File size: 5,950 Bytes
8bacc03 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch
from torch.nn import functional as F
from icecream import ic
from functools import partial
from .modeling import ImageEncoderViT, MaskDecoder, PromptEncoder, Sam, TwoWayTransformer
def build_sam_vit_h(image_size, num_classes, pixel_mean=[123.675, 116.28, 103.53], pixel_std=[58.395, 57.12, 57.375],
checkpoint=None):
return _build_sam(
encoder_embed_dim=1280,
encoder_depth=32,
encoder_num_heads=16,
encoder_global_attn_indexes=[7, 15, 23, 31],
checkpoint=checkpoint,
num_classes=num_classes,
image_size=image_size,
pixel_mean=pixel_mean,
pixel_std=pixel_std
)
build_sam = build_sam_vit_h
def build_sam_vit_l(image_size, num_classes, pixel_mean=[123.675, 116.28, 103.53], pixel_std=[58.395, 57.12, 57.375],
checkpoint=None):
return _build_sam(
encoder_embed_dim=1024,
encoder_depth=24,
encoder_num_heads=16,
encoder_global_attn_indexes=[5, 11, 17, 23],
checkpoint=checkpoint,
num_classes=num_classes,
image_size=image_size,
pixel_mean=pixel_mean,
pixel_std=pixel_std
)
def build_sam_vit_b(image_size, num_classes, pixel_mean=[123.675, 116.28, 103.53], pixel_std=[58.395, 57.12, 57.375],
checkpoint=None):
return _build_sam(
encoder_embed_dim=768,
encoder_depth=12,
encoder_num_heads=12,
encoder_global_attn_indexes=[2, 5, 8, 11],
# adopt global attention at [3, 6, 9, 12] transform layer, else window attention layer
checkpoint=checkpoint,
num_classes=num_classes,
image_size=image_size,
pixel_mean=pixel_mean,
pixel_std=pixel_std
)
sam_model_registry = {
"default": build_sam_vit_h,
"vit_h": build_sam_vit_h,
"vit_l": build_sam_vit_l,
"vit_b": build_sam_vit_b,
}
def _build_sam(
encoder_embed_dim,
encoder_depth,
encoder_num_heads,
encoder_global_attn_indexes,
num_classes,
image_size,
pixel_mean,
pixel_std,
checkpoint=None,
):
prompt_embed_dim = 256
image_size = image_size
vit_patch_size = 16
image_embedding_size = image_size // vit_patch_size # Divide by 16 here
sam = Sam(
image_encoder=ImageEncoderViT(
depth=encoder_depth,
embed_dim=encoder_embed_dim,
img_size=image_size,
mlp_ratio=4,
norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
num_heads=encoder_num_heads,
patch_size=vit_patch_size,
qkv_bias=True,
use_rel_pos=True,
global_attn_indexes=encoder_global_attn_indexes,
window_size=14,
out_chans=prompt_embed_dim,
),
prompt_encoder=PromptEncoder(
embed_dim=prompt_embed_dim,
image_embedding_size=(image_embedding_size, image_embedding_size),
input_image_size=(image_size, image_size),
mask_in_chans=16,
),
mask_decoder=MaskDecoder(
# num_multimask_outputs=3,
num_multimask_outputs=num_classes,
transformer=TwoWayTransformer(
depth=2,
embedding_dim=prompt_embed_dim,
mlp_dim=2048,
num_heads=8,
),
transformer_dim=prompt_embed_dim,
iou_head_depth=3,
iou_head_hidden_dim=256,
),
# pixel_mean=[123.675, 116.28, 103.53],
# pixel_std=[58.395, 57.12, 57.375],
pixel_mean=pixel_mean,
pixel_std=pixel_std
)
# sam.eval()
sam.train()
if checkpoint is not None:
with open(checkpoint, "rb") as f:
state_dict = torch.load(f)
try:
sam.load_state_dict(state_dict)
except:
new_state_dict = load_from(sam, state_dict, image_size, vit_patch_size)
sam.load_state_dict(new_state_dict)
return sam, image_embedding_size
def load_from(sam, state_dict, image_size, vit_patch_size):
sam_dict = sam.state_dict()
except_keys = ['mask_tokens', 'output_hypernetworks_mlps', 'iou_prediction_head']
new_state_dict = {k: v for k, v in state_dict.items() if
k in sam_dict.keys() and except_keys[0] not in k and except_keys[1] not in k and except_keys[2] not in k}
pos_embed = new_state_dict['image_encoder.pos_embed']
token_size = int(image_size // vit_patch_size)
if pos_embed.shape[1] != token_size:
# resize pos embedding, which may sacrifice the performance, but I have no better idea
pos_embed = pos_embed.permute(0, 3, 1, 2) # [b, c, h, w]
pos_embed = F.interpolate(pos_embed, (token_size, token_size), mode='bilinear', align_corners=False)
pos_embed = pos_embed.permute(0, 2, 3, 1) # [b, h, w, c]
new_state_dict['image_encoder.pos_embed'] = pos_embed
rel_pos_keys = [k for k in sam_dict.keys() if 'rel_pos' in k]
global_rel_pos_keys = [k for k in rel_pos_keys if '2' in k or '5' in k or '8' in k or '11' in k]
for k in global_rel_pos_keys:
rel_pos_params = new_state_dict[k]
h, w = rel_pos_params.shape
rel_pos_params = rel_pos_params.unsqueeze(0).unsqueeze(0)
rel_pos_params = F.interpolate(rel_pos_params, (token_size * 2 - 1, w), mode='bilinear', align_corners=False)
new_state_dict[k] = rel_pos_params[0, 0, ...]
sam_dict.update(new_state_dict)
return sam_dict
|