14)
+  - oxford_flowers102
+  - pcam
+  - fer2013
+  - oxford-iiit-pet
+  - stl10
+  - cifar100
+  # additional 6 tasks in the TALL mask paper (TALL 20)
+  - cifar10
+  - food101
+  - fashion_mnist
+  - emnist_letters
+  - kmnist
+  - rendered-sst2
+

fusion_bench_config/dataset/image_classification/test/cifar10.yaml

@@ -1,4 +1,4 @@
-dtd:
+cifar10:
   _target_: datasets.load_dataset
   path: tanganke/cifar10
   split: test

fusion_bench_config/dataset/image_classification/test/cifar100.yaml

@@ -1,4 +1,4 @@
-dtd:
+cifar100:
   _target_: datasets.load_dataset
   path: tanganke/cifar100
   split: test

fusion_bench_config/dataset/image_classification/test/cub-200-2011.yaml

@@ -0,0 +1,4 @@
+cub-200-2011:
+  _target_: datasets.load_dataset
+  path: Donghyun99/CUB-200-2011
+  split: test

fusion_bench_config/dataset/image_classification/test/emnist_letters.yaml

@@ -0,0 +1,5 @@
+emnist_letters:
+  _target_: datasets.load_dataset
+  path: tanganke/emnist_letters
+  split: test
+

fusion_bench_config/dataset/image_classification/test/emnist_mnist.yaml

@@ -0,0 +1,4 @@
+emnist_mnist:
+  _target_: datasets.load_dataset
+  path: tanganke/emnist_mnist
+  split: test

fusion_bench_config/dataset/image_classification/test/fashion_mnist.yaml

@@ -0,0 +1,4 @@
+fashion_mnist:
+  _target_: datasets.load_dataset
+  path: zalando-datasets/fashion_mnist
+  split: test

fusion_bench_config/dataset/image_classification/test/fer2013.yaml

@@ -0,0 +1,3 @@
+fer2013:
+  _target_: fusion_bench.dataset.fer2013.load_fer2013
+  split: test

fusion_bench_config/dataset/image_classification/test/food101.yaml

@@ -0,0 +1,4 @@
+food101:
+  _target_: datasets.load_dataset
+  path: ethz/food101
+  split: validation

fusion_bench_config/dataset/image_classification/test/kmnist.yaml

@@ -0,0 +1,4 @@
+kmnist:
+  _target_: datasets.load_dataset
+  path: tanganke/kmnist
+  split: test

fusion_bench_config/dataset/image_classification/test/mango-leaf-disease.yaml

@@ -0,0 +1,4 @@
+mango-leaf-disease:
+  _target_: datasets.load_dataset
+  path: AfiqN/mango-leaf-disease
+  split: test

fusion_bench_config/dataset/image_classification/test/oxford-iiit-pet.yaml

@@ -0,0 +1,4 @@
+oxford-iiit-pet:
+  _target_: datasets.load_dataset
+  path: timm/oxford-iiit-pet
+  split: test

fusion_bench_config/dataset/image_classification/test/oxford_flowers102.yaml

@@ -0,0 +1,4 @@
+oxford_flowers102:
+  _target_: datasets.load_dataset
+  path: dpdl-benchmark/oxford_flowers102
+  split: test

fusion_bench_config/dataset/image_classification/test/pcam.yaml

@@ -0,0 +1,4 @@
+pcam:
+  _target_: datasets.load_dataset
+  path: 1aurent/PatchCamelyon
+  split: test

fusion_bench_config/dataset/image_classification/test/rendered-sst2.yaml

@@ -0,0 +1,4 @@
+rendered-sst2:
+  _target_: datasets.load_dataset
+  path: nateraw/rendered-sst2
+  split: test

fusion_bench_config/dataset/image_classification/test/stl10.yaml

@@ -0,0 +1,4 @@
+stl10:
+  _target_: datasets.load_dataset
+  path: tanganke/stl10
+  split: test

fusion_bench_config/dataset/image_classification/train/TALL14.yaml

@@ -0,0 +1,20 @@
+# The 14 task used in the paper: 
+# Wang et al. Localizing Task Information for Improved Model Merging and Compression
+# http://arxiv.org/abs/2405.07813
+defaults:
+  # eight tasks in the task arithmetic paper
+  - sun397
+  - stanford-cars
+  - resisc45
+  - eurosat
+  - svhn
+  - gtsrb
+  - mnist
+  - dtd
+  # additional 6 tasks in the TALL mask paper
+  - oxford_flowers102
+  - pcam
+  - fer2013
+  - oxford-iiit-pet
+  - stl10
+  - cifar100

fusion_bench_config/dataset/image_classification/train/TALL20.yaml

@@ -0,0 +1,28 @@
+# The 20 task used in the paper: 
+# Wang et al. Localizing Task Information for Improved Model Merging and Compression
+# http://arxiv.org/abs/2405.07813
+defaults:
+  # eight tasks in the task arithmetic paper
+  - sun397
+  - stanford-cars
+  - resisc45
+  - eurosat
+  - svhn
+  - gtsrb
+  - mnist
+  - dtd
+  # additional 6 tasks in the TALL mask paper (TALL 14)
+  - oxford_flowers102
+  - pcam
+  - fer2013
+  - oxford-iiit-pet
+  - stl10
+  - cifar100
+  # additional 6 tasks in the TALL mask paper (TALL 20)
+  - cifar10
+  - food101
+  - fashion_mnist
+  - emnist_letters
+  - kmnist
+  - rendered-sst2
+

fusion_bench_config/dataset/image_classification/train/cifar10.yaml

@@ -1,4 +1,4 @@
-dtd:
+cifar10:
   _target_: datasets.load_dataset
   path: tanganke/cifar10
   split: train

fusion_bench_config/dataset/image_classification/train/cifar100.yaml

@@ -1,4 +1,4 @@
-dtd:
+cifar100:
   _target_: datasets.load_dataset
   path: tanganke/cifar100
   split: train

fusion_bench_config/dataset/image_classification/train/cub-200-2011.yaml

@@ -0,0 +1,4 @@
+cub-200-2011:
+  _target_: datasets.load_dataset
+  path: Donghyun99/CUB-200-2011
+  split: train

fusion_bench_config/dataset/image_classification/train/emnist_letters.yaml

@@ -0,0 +1,4 @@
+emnist_letters:
+  _target_: datasets.load_dataset
+  path: tanganke/emnist_letters
+  split: train

fusion_bench_config/dataset/image_classification/train/emnist_mnist.yaml

@@ -0,0 +1,4 @@
+emnist_mnist:
+  _target_: datasets.load_dataset
+  path: tanganke/emnist_mnist
+  split: train

fusion_bench_config/dataset/image_classification/train/fashion_mnist.yaml

@@ -0,0 +1,4 @@
+fashion_mnist:
+  _target_: datasets.load_dataset
+  path: zalando-datasets/fashion_mnist
+  split: train

fusion_bench_config/dataset/image_classification/train/fer2013.yaml

@@ -0,0 +1,3 @@
+fer2013:
+  _target_: fusion_bench.dataset.fer2013.load_fer2013
+  split: train

fusion_bench_config/dataset/image_classification/train/food101.yaml

@@ -0,0 +1,4 @@
+food101:
+  _target_: datasets.load_dataset
+  path: ethz/food101
+  split: train

fusion_bench_config/dataset/image_classification/train/kmnist.yaml

@@ -0,0 +1,4 @@
+kmnist:
+  _target_: datasets.load_dataset
+  path: tanganke/kmnist
+  split: train

fusion_bench_config/dataset/image_classification/train/mango-leaf-disease.yaml

@@ -0,0 +1,4 @@
+mango-leaf-disease:
+  _target_: datasets.load_dataset
+  path: AfiqN/mango-leaf-disease
+  split: train

fusion_bench_config/dataset/image_classification/train/oxford-iiit-pet.yaml

@@ -0,0 +1,4 @@
+oxford-iiit-pet:
+  _target_: datasets.load_dataset
+  path: timm/oxford-iiit-pet
+  split: train

fusion_bench_config/dataset/image_classification/train/oxford_flowers102.yaml

@@ -0,0 +1,4 @@
+oxford_flowers102:
+  _target_: datasets.load_dataset
+  path: dpdl-benchmark/oxford_flowers102
+  split: train

fusion_bench_config/dataset/image_classification/train/pcam.yaml

@@ -0,0 +1,4 @@
+pcam:
+  _target_: datasets.load_dataset
+  path: 1aurent/PatchCamelyon
+  split: train

fusion_bench_config/dataset/image_classification/train/rendered-sst2.yaml

@@ -0,0 +1,4 @@
+rendered-sst2:
+  _target_: datasets.load_dataset
+  path: nateraw/rendered-sst2
+  split: train

fusion_bench_config/dataset/image_classification/train/stl10.yaml

@@ -0,0 +1,4 @@
+stl10:
+  _target_: datasets.load_dataset
+  path: tanganke/stl10
+  split: train

fusion_bench_config/dataset/llm_sft/alpaca_cleaned.yaml

@@ -0,0 +1,6 @@
+alpaca-cleaned:
+  _target_: fusion_bench.dataset.llama.alpaca.load_tokenized_alpaca_dataset
+  tokenizer: ???
+  path: "yahma/alpaca-cleaned"
+  split: train
+  cache_path: null

fusion_bench_config/dataset/llm_sft/ultrachat_200k.yaml

@@ -0,0 +1,3 @@
+ultrachat-200k:
+  _target_: fusion_bench.dataset.ultrachat.load_tokenized_ultrachat_200k
+  tokenizer: ???

fusion_bench_config/fabric/llama_peft_fsdp.yaml

@@ -0,0 +1,16 @@
+defaults:
+  - loggers: tensorboard_logger
+  - strategy: llama_peft_fsdp
+  - _self_
+
+_target_: lightning.Fabric
+_recursive_: true
+# Number of devices to train on (``int``), which GPUs to train on (``list`` or ``str``), or ``"auto"``.
+# The value applies per node.
+devices: auto
+# The hardware to run on. Possible choices are:
+# ``"cpu"``, ``"cuda"``, ``"mps"``, ``"gpu"``, ``"tpu"``, ``"auto"``.
+# for example: fabric.accelerator=cpu
+accelerator: auto
+# reference to the precision policy: https://lightning.ai/docs/fabric/stable/api/fabric_args.html#precision
+precision: bf16-true

fusion_bench_config/fabric/loggers/wandb_logger.yaml

@@ -0,0 +1,2 @@
+# https://lightning.ai/docs/fabric/2.4.0/guide/loggers/wandb.html#weights-and-biases
+_target_: wandb.integration.lightning.fabric.WandbLogger

fusion_bench_config/fabric/strategy/deepspeed.yaml

@@ -0,0 +1,10 @@
+# https://lightning.ai/docs/fabric/2.4.0/api/generated/lightning.fabric.strategies.DeepSpeedStrategy.html#deepspeedstrategy
+_target_: lightning.fabric.strategies.DeepSpeedStrategy
+
+accelerator: null
+zero_optimization: true
+stage: 2
+offload_optimizer: false
+offload_parameters: false
+offload_params_device: "cpu"
+offload_optimizer_device: "cpu"

fusion_bench_config/fabric/strategy/llama_peft_fsdp.yaml

@@ -0,0 +1,9 @@
+_target_: lightning.fabric.strategies.FSDPStrategy
+sharding_strategy: FULL_SHARD
+state_dict_type: full # Save a single, consolidated checkpoint file
+cpu_offload: false
+auto_wrap_policy:
+  _target_: fusion_bench.mixins.lightning_fabric.get_size_based_auto_wrap_policy
+activation_checkpointing_policy: ${.auto_wrap_policy}
+# limit_all_gathers: true
+

fusion_bench_config/fabric_model_fusion.yaml

@@ -11,7 +11,7 @@ _target_: fusion_bench.programs.FabricModelFusionProgram
 _recursive_: false
 fast_dev_run: false # Run a single batch of data to test the model or method
 # Run the script without actually running the experiment, use with `print_config=true`. 
-# You can also use `--cfg` or `-c` to  show the configuration instead of runing.
+# You can also use `--cfg` or `-c` to show the configuration instead of running.
 dry_run: false
 print_config: true # Print the configuration to the console
 merged_model_save_path: null # path to save the merged model, use "{log_dir}" to refer to the logger directory, for example `merged_model_save_path=\{log_dir\}/merged_model`

fusion_bench_config/llama_full_finetune.yaml

@@ -0,0 +1,19 @@
+defaults:
+  - hydra: default
+  - fabric: llama_fsdp
+  # --- Model, Method, Task ---
+  - method: lm_finetune/fullfinetune_sft.yaml
+  - modelpool: CausalLMPool/llama_alpaca_cleaned.yaml
+  - taskpool: dummy
+  - _self_
+
+_target_: fusion_bench.programs.FabricModelFusionProgram
+_recursive_: false
+
+fast_dev_run: false # Run a single batch of data to test the model or method
+# Run the script without actually running the experiment, use with `print_config=true`.
+# You can also use `--cfg` or `-c` to show the configuration instead of running.
+dry_run: false
+print_config: true # Print the configuration to the console
+report_save_path: null # path to save the result report
+print_function_call: true # set to false if you don't want to print the details of instantiate calls

fusion_bench_config/method/lm_finetune/bradley_terry_rm.yaml

@@ -0,0 +1,47 @@
+_target_: fusion_bench.method.BradleyTerryRewardModeling
+_recursive_: False
+
+optimizer:
+  _target_: torch.optim.AdamW
+  lr: 1e-5
+  weight_decay: 0.01
+  fused: null
+
+lr_scheduler:
+  _target_: fusion_bench.optim.lr_scheduler.CosineDecayWithWarmup
+  T_max: _T_max_ # this will be replaced by the expected number of training steps
+  init_lr: 0
+  warmup_steps: 100
+  max_lr: ${..optimizer.lr}
+  min_lr: 1e-6
+
+dataloader_kwargs:
+  # per-gpu batch size
+  batch_size: 1
+  num_workers: 0
+  pin_memory: True
+
+# Training hyperparameters
+# if max_epochs=-1, max_steps will be used to determine the number of training steps
+max_epochs: 3
+max_steps: -1
+max_steps_per_epoch: -1
+accumulate_grad_batches: 1
+lr_scheduler_interval: step
+lr_scheduler_frequency: 1
+# Checkpointing may be done by epoch or step, and at the end of training
+# `checkpoint_save_interval` can be 'epoch' or 'step'
+checkpoint_save_interval: epoch
+checkpoint_save_frequency: 1
+# Whether to use gradient clipping, and if so, the value and algorithm
+gradient_clip_val: null
+gradient_clip_algorithm: norm
+save_optimizer_state: false
+# save_full_model must be true when using shared FSDP
+save_full_model: true
+# save_ckpt_type can be 'hf' or 'lightning'
+save_ckpt_type: lightning
+# Path to checkpoint to load from, used for resuming training
+ckpt_path: null
+max_length: 4096
+fix_token_embedding: true

fusion_bench_config/method/lm_finetune/fullfinetune_sft.yaml

@@ -3,14 +3,17 @@ _recursive_: False
 
 optimizer:
   _target_: torch.optim.AdamW
-  fused: True
+  lr: 1e-5
   weight_decay: 0.01
-  lr: 5e-5
+  fused: null
 
 lr_scheduler:
-  _target_: torch.optim.lr_scheduler.CosineAnnealingLR
+  _target_: fusion_bench.optim.lr_scheduler.CosineDecayWithWarmup
   T_max: _T_max_ # this will be replaced by the expected number of training steps
-  eta_min: 1e-6
+  init_lr: 0
+  warmup_steps: 100
+  max_lr: ${..optimizer.lr}
+  min_lr: 1e-6
 
 dataloader_kwargs:
   # per-gpu batch size
@@ -36,5 +39,9 @@ gradient_clip_algorithm: norm
 save_optimizer_state: false
 # save_full_model must be true when using shared FSDP
 save_full_model: true
+# save_ckpt_type can be 'hf' or 'lightning'
+save_ckpt_type: lightning
 # Path to checkpoint to load from, used for resuming training
 ckpt_path: null
+max_length: 4096
+fix_token_embedding: true

fusion_bench_config/method/lm_finetune/peftfinetune_sft.yaml

@@ -3,9 +3,9 @@ _recursive_: False
 
 optimizer:
   _target_: torch.optim.AdamW
-  fused: True
+  lr: 1e-4
   weight_decay: 0.01
-  lr: 5e-5
+  fused: null
 
 lr_scheduler:
   _target_: torch.optim.lr_scheduler.CosineAnnealingLR
@@ -56,6 +56,8 @@ gradient_clip_algorithm: norm
 save_optimizer_state: false
 # save_full_model must be true when using shared FSDP
 save_full_model: false
+# save_ckpt_type can be 'peft' or 'lightning'
+save_ckpt_type: lightning
 # Path to checkpoint to load from, used for resuming training
 ckpt_path: null
 max_length: 4096

fusion_bench_config/method/surgery/adamerging_surgery.yaml

@@ -0,0 +1,27 @@
+# this option can be "clip_task_wise_adamerging"
+name: clip_layer_wise_adamerging_surgery
+# this weights can be a list of float, or a string that points to a *.np, *.pt file containing the weights
+# if weights is specified, skip the test-time adaptation training
+weights: null
+# learning rate
+optimizer: adam
+lr: 1e-3
+init_values: 0.3
+# 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
+# this is overrided by `fabric.devices` if launched from the `fusion_bench` CLI.
+devices: 1
+batch_size: 16
+num_workers: 8
+max_steps: 1000
+fast_dev_run: ${fast_dev_run}
+# the path for saving the merging weights
+save_merging_weights: 'merging_weights.pt'
+cache_dir: outputs
+
+# parameters of Surgery
+eval_iterations: 200
+surgery_steps: 1000

fusion_bench_config/method/task_singular_vector/TaskSingularVectorMerging.yaml

@@ -0,0 +1,2 @@
+_target_: fusion_bench.method.TaskSingularVectorMerging
+remove_keys: null

fusion_bench_config/model/clip-vit/README.md

@@ -0,0 +1,38 @@
+This folder contains the configuration for the CLIP-ViT models (managed by `fusion_bench.modelpool.CLIPVisionModelPool`).
+
+## Expected Configuration
+
+### Detailed Configuration
+
+
+```yaml
+${name_of_model}:
+  _target_: ${function_to_load_model}
+  ... # arguments to pass to the function
+```
+
+For example, to load the pre-trained CLIP-ViT-B/16 model, you can use the following configuration:
+
+```yaml
+_pretrained_: # `_pretrained_` is a special key in FusionBench that indicates the model is pre-trained
+  _target_: transformers.CLIPVisionModel.from_pretrained
+  pretrained_model_name_or_path: openai/clip-vit-base-patch16
+```
+
+In this case, calling `modelpool.load_model("_pretrained_")` will return a `transformers.CLIPVisionModel` instance, which is equivalent to call `transformers.CLIPVisionModel.from_pretrained("openai/clip-vit-base-patch16")`.
+
+The detailed configuration is more flexible and can be used when you need to pass additional arguments to the `from_pretrained` function or call custom functions to load and preprocess the model.
+
+### Simplified Configuration
+
+```yaml
+${name_of_model}: ${pretrained_model_name_or_path}
+```
+
+This is a simplified configuration that is equivalent to the detailed configuration.
+
+For example, to load the pre-trained CLIP-ViT-B/16 model, you can use the following configuration:
+
+```yaml
+_pretrained_: openai/clip-vit-base-patch16
+```

fusion_bench_config/model/clip-vit/clip-vit-base-patch16.yaml

@@ -1,3 +1 @@
-_pretrained_:
-  _target_: transformers.CLIPVisionModel.from_pretrained
-  pretrained_model_name_or_path: openai/clip-vit-base-patch16
+_pretrained_: openai/clip-vit-base-patch16

fusion_bench_config/model/clip-vit/clip-vit-base-patch16_TALL14.yaml

@@ -0,0 +1,22 @@
+# The 14 task used in the paper: 
+# Wang et al. Localizing Task Information for Improved Model Merging and Compression
+# http://arxiv.org/abs/2405.07813
+defaults:
+  # pre-trained model
+  - clip-vit-base-patch16
+  # eight tasks in the task arithmetic paper
+  - clip-vit-base-patch16_sun397
+  - clip-vit-base-patch16_stanford-cars
+  - clip-vit-base-patch16_resisc45
+  - clip-vit-base-patch16_eurosat
+  - clip-vit-base-patch16_svhn
+  - clip-vit-base-patch16_gtsrb
+  - clip-vit-base-patch16_mnist
+  - clip-vit-base-patch16_dtd
+  # additional 6 tasks in the TALL mask paper
+  - clip-vit-base-patch16_oxford_flowers102
+  - clip-vit-base-patch16_pcam
+  - clip-vit-base-patch16_fer2013
+  - clip-vit-base-patch16_oxford-iiit-pet
+  - clip-vit-base-patch16_stl10
+  - clip-vit-base-patch16_cifar100

fusion_bench_config/model/clip-vit/clip-vit-base-patch16_TALL20.yaml

@@ -0,0 +1,29 @@
+# The 20 task used in the paper: 
+# Wang et al. Localizing Task Information for Improved Model Merging and Compression
+# http://arxiv.org/abs/2405.07813
+defaults:
+  # pre-trained model
+  - clip-vit-base-patch16
+  # eight tasks in the task arithmetic paper
+  - clip-vit-base-patch16_sun397
+  - clip-vit-base-patch16_stanford-cars
+  - clip-vit-base-patch16_resisc45
+  - clip-vit-base-patch16_eurosat
+  - clip-vit-base-patch16_svhn
+  - clip-vit-base-patch16_gtsrb
+  - clip-vit-base-patch16_mnist
+  - clip-vit-base-patch16_dtd
+  # additional 6 tasks in the TALL mask paper (TALL 14)
+  - clip-vit-base-patch16_oxford_flowers102
+  - clip-vit-base-patch16_pcam
+  - clip-vit-base-patch16_fer2013
+  - clip-vit-base-patch16_oxford-iiit-pet
+  - clip-vit-base-patch16_stl10
+  - clip-vit-base-patch16_cifar100
+  # additional 6 tasks in the TALL mask paper (TALL 20)
+  - clip-vit-base-patch16_cifar10
+  - clip-vit-base-patch16_food101
+  - clip-vit-base-patch16_fashion_mnist
+  - clip-vit-base-patch16_emnist_letters
+  - clip-vit-base-patch16_kmnist
+  - clip-vit-base-patch16_rendered-sst2

fusion_bench_config/model/clip-vit/clip-vit-base-patch16_cifar10.yaml

@@ -0,0 +1 @@
+cifar10: tanganke/clip-vit-base-patch16_cifar10

fusion_bench_config/model/clip-vit/clip-vit-base-patch16_cifar100.yaml

@@ -0,0 +1 @@
+cifar100: tanganke/clip-vit-base-patch16_cifar100

fusion_bench_config/model/clip-vit/clip-vit-base-patch16_dtd.yaml

@@ -1,3 +1 @@
-dtd:
-  _target_: transformers.CLIPVisionModel.from_pretrained
-  pretrained_model_name_or_path: tanganke/clip-vit-base-patch16_dtd
+dtd: tanganke/clip-vit-base-patch16_dtd

fusion_bench_config/model/clip-vit/clip-vit-base-patch16_emnist_letters.yaml

@@ -0,0 +1 @@
+emnist_letters: tanganke/clip-vit-base-patch16_emnist_letters

fusion_bench_config/model/clip-vit/clip-vit-base-patch16_eurosat.yaml

@@ -1,3 +1 @@
-eurosat:
-  _target_: transformers.CLIPVisionModel.from_pretrained
-  pretrained_model_name_or_path: tanganke/clip-vit-base-patch16_eurosat
+eurosat: tanganke/clip-vit-base-patch16_eurosat