HongChuang/ultralytics-ascend
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 3

ultralytics 8.2.73 Meta SAM2 Refactor (#14867)

Browse source 
Signed-off-by: Glenn Jocher <glenn.jocher@ultralytics.com>
Co-authored-by: UltralyticsAssistant <web@ultralytics.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>
This commit is contained in:
Laughing 2024-08-05 08:53:45 +08:00 • committed by GitHub
parent bea4c93278
commit 5d9046abda
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
44 changed files with 4542 additions and 3624 deletions
Download patch file Download diff file Expand all files Collapse all files

209
ultralytics/models/sam/build.py
View file

@ -13,14 +13,15 @@ import torch
from ultralytics.utils.downloads import attempt_download_asset
from .modules.decoders import MaskDecoder
from .modules.encoders import ImageEncoderViT, PromptEncoder
from .modules.sam import SAMModel
from .modules.encoders import FpnNeck, Hiera, ImageEncoder, ImageEncoderViT, MemoryEncoder, PromptEncoder
from .modules.memory_attention import MemoryAttention, MemoryAttentionLayer
from .modules.sam import SAM2Model, SAMModel
from .modules.tiny_encoder import TinyViT
from .modules.transformer import TwoWayTransformer
def build_sam_vit_h(checkpoint=None):
"""Build and return a Segment Anything Model (SAM) h-size model."""
"""Builds and returns a Segment Anything Model (SAM) h-size model with specified encoder parameters."""
return _build_sam(
encoder_embed_dim=1280,
encoder_depth=32,
@ -31,7 +32,7 @@ def build_sam_vit_h(checkpoint=None):
def build_sam_vit_l(checkpoint=None):
"""Build and return a Segment Anything Model (SAM) l-size model."""
"""Builds and returns a Segment Anything Model (SAM) l-size model with specified encoder parameters."""
return _build_sam(
encoder_embed_dim=1024,
encoder_depth=24,
@ -42,7 +43,7 @@ def build_sam_vit_l(checkpoint=None):
def build_sam_vit_b(checkpoint=None):
"""Build and return a Segment Anything Model (SAM) b-size model."""
"""Constructs and returns a Segment Anything Model (SAM) with b-size architecture and optional checkpoint."""
return _build_sam(
encoder_embed_dim=768,
encoder_depth=12,
@ -53,7 +54,7 @@ def build_sam_vit_b(checkpoint=None):
def build_mobile_sam(checkpoint=None):
"""Build and return Mobile Segment Anything Model (Mobile-SAM)."""
"""Builds and returns a Mobile Segment Anything Model (Mobile-SAM) for efficient image segmentation."""
return _build_sam(
encoder_embed_dim=[64, 128, 160, 320],
encoder_depth=[2, 2, 6, 2],
@ -64,10 +65,85 @@ def build_mobile_sam(checkpoint=None):
)
def build_sam2_t(checkpoint=None):
"""Builds and returns a Segment Anything Model 2 (SAM2) tiny-size model with specified architecture parameters."""
return _build_sam2(
encoder_embed_dim=96,
encoder_stages=[1, 2, 7, 2],
encoder_num_heads=1,
encoder_global_att_blocks=[5, 7, 9],
encoder_window_spec=[8, 4, 14, 7],
encoder_backbone_channel_list=[768, 384, 192, 96],
checkpoint=checkpoint,
)
def build_sam2_s(checkpoint=None):
"""Builds and returns a small-size Segment Anything Model (SAM2) with specified architecture parameters."""
return _build_sam2(
encoder_embed_dim=96,
encoder_stages=[1, 2, 11, 2],
encoder_num_heads=1,
encoder_global_att_blocks=[7, 10, 13],
encoder_window_spec=[8, 4, 14, 7],
encoder_backbone_channel_list=[768, 384, 192, 96],
checkpoint=checkpoint,
)
def build_sam2_b(checkpoint=None):
"""Builds and returns a SAM2 base-size model with specified architecture parameters."""
return _build_sam2(
encoder_embed_dim=112,
encoder_stages=[2, 3, 16, 3],
encoder_num_heads=2,
encoder_global_att_blocks=[12, 16, 20],
encoder_window_spec=[8, 4, 14, 7],
encoder_window_spatial_size=[14, 14],
encoder_backbone_channel_list=[896, 448, 224, 112],
checkpoint=checkpoint,
)
def build_sam2_l(checkpoint=None):
"""Builds and returns a large-size Segment Anything Model (SAM2) with specified architecture parameters."""
return _build_sam2(
encoder_embed_dim=144,
encoder_stages=[2, 6, 36, 4],
encoder_num_heads=2,
encoder_global_att_blocks=[23, 33, 43],
encoder_window_spec=[8, 4, 16, 8],
encoder_backbone_channel_list=[1152, 576, 288, 144],
checkpoint=checkpoint,
)
def _build_sam(
encoder_embed_dim, encoder_depth, encoder_num_heads, encoder_global_attn_indexes, checkpoint=None, mobile_sam=False
encoder_embed_dim,
encoder_depth,
encoder_num_heads,
encoder_global_attn_indexes,
checkpoint=None,
mobile_sam=False,
):
"""Builds the selected SAM model architecture."""
"""
Builds a Segment Anything Model (SAM) with specified encoder parameters.
Args:
encoder_embed_dim (int | List[int]): Embedding dimension for the encoder.
encoder_depth (int | List[int]): Depth of the encoder.
encoder_num_heads (int | List[int]): Number of attention heads in the encoder.
encoder_global_attn_indexes (List[int] | None): Indexes for global attention in the encoder.
checkpoint (str | None): Path to the model checkpoint file.
mobile_sam (bool): Whether to build a Mobile-SAM model.
Returns:
(SAMModel): A Segment Anything Model instance with the specified architecture.
Examples:
>>> sam = _build_sam(768, 12, 12, [2, 5, 8, 11])
>>> sam = _build_sam([64, 128, 160, 320], [2, 2, 6, 2], [2, 4, 5, 10], None, mobile_sam=True)
"""
prompt_embed_dim = 256
image_size = 1024
vit_patch_size = 16
@ -139,16 +215,131 @@ def _build_sam(
return sam
def _build_sam2(
encoder_embed_dim=1280,
encoder_stages=[2, 6, 36, 4],
encoder_num_heads=2,
encoder_global_att_blocks=[7, 15, 23, 31],
encoder_backbone_channel_list=[1152, 576, 288, 144],
encoder_window_spatial_size=[7, 7],
encoder_window_spec=[8, 4, 16, 8],
checkpoint=None,
):
"""
Builds and returns a Segment Anything Model 2 (SAM2) with specified architecture parameters.
Args:
encoder_embed_dim (int): Embedding dimension for the encoder.
encoder_stages (List[int]): Number of blocks in each stage of the encoder.
encoder_num_heads (int): Number of attention heads in the encoder.
encoder_global_att_blocks (List[int]): Indices of global attention blocks in the encoder.
encoder_backbone_channel_list (List[int]): Channel dimensions for each level of the encoder backbone.
encoder_window_spatial_size (List[int]): Spatial size of the window for position embeddings.
encoder_window_spec (List[int]): Window specifications for each stage of the encoder.
checkpoint (str | None): Path to the checkpoint file for loading pre-trained weights.
Returns:
(SAM2Model): A configured and initialized SAM2 model.
Examples:
>>> sam2_model = _build_sam2(encoder_embed_dim=96, encoder_stages=[1, 2, 7, 2])
>>> sam2_model.eval()
"""
image_encoder = ImageEncoder(
trunk=Hiera(
embed_dim=encoder_embed_dim,
num_heads=encoder_num_heads,
stages=encoder_stages,
global_att_blocks=encoder_global_att_blocks,
window_pos_embed_bkg_spatial_size=encoder_window_spatial_size,
window_spec=encoder_window_spec,
),
neck=FpnNeck(
d_model=256,
backbone_channel_list=encoder_backbone_channel_list,
fpn_top_down_levels=[2, 3],
fpn_interp_model="nearest",
),
scalp=1,
)
memory_attention = MemoryAttention(d_model=256, pos_enc_at_input=True, num_layers=4, layer=MemoryAttentionLayer())
memory_encoder = MemoryEncoder(out_dim=64)
sam2 = SAM2Model(
image_encoder=image_encoder,
memory_attention=memory_attention,
memory_encoder=memory_encoder,
num_maskmem=7,
image_size=1024,
sigmoid_scale_for_mem_enc=20.0,
sigmoid_bias_for_mem_enc=-10.0,
use_mask_input_as_output_without_sam=True,
directly_add_no_mem_embed=True,
use_high_res_features_in_sam=True,
multimask_output_in_sam=True,
iou_prediction_use_sigmoid=True,
use_obj_ptrs_in_encoder=True,
add_tpos_enc_to_obj_ptrs=True,
only_obj_ptrs_in_the_past_for_eval=True,
pred_obj_scores=True,
pred_obj_scores_mlp=True,
fixed_no_obj_ptr=True,
multimask_output_for_tracking=True,
use_multimask_token_for_obj_ptr=True,
multimask_min_pt_num=0,
multimask_max_pt_num=1,
use_mlp_for_obj_ptr_proj=True,
compile_image_encoder=False,
sam_mask_decoder_extra_args=dict(
dynamic_multimask_via_stability=True,
dynamic_multimask_stability_delta=0.05,
dynamic_multimask_stability_thresh=0.98,
),
)
if checkpoint is not None:
checkpoint = attempt_download_asset(checkpoint)
with open(checkpoint, "rb") as f:
state_dict = torch.load(f)["model"]
sam2.load_state_dict(state_dict)
sam2.eval()
return sam2
sam_model_map = {
"sam_h.pt": build_sam_vit_h,
"sam_l.pt": build_sam_vit_l,
"sam_b.pt": build_sam_vit_b,
"mobile_sam.pt": build_mobile_sam,
"sam2_t.pt": build_sam2_t,
"sam2_s.pt": build_sam2_s,
"sam2_b.pt": build_sam2_b,
"sam2_l.pt": build_sam2_l,
}
def build_sam(ckpt="sam_b.pt"):
"""Build a SAM model specified by ckpt."""
"""
Builds and returns a Segment Anything Model (SAM) based on the provided checkpoint.
Args:
ckpt (str | Path): Path to the checkpoint file or name of a pre-defined SAM model.
Returns:
(SAMModel | SAM2Model): A configured and initialized SAM or SAM2 model instance.
Raises:
FileNotFoundError: If the provided checkpoint is not a supported SAM model.
Examples:
>>> sam_model = build_sam("sam_b.pt")
>>> sam_model = build_sam("path/to/custom_checkpoint.pt")
Notes:
Supported pre-defined models include:
- SAM: 'sam_h.pt', 'sam_l.pt', 'sam_b.pt', 'mobile_sam.pt'
- SAM2: 'sam2_t.pt', 'sam2_s.pt', 'sam2_b.pt', 'sam2_l.pt'
"""
model_builder = None
ckpt = str(ckpt) # to allow Path ckpt types
for k in sam_model_map.keys():