fusion-bench 0.2.30__tar.gz → 0.2.31__tar.gz

{fusion_bench-0.2.30 → fusion_bench-0.2.31}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fusion-bench
-Version: 0.2.30
+Version: 0.2.31
 Summary: A Comprehensive Benchmark of Deep Model Fusion
 Author-email: Anke Tang <tang.anke@foxmail.com>
 Project-URL: Repository, https://github.com/tanganke/fusion_bench

{fusion_bench-0.2.30 → fusion_bench-0.2.31}/fusion_bench/constants/runtime.py

@@ -89,7 +89,10 @@ class RuntimeConstants:
             self._initialized = True
 
     debug = False
-    """Global debug flag for enabling verbose logging and debugging features."""
+    """
+    Global debug flag for enabling verbose logging and debugging features.
+    Use `RuntimeConstants().debug` instead of `RuntimeConstants.debug`
+    """
 
     @property
     def cache_dir(self) -> Path:

{fusion_bench-0.2.30 → fusion_bench-0.2.31}/fusion_bench/method/classification/image_classification_finetune.py

@@ -173,6 +173,7 @@ class ImageClassificationFineTuning(BaseAlgorithm):
                 ),
             },
         )
+        lit_module.train()
 
         log_dir = (
             self._program.path.log_dir

fusion_bench-0.2.31/fusion_bench/method/concrete_subspace/clip_concrete_tsvm.py

@@ -0,0 +1,285 @@
+import logging
+import os
+from copy import deepcopy
+from typing import TYPE_CHECKING, Dict, Iterable, List, Literal, Optional
+
+import torch
+from omegaconf import DictConfig
+from tqdm import tqdm
+
+from fusion_bench import (
+    BaseAlgorithm,
+    OpenCLIPClassificationMixin,
+    OpenCLIPVisionModelPool,
+    SimpleProfilerMixin,
+    StateDictType,
+    auto_register_config,
+    get_rankzero_logger,
+    instantiate,
+)
+from fusion_bench.method.adamerging.entropy_loss import entropy_loss
+from fusion_bench.method.task_singular_vector import TaskSingularVectorMerging
+from fusion_bench.method.task_singular_vector.utils import (
+    TSVM_utils,
+    check_parameterNamesMatch,
+    check_state_dicts_equal,
+    state_dict_to_vector,
+    vector_to_state_dict,
+)
+from fusion_bench.models.masks import MaskModel, mask_sparsity
+from fusion_bench.models.open_clip import (
+    ClassificationHead,
+    ImageClassifier,
+    ImageEncoder,
+)
+from fusion_bench.models.wrappers.task_wise_fusion import (
+    TaskWiseMergedModel,
+    get_task_wise_weights,
+)
+from fusion_bench.utils.devices import clear_cuda_cache
+from fusion_bench.utils.dtype import parse_dtype
+from fusion_bench.utils.parameters import print_parameters, print_trainable_parameters
+from fusion_bench.utils.rich_utils import print_config_yaml
+from fusion_bench.utils.state_dict_arithmetic import (
+    _validate_state_dict_same_keys,
+    state_dict_add,
+    state_dict_hadamard_product,
+    state_dict_mul,
+    state_dict_sub,
+)
+
+log = get_rankzero_logger(__name__)
+
+
+@auto_register_config
+class ConcreteTSVMForOpenCLIP(
+    OpenCLIPClassificationMixin,
+    SimpleProfilerMixin,
+    BaseAlgorithm,
+):
+    def __init__(
+        self,
+        dataloader_kwargs: DictConfig,
+        optimizer: DictConfig,
+        lr_scheduler: DictConfig,
+        max_steps: int,
+        save_interval: int,
+        initial_logits: float,
+        temperature: float,
+        eval_mask_type: Literal["continuous", "discrete"],
+        mask_checkpoint: Optional[str],
+        merge_dtype: str,
+        clamp_weights: bool,
+        tie_weights: bool,
+        strict: bool,
+        skip_training: bool,
+        # === TSVM parameters ===
+        exclude_keys: Optional[List[str]],
+        alpha: float,
+        return_single_task_models: bool = True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        if not return_single_task_models:
+            log.warning("return_single_task_models is forced to be True here.")
+            self.return_single_task_models = True
+
+    @torch.no_grad()
+    def setup_models(self):
+        """
+        load the pre-trained model, task vectors, and construct the mask model.
+        """
+        merge_dtype = parse_dtype(self.merge_dtype)
+        modelpool = self.modelpool
+
+        # load the pre-trained model
+        pretrained_model = modelpool.load_pretrained_model()
+        self.set_clip_processor(stage="test", processor=pretrained_model.val_preprocess)
+
+        # constrcute mask model
+        mask_model = MaskModel(
+            pretrained_model, ignore_untrained_params=True, parameter_type="logits"
+        )
+        if merge_dtype is not None:
+            mask_model.to(merge_dtype)
+        mask_model.fill_(self.initial_logits)
+
+        if self.fabric.is_global_zero:
+            print("summary of mask model:")
+            print_parameters(mask_model)
+
+        if self.fabric.is_global_zero:
+            tsvm_algo = TaskSingularVectorMerging(
+                alpha=self.alpha,
+                exclude_keys=self.exclude_keys,
+                return_single_task_models=self.return_single_task_models,
+            )
+            tsvm_algo._fabric_instance = self.fabric
+            models = tsvm_algo.run(modelpool)
+
+            finetuned_models = [models[name] for name in modelpool.model_names]
+
+            task_wise_weight = get_task_wise_weights(
+                num_models=len(modelpool.model_names),
+                init_values=self.alpha,
+            )
+
+            # create a wrapped model
+            module = TaskWiseMergedModel(
+                task_wise_weight=task_wise_weight,
+                pretrained_model=pretrained_model,
+                finetuned_models=finetuned_models,
+                clamp_weights=self.clamp_weights,
+                tie_weights=self.tie_weights,
+                strict=self.strict,
+                task_vector_dtype=merge_dtype,
+            )
+            module = module.to(dtype=merge_dtype)
+
+            print("trainable parameter summary of merged model (TaskWiseMergedModel):")
+            print_trainable_parameters(module)
+        else:
+            module = None
+
+        with torch.no_grad():
+            self.fabric.barrier()
+            module = self.fabric.broadcast(module, src=0)
+
+        return module, mask_model
+
+    def train_mask(self, module: TaskWiseMergedModel, mask_model: MaskModel):
+        """
+        Train the mask model using the provided module.
+
+        This method configures the optimizer, sets up the mask model, and performs test-time adaptation to train the mask model.
+
+        Args:
+            module (TaskWiseMergedModel): The wrapped model with task-wise weights.
+            mask_model (MaskModel): The mask model to be trained.
+        """
+        config = self.config
+        merge_dtype = parse_dtype(self.merge_dtype)
+        log.info(f"Using merge dtype: {merge_dtype}")
+
+        optimizer: "torch.optim.Optimizer" = instantiate(
+            self.optimizer,
+            params=filter(lambda p: p.requires_grad, mask_model.parameters()),
+        )
+        print(f"{optimizer=}")
+        if self.lr_scheduler is not None:
+            lr_scheduler = instantiate(
+                self.lr_scheduler,
+                optimizer=optimizer,
+            )
+            print(f"{lr_scheduler=}")
+        else:
+            lr_scheduler = None
+
+        log.info("Setup models and optimizer with Fabric.")
+        mask_model, optimizer = self.fabric.setup(mask_model, optimizer)
+
+        log.info("Move the merged module to the correct device and disable gradients.")
+        module.requires_grad_(False)
+        module.to(mask_model.device)
+
+        mask_model.train()
+        optimizer.zero_grad()
+        for step_idx in (
+            pbar := tqdm(
+                range(self.config.max_steps if not self.is_debug_mode else 5),
+                ("[DEBUG MODE] " if self.is_debug_mode else "")
+                + "Concrete TSVM Test-Time Adaptation",
+                dynamic_ncols=True,
+                disable=not self.fabric.is_global_zero,
+            )
+        ):
+            metrics = {}
+            # sample a shared mask and merge weights
+            with self.profile("sample mask"):
+                mask = mask_model.sample_mask(
+                    mask_type="continuous", temperature=config.temperature
+                )
+                metrics["train/sparsity"] = mask_sparsity(mask)
+            with self.profile("merge weights"):
+                # rescale mask
+                for name, m in mask.items():
+                    mask[name] = m / torch.mean(m)
+                module.merge_weights(task_vector_mask=mask)
+
+            # ------ inner optimization goes here ------
+            # NOTE:
+            #   Because the algorithmic parameters of TSVM are assumed to be chosen on a validation test
+            #   set, we do not need to perform inner optimization here. So here we skip the inner optimization step.
+            # ------------------------------------------
+
+            total_loss = None
+            for task in self.modelpool.model_names:
+                with self.profile("data loading"):
+                    batch = next(self.get_shuffled_test_loader_iter(task))
+                    # NOTE: The labels are not allowed to be used during test-time adaptation
+                    images = batch[0].to(dtype=merge_dtype)
+                with self.profile("forward pass"):
+                    logits = self.compute_logits(module, images, task)
+                    loss = entropy_loss(logits)
+                    total_loss = loss if total_loss is None else total_loss + loss
+
+            with self.profile("compute grad"):
+                self.fabric.backward(total_loss)
+
+            with self.profile("optimizer step"):
+                optimizer.step()
+                optimizer.zero_grad()
+
+                if lr_scheduler is not None:
+                    lr_scheduler.step()
+
+            metrics.update({"train/loss": loss.item()})
+            self.fabric.log_dict(metrics, step=step_idx)
+            pbar.set_postfix(metrics)
+
+            if (step_idx + 1) % self.config.save_interval == 0:
+                with self.profiler.profile("save checkpoint"):
+                    save_dir = os.path.join(self.fabric.logger.log_dir, "checkpoints")
+                    if not os.path.exists(save_dir):
+                        os.makedirs(save_dir, exist_ok=True)
+                    save_path = os.path.join(save_dir, f"mask_steps_{step_idx}.pt")
+                    print(f"saving checkpoint to {save_path}")
+                    state = {"model": mask_model}
+                    self.fabric.save(save_path, state)
+
+                    # Create or update a symbolic link to the latest checkpoint
+                    if self.fabric.is_global_zero:
+                        symlink_path = os.path.join(save_dir, "latest_checkpoint.pt")
+                        if os.path.exists(symlink_path):
+                            os.remove(symlink_path)
+                        os.link(os.path.abspath(save_path), symlink_path)
+
+                self.print_profile_summary()
+
+    def run(self, modelpool: OpenCLIPVisionModelPool):
+        self.modelpool = modelpool
+        merge_dtype = parse_dtype(self.merge_dtype)
+
+        with self.profile("setup models"):
+            module, mask_model = self.setup_models()
+            self.setup_zero_shot_classification_head(freeze=True, dtype=merge_dtype)
+
+        if self.mask_checkpoint is None:
+            if not self.skip_training:
+                clear_cuda_cache()
+                self.train_mask(module, mask_model=mask_model)
+        else:
+            if self.fabric.is_global_zero:
+                print("loading mask from checkpoint", self.mask_checkpoint)
+            self.fabric.load(self.mask_checkpoint, {"model": mask_model})
+
+        with torch.no_grad():
+            clear_cuda_cache()
+            mask = mask_model.sample_mask(
+                mask_type=self.eval_mask_type, temperature=self.temperature
+            )
+            # rescale mask
+            for name, m in mask.items():
+                mask[name] = m / torch.mean(m)
+            model = module.merge_and_unload(mask)
+        return model.to(dtype=torch.float32)

{fusion_bench-0.2.30 → fusion_bench-0.2.31}/fusion_bench/method/task_singular_vector/TSVM.py

@@ -249,12 +249,13 @@ class TaskSingularVectorMerging(BaseAlgorithm, LightningFabricMixin):
         # - SVD finds the principal components (most important directions)
         # - Task vectors are reconstructed using only the most significant components
         # - The reconstructed vectors are merged (summed) to create a unified task vector
-        new_merged_tv = TSVM_utils.compute_and_sum_svd_mem_reduction(
-            task_vectors,
-            exclude_keys=self.exclude_keys,  # Skip certain parameters from SVD
-            accelerator=accelerator,  # Use GPU if available
-            return_single_task_models=self.return_single_task_models,
-        )
+        with torch.no_grad():
+            new_merged_tv = TSVM_utils.compute_and_sum_svd_mem_reduction(
+                task_vectors,
+                exclude_keys=self.exclude_keys,  # Skip certain parameters from SVD
+                accelerator=accelerator,  # Use GPU if available
+                return_single_task_models=self.return_single_task_models,
+            )
 
         # Handle the case where individual transformed task vectors are also returned
         if self.return_single_task_models:

{fusion_bench-0.2.30 → fusion_bench-0.2.31}/fusion_bench/method/task_singular_vector/utils/TSVM_utils.py

@@ -311,7 +311,6 @@ def compute_and_sum_svd_mem_reduction_lossless_eigen(
 
 ###############
 #### TSV Merge Orthogonalization
-@torch.no_grad()
 def compute_and_sum_svd_mem_reduction(
     task_vectors: List[StateDictType],
     exclude_keys: Optional[List[str]] = None,

{fusion_bench-0.2.30 → fusion_bench-0.2.31}/fusion_bench/mixins/lightning_fabric.py

@@ -10,6 +10,7 @@ from lightning.fabric.loggers import TensorBoardLogger
 from lightning.fabric.utilities.rank_zero import rank_zero_only
 from omegaconf import DictConfig, OmegaConf
 
+from fusion_bench.constants import RuntimeConstants
 from fusion_bench.utils import import_object
 from fusion_bench.utils.instantiate_utils import instantiate
 
@@ -206,14 +207,7 @@ class LightningFabricMixin:
         Returns:
             bool: True if fast_dev_run is enabled, False otherwise.
         """
-        if hasattr(self, "config") and self.config.get("fast_dev_run", False):
-            return True
-        elif hasattr(self, "_program") and self._program.config.get(
-            "fast_dev_run", False
-        ):
-            return True
-        else:
-            return False
+        return RuntimeConstants().debug
 
     def log(self, name: str, value: Any, step: Optional[int] = None):
         """

fusion_bench-0.2.31/fusion_bench/mixins/openclip_classification.py

@@ -0,0 +1,165 @@
+import functools
+import logging
+from typing import TYPE_CHECKING, Callable, Dict, Iterator, List, Literal, Optional
+
+import torch
+from omegaconf import DictConfig
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+
+from fusion_bench.dataset.clip_dataset import CLIPDataset
+from fusion_bench.mixins import LightningFabricMixin
+from fusion_bench.modelpool import OpenCLIPVisionModelPool
+from fusion_bench.models.open_clip import (
+    ClassificationHead,
+    ImageClassifier,
+    ImageEncoder,
+)
+from fusion_bench.utils.data import InfiniteDataLoader
+
+log = logging.getLogger(__name__)
+
+
+class OpenCLIPClassificationMixin(LightningFabricMixin):
+
+    _train_processor = None
+    _test_processor = None
+    dataloader_kwargs: DictConfig
+    modelpool: OpenCLIPVisionModelPool
+    zero_shot_heads: Dict[str, ClassificationHead] = {}
+
+    def _init_processor(self, encoder: Optional["ImageEncoder"] = None):
+        """
+        Initialize the CLIP processors for training and testing.
+        """
+        if encoder is None:
+            encoder: "ImageEncoder" = self.modelpool.load_pretrained_or_first_model()
+        self._train_processor = encoder.train_preprocess
+        self._test_processor = encoder.val_preprocess
+        return self._train_processor, self._test_processor
+
+    def get_clip_processor(self, stage: Literal["train", "test"]):
+        """
+        Get the CLIP processor, loading it from the model pool if necessary.
+
+        Returns:
+            CLIPProcessor: The CLIP processor for image and text preprocessing.
+
+        Raises:
+            AssertionError: If the model pool is not set.
+        """
+        if stage == "train":
+            if self._train_processor is None:
+                self._init_processor()
+            return self._train_processor
+        elif stage == "test":
+            if self._test_processor is None:
+                self._init_processor()
+            return self._test_processor
+        else:
+            raise ValueError(f"Invalid stage: {stage}")
+
+    def setup_zero_shot_classification_head(
+        self,
+        task_names: Optional[List[str]] = None,
+        freeze: bool = True,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        # check task names consistency across processes
+        _task_names = self.fabric.broadcast(task_names, src=0)
+        if not self.fabric.is_global_zero and task_names != _task_names:
+            raise ValueError("The `task_names` must be the same across all processes.")
+
+        for task in tqdm(
+            self.modelpool.model_names if task_names is None else task_names,
+            "Setting up zero-shot classification head",
+            disable=not self.fabric.is_global_zero,
+        ):
+            head = self.modelpool.load_classification_head(task)
+            if freeze:
+                head.requires_grad_(False)
+            if dtype is not None:
+                head = head.to(dtype=dtype)
+            self.zero_shot_heads[task] = self.to_device(head)
+
+    def set_clip_processor(self, stage: Literal["train", "test"], processor: Callable):
+        """
+        Set the CLIP processor for a specific stage.
+
+        Args:
+            stage (Literal["train", "test"]): The stage for which to set the processor.
+            processor (Callable): The CLIP processor to set.
+        """
+        if stage == "train":
+            self._train_processor = processor
+        elif stage == "test":
+            self._test_processor = processor
+        else:
+            raise ValueError(f"Invalid stage: {stage}")
+
+    @functools.cache
+    def get_shuffled_test_loader_iter(
+        self,
+        task: str,
+        batch_size: Optional[int] = None,
+        num_workers: Optional[int] = None,
+        **loader_kwargs,
+    ) -> Iterator:
+        """
+        Get an iterator for a shuffled test DataLoader.
+
+        This method creates a DataLoader for the test dataset of the specified task,
+        with shuffling enabled. It allows for optional customization of batch size,
+        number of workers, and other DataLoader keyword arguments.
+
+        Args:
+            task (str): The task identifier for which the test dataset is to be loaded.
+            batch_size (Optional[int]): The batch size to use for the DataLoader. If None, the default batch size is used.
+            num_workers (Optional[int]): The number of worker processes to use for data loading. If None, the default number of workers is used.
+            **loader_kwargs: Additional keyword arguments to pass to the DataLoader.
+
+        Returns:
+            Iterator: An iterator over the shuffled test DataLoader.
+        """
+        # get dataloader kwargs
+        dataloader_kwargs = self.dataloader_kwargs.copy()
+        dataloader_kwargs["shuffle"] = True
+        if batch_size is not None:
+            dataloader_kwargs["batch_size"] = batch_size
+        if num_workers is not None:
+            dataloader_kwargs["num_workers"] = num_workers
+        dataloader_kwargs.update(loader_kwargs)
+
+        # get the test dataset
+        clip_dataset = CLIPDataset(
+            self.modelpool.load_test_dataset(task),
+            processor=self.get_clip_processor(stage="test"),
+        )
+        # create the dataloader
+        loader = DataLoader(clip_dataset, **dataloader_kwargs)
+        loader = self.fabric.setup_dataloaders(loader)
+        return iter(InfiniteDataLoader(loader))
+
+    def compute_logits(
+        self,
+        module: ImageClassifier,
+        images,
+        task: str,
+    ):
+        """
+        Compute the logits for a batch of images using the provided module and task.
+
+        Args:
+            module (ImageClassifier): The image classification module to use for computing logits.
+            images (torch.Tensor): The batch of images for which to compute logits.
+            task (str): The task identifier to specify which classification head to use.
+
+        Returns:
+            torch.Tensor: The computed logits for the input images.
+        """
+        if len(self.zero_shot_heads) == 0:
+            self.setup_zero_shot_classification_head()
+        task_head = self.zero_shot_heads[task]
+        features = module(images)
+        logits = task_head(features)
+        return logits

{fusion_bench-0.2.30 → fusion_bench-0.2.31}/fusion_bench/modelpool/base_pool.py

@@ -7,6 +7,7 @@ from omegaconf import DictConfig, OmegaConf, UnsupportedValueType
 from torch import nn
 from torch.utils.data import Dataset
 
+from fusion_bench import TorchModelType
 from fusion_bench.mixins import BaseYAMLSerializable, HydraConfigMixin
 from fusion_bench.utils import (
     ValidationError,

{fusion_bench-0.2.30 → fusion_bench-0.2.31}/fusion_bench/modelpool/openclip_vision/modelpool.py

@@ -1,7 +1,7 @@
 import logging
 import pickle
 import sys
-from typing import Callable, Optional, Union, cast
+from typing import Callable, Optional, Union, cast, override
 
 import torch
 from datasets import load_dataset
@@ -41,8 +41,8 @@ def _check_and_redirect_open_clip_modeling():
             )
 
     try:
-        import src
-        import src.modeling
+        import src  # type: ignore
+        import src.modeling  # type: ignore
     except ImportError:
         if "src" not in sys.modules:
             # redirect the import of `src` to `fusion_bench.models.open_clip`
@@ -114,6 +114,7 @@ class OpenCLIPVisionModelPool(BaseModelPool):
             self._test_processor = encoder.val_preprocess
         return self._test_processor
 
+    @override
     def load_model(
         self, model_name_or_config: Union[str, DictConfig], *args, **kwargs
     ) -> ImageEncoder:
@@ -210,6 +211,8 @@ class OpenCLIPVisionModelPool(BaseModelPool):
         - A string, which is the path to the model checkpoint in pickle format. Load directly using `torch.load`.
         - Default, load the model using `instantiate` from hydra.
         """
+        if self._classification_heads is None:
+            raise ValueError("No classification heads are defined in the model pool.")
         if (
             isinstance(model_name_or_config, str)
             and model_name_or_config in self._classification_heads
@@ -222,6 +225,8 @@ class OpenCLIPVisionModelPool(BaseModelPool):
         return head
 
     def load_train_dataset(self, dataset_name: str, *args, **kwargs):
+        if self._train_datasets is None:
+            raise ValueError("No train datasets are defined in the model pool.")
         dataset_config = self._train_datasets[dataset_name]
         if isinstance(dataset_config, str):
             log.info(
@@ -233,6 +238,8 @@ class OpenCLIPVisionModelPool(BaseModelPool):
         return dataset
 
     def load_val_dataset(self, dataset_name: str, *args, **kwargs):
+        if self._val_datasets is None:
+            raise ValueError("No val datasets are defined in the model pool.")
         dataset_config = self._val_datasets[dataset_name]
         if isinstance(dataset_config, str):
             log.info(
@@ -244,6 +251,8 @@ class OpenCLIPVisionModelPool(BaseModelPool):
         return dataset
 
     def load_test_dataset(self, dataset_name: str, *args, **kwargs):
+        if self._test_datasets is None:
+            raise ValueError("No test datasets are defined in the model pool.")
         dataset_config = self._test_datasets[dataset_name]
         if isinstance(dataset_config, str):
             log.info(