fusion-bench 0.2.10__tar.gz → 0.2.11__tar.gz

{fusion_bench-0.2.10 → fusion_bench-0.2.11}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: fusion_bench
-Version: 0.2.10
+Version: 0.2.11
 Summary: A Comprehensive Benchmark of Deep Model Fusion
 Author-email: Anke Tang <tang.anke@foxmail.com>
 License: MIT License

{fusion_bench-0.2.10 → fusion_bench-0.2.11}/fusion_bench/compat/method/__init__.py

@@ -20,12 +20,17 @@ class AlgorithmFactory:
         # model merging methods
         "clip_task_wise_adamerging": ".adamerging.clip_task_wise_adamerging.CLIPTaskWiseAdaMergingAlgorithm",
         "clip_layer_wise_adamerging": ".adamerging.clip_layer_wise_adamerging.CLIPLayerWiseAdaMergingAlgorithm",
+        "clip_layer_wise_adamerging_doge_ta": ".DOGE_TA.clip_layer_wise_adamerging.CLIPLayerWiseAdaMergingAlgorithm",
         "singular_projection_merging": "fusion_bench.method.smile_upscaling.singular_projection_merging.SingularProjectionMergingAlgorithm",
         "clip_layer_wise_adamerging_surgery": ".surgery.clip_layer_wise_adamerging_surgery.CLIPLayerWiseAdaMergingSurgeryAlgorithm",
         # plug-and-play model merging methods
         "clip_concrete_task_arithmetic": ".concrete_subspace.clip_concrete_task_arithmetic.ConcreteTaskArithmeticAlgorithmForCLIP",
         "clip_concrete_task_wise_adamerging": ".concrete_subspace.clip_concrete_adamerging.ConcreteTaskWiseAdaMergingForCLIP",
         "clip_concrete_layer_wise_adamerging": ".concrete_subspace.clip_concrete_adamerging.ConcreteLayerWiseAdaMergingForCLIP",
+        "clip_post_defense_AWM": ".concrete_subspace.clip_post_defense.PostDefenseAWMAlgorithmForCLIP",
+        "clip_post_defense_SAU": ".concrete_subspace.clip_post_defense.PostDefenseSAUAlgorithmForCLIP",
+        "clip_safe_concrete_layer_wise_adamerging": ".concrete_subspace.clip_safe_concrete_adamerging.ConcreteSafeLayerWiseAdaMergingForCLIP",
+        "clip_safe_concrete_task_wise_adamerging": ".concrete_subspace.clip_safe_concrete_adamerging.ConcreteSafeTaskWiseAdaMergingForCLIP",
         # model mixing methods
         "clip_weight_ensembling_moe": ".we_moe.clip_we_moe.CLIPWeightEnsemblingMoEAlgorithm",
         "sparse_clip_weight_ensembling_moe": "fusion_bench.method.SparseCLIPWeightEnsemblingMoEAlgorithm",

fusion_bench-0.2.11/fusion_bench/method/DOGE_TA/DOGE_TA.py

@@ -0,0 +1,364 @@
+R"""
+This script contains the general implementation of Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent.
+
+https://arxiv.org/abs/2501.01230
+
+Example Usage:
+
+```bash
+fusion_bench \
+    method=DOGE_TA/DOGE_TA \
+    modelpool=CLIPVisionModelPool/clip-vit-base-patch32_TA8_model_only \
+    taskpool=CLIPVisionModelTaskPool/clip-vit-classification_TA8
+
+fusion_bench \
+    method=adamerging \
+    method.name=clip_layer_wise_adamerging_doge_ta \
+    modelpool=CLIPVisionModelPool/clip-vit-base-patch32_TA8 \
+    taskpool=CLIPVisionModelTaskPool/clip-vit-classification_TA8
+```
+"""
+
+import copy
+import logging
+import time
+from collections import OrderedDict
+from copy import deepcopy
+from functools import reduce
+from typing import Dict, List, Mapping, TypeVar, Union  # noqa: F401
+
+import lightning as L
+import torch
+from torch import nn
+
+from fusion_bench.method.base_algorithm import BaseAlgorithm
+from fusion_bench.mixins.lightning_fabric import LightningFabricMixin
+from fusion_bench.mixins.simple_profiler import SimpleProfilerMixin
+from fusion_bench.modelpool import BaseModelPool
+from fusion_bench.utils.state_dict_arithmetic import (
+    state_dict_add,
+    state_dict_mul,
+    state_dict_sub,
+)
+from fusion_bench.utils.type import StateDictType
+
+log = logging.getLogger(__name__)
+
+
+class DOGE_TA_Algorithm(
+    BaseAlgorithm,
+    SimpleProfilerMixin,
+    LightningFabricMixin,
+):
+    """
+    Task Arithmetic Algorithm for model fusion with learnable delta.
+
+    This class extends the Task Arithmetic method to include a learnable delta
+    for task vectors, optimized to maximize cosine similarity among the task vectors.
+
+    Attributes:
+        scaling_factor (int): The factor by which the task vectors will be scaled before merging.
+        delta (StateDictType): A learnable parameter to adjust task vectors, initialized as zeros.
+    """
+
+    _config_mapping = BaseAlgorithm._config_mapping | {
+        "subspace": "subspace",
+        "K": "K",
+        "lamda": "lamda",
+    }
+
+    def __init__(self, subspace, K, lamda):
+        self.delta = None  # Initialize delta as None; will be set during run
+        self.subspace = subspace
+        self.K = K
+        self.lamda = lamda
+        super().__init__()
+
+    @property
+    def device(self) -> torch.device:
+        return self.fabric.device
+
+    @torch.no_grad()
+    def compute_task_vectors(
+        self, modelpool: BaseModelPool, pretrained_model: nn.Module
+    ) -> List[StateDictType]:
+        """
+        Computes task vectors for each model in the model pool relative to the pretrained model.
+        """
+        task_vectors = []
+        pretrained_sd = pretrained_model.state_dict(keep_vars=True)
+        filtered_keys = [
+            k
+            for k in pretrained_sd.keys()
+            if ("encoder" in k and "layer_norm" not in k and "weight" in k)
+        ]  # Flan T5: "layer_norm" not in k and ("q.weight" in k or "v.weight" in k)
+
+        for model_name in modelpool.model_names:
+            model = modelpool.load_model(model_name)
+            model_sd = model.state_dict(keep_vars=True)
+
+            filtered_task_vector = {
+                k: (model_sd[k] - pretrained_sd[k]) for k in filtered_keys
+            }
+            task_vectors.append(filtered_task_vector)
+
+        return task_vectors
+
+    def taskvector_loss(self, layer_vectors, layer_delta, layer_lamdas) -> torch.Tensor:
+        """
+        Computes the loss based on delta and task vectors for a specific layer.
+        """
+        total_loss = 0.0
+
+        layer_vectors_scale = layer_vectors * layer_lamdas.view(-1, 1, 1)
+        sum_over_num_vectors = layer_vectors_scale.sum(dim=0)
+
+        layer_delta_scale = layer_delta.unsqueeze(0) * layer_lamdas.view(-1, 1, 1)
+        sum_over_delta = layer_delta_scale.sum(dim=0)
+
+        # Iterate through each vector and calculate the loss one by one
+        for v_j in layer_vectors:
+            part1 = -v_j * sum_over_num_vectors
+            part2 = -v_j * sum_over_delta
+            part3 = v_j * v_j
+
+            expression = part1 + part2 + part3
+            layer_loss = expression.sum(dim=1).pow(2).sum()
+
+            # Cumulative total loss
+            total_loss += layer_loss
+        return total_loss
+
+    @torch.enable_grad()
+    def optimize_delta(self, task_vectors: List[StateDictType]) -> None:
+        """
+        Optimizes the delta based on the loss of task vectors.
+        """
+        if self.delta is None:
+            self.delta = {
+                k: nn.Parameter(torch.zeros_like(v, device=self.device).detach())
+                for k, v in task_vectors[0].items()
+            }
+
+        optimizer = torch.optim.Adam(self.delta.values(), lr=1e-4)
+        initial_mem = torch.cuda.memory_allocated()
+        start_time = time.time()
+        for layer_name in task_vectors[0].keys():
+            layer_vectors = torch.stack([vec[layer_name] for vec in task_vectors]).to(
+                self.device
+            )
+            layer_lamdas = torch.stack(
+                [lamdas[layer_name] for lamdas in self.lamdas]
+            ).to(self.device)
+            for _ in range(400):
+                optimizer.zero_grad()
+                loss = self.taskvector_loss(
+                    layer_vectors, self.delta[layer_name], layer_lamdas
+                )
+                self.fabric.backward(loss)
+                grad_proj = (
+                    self.projection[layer_name] @ self.delta[layer_name].grad.detach()
+                )
+                self.delta[layer_name].grad.data = self.delta[
+                    layer_name
+                ].grad.data.sub_(grad_proj)
+                optimizer.step()
+                self.delta[layer_name].grad = None
+        end_time = time.time()
+        print(f"Running time: {end_time - start_time} s")
+        final_mem = torch.cuda.memory_allocated()
+        print(f"Memory usage: {(final_mem - initial_mem) / (1024 ** 2)} MB")
+        print("Optimization completed.")
+
+    @torch.no_grad()
+    def run(self, modelpool: Union[BaseModelPool, Dict[str, nn.Module]]):
+        """
+        Runs the Algorithm with learnable delta to fuse models in the given model pool.
+
+        Args:
+            modelpool (Union[BaseModelPool, Dict[str, nn.Module]]): The pool of models to fuse.
+
+        Returns:
+            nn.Module: The pre-trained model with the merged task vectors after optimizing delta.
+        """
+        if not isinstance(modelpool, BaseModelPool):
+            modelpool = BaseModelPool(modelpool)
+
+        log.info("Fusing models using DOGE_TA with learnable delta.")
+        with self.profile("load model"):
+            pretrained_model = modelpool.load_model("_pretrained_")
+
+        task_vectors = self.compute_task_vectors(modelpool, pretrained_model)
+
+        self.lamdas = self.compute_layer_lamdas(task_vectors)
+        self.projection = {}
+        for layer_name in task_vectors[0].keys():
+            for i, vector in enumerate(task_vectors):
+                layer_vector = vector[layer_name].to(self.device)
+                u, s, v = torch.linalg.svd(layer_vector, full_matrices=False)
+                if i == 0:
+                    print(f"Computed SVD for {layer_name}...")
+                    sum_u = torch.zeros_like(u, device=layer_vector.device)
+                    sum_s = torch.zeros_like(s, device=layer_vector.device)
+                    sum_v = torch.zeros_like(v, device=layer_vector.device)
+
+                reduced_index_s = int(s.shape[0] / len(task_vectors))
+
+                # select only the first reduced_index_s columns of u and place them
+                sum_u[:, i * reduced_index_s : (i + 1) * reduced_index_s] = u[
+                    :, :reduced_index_s
+                ]
+                sum_s[i * reduced_index_s : (i + 1) * reduced_index_s] = s[
+                    :reduced_index_s
+                ]
+                # select only the first reduced_index_s rows of v and place them
+                sum_v[i * reduced_index_s : (i + 1) * reduced_index_s, :] = v[
+                    :reduced_index_s, :
+                ]
+            u_u, s_u, v_u = torch.linalg.svd(sum_u, full_matrices=False)
+            layer_proj = torch.matmul(
+                u_u[:, : int(s.shape[0] / self.config.subspace)],
+                u_u[:, : int(s.shape[0] / self.config.subspace)].T,
+            )
+            self.projection[layer_name] = layer_proj
+
+        self.optimize_delta(task_vectors)
+
+        del self.projection
+        self.delta = {key: param.detach().cpu() for key, param in self.delta.items()}
+        self.lamdas = [
+            {key: param.cpu() for key, param in lamdas.items()}
+            for lamdas in self.lamdas
+        ]
+        task_vectors = [
+            {k: v.cpu() for k, v in task_vector.items()} for task_vector in task_vectors
+        ]
+        flat_vectors = []
+        vector_masks = []
+        for idx, task_vector in enumerate(task_vectors):
+            flat_vector = self.state_dict_to_vector(task_vector)
+            vector_mask = self.topk_values_mask(flat_vector, K=self.config.K)
+            flat_vectors.append(flat_vector)
+            vector_masks.append(vector_mask)
+        flat_delta = self.state_dict_to_vector(self.delta)
+
+        adjusted_vectors = [
+            self.vector_to_state_dict(
+                (flat_vector + flat_delta) * vector_mask, self.delta
+            )
+            for flat_vector, vector_mask in zip(flat_vectors, vector_masks)
+        ]
+
+        for layer_name in adjusted_vectors[0].keys():
+            layer_vectors = torch.stack(
+                [vec[layer_name] for vec in adjusted_vectors], dim=0
+            )
+            layer_lamdas = torch.stack(
+                [lamdas[layer_name] for lamdas in self.lamdas], dim=0
+            )
+            layer_vectors_scale = layer_vectors * layer_lamdas.view(-1, 1, 1)
+            task_vectors[0][layer_name] = layer_vectors_scale.sum(dim=0)
+
+        final_state_dict = {}
+        pretrained_sd = pretrained_model.state_dict(keep_vars=True)
+        for k, v in pretrained_sd.items():
+            if k in task_vectors[0]:
+                final_state_dict[k] = v + task_vectors[0][k]
+            else:
+                final_state_dict[k] = v
+
+        pretrained_model.load_state_dict(final_state_dict)
+
+        self.print_profile_summary()
+        return pretrained_model
+
+    def compute_lamdas(self, vectors: List[StateDictType]) -> torch.Tensor:
+        lamdas = []
+        for vec in vectors:
+            norm_vec = torch.norm(
+                torch.cat([param.flatten() for param in vec.values()])
+            )
+            # norm_vec = sum([torch.norm(param) for param in vec.values()])
+            lamdas.append(self.config.lamda / norm_vec)
+        print(lamdas)
+        return lamdas
+
+    def compute_layer_lamdas(self, vectors: List[StateDictType]) -> torch.Tensor:
+        lamdas = []
+        for vec in vectors:
+            tmp = {}
+            for layer_name in vec.keys():
+                norm_vec = torch.norm(vec[layer_name])
+                tmp[layer_name] = self.config.lamda / norm_vec
+            lamdas.append(tmp)
+        return lamdas
+
+    def topk_values_mask(self, M, K):
+        if K > 1:
+            K /= 100
+
+        original_shape = M.shape
+        if M.dim() == 1:
+            M = M.unsqueeze(0)
+
+        n, d = M.shape
+        k = int(d * K)
+        k = d - k  # Keep top k elements instead of bottom k elements
+
+        # Find the k-th smallest element by magnitude for each row
+        kth_values, _ = M.abs().kthvalue(k, dim=1, keepdim=True)
+        # Create a mask tensor with True for the top k elements in each row
+        mask = M.abs() >= kth_values
+        final_mask = mask.squeeze() if original_shape == M.squeeze().shape else mask
+
+        return final_mask
+
+    def state_dict_to_vector(self, state_dict, remove_keys=[]):
+        """
+        Convert a state dictionary to a vector, removing specified keys.
+
+        Args:
+            state_dict (dict): The state dictionary to convert.
+            remove_keys (list): List of keys to remove from the state dictionary.
+
+        Returns:
+            Tensor: A vector representation of the state dictionary.
+        """
+        shared_state_dict = copy.deepcopy(state_dict)
+        for key in remove_keys:
+            if key in shared_state_dict:
+                del shared_state_dict[key]
+        sorted_shared_state_dict = OrderedDict(sorted(shared_state_dict.items()))
+        return nn.utils.parameters_to_vector(
+            [value.reshape(-1) for key, value in sorted_shared_state_dict.items()]
+        )
+
+    def vector_to_state_dict(self, vector, state_dict, remove_keys=[]):
+        """
+        Convert a vector back to a state dictionary, removing specified keys.
+
+        Args:
+            vector (Tensor): The vector to convert.
+            state_dict (dict): The reference state dictionary.
+            remove_keys (list): List of keys to remove from the state dictionary.
+
+        Returns:
+            dict: A state dictionary representation of the vector.
+        """
+        # create a reference dict to define the order of the vector
+        reference_dict = copy.deepcopy(state_dict)
+        for key in remove_keys:
+            if key in reference_dict:
+                del reference_dict[key]
+        sorted_reference_dict = OrderedDict(sorted(reference_dict.items()))
+
+        # create a shared state dict using the reference dict
+        nn.utils.vector_to_parameters(vector, sorted_reference_dict.values())
+
+        # add back the encoder and decoder embedding weights.
+        if "transformer.shared.weight" in sorted_reference_dict:
+            for key in remove_keys:
+                sorted_reference_dict[key] = sorted_reference_dict[
+                    "transformer.shared.weight"
+                ]
+        return sorted_reference_dict

fusion_bench-0.2.11/fusion_bench/method/DOGE_TA/__init__.py

@@ -0,0 +1,2 @@
+# flake8: noqa F401
+from .DOGE_TA import DOGE_TA_Algorithm

fusion_bench-0.2.11/fusion_bench/method/DOGE_TA/layer_wise_adamerging.py

@@ -0,0 +1,250 @@
+import logging
+import os
+from abc import abstractmethod
+from typing import TYPE_CHECKING, Any, List, Mapping, TypeVar, Union, cast  # noqa: F401
+
+import torch
+from lightning.fabric.utilities.rank_zero import rank_zero_only
+from omegaconf import DictConfig
+from torch import Tensor, nn
+from torch.utils.data import DataLoader
+from tqdm.autonotebook import tqdm
+
+from fusion_bench.compat.method import ModelFusionAlgorithm
+from fusion_bench.compat.modelpool import ModelPool
+from fusion_bench.method.adamerging.entropy_loss import entropy_loss
+from fusion_bench.method.adamerging.utils import get_memory_usage
+from fusion_bench.mixins.lightning_fabric import LightningFabricMixin
+from fusion_bench.mixins.simple_profiler import SimpleProfilerMixin
+from fusion_bench.models.wrappers.layer_wise_fusion_doge_ta import (
+    LayerWiseMergedModel,
+    get_layer_wise_weights,
+)
+from fusion_bench.utils.data import load_tensor_from_file
+from fusion_bench.utils.type import TorchModelType
+
+if TYPE_CHECKING:
+    from fusion_bench.programs.fabric_fusion_program import FabricModelFusionProgram
+
+log = logging.getLogger(__name__)
+
+
+class LayerWiseAdaMergingAlgorithm(
+    ModelFusionAlgorithm,
+    LightningFabricMixin,
+    SimpleProfilerMixin,
+):
+    _program: "FabricModelFusionProgram"
+    """The program that this algorithm is running on."""
+
+    """
+    Implements the Layer-Wise AdaMerging Algorithm.
+
+    This class merges the layers of a pretrained model with those of several fine-tuned models.
+    The merging is controlled by layer-wise weights, which can be initialized based on a provided configuration or loaded from a file.
+    """
+
+    def __init__(self, algorithm_config: DictConfig):
+        """
+        Initialize the LayerWiseAdaMergingAlgorithm with the given configuration.
+
+        Args:
+            algorithm_config (DictConfig): The configuration for the algorithm.
+        """
+        super().__init__(algorithm_config)
+
+    def construct_layer_wise_merged_model(self, modelpool: "ModelPool"):
+        """
+        Constructs a wrapped layer-wise merged model from model pool.
+
+        This method creates a new wrapped model by merging the layers of a pretrained model with those of several fine-tuned models.
+        The merging is controlled by layer-wise weights, which is a `torch.Tensor` of the shape `(num_models, num_layers)`.
+        The merging weights can be initialized based on a provided configuration or loaded from a file.
+
+        Args:
+            modelpool (ModelPool): An object containing the pretrained model and fine-tuned models to be merged.
+
+        Returns:
+            LayerWiseMergedModel: An instance of the merged model with layer-wise weights applied.
+        """
+        pretrained_model = modelpool.load_model("_pretrained_")
+        finetuned_models = [
+            modelpool.load_model(name) for name in modelpool.model_names
+        ]
+
+        # initialize layer-wise weights using the provided configuration `init_values` or load from file if `weights` is provided
+        if self.config.weights is None:
+            layer_wise_weight = get_layer_wise_weights(
+                num_models=len(modelpool.model_names),
+                num_layers=len(
+                    tuple(
+                        filter(lambda p: p.requires_grad, pretrained_model.parameters())
+                    )
+                ),
+                init_values=self.config.init_values,
+            )
+        else:
+            if isinstance(self.config.weights, str):
+                # self.config.weights is a path to a saved tensor
+                layer_wise_weight = load_tensor_from_file(self.config.weights)
+            else:
+                raise ValueError(f"Unsupported weights format: {self.config.weights}")
+
+        module = LayerWiseMergedModel(
+            layer_wise_weight=layer_wise_weight,
+            pretrained_model=pretrained_model,
+            finetuned_models=finetuned_models,
+            clamp_weights=self.config.clamp_weights,
+            tie_weights=self.config.tie_weights,
+            strict=self.config.strict,
+        )
+        print(f"{layer_wise_weight.size()=}, {layer_wise_weight.numel()=}")
+        return module
+
+    @rank_zero_only
+    def save_merging_weights(self, file_path: str, merging_weights: torch.Tensor):
+        """
+        Save the merging weights to a file.
+
+        Args:
+            file_path (str): The path to save the merging weights.
+            merging_weights (torch.Tensor): The merging weights to save.
+        """
+        if self.fabric.is_global_zero and self.config.get(
+            "save_merging_weights", False
+        ):
+            if isinstance(file_path, str) and not file_path.startswith(("/", ".")):
+                # if the file path is not absolute or relative to current working directory, save it in the log directory
+                save_path = os.path.join(self.log_dir, file_path)
+            else:
+                save_path = file_path
+            log.info(f"saving merging weights to {save_path}.")
+            if os.path.dirname(save_path):
+                os.makedirs(os.path.dirname(save_path), exist_ok=True)
+            torch.save(merging_weights.detach().cpu(), save_path)
+
+    def run(self, modelpool: ModelPool, **kwargs):
+        """
+        Run the Layer-Wise AdaMerging Algorithm.
+
+        This method constructs the wrapped model and performs test-time adaptation if necessary.
+
+        Args:
+            modelpool (ModelPool): The model pool containing the pretrained and fine-tuned models.
+
+        Returns:
+            LayerWiseMergedModel: The merged model after test-time adaptation.
+        """
+        log.info("Fusing models using layer-wise adaptive merging.")
+        self.modelpool = modelpool
+        self.log_hyperparams(self.config)
+
+        with self.profile("construct the wrapped model"):
+            module = self.construct_layer_wise_merged_model(modelpool)
+
+        if self.config.weights is not None:
+            # skip the test-time adaptation
+            return module.merge_and_unload()
+        else:
+            with self.profile("test-time adaptation"):
+                module = self.test_time_adaptation(module)
+            if self.config.get("save_merging_weights", False):
+                self.save_merging_weights(
+                    self.config.save_merging_weights, module.merge_weight
+                )
+            return module.merge_and_unload()
+
+    def on_test_time_adaptation_start(self):
+        """
+        Something to do before the test-time adaptation starts. Such as setting up the task-specific heads.
+        """
+        pass
+
+    @abstractmethod
+    def get_shuffled_test_loader_iter(self, task: str) -> DataLoader:
+        """
+        Loader of test dataset for test-time adaptation. labels are not needed.
+
+        Args:
+            task (str): The name of the task.
+
+        Returns:
+            DataLoader: The data loader for the test dataset.
+        """
+        pass
+
+    @abstractmethod
+    def compute_logits(self, module, images: Tensor, task: str) -> Tensor:
+        """
+        Compute the logits for the given images and task.
+
+        Args:
+            module: The model module.
+            images (Tensor): The input images.
+            task (str): The name of the task.
+
+        Returns:
+            Tensor: The computed logits.
+        """
+        pass
+
+    def test_time_adaptation(self, module: "LayerWiseMergedModel[TorchModelType]"):
+        """
+        Perform test-time adaptation on the merged model.
+
+        This method adapts the merging weights during test-time to improve performance.
+
+        Args:
+            module (LayerWiseMergedModel): The merged model.
+
+        Returns:
+            LayerWiseMergedModel: The adapted merged model.
+        """
+        self.on_test_time_adaptation_start()
+
+        # configure optimizer
+        if self.config.optimizer == "adam":
+            optimizer = torch.optim.Adam([module.merge_weight], lr=self.config.lr)
+            print(f"{optimizer=}")
+            module, optimizer = self.fabric.setup(module, optimizer)
+        else:
+            raise ValueError(f"Unsupported optimizer: {self.config.optimizer}")
+
+        module.train()
+        module.merge_weights()
+        for step_idx in (
+            pbar := tqdm(
+                range(self.config.max_steps if not self.is_debug_mode else 1),
+                ("[DEBUG MODE] " if self.is_debug_mode else "")
+                + "AdaMerging Test-time adaptation",
+                dynamic_ncols=True,
+            )
+        ):
+            # default behavior for first-order optimizers
+            for task in self.modelpool.model_names:
+                with self.profile("data loading"):
+                    batch = next(self.get_shuffled_test_loader_iter(task))
+                with self.profile("forward pass"):
+                    logits = self.compute_logits(module, batch[0], task)
+                    loss = entropy_loss(logits)
+                with self.profile("backward pass"):
+                    self.fabric.backward(loss, retain_graph=True)
+
+            with self.profile("optimizer step"):
+                optimizer.step()
+                optimizer.zero_grad()
+            with self.profile("merging weights"):
+                module.merge_weights()
+
+            metrics = {
+                "train/loss": loss.item(),
+                "train/weight_max": module.merge_weight.max().item(),
+                "train/weight_min": module.merge_weight.min().item(),
+                "train/weight_mean": module.merge_weight.mean().item(),
+            }
+            self.fabric.log_dict(metrics, step=step_idx)
+            pbar.set_postfix(metrics)
+
+        log.info(get_memory_usage(f"after adamerging, the memory usage of GPU is:"))
+        self.print_profile_summary()
+        return module

{fusion_bench-0.2.10 → fusion_bench-0.2.11}/fusion_bench/method/__init__.py

@@ -53,6 +53,7 @@ _import_structure = {
         "PWEMoExactParetoOptimalForCLIP",
     ],
     "ada_svd": ["AdaSVDMergingForCLIPVisionModel"],
+    "DOGE_TA": ["DOGE_TA_Algorithm"],
     "task_singular_vector": ["TaskSingularVectorMerging"],
     "isotropic_merging": [
         "ISO_C_Merge",  # alias
@@ -66,6 +67,10 @@ _import_structure = {
         "ConcreteTaskArithmeticAlgorithmForCLIP",
         "ConcreteTaskWiseAdaMergingForCLIP",
         "ConcreteLayerWiseAdaMergingForCLIP",
+        "ConcreteSafeLayerWiseAdaMergingForCLIP",
+        "ConcreteSafeTaskWiseAdaMergingForCLIP",
+        "PostDefenseAWMAlgorithmForCLIP",
+        "PostDefenseSAUAlgorithmForCLIP",
     ],
     # model mixing methods
     "depth_upscaling": ["DepthUpscalingAlgorithm", "DepthUpscalingForLlama"],
@@ -113,12 +118,17 @@ if TYPE_CHECKING:
     )
     from .concrete_subspace import (
         ConcreteLayerWiseAdaMergingForCLIP,
+        ConcreteSafeLayerWiseAdaMergingForCLIP,
+        ConcreteSafeTaskWiseAdaMergingForCLIP,
         ConcreteTaskArithmeticAlgorithmForCLIP,
         ConcreteTaskWiseAdaMergingForCLIP,
+        PostDefenseAWMAlgorithmForCLIP,
+        PostDefenseSAUAlgorithmForCLIP,
     )
     from .dare import DareSimpleAverage, DareTaskArithmetic, DareTiesMerging
     from .dawe import DataAdaptiveWeightEnsemblingForCLIP
     from .depth_upscaling import DepthUpscalingAlgorithm, DepthUpscalingForLlama
+    from .DOGE_TA import DOGE_TA_Algorithm
     from .dummy import DummyAlgorithm
     from .ensemble import (
         MaxModelPredictorAlgorithm,