Upload folder stage1_qwen25_token_merging to stage_1/token_merging/stage1_qwen25_token_merging

stage_1/token_merging/stage1_qwen25_token_merging/projector/projector.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e04644f58e7f4ccdd63fd5a07319846667c81c5b8071b7709f5b804774c87840
+size 40384848

stage_1/token_merging/stage1_qwen25_token_merging/token_merger/merger.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08d3c4692384bad2a12168babda110018e9441148a8b2db397bb84cd8255f631
+size 20784

stage_1/token_merging/stage1_qwen25_token_merging/xtuner_config.py

@@ -0,0 +1,250 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.visualization import Visualizer, WandbVisBackend
+from mmengine.optim import AmpOptimWrapper, ConstantLR, LinearLR, CosineAnnealingLR
+from torch.optim import AdamW
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig, CLIPImageProcessor,
+                          CLIPVisionModel)
+ 
+from xtuner.dataset import LLaVADataset
+from xtuner.dataset.collate_fns import default_collate_fn
+from xtuner.dataset.map_fns import llava_map_fn, template_map_fn_factory
+from xtuner.engine.hooks import DatasetInfoHook, EvaluateChatHook, HFCheckpointHook, ThroughputHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model.llava_no_longnet import LLaVAModel
+from xtuner.utils import PROMPT_TEMPLATE
+# from xtuner.model.torchscale.model.create_longnet_for_training import create_longvit_model_fast as create_longnet_vit
+# from xtuner.model.torchscale.model.LongNetVit import gigapath_slide_enc3l1536d
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+llm_name_or_path = 'Qwen/Qwen2.5-7B-Instruct'
+# Data
+data_path = '/mnt/bn/yuxuanwang/meilong/code/projects/efficient_foundation_wsi_llava/merged_dataset/stage1_morph2.json'
+image_path_list = None
+
+prompt_template = PROMPT_TEMPLATE.qwen_chat
+
+# 长序列：保持 per_image_length == sample_num
+max_length = 15836
+per_image_length = 10240
+sample_type = 'wsi'  # 'wsi' or 'image'
+
+# Scheduler & Optimizer (epoch-based)
+batch_size = 1
+accumulative_counts = 400 # 5 * 400 =  2000
+dataloader_num_workers = 5
+seed = 2025
+optim_type = AdamW
+lr = 1e-3
+betas = (0.9, 0.999)
+weight_decay = 0.0            # 适度WD抑制漂移
+max_norm = 1                   # 更紧的梯度裁剪
+
+# 以 epoch 为主
+max_epochs = 2
+warmup_ratio = 0.05            # 预热占比（相对 max_iters）
+
+# Save
+save_steps = 5120
+save_total_limit = 8 # Maximum checkpoints to keep (-1 means unlimited)
+
+# Evaluate the generation performance during the training
+evaluation_freq = 512
+SYSTEM = ''
+evaluation_images = '/mnt/bn/xudong-va/meilong/datasets/Token_Compression/skcm_224x224_b20_t15/h5_files/TCGA-EB-A5UN-06Z-00-DX1.h5'
+evaluation_inputs = ['Are the tumor cells organized in a lobulated pattern within the slide?']
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=llm_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+bnb = dict(
+    type=BitsAndBytesConfig,
+    load_in_4bit=True,
+    load_in_8bit=False,
+    llm_int8_threshold=6.0,
+    llm_int8_has_fp16_weight=False,
+    bnb_4bit_compute_dtype=torch.bfloat16,
+    bnb_4bit_use_double_quant=True,
+    bnb_4bit_quant_type="nf4",
+)
+
+model = dict(
+    type=LLaVAModel,
+    freeze_llm=True,
+    train_stage='1',
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=llm_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.bfloat16,
+        attn_implementation='flash_attention_2',
+        quantization_config=bnb
+    ),
+
+    max_position_embeddings = None, # original 32000 + 
+    enable_token_merge = True,
+    # 建议：前期用 resampler 更稳（也更省显存）；如不需要可改回 False
+    use_perceiver_resampler=False,
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+llava_dataset = dict(
+    type=LLaVADataset,
+    data_path=data_path,
+    image_folder='',
+    image_path_list=image_path_list,
+    tokenizer=tokenizer,
+    dataset_map_fn=llava_map_fn,
+    template_map_fn=dict(type=template_map_fn_factory, template=prompt_template),
+    max_length=max_length,
+    per_image_length=per_image_length,
+    pad_image_to_square=False,
+    sample_num=per_image_length,        
+    image_feature_prefix='/mnt/bn/xudong-va/meilong/datasets/Token_Compression',
+    image_feature_suffix='.h5',
+    identifier='_224x224_b20_t15',
+    unwanted_prefix_csv='/mnt/bn/yuxuanwang/meilong/code/projects/efficient_foundation_wsi_llava/merged_dataset/missing_slides3.csv',
+    sample_strategy='linspace', #use linspace
+)
+
+
+# cying: add: per_image_length=per_image_length,
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    pin_memory=True,
+    persistent_workers=True, 
+    prefetch_factor=4,
+    dataset=llava_dataset,
+    sampler=dict(type=DefaultSampler, shuffle=True),
+    collate_fn=dict(type=default_collate_fn)
+)
+
+
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    paramwise_cfg = dict(
+    norm_decay_mult=0.0,
+    bias_decay_mult=0.0,
+        
+        ),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=True),  # 遇到 NaN 立刻报错
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='bfloat16',
+)
+
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=0.01,         # 从 1% 的 lr 慢启动
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True  # 按 iter 计算
+    ),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True
+    )
+]
+
+
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    dict(
+        type=EvaluateChatHook,
+        tokenizer=tokenizer,
+        every_n_iters=evaluation_freq,
+        evaluation_inputs=evaluation_inputs,
+        evaluation_images=evaluation_images,
+        system=SYSTEM,
+        prompt_template=prompt_template),
+    dict(type = ThroughputHook)
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+
+visualizer = dict(
+   type=Visualizer,
+   vis_backends=[
+       dict(type=WandbVisBackend, init_kwargs=dict(project='stage1_no_longnet_with_language_guide2'))])
+visualizer = None
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+ 
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(
+    by_epoch=False,
+    window_size=1,
+    mean_pattern=r".*(loss|time|data_time|grad_norm|tflops).*",
+)