Add files using upload-large-folder tool

MaskTokenizer/projects/vlm/vq_sam2/configs/a100_qwen25vl/qwen25vl-7b_vqsam2_refseg.py

@@ -0,0 +1,254 @@
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from peft import LoraConfig
+
+from xtuner.dataset import ConcatDataset
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.runner import TrainLoop
+
+import torch
+from transformers import (
+    Qwen2_5_VLForConditionalGeneration,
+    AutoTokenizer,
+    AutoProcessor,
+)
+
+from projects.transformers.vq_sam2 import VQ_SAM2, VQ_SAM2Config, SAM2Config
+from projects.vlm.vq_sam2.models import QWEN25VL_VQSAM2Model, DirectResize
+
+from projects.vlm.vq_sam2.datasets_qwen25vl import LazySupervisedDataset, qwen25vl_vqsam2_collate_fn
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+qwen25vl_path = "./work_dirs/qwen25vl_7b_mask_loss_refseg_only_v2/hf_ckpt"
+vqsam2_pretrained_pth = "pretrained_weights/iter_17923_resampled_256x4.pth"
+sam2_path = "pretrained_weights/sam2.1_hiera_large.pt"
+pretrained_pth = None
+
+work_dir = "work_dirs/qwen25vl_7b_mask_loss_refseg_only_v3"
+
+# Scheduler & Optimizer
+batch_size = 4  # per_device
+accumulative_counts = 2
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+lr = 4e-5
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.05
+
+# Save
+save_steps = 1000
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+model_max_length=8192
+
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+sam2_config = dict(
+    type=SAM2Config,
+    cfg_path="sam2.1_hiera_l.yaml",
+    ckpt_path=sam2_path,
+)
+
+vq_sam2_config = dict(
+    type=VQ_SAM2Config,
+    sam2_config=sam2_config,
+    codebook_size=256,
+    codebook_depth=4,
+    shared_codebook=False,
+    latent_dim=256,
+    loss_sample_points=True,
+    vq_loss_weight=1.0,
+)
+
+model = dict(
+    type=QWEN25VL_VQSAM2Model,
+    qwen25vl_hf_model=dict(
+        type=Qwen2_5_VLForConditionalGeneration.from_pretrained,
+        pretrained_model_name_or_path=qwen25vl_path,
+        attn_implementation="flash_attention_2",
+        torch_dtype=torch.bfloat16,
+        trust_remote_code=True,
+    ),
+    vqsam2_hf_model=dict(
+        type=VQ_SAM2,
+        config=vq_sam2_config,
+    ),
+    tokenizer=dict(
+        type=AutoTokenizer.from_pretrained,
+        pretrained_model_name_or_path=qwen25vl_path,
+        cache_dir="./cache",
+        model_max_length=model_max_length,
+        padding_side="right",
+        use_fast=False,
+    ),
+    preprocessor=dict(
+        type=AutoProcessor.from_pretrained,
+        pretrained_model_name_or_path=qwen25vl_path,
+    ),
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        bias='none',
+        task_type='CAUSAL_LM'
+    ),
+    vqsam2_pretrained_weights=vqsam2_pretrained_pth,
+    pretrained_pth=pretrained_pth,
+    freeze_sam2_decoder=True,
+    repa_loss=False,
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+
+sam2_image_processor = dict(
+    type=DirectResize,
+    target_length=1024,
+)
+
+standard_dataset = dict(
+    type=LazySupervisedDataset,
+    tokenizer=dict(
+        type=AutoTokenizer.from_pretrained,
+        pretrained_model_name_or_path=qwen25vl_path,
+        cache_dir="./cache",
+        model_max_length=model_max_length,
+        padding_side="right",
+        use_fast=False,
+    ),
+    data_args=dict(
+        model_type="qwen2.5vl",
+        dataset_use="mask_generation_refseg%500",
+        max_pixels=2048*28*28,
+        min_pixels=4*28*28,
+        video_max_total_pixels=576*28*28,
+        video_min_total_pixels=144*28*28,
+        image_processor=dict(
+            type=AutoProcessor.from_pretrained,
+            pretrained_model_name_or_path=qwen25vl_path,
+        ),
+    ),
+    sam_preprocessor=sam2_image_processor,
+)
+
+
+train_dataset = dict(
+    type=ConcatDataset, datasets=[
+        standard_dataset,
+    ]
+)
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=qwen25vl_vqsam2_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),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='bfloat16'
+)
+
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    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),
+]
+
+# 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,
+        save_optimizer=False,
+        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'),
+)
+
+# set visualizer
+# visualizer = None
+from mmengine.visualization import Visualizer, TensorboardVisBackend
+visualizer = dict(type=Visualizer, vis_backends=[dict(type=TensorboardVisBackend)])
+
+# 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)

MaskTokenizer/projects/vlm/vq_sam2/configs/a100_qwen25vl/qwen25vl_vqsam2_refseg_debug.py

@@ -0,0 +1,252 @@
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from peft import LoraConfig
+
+from xtuner.dataset import ConcatDataset
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.runner import TrainLoop
+
+import torch
+from transformers import (
+    Qwen2_5_VLForConditionalGeneration,
+    AutoTokenizer,
+    AutoProcessor,
+)
+
+from projects.transformers.vq_sam2 import VQ_SAM2, VQ_SAM2Config, SAM2Config
+from projects.vlm.vq_sam2.models import QWEN25VL_VQSAM2Model, DirectResize
+
+from projects.vlm.vq_sam2.datasets_qwen25vl import LazySupervisedDataset, qwen25vl_vqsam2_collate_fn
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+qwen25vl_path = "pretrained_weights/qwen2_5_vl_vq_sam2_3b_256x4"
+vqsam2_pretrained_pth = "pretrained_weights/iter_17923_resampled_256x4.pth"
+sam2_path = "pretrained_weights/sam2.1_hiera_large.pt"
+pretrained_pth = "./work_dirs/qwen25vl_3b_mask_loss_refseg_only_v4/iter_10000.pth"
+
+work_dir = "work_dirs/qwen25vl_3b_mask_loss_refseg_only_v4_debug"
+
+# Scheduler & Optimizer
+batch_size = 4  # per_device
+accumulative_counts = 2
+dataloader_num_workers = 4
+max_epochs = 1
+optim_type = AdamW
+lr = 4e-5
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.05
+
+# Save
+save_steps = 1000
+save_total_limit = -1  # Maximum checkpoints to keep (-1 means unlimited)
+
+model_max_length=8192
+
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+sam2_config = dict(
+    type=SAM2Config,
+    cfg_path="sam2.1_hiera_l.yaml",
+    ckpt_path=sam2_path,
+)
+
+vq_sam2_config = dict(
+    type=VQ_SAM2Config,
+    sam2_config=sam2_config,
+    codebook_size=256,
+    codebook_depth=4,
+    shared_codebook=False,
+    latent_dim=256,
+    loss_sample_points=True,
+    vq_loss_weight=1.0,
+)
+
+model = dict(
+    type=QWEN25VL_VQSAM2Model,
+    qwen25vl_hf_model=dict(
+        type=Qwen2_5_VLForConditionalGeneration.from_pretrained,
+        pretrained_model_name_or_path=qwen25vl_path,
+        attn_implementation="flash_attention_2",
+        torch_dtype=torch.bfloat16,
+        trust_remote_code=True,
+    ),
+    vqsam2_hf_model=dict(
+        type=VQ_SAM2,
+        config=vq_sam2_config,
+    ),
+    tokenizer=dict(
+        type=AutoTokenizer.from_pretrained,
+        pretrained_model_name_or_path=qwen25vl_path,
+        cache_dir="./cache",
+        model_max_length=model_max_length,
+        padding_side="right",
+        use_fast=False,
+    ),
+    preprocessor=dict(
+        type=AutoProcessor.from_pretrained,
+        pretrained_model_name_or_path=qwen25vl_path,
+    ),
+    llm_lora=dict(
+        type=LoraConfig,
+        r=128,
+        lora_alpha=256,
+        lora_dropout=0.05,
+        bias='none',
+        task_type='CAUSAL_LM'
+    ),
+    vqsam2_pretrained_weights=vqsam2_pretrained_pth,
+    pretrained_pth=pretrained_pth,
+    freeze_sam2_decoder=True,
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+
+sam2_image_processor = dict(
+    type=DirectResize,
+    target_length=1024,
+)
+
+standard_dataset = dict(
+    type=LazySupervisedDataset,
+    tokenizer=dict(
+        type=AutoTokenizer.from_pretrained,
+        pretrained_model_name_or_path=qwen25vl_path,
+        cache_dir="./cache",
+        model_max_length=model_max_length,
+        padding_side="right",
+        use_fast=False,
+    ),
+    data_args=dict(
+        model_type="qwen2.5vl",
+        dataset_use="mask_generation_visual7w%500",
+        max_pixels=2048*28*28,
+        min_pixels=4*28*28,
+        video_max_total_pixels=576*28*28,
+        video_min_total_pixels=144*28*28,
+        image_processor=dict(
+            type=AutoProcessor.from_pretrained,
+            pretrained_model_name_or_path=qwen25vl_path,
+        ),
+    ),
+    sam_preprocessor=sam2_image_processor,
+)
+
+
+train_dataset = dict(
+    type=ConcatDataset, datasets=[
+        standard_dataset,
+    ]
+)
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=qwen25vl_vqsam2_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),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='bfloat16'
+)
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    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),
+]
+
+# 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,
+        save_optimizer=False,
+        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'),
+)
+
+# set visualizer
+# visualizer = None
+from mmengine.visualization import Visualizer, TensorboardVisBackend
+visualizer = dict(type=Visualizer, vis_backends=[dict(type=TensorboardVisBackend)])
+
+# 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)

MaskTokenizer/projects/vlm/vq_sam2/configs/a100_train/vq_sam2_codebookx4depthx1024sizex256dimxunsharex1MT_datasetxsa1bxcoconutx1.py

@@ -0,0 +1,212 @@
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+
+from xtuner.dataset import ConcatDataset
+from xtuner.dataset.samplers import LengthGroupedSampler
+from xtuner.engine.runner import TrainLoop
+
+from projects.transformers.vq_sam2 import VQ_SAM2, VQ_SAM2Config, SAM2Config
+from projects.vlm.vq_sam2.models import VQ_SAM2Model, DirectResize
+
+from projects.vlm.vq_sam2.datasets import CoCoPanoSegDataset, SA1BDataset, COCONUTDataset
+from projects.vlm.vq_sam2.datasets import vq_sam2_collate_fn
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+sam2_path = "pretrained_weights/sam2.1_hiera_large.pt"
+pretrained_pth = None
+
+# Scheduler & Optimizer
+batch_size = 16  # per_device
+accumulative_counts = 1
+dataloader_num_workers = 16
+max_epochs = 1
+optim_type = AdamW
+lr = 2e-5
+betas = (0.9, 0.999)
+weight_decay = 0.05
+max_norm = 1  # grad clip
+warmup_ratio = 0.05
+
+# Save
+save_steps = 3000
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+sam2_config = dict(
+    type=SAM2Config,
+    cfg_path="sam2.1_hiera_l.yaml",
+    ckpt_path=sam2_path,
+)
+
+vq_sam2_config = dict(
+    type=VQ_SAM2Config,
+    sam2_config=sam2_config,
+    codebook_size=1024,
+    codebook_depth=4,
+    shared_codebook=False,
+    latent_dim=256,
+    loss_sample_points=True,
+    vq_loss_weight=1.0,
+)
+
+model = dict(
+    type=VQ_SAM2Model,
+    hf_model=dict(
+        type=VQ_SAM2,
+        config=vq_sam2_config,
+    ),
+    sam2_pretrained_weights=sam2_path,
+    pretrained_pth=pretrained_pth,
+    freeze_sam2_decoder=False,
+)
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+
+sam2_image_processor = dict(
+    type=DirectResize,
+    target_length=1024,
+)
+
+DATA_ROOT = ''
+sam_info_json = "./data/sam_info.json"
+coconut_info_json = "./data/coconut_segments.json"
+
+sa1b_dataset = dict(
+    type=SA1BDataset,
+    image_folder=DATA_ROOT,
+    preprocessor=sam2_image_processor,
+    multi_targets=False,
+    repeats=1.0,
+    fast_load=True,
+    sam_info_json=sam_info_json,
+    scan_record_folder='./left_sa1b_indices/vq_sam2_codebookx4depthx1024sizex256dimxunsharex1MT_datasetxsa1bxcoconutx1/'
+)
+
+coconut_dataset = dict(
+    type=COCONUTDataset,
+    image_folder=DATA_ROOT,
+    preprocessor=sam2_image_processor,
+    repeats=1.0,
+    coconut_info_json=coconut_info_json,
+)
+
+
+
+
+train_dataset = dict(
+    type=ConcatDataset, datasets=[
+        sa1b_dataset, coconut_dataset,
+    ]
+)
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=train_dataset,
+    sampler=dict(
+        type=LengthGroupedSampler,
+        length_property='modality_length',
+        per_device_batch_size=batch_size * accumulative_counts),
+    collate_fn=dict(type=vq_sam2_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),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float32'
+)
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    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),
+]
+
+# 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,
+        save_optimizer=False,
+        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'),
+)
+
+# set visualizer
+# visualizer = None
+from mmengine.visualization import Visualizer, TensorboardVisBackend
+visualizer = dict(type=Visualizer, vis_backends=[dict(type=TensorboardVisBackend)])
+
+# 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)