Spaces:
Sleeping
Sleeping
Upload 40 files
Browse files- LICENSE +201 -0
- MANIFEST.in +1 -0
- app.py +121 -0
- metaseg/__init__.py +12 -0
- metaseg/app.py +121 -0
- metaseg/automatic_mask_generator.py +368 -0
- metaseg/build_sam.py +107 -0
- metaseg/demo.py +112 -0
- metaseg/modeling/__init__.py +11 -0
- metaseg/modeling/__pycache__/__init__.cpython-310.pyc +0 -0
- metaseg/modeling/__pycache__/common.cpython-310.pyc +0 -0
- metaseg/modeling/__pycache__/image_encoder.cpython-310.pyc +0 -0
- metaseg/modeling/__pycache__/mask_decoder.cpython-310.pyc +0 -0
- metaseg/modeling/__pycache__/prompt_encoder.cpython-310.pyc +0 -0
- metaseg/modeling/__pycache__/sam.cpython-310.pyc +0 -0
- metaseg/modeling/__pycache__/transformer.cpython-310.pyc +0 -0
- metaseg/modeling/common.py +43 -0
- metaseg/modeling/image_encoder.py +389 -0
- metaseg/modeling/mask_decoder.py +169 -0
- metaseg/modeling/prompt_encoder.py +212 -0
- metaseg/modeling/sam.py +174 -0
- metaseg/modeling/transformer.py +232 -0
- metaseg/predictor.py +264 -0
- metaseg/utils/__init__.py +5 -0
- metaseg/utils/__pycache__/__init__.cpython-310.pyc +0 -0
- metaseg/utils/__pycache__/amg.cpython-310.pyc +0 -0
- metaseg/utils/__pycache__/file.cpython-310.pyc +0 -0
- metaseg/utils/__pycache__/transforms.cpython-310.pyc +0 -0
- metaseg/utils/amg.py +330 -0
- metaseg/utils/file.py +32 -0
- metaseg/utils/onnx.py +138 -0
- metaseg/utils/transforms.py +92 -0
- pyproject.toml +6 -0
- requirements.txt +10 -0
- scripts/amg.py +233 -0
- scripts/code_format.sh +2 -0
- scripts/export_onnx_model.py +198 -0
- scripts/package.sh +2 -0
- setup.cfg +11 -0
- setup.py +58 -0
LICENSE
ADDED
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Apache License
|
2 |
+
Version 2.0, January 2004
|
3 |
+
http://www.apache.org/licenses/
|
4 |
+
|
5 |
+
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
6 |
+
|
7 |
+
1. Definitions.
|
8 |
+
|
9 |
+
"License" shall mean the terms and conditions for use, reproduction,
|
10 |
+
and distribution as defined by Sections 1 through 9 of this document.
|
11 |
+
|
12 |
+
"Licensor" shall mean the copyright owner or entity authorized by
|
13 |
+
the copyright owner that is granting the License.
|
14 |
+
|
15 |
+
"Legal Entity" shall mean the union of the acting entity and all
|
16 |
+
other entities that control, are controlled by, or are under common
|
17 |
+
control with that entity. For the purposes of this definition,
|
18 |
+
"control" means (i) the power, direct or indirect, to cause the
|
19 |
+
direction or management of such entity, whether by contract or
|
20 |
+
otherwise, or (ii) ownership of fifty percent (50%) or more of the
|
21 |
+
outstanding shares, or (iii) beneficial ownership of such entity.
|
22 |
+
|
23 |
+
"You" (or "Your") shall mean an individual or Legal Entity
|
24 |
+
exercising permissions granted by this License.
|
25 |
+
|
26 |
+
"Source" form shall mean the preferred form for making modifications,
|
27 |
+
including but not limited to software source code, documentation
|
28 |
+
source, and configuration files.
|
29 |
+
|
30 |
+
"Object" form shall mean any form resulting from mechanical
|
31 |
+
transformation or translation of a Source form, including but
|
32 |
+
not limited to compiled object code, generated documentation,
|
33 |
+
and conversions to other media types.
|
34 |
+
|
35 |
+
"Work" shall mean the work of authorship, whether in Source or
|
36 |
+
Object form, made available under the License, as indicated by a
|
37 |
+
copyright notice that is included in or attached to the work
|
38 |
+
(an example is provided in the Appendix below).
|
39 |
+
|
40 |
+
"Derivative Works" shall mean any work, whether in Source or Object
|
41 |
+
form, that is based on (or derived from) the Work and for which the
|
42 |
+
editorial revisions, annotations, elaborations, or other modifications
|
43 |
+
represent, as a whole, an original work of authorship. For the purposes
|
44 |
+
of this License, Derivative Works shall not include works that remain
|
45 |
+
separable from, or merely link (or bind by name) to the interfaces of,
|
46 |
+
the Work and Derivative Works thereof.
|
47 |
+
|
48 |
+
"Contribution" shall mean any work of authorship, including
|
49 |
+
the original version of the Work and any modifications or additions
|
50 |
+
to that Work or Derivative Works thereof, that is intentionally
|
51 |
+
submitted to Licensor for inclusion in the Work by the copyright owner
|
52 |
+
or by an individual or Legal Entity authorized to submit on behalf of
|
53 |
+
the copyright owner. For the purposes of this definition, "submitted"
|
54 |
+
means any form of electronic, verbal, or written communication sent
|
55 |
+
to the Licensor or its representatives, including but not limited to
|
56 |
+
communication on electronic mailing lists, source code control systems,
|
57 |
+
and issue tracking systems that are managed by, or on behalf of, the
|
58 |
+
Licensor for the purpose of discussing and improving the Work, but
|
59 |
+
excluding communication that is conspicuously marked or otherwise
|
60 |
+
designated in writing by the copyright owner as "Not a Contribution."
|
61 |
+
|
62 |
+
"Contributor" shall mean Licensor and any individual or Legal Entity
|
63 |
+
on behalf of whom a Contribution has been received by Licensor and
|
64 |
+
subsequently incorporated within the Work.
|
65 |
+
|
66 |
+
2. Grant of Copyright License. Subject to the terms and conditions of
|
67 |
+
this License, each Contributor hereby grants to You a perpetual,
|
68 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
69 |
+
copyright license to reproduce, prepare Derivative Works of,
|
70 |
+
publicly display, publicly perform, sublicense, and distribute the
|
71 |
+
Work and such Derivative Works in Source or Object form.
|
72 |
+
|
73 |
+
3. Grant of Patent License. Subject to the terms and conditions of
|
74 |
+
this License, each Contributor hereby grants to You a perpetual,
|
75 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
76 |
+
(except as stated in this section) patent license to make, have made,
|
77 |
+
use, offer to sell, sell, import, and otherwise transfer the Work,
|
78 |
+
where such license applies only to those patent claims licensable
|
79 |
+
by such Contributor that are necessarily infringed by their
|
80 |
+
Contribution(s) alone or by combination of their Contribution(s)
|
81 |
+
with the Work to which such Contribution(s) was submitted. If You
|
82 |
+
institute patent litigation against any entity (including a
|
83 |
+
cross-claim or counterclaim in a lawsuit) alleging that the Work
|
84 |
+
or a Contribution incorporated within the Work constitutes direct
|
85 |
+
or contributory patent infringement, then any patent licenses
|
86 |
+
granted to You under this License for that Work shall terminate
|
87 |
+
as of the date such litigation is filed.
|
88 |
+
|
89 |
+
4. Redistribution. You may reproduce and distribute copies of the
|
90 |
+
Work or Derivative Works thereof in any medium, with or without
|
91 |
+
modifications, and in Source or Object form, provided that You
|
92 |
+
meet the following conditions:
|
93 |
+
|
94 |
+
(a) You must give any other recipients of the Work or
|
95 |
+
Derivative Works a copy of this License; and
|
96 |
+
|
97 |
+
(b) You must cause any modified files to carry prominent notices
|
98 |
+
stating that You changed the files; and
|
99 |
+
|
100 |
+
(c) You must retain, in the Source form of any Derivative Works
|
101 |
+
that You distribute, all copyright, patent, trademark, and
|
102 |
+
attribution notices from the Source form of the Work,
|
103 |
+
excluding those notices that do not pertain to any part of
|
104 |
+
the Derivative Works; and
|
105 |
+
|
106 |
+
(d) If the Work includes a "NOTICE" text file as part of its
|
107 |
+
distribution, then any Derivative Works that You distribute must
|
108 |
+
include a readable copy of the attribution notices contained
|
109 |
+
within such NOTICE file, excluding those notices that do not
|
110 |
+
pertain to any part of the Derivative Works, in at least one
|
111 |
+
of the following places: within a NOTICE text file distributed
|
112 |
+
as part of the Derivative Works; within the Source form or
|
113 |
+
documentation, if provided along with the Derivative Works; or,
|
114 |
+
within a display generated by the Derivative Works, if and
|
115 |
+
wherever such third-party notices normally appear. The contents
|
116 |
+
of the NOTICE file are for informational purposes only and
|
117 |
+
do not modify the License. You may add Your own attribution
|
118 |
+
notices within Derivative Works that You distribute, alongside
|
119 |
+
or as an addendum to the NOTICE text from the Work, provided
|
120 |
+
that such additional attribution notices cannot be construed
|
121 |
+
as modifying the License.
|
122 |
+
|
123 |
+
You may add Your own copyright statement to Your modifications and
|
124 |
+
may provide additional or different license terms and conditions
|
125 |
+
for use, reproduction, or distribution of Your modifications, or
|
126 |
+
for any such Derivative Works as a whole, provided Your use,
|
127 |
+
reproduction, and distribution of the Work otherwise complies with
|
128 |
+
the conditions stated in this License.
|
129 |
+
|
130 |
+
5. Submission of Contributions. Unless You explicitly state otherwise,
|
131 |
+
any Contribution intentionally submitted for inclusion in the Work
|
132 |
+
by You to the Licensor shall be under the terms and conditions of
|
133 |
+
this License, without any additional terms or conditions.
|
134 |
+
Notwithstanding the above, nothing herein shall supersede or modify
|
135 |
+
the terms of any separate license agreement you may have executed
|
136 |
+
with Licensor regarding such Contributions.
|
137 |
+
|
138 |
+
6. Trademarks. This License does not grant permission to use the trade
|
139 |
+
names, trademarks, service marks, or product names of the Licensor,
|
140 |
+
except as required for reasonable and customary use in describing the
|
141 |
+
origin of the Work and reproducing the content of the NOTICE file.
|
142 |
+
|
143 |
+
7. Disclaimer of Warranty. Unless required by applicable law or
|
144 |
+
agreed to in writing, Licensor provides the Work (and each
|
145 |
+
Contributor provides its Contributions) on an "AS IS" BASIS,
|
146 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
|
147 |
+
implied, including, without limitation, any warranties or conditions
|
148 |
+
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
|
149 |
+
PARTICULAR PURPOSE. You are solely responsible for determining the
|
150 |
+
appropriateness of using or redistributing the Work and assume any
|
151 |
+
risks associated with Your exercise of permissions under this License.
|
152 |
+
|
153 |
+
8. Limitation of Liability. In no event and under no legal theory,
|
154 |
+
whether in tort (including negligence), contract, or otherwise,
|
155 |
+
unless required by applicable law (such as deliberate and grossly
|
156 |
+
negligent acts) or agreed to in writing, shall any Contributor be
|
157 |
+
liable to You for damages, including any direct, indirect, special,
|
158 |
+
incidental, or consequential damages of any character arising as a
|
159 |
+
result of this License or out of the use or inability to use the
|
160 |
+
Work (including but not limited to damages for loss of goodwill,
|
161 |
+
work stoppage, computer failure or malfunction, or any and all
|
162 |
+
other commercial damages or losses), even if such Contributor
|
163 |
+
has been advised of the possibility of such damages.
|
164 |
+
|
165 |
+
9. Accepting Warranty or Additional Liability. While redistributing
|
166 |
+
the Work or Derivative Works thereof, You may choose to offer,
|
167 |
+
and charge a fee for, acceptance of support, warranty, indemnity,
|
168 |
+
or other liability obligations and/or rights consistent with this
|
169 |
+
License. However, in accepting such obligations, You may act only
|
170 |
+
on Your own behalf and on Your sole responsibility, not on behalf
|
171 |
+
of any other Contributor, and only if You agree to indemnify,
|
172 |
+
defend, and hold each Contributor harmless for any liability
|
173 |
+
incurred by, or claims asserted against, such Contributor by reason
|
174 |
+
of your accepting any such warranty or additional liability.
|
175 |
+
|
176 |
+
END OF TERMS AND CONDITIONS
|
177 |
+
|
178 |
+
APPENDIX: How to apply the Apache License to your work.
|
179 |
+
|
180 |
+
To apply the Apache License to your work, attach the following
|
181 |
+
boilerplate notice, with the fields enclosed by brackets "[]"
|
182 |
+
replaced with your own identifying information. (Don't include
|
183 |
+
the brackets!) The text should be enclosed in the appropriate
|
184 |
+
comment syntax for the file format. We also recommend that a
|
185 |
+
file or class name and description of purpose be included on the
|
186 |
+
same "printed page" as the copyright notice for easier
|
187 |
+
identification within third-party archives.
|
188 |
+
|
189 |
+
Copyright [yyyy] [name of copyright owner]
|
190 |
+
|
191 |
+
Licensed under the Apache License, Version 2.0 (the "License");
|
192 |
+
you may not use this file except in compliance with the License.
|
193 |
+
You may obtain a copy of the License at
|
194 |
+
|
195 |
+
http://www.apache.org/licenses/LICENSE-2.0
|
196 |
+
|
197 |
+
Unless required by applicable law or agreed to in writing, software
|
198 |
+
distributed under the License is distributed on an "AS IS" BASIS,
|
199 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
200 |
+
See the License for the specific language governing permissions and
|
201 |
+
limitations under the License.
|
MANIFEST.in
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
include requirements.txt
|
app.py
ADDED
@@ -0,0 +1,121 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
|
3 |
+
from metaseg import SegAutoMaskGenerator
|
4 |
+
|
5 |
+
|
6 |
+
def image_app():
|
7 |
+
with gr.Blocks():
|
8 |
+
with gr.Row():
|
9 |
+
with gr.Column():
|
10 |
+
seg_automask_image_file = gr.Image(type="filepath").style(height=260)
|
11 |
+
|
12 |
+
with gr.Row():
|
13 |
+
with gr.Column():
|
14 |
+
seg_automask_image_model_type = gr.Dropdown(
|
15 |
+
choices=[
|
16 |
+
"vit_h",
|
17 |
+
"vit_l",
|
18 |
+
"vit_b",
|
19 |
+
],
|
20 |
+
value="vit_l",
|
21 |
+
label="Model Type",
|
22 |
+
)
|
23 |
+
|
24 |
+
seg_automask_image_points_per_side = gr.Slider(
|
25 |
+
minimum=0,
|
26 |
+
maximum=32,
|
27 |
+
step=2,
|
28 |
+
value=16,
|
29 |
+
label="Points per Side",
|
30 |
+
)
|
31 |
+
|
32 |
+
seg_automask_image_points_per_batch = gr.Slider(
|
33 |
+
minimum=0,
|
34 |
+
maximum=64,
|
35 |
+
step=2,
|
36 |
+
value=64,
|
37 |
+
label="Points per Batch",
|
38 |
+
)
|
39 |
+
|
40 |
+
seg_automask_image_predict = gr.Button(value="Generator")
|
41 |
+
|
42 |
+
with gr.Column():
|
43 |
+
output_image = gr.Image()
|
44 |
+
|
45 |
+
seg_automask_image_predict.click(
|
46 |
+
fn=SegAutoMaskGenerator().save_image,
|
47 |
+
inputs=[
|
48 |
+
seg_automask_image_file,
|
49 |
+
seg_automask_image_model_type,
|
50 |
+
seg_automask_image_points_per_side,
|
51 |
+
seg_automask_image_points_per_batch,
|
52 |
+
],
|
53 |
+
outputs=[output_image],
|
54 |
+
)
|
55 |
+
|
56 |
+
|
57 |
+
def video_app():
|
58 |
+
with gr.Blocks():
|
59 |
+
with gr.Row():
|
60 |
+
with gr.Column():
|
61 |
+
seg_automask_video_file = gr.Video().style(height=260)
|
62 |
+
|
63 |
+
with gr.Row():
|
64 |
+
with gr.Column():
|
65 |
+
seg_automask_video_model_type = gr.Dropdown(
|
66 |
+
choices=[
|
67 |
+
"vit_h",
|
68 |
+
"vit_l",
|
69 |
+
"vit_b",
|
70 |
+
],
|
71 |
+
value="vit_l",
|
72 |
+
label="Model Type",
|
73 |
+
)
|
74 |
+
|
75 |
+
seg_automask_video_points_per_side = gr.Slider(
|
76 |
+
minimum=0,
|
77 |
+
maximum=32,
|
78 |
+
step=2,
|
79 |
+
value=16,
|
80 |
+
label="Points per Side",
|
81 |
+
)
|
82 |
+
seg_automask_video_points_per_batch = gr.Slider(
|
83 |
+
minimum=0,
|
84 |
+
maximum=64,
|
85 |
+
step=2,
|
86 |
+
value=64,
|
87 |
+
label="Points per Batch",
|
88 |
+
)
|
89 |
+
|
90 |
+
seg_automask_video_predict = gr.Button(value="Generator")
|
91 |
+
with gr.Column():
|
92 |
+
output_video = gr.Video()
|
93 |
+
|
94 |
+
seg_automask_video_predict.click(
|
95 |
+
fn=SegAutoMaskGenerator().save_image,
|
96 |
+
inputs=[
|
97 |
+
seg_automask_video_file,
|
98 |
+
seg_automask_video_model_type,
|
99 |
+
seg_automask_video_points_per_side,
|
100 |
+
seg_automask_video_points_per_batch,
|
101 |
+
],
|
102 |
+
outputs=[output_video],
|
103 |
+
)
|
104 |
+
|
105 |
+
|
106 |
+
def metaseg_app():
|
107 |
+
app = gr.Blocks()
|
108 |
+
with app:
|
109 |
+
with gr.Row():
|
110 |
+
with gr.Column():
|
111 |
+
with gr.Tab("Image"):
|
112 |
+
image_app()
|
113 |
+
with gr.Tab("Video"):
|
114 |
+
video_app()
|
115 |
+
|
116 |
+
app.queue(concurrency_count=2)
|
117 |
+
app.launch(debug=True, enable_queue=True)
|
118 |
+
|
119 |
+
|
120 |
+
if __name__ == "__main__":
|
121 |
+
metaseg_app()
|
metaseg/__init__.py
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from metaseg.automatic_mask_generator import SamAutomaticMaskGenerator
|
8 |
+
from metaseg.build_sam import build_sam, build_sam_vit_b, build_sam_vit_h, build_sam_vit_l, sam_model_registry
|
9 |
+
from metaseg.demo import SegAutoMaskGenerator
|
10 |
+
from metaseg.predictor import SamPredictor
|
11 |
+
|
12 |
+
__version__ = "0.2.3"
|
metaseg/app.py
ADDED
@@ -0,0 +1,121 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
|
3 |
+
from metaseg import SegAutoMaskGenerator
|
4 |
+
|
5 |
+
|
6 |
+
def image_app():
|
7 |
+
with gr.Blocks():
|
8 |
+
with gr.Row():
|
9 |
+
with gr.Column():
|
10 |
+
seg_automask_image_file = gr.Image(type="filepath").style(height=260)
|
11 |
+
|
12 |
+
with gr.Row():
|
13 |
+
with gr.Column():
|
14 |
+
seg_automask_image_model_type = gr.Dropdown(
|
15 |
+
choices=[
|
16 |
+
"vit_h",
|
17 |
+
"vit_l",
|
18 |
+
"vit_b",
|
19 |
+
],
|
20 |
+
value="vit_l",
|
21 |
+
label="Model Type",
|
22 |
+
)
|
23 |
+
|
24 |
+
seg_automask_image_points_per_side = gr.Slider(
|
25 |
+
minimum=0,
|
26 |
+
maximum=32,
|
27 |
+
step=2,
|
28 |
+
value=16,
|
29 |
+
label="Points per Side",
|
30 |
+
)
|
31 |
+
|
32 |
+
seg_automask_image_points_per_batch = gr.Slider(
|
33 |
+
minimum=0,
|
34 |
+
maximum=64,
|
35 |
+
step=2,
|
36 |
+
value=64,
|
37 |
+
label="Points per Batch",
|
38 |
+
)
|
39 |
+
|
40 |
+
seg_automask_image_predict = gr.Button(value="Generator")
|
41 |
+
|
42 |
+
with gr.Column():
|
43 |
+
output_image = gr.Image()
|
44 |
+
|
45 |
+
seg_automask_image_predict.click(
|
46 |
+
fn=SegAutoMaskGenerator().save_image,
|
47 |
+
inputs=[
|
48 |
+
seg_automask_image_file,
|
49 |
+
seg_automask_image_model_type,
|
50 |
+
seg_automask_image_points_per_side,
|
51 |
+
seg_automask_image_points_per_batch,
|
52 |
+
],
|
53 |
+
outputs=[output_image],
|
54 |
+
)
|
55 |
+
|
56 |
+
|
57 |
+
def video_app():
|
58 |
+
with gr.Blocks():
|
59 |
+
with gr.Row():
|
60 |
+
with gr.Column():
|
61 |
+
seg_automask_video_file = gr.Video().style(height=260)
|
62 |
+
|
63 |
+
with gr.Row():
|
64 |
+
with gr.Column():
|
65 |
+
seg_automask_video_model_type = gr.Dropdown(
|
66 |
+
choices=[
|
67 |
+
"vit_h",
|
68 |
+
"vit_l",
|
69 |
+
"vit_b",
|
70 |
+
],
|
71 |
+
value="vit_l",
|
72 |
+
label="Model Type",
|
73 |
+
)
|
74 |
+
|
75 |
+
seg_automask_video_points_per_side = gr.Slider(
|
76 |
+
minimum=0,
|
77 |
+
maximum=32,
|
78 |
+
step=2,
|
79 |
+
value=16,
|
80 |
+
label="Points per Side",
|
81 |
+
)
|
82 |
+
seg_automask_video_points_per_batch = gr.Slider(
|
83 |
+
minimum=0,
|
84 |
+
maximum=64,
|
85 |
+
step=2,
|
86 |
+
value=64,
|
87 |
+
label="Points per Batch",
|
88 |
+
)
|
89 |
+
|
90 |
+
seg_automask_video_predict = gr.Button(value="Generator")
|
91 |
+
with gr.Column():
|
92 |
+
output_video = gr.Video()
|
93 |
+
|
94 |
+
seg_automask_video_predict.click(
|
95 |
+
fn=SegAutoMaskGenerator().save_image,
|
96 |
+
inputs=[
|
97 |
+
seg_automask_video_file,
|
98 |
+
seg_automask_video_model_type,
|
99 |
+
seg_automask_video_points_per_side,
|
100 |
+
seg_automask_video_points_per_batch,
|
101 |
+
],
|
102 |
+
outputs=[output_video],
|
103 |
+
)
|
104 |
+
|
105 |
+
|
106 |
+
def metaseg_app():
|
107 |
+
app = gr.Blocks()
|
108 |
+
with app:
|
109 |
+
with gr.Row():
|
110 |
+
with gr.Column():
|
111 |
+
with gr.Tab("Image"):
|
112 |
+
image_app()
|
113 |
+
with gr.Tab("Video"):
|
114 |
+
video_app()
|
115 |
+
|
116 |
+
app.queue(concurrency_count=2)
|
117 |
+
app.launch(debug=True, enable_queue=True)
|
118 |
+
|
119 |
+
|
120 |
+
if __name__ == "__main__":
|
121 |
+
metaseg_app()
|
metaseg/automatic_mask_generator.py
ADDED
@@ -0,0 +1,368 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from typing import Any, Dict, List, Optional, Tuple
|
8 |
+
|
9 |
+
import numpy as np
|
10 |
+
import torch
|
11 |
+
from torchvision.ops.boxes import batched_nms, box_area # type: ignore
|
12 |
+
|
13 |
+
from metaseg.modeling import Sam
|
14 |
+
from metaseg.predictor import SamPredictor
|
15 |
+
from metaseg.utils.amg import (
|
16 |
+
MaskData,
|
17 |
+
area_from_rle,
|
18 |
+
batch_iterator,
|
19 |
+
batched_mask_to_box,
|
20 |
+
box_xyxy_to_xywh,
|
21 |
+
build_all_layer_point_grids,
|
22 |
+
calculate_stability_score,
|
23 |
+
coco_encode_rle,
|
24 |
+
generate_crop_boxes,
|
25 |
+
is_box_near_crop_edge,
|
26 |
+
mask_to_rle_pytorch,
|
27 |
+
remove_small_regions,
|
28 |
+
rle_to_mask,
|
29 |
+
uncrop_boxes_xyxy,
|
30 |
+
uncrop_masks,
|
31 |
+
uncrop_points,
|
32 |
+
)
|
33 |
+
|
34 |
+
|
35 |
+
class SamAutomaticMaskGenerator:
|
36 |
+
def __init__(
|
37 |
+
self,
|
38 |
+
model: Sam,
|
39 |
+
points_per_side: Optional[int] = 32,
|
40 |
+
points_per_batch: int = 64,
|
41 |
+
pred_iou_thresh: float = 0.88,
|
42 |
+
stability_score_thresh: float = 0.95,
|
43 |
+
stability_score_offset: float = 1.0,
|
44 |
+
box_nms_thresh: float = 0.7,
|
45 |
+
crop_n_layers: int = 0,
|
46 |
+
crop_nms_thresh: float = 0.7,
|
47 |
+
crop_overlap_ratio: float = 512 / 1500,
|
48 |
+
crop_n_points_downscale_factor: int = 1,
|
49 |
+
point_grids: Optional[List[np.ndarray]] = None,
|
50 |
+
min_mask_region_area: int = 0,
|
51 |
+
output_mode: str = "binary_mask",
|
52 |
+
) -> None:
|
53 |
+
"""
|
54 |
+
Using a SAM model, generates masks for the entire image.
|
55 |
+
Generates a grid of point prompts over the image, then filters
|
56 |
+
low quality and duplicate masks. The default settings are chosen
|
57 |
+
for SAM with a ViT-H backbone.
|
58 |
+
|
59 |
+
Arguments:
|
60 |
+
model (Sam): The SAM model to use for mask prediction.
|
61 |
+
points_per_side (int or None): The number of points to be sampled
|
62 |
+
along one side of the image. The total number of points is
|
63 |
+
points_per_side**2. If None, 'point_grids' must provide explicit
|
64 |
+
point sampling.
|
65 |
+
points_per_batch (int): Sets the number of points run simultaneously
|
66 |
+
by the model. Higher numbers may be faster but use more GPU memory.
|
67 |
+
pred_iou_thresh (float): A filtering threshold in [0,1], using the
|
68 |
+
model's predicted mask quality.
|
69 |
+
stability_score_thresh (float): A filtering threshold in [0,1], using
|
70 |
+
the stability of the mask under changes to the cutoff used to binarize
|
71 |
+
the model's mask predictions.
|
72 |
+
stability_score_offset (float): The amount to shift the cutoff when
|
73 |
+
calculated the stability score.
|
74 |
+
box_nms_thresh (float): The box IoU cutoff used by non-maximal
|
75 |
+
suppression to filter duplicate masks.
|
76 |
+
crops_n_layers (int): If >0, mask prediction will be run again on
|
77 |
+
crops of the image. Sets the number of layers to run, where each
|
78 |
+
layer has 2**i_layer number of image crops.
|
79 |
+
crops_nms_thresh (float): The box IoU cutoff used by non-maximal
|
80 |
+
suppression to filter duplicate masks between different crops.
|
81 |
+
crop_overlap_ratio (float): Sets the degree to which crops overlap.
|
82 |
+
In the first crop layer, crops will overlap by this fraction of
|
83 |
+
the image length. Later layers with more crops scale down this overlap.
|
84 |
+
crop_n_points_downscale_factor (int): The number of points-per-side
|
85 |
+
sampled in layer n is scaled down by crop_n_points_downscale_factor**n.
|
86 |
+
point_grids (list(np.ndarray) or None): A list over explicit grids
|
87 |
+
of points used for sampling, normalized to [0,1]. The nth grid in the
|
88 |
+
list is used in the nth crop layer. Exclusive with points_per_side.
|
89 |
+
min_mask_region_area (int): If >0, postprocessing will be applied
|
90 |
+
to remove disconnected regions and holes in masks with area smaller
|
91 |
+
than min_mask_region_area. Requires opencv.
|
92 |
+
output_mode (str): The form masks are returned in. Can be 'binary_mask',
|
93 |
+
'uncompressed_rle', or 'coco_rle'. 'coco_rle' requires pycocotools.
|
94 |
+
For large resolutions, 'binary_mask' may consume large amounts of
|
95 |
+
memory.
|
96 |
+
"""
|
97 |
+
|
98 |
+
assert (points_per_side is None) != (
|
99 |
+
point_grids is None
|
100 |
+
), "Exactly one of points_per_side or point_grid must be provided."
|
101 |
+
if points_per_side is not None:
|
102 |
+
self.point_grids = build_all_layer_point_grids(
|
103 |
+
points_per_side,
|
104 |
+
crop_n_layers,
|
105 |
+
crop_n_points_downscale_factor,
|
106 |
+
)
|
107 |
+
elif point_grids is not None:
|
108 |
+
self.point_grids = point_grids
|
109 |
+
else:
|
110 |
+
raise ValueError("Can't have both points_per_side and point_grid be None.")
|
111 |
+
|
112 |
+
assert output_mode in [
|
113 |
+
"binary_mask",
|
114 |
+
"uncompressed_rle",
|
115 |
+
"coco_rle",
|
116 |
+
], f"Unknown output_mode {output_mode}."
|
117 |
+
if output_mode == "coco_rle":
|
118 |
+
from pycocotools import mask as mask_utils # type: ignore # noqa: F401
|
119 |
+
|
120 |
+
if min_mask_region_area > 0:
|
121 |
+
import cv2 # type: ignore # noqa: F401
|
122 |
+
|
123 |
+
self.predictor = SamPredictor(model)
|
124 |
+
self.points_per_batch = points_per_batch
|
125 |
+
self.pred_iou_thresh = pred_iou_thresh
|
126 |
+
self.stability_score_thresh = stability_score_thresh
|
127 |
+
self.stability_score_offset = stability_score_offset
|
128 |
+
self.box_nms_thresh = box_nms_thresh
|
129 |
+
self.crop_n_layers = crop_n_layers
|
130 |
+
self.crop_nms_thresh = crop_nms_thresh
|
131 |
+
self.crop_overlap_ratio = crop_overlap_ratio
|
132 |
+
self.crop_n_points_downscale_factor = crop_n_points_downscale_factor
|
133 |
+
self.min_mask_region_area = min_mask_region_area
|
134 |
+
self.output_mode = output_mode
|
135 |
+
|
136 |
+
@torch.no_grad()
|
137 |
+
def generate(self, image: np.ndarray) -> List[Dict[str, Any]]:
|
138 |
+
"""
|
139 |
+
Generates masks for the given image.
|
140 |
+
|
141 |
+
Arguments:
|
142 |
+
image (np.ndarray): The image to generate masks for, in HWC uint8 format.
|
143 |
+
|
144 |
+
Returns:
|
145 |
+
list(dict(str, any)): A list over records for masks. Each record is
|
146 |
+
a dict containing the following keys:
|
147 |
+
segmentation (dict(str, any) or np.ndarray): The mask. If
|
148 |
+
output_mode='binary_mask', is an array of shape HW. Otherwise,
|
149 |
+
is a dictionary containing the RLE.
|
150 |
+
bbox (list(float)): The box around the mask, in XYWH format.
|
151 |
+
area (int): The area in pixels of the mask.
|
152 |
+
predicted_iou (float): The model's own prediction of the mask's
|
153 |
+
quality. This is filtered by the pred_iou_thresh parameter.
|
154 |
+
point_coords (list(list(float))): The point coordinates input
|
155 |
+
to the model to generate this mask.
|
156 |
+
stability_score (float): A measure of the mask's quality. This
|
157 |
+
is filtered on using the stability_score_thresh parameter.
|
158 |
+
crop_box (list(float)): The crop of the image used to generate
|
159 |
+
the mask, given in XYWH format.
|
160 |
+
"""
|
161 |
+
|
162 |
+
# Generate masks
|
163 |
+
mask_data = self._generate_masks(image)
|
164 |
+
|
165 |
+
# Filter small disconnected regions and holes in masks
|
166 |
+
if self.min_mask_region_area > 0:
|
167 |
+
mask_data = self.postprocess_small_regions(
|
168 |
+
mask_data,
|
169 |
+
self.min_mask_region_area,
|
170 |
+
max(self.box_nms_thresh, self.crop_nms_thresh),
|
171 |
+
)
|
172 |
+
|
173 |
+
# Encode masks
|
174 |
+
if self.output_mode == "coco_rle":
|
175 |
+
mask_data["segmentations"] = [coco_encode_rle(rle) for rle in mask_data["rles"]]
|
176 |
+
elif self.output_mode == "binary_mask":
|
177 |
+
mask_data["segmentations"] = [rle_to_mask(rle) for rle in mask_data["rles"]]
|
178 |
+
else:
|
179 |
+
mask_data["segmentations"] = mask_data["rles"]
|
180 |
+
|
181 |
+
# Write mask records
|
182 |
+
curr_anns = []
|
183 |
+
for idx in range(len(mask_data["segmentations"])):
|
184 |
+
ann = {
|
185 |
+
"segmentation": mask_data["segmentations"][idx],
|
186 |
+
"area": area_from_rle(mask_data["rles"][idx]),
|
187 |
+
"bbox": box_xyxy_to_xywh(mask_data["boxes"][idx]).tolist(),
|
188 |
+
"predicted_iou": mask_data["iou_preds"][idx].item(),
|
189 |
+
"point_coords": [mask_data["points"][idx].tolist()],
|
190 |
+
"stability_score": mask_data["stability_score"][idx].item(),
|
191 |
+
"crop_box": box_xyxy_to_xywh(mask_data["crop_boxes"][idx]).tolist(),
|
192 |
+
}
|
193 |
+
curr_anns.append(ann)
|
194 |
+
|
195 |
+
return curr_anns
|
196 |
+
|
197 |
+
def _generate_masks(self, image: np.ndarray) -> MaskData:
|
198 |
+
orig_size = image.shape[:2]
|
199 |
+
crop_boxes, layer_idxs = generate_crop_boxes(orig_size, self.crop_n_layers, self.crop_overlap_ratio)
|
200 |
+
|
201 |
+
# Iterate over image crops
|
202 |
+
data = MaskData()
|
203 |
+
for crop_box, layer_idx in zip(crop_boxes, layer_idxs):
|
204 |
+
crop_data = self._process_crop(image, crop_box, layer_idx, orig_size)
|
205 |
+
data.cat(crop_data)
|
206 |
+
|
207 |
+
# Remove duplicate masks between crops
|
208 |
+
if len(crop_boxes) > 1:
|
209 |
+
# Prefer masks from smaller crops
|
210 |
+
scores = 1 / box_area(data["crop_boxes"])
|
211 |
+
scores = scores.to(data["boxes"].device)
|
212 |
+
keep_by_nms = batched_nms(
|
213 |
+
data["boxes"].float(),
|
214 |
+
scores,
|
215 |
+
torch.zeros(len(data["boxes"])), # categories
|
216 |
+
iou_threshold=self.crop_nms_thresh,
|
217 |
+
)
|
218 |
+
data.filter(keep_by_nms)
|
219 |
+
|
220 |
+
data.to_numpy()
|
221 |
+
return data
|
222 |
+
|
223 |
+
def _process_crop(
|
224 |
+
self,
|
225 |
+
image: np.ndarray,
|
226 |
+
crop_box: List[int],
|
227 |
+
crop_layer_idx: int,
|
228 |
+
orig_size: Tuple[int, ...],
|
229 |
+
) -> MaskData:
|
230 |
+
# Crop the image and calculate embeddings
|
231 |
+
x0, y0, x1, y1 = crop_box
|
232 |
+
cropped_im = image[y0:y1, x0:x1, :]
|
233 |
+
cropped_im_size = cropped_im.shape[:2]
|
234 |
+
self.predictor.set_image(cropped_im)
|
235 |
+
|
236 |
+
# Get points for this crop
|
237 |
+
points_scale = np.array(cropped_im_size)[None, ::-1]
|
238 |
+
points_for_image = self.point_grids[crop_layer_idx] * points_scale
|
239 |
+
|
240 |
+
# Generate masks for this crop in batches
|
241 |
+
data = MaskData()
|
242 |
+
for (points,) in batch_iterator(self.points_per_batch, points_for_image):
|
243 |
+
batch_data = self._process_batch(points, cropped_im_size, crop_box, orig_size)
|
244 |
+
data.cat(batch_data)
|
245 |
+
del batch_data
|
246 |
+
self.predictor.reset_image()
|
247 |
+
|
248 |
+
# Remove duplicates within this crop.
|
249 |
+
keep_by_nms = batched_nms(
|
250 |
+
data["boxes"].float(),
|
251 |
+
data["iou_preds"],
|
252 |
+
torch.zeros(len(data["boxes"])), # categories
|
253 |
+
iou_threshold=self.box_nms_thresh,
|
254 |
+
)
|
255 |
+
data.filter(keep_by_nms)
|
256 |
+
|
257 |
+
# Return to the original image frame
|
258 |
+
data["boxes"] = uncrop_boxes_xyxy(data["boxes"], crop_box)
|
259 |
+
data["points"] = uncrop_points(data["points"], crop_box)
|
260 |
+
data["crop_boxes"] = torch.tensor([crop_box for _ in range(len(data["rles"]))])
|
261 |
+
|
262 |
+
return data
|
263 |
+
|
264 |
+
def _process_batch(
|
265 |
+
self,
|
266 |
+
points: np.ndarray,
|
267 |
+
im_size: Tuple[int, ...],
|
268 |
+
crop_box: List[int],
|
269 |
+
orig_size: Tuple[int, ...],
|
270 |
+
) -> MaskData:
|
271 |
+
orig_h, orig_w = orig_size
|
272 |
+
|
273 |
+
# Run model on this batch
|
274 |
+
transformed_points = self.predictor.transform.apply_coords(points, im_size)
|
275 |
+
in_points = torch.as_tensor(transformed_points, device=self.predictor.device)
|
276 |
+
in_labels = torch.ones(in_points.shape[0], dtype=torch.int, device=in_points.device)
|
277 |
+
masks, iou_preds, _ = self.predictor.predict_torch(
|
278 |
+
in_points[:, None, :],
|
279 |
+
in_labels[:, None],
|
280 |
+
multimask_output=True,
|
281 |
+
return_logits=True,
|
282 |
+
)
|
283 |
+
|
284 |
+
# Serialize predictions and store in MaskData
|
285 |
+
data = MaskData(
|
286 |
+
masks=masks.flatten(0, 1),
|
287 |
+
iou_preds=iou_preds.flatten(0, 1),
|
288 |
+
points=torch.as_tensor(points.repeat(masks.shape[1], axis=0)),
|
289 |
+
)
|
290 |
+
del masks
|
291 |
+
|
292 |
+
# Filter by predicted IoU
|
293 |
+
if self.pred_iou_thresh > 0.0:
|
294 |
+
keep_mask = data["iou_preds"] > self.pred_iou_thresh
|
295 |
+
data.filter(keep_mask)
|
296 |
+
|
297 |
+
# Calculate stability score
|
298 |
+
data["stability_score"] = calculate_stability_score(
|
299 |
+
data["masks"], self.predictor.model.mask_threshold, self.stability_score_offset
|
300 |
+
)
|
301 |
+
if self.stability_score_thresh > 0.0:
|
302 |
+
keep_mask = data["stability_score"] >= self.stability_score_thresh
|
303 |
+
data.filter(keep_mask)
|
304 |
+
|
305 |
+
# Threshold masks and calculate boxes
|
306 |
+
data["masks"] = data["masks"] > self.predictor.model.mask_threshold
|
307 |
+
data["boxes"] = batched_mask_to_box(data["masks"])
|
308 |
+
|
309 |
+
# Filter boxes that touch crop boundaries
|
310 |
+
keep_mask = ~is_box_near_crop_edge(data["boxes"], crop_box, [0, 0, orig_w, orig_h])
|
311 |
+
if not torch.all(keep_mask):
|
312 |
+
data.filter(keep_mask)
|
313 |
+
|
314 |
+
# Compress to RLE
|
315 |
+
data["masks"] = uncrop_masks(data["masks"], crop_box, orig_h, orig_w)
|
316 |
+
data["rles"] = mask_to_rle_pytorch(data["masks"])
|
317 |
+
del data["masks"]
|
318 |
+
|
319 |
+
return data
|
320 |
+
|
321 |
+
@staticmethod
|
322 |
+
def postprocess_small_regions(mask_data: MaskData, min_area: int, nms_thresh: float) -> MaskData:
|
323 |
+
"""
|
324 |
+
Removes small disconnected regions and holes in masks, then reruns
|
325 |
+
box NMS to remove any new duplicates.
|
326 |
+
|
327 |
+
Edits mask_data in place.
|
328 |
+
|
329 |
+
Requires open-cv as a dependency.
|
330 |
+
"""
|
331 |
+
if len(mask_data["rles"]) == 0:
|
332 |
+
return mask_data
|
333 |
+
|
334 |
+
# Filter small disconnected regions and holes
|
335 |
+
new_masks = []
|
336 |
+
scores = []
|
337 |
+
for rle in mask_data["rles"]:
|
338 |
+
mask = rle_to_mask(rle)
|
339 |
+
|
340 |
+
mask, changed = remove_small_regions(mask, min_area, mode="holes")
|
341 |
+
unchanged = not changed
|
342 |
+
mask, changed = remove_small_regions(mask, min_area, mode="islands")
|
343 |
+
unchanged = unchanged and not changed
|
344 |
+
|
345 |
+
new_masks.append(torch.as_tensor(mask).unsqueeze(0))
|
346 |
+
# Give score=0 to changed masks and score=1 to unchanged masks
|
347 |
+
# so NMS will prefer ones that didn't need postprocessing
|
348 |
+
scores.append(float(unchanged))
|
349 |
+
|
350 |
+
# Recalculate boxes and remove any new duplicates
|
351 |
+
masks = torch.cat(new_masks, dim=0)
|
352 |
+
boxes = batched_mask_to_box(masks)
|
353 |
+
keep_by_nms = batched_nms(
|
354 |
+
boxes.float(),
|
355 |
+
torch.as_tensor(scores),
|
356 |
+
torch.zeros(len(boxes)), # categories
|
357 |
+
iou_threshold=nms_thresh,
|
358 |
+
)
|
359 |
+
|
360 |
+
# Only recalculate RLEs for masks that have changed
|
361 |
+
for i_mask in keep_by_nms:
|
362 |
+
if scores[i_mask] == 0.0:
|
363 |
+
mask_torch = masks[i_mask].unsqueeze(0)
|
364 |
+
mask_data["rles"][i_mask] = mask_to_rle_pytorch(mask_torch)[0]
|
365 |
+
mask_data["boxes"][i_mask] = boxes[i_mask] # update res directly
|
366 |
+
mask_data.filter(keep_by_nms)
|
367 |
+
|
368 |
+
return mask_data
|
metaseg/build_sam.py
ADDED
@@ -0,0 +1,107 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from functools import partial
|
8 |
+
|
9 |
+
import torch
|
10 |
+
|
11 |
+
from metaseg.modeling import ImageEncoderViT, MaskDecoder, PromptEncoder, Sam, TwoWayTransformer
|
12 |
+
|
13 |
+
|
14 |
+
def build_sam_vit_h(checkpoint=None):
|
15 |
+
return _build_sam(
|
16 |
+
encoder_embed_dim=1280,
|
17 |
+
encoder_depth=32,
|
18 |
+
encoder_num_heads=16,
|
19 |
+
encoder_global_attn_indexes=[7, 15, 23, 31],
|
20 |
+
checkpoint=checkpoint,
|
21 |
+
)
|
22 |
+
|
23 |
+
|
24 |
+
build_sam = build_sam_vit_h
|
25 |
+
|
26 |
+
|
27 |
+
def build_sam_vit_l(checkpoint=None):
|
28 |
+
return _build_sam(
|
29 |
+
encoder_embed_dim=1024,
|
30 |
+
encoder_depth=24,
|
31 |
+
encoder_num_heads=16,
|
32 |
+
encoder_global_attn_indexes=[5, 11, 17, 23],
|
33 |
+
checkpoint=checkpoint,
|
34 |
+
)
|
35 |
+
|
36 |
+
|
37 |
+
def build_sam_vit_b(checkpoint=None):
|
38 |
+
return _build_sam(
|
39 |
+
encoder_embed_dim=768,
|
40 |
+
encoder_depth=12,
|
41 |
+
encoder_num_heads=12,
|
42 |
+
encoder_global_attn_indexes=[2, 5, 8, 11],
|
43 |
+
checkpoint=checkpoint,
|
44 |
+
)
|
45 |
+
|
46 |
+
|
47 |
+
sam_model_registry = {
|
48 |
+
"default": build_sam,
|
49 |
+
"vit_h": build_sam,
|
50 |
+
"vit_l": build_sam_vit_l,
|
51 |
+
"vit_b": build_sam_vit_b,
|
52 |
+
}
|
53 |
+
|
54 |
+
|
55 |
+
def _build_sam(
|
56 |
+
encoder_embed_dim,
|
57 |
+
encoder_depth,
|
58 |
+
encoder_num_heads,
|
59 |
+
encoder_global_attn_indexes,
|
60 |
+
checkpoint=None,
|
61 |
+
):
|
62 |
+
prompt_embed_dim = 256
|
63 |
+
image_size = 1024
|
64 |
+
vit_patch_size = 16
|
65 |
+
image_embedding_size = image_size // vit_patch_size
|
66 |
+
sam = Sam(
|
67 |
+
image_encoder=ImageEncoderViT(
|
68 |
+
depth=encoder_depth,
|
69 |
+
embed_dim=encoder_embed_dim,
|
70 |
+
img_size=image_size,
|
71 |
+
mlp_ratio=4,
|
72 |
+
norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
|
73 |
+
num_heads=encoder_num_heads,
|
74 |
+
patch_size=vit_patch_size,
|
75 |
+
qkv_bias=True,
|
76 |
+
use_rel_pos=True,
|
77 |
+
global_attn_indexes=encoder_global_attn_indexes,
|
78 |
+
window_size=14,
|
79 |
+
out_chans=prompt_embed_dim,
|
80 |
+
),
|
81 |
+
prompt_encoder=PromptEncoder(
|
82 |
+
embed_dim=prompt_embed_dim,
|
83 |
+
image_embedding_size=(image_embedding_size, image_embedding_size),
|
84 |
+
input_image_size=(image_size, image_size),
|
85 |
+
mask_in_chans=16,
|
86 |
+
),
|
87 |
+
mask_decoder=MaskDecoder(
|
88 |
+
num_multimask_outputs=3,
|
89 |
+
transformer=TwoWayTransformer(
|
90 |
+
depth=2,
|
91 |
+
embedding_dim=prompt_embed_dim,
|
92 |
+
mlp_dim=2048,
|
93 |
+
num_heads=8,
|
94 |
+
),
|
95 |
+
transformer_dim=prompt_embed_dim,
|
96 |
+
iou_head_depth=3,
|
97 |
+
iou_head_hidden_dim=256,
|
98 |
+
),
|
99 |
+
pixel_mean=[123.675, 116.28, 103.53],
|
100 |
+
pixel_std=[58.395, 57.12, 57.375],
|
101 |
+
)
|
102 |
+
sam.eval()
|
103 |
+
if checkpoint is not None:
|
104 |
+
with open(checkpoint, "rb") as f:
|
105 |
+
state_dict = torch.load(f)
|
106 |
+
sam.load_state_dict(state_dict)
|
107 |
+
return sam
|
metaseg/demo.py
ADDED
@@ -0,0 +1,112 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional
|
2 |
+
|
3 |
+
import cv2
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
|
7 |
+
from metaseg import SamAutomaticMaskGenerator, sam_model_registry
|
8 |
+
from metaseg.utils.file import download_model
|
9 |
+
|
10 |
+
|
11 |
+
class SegAutoMaskGenerator:
|
12 |
+
def __init__(self):
|
13 |
+
self.model = None
|
14 |
+
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
15 |
+
|
16 |
+
def load_model(self, model_type):
|
17 |
+
if self.model is None:
|
18 |
+
model_path = download_model(model_type)
|
19 |
+
model = sam_model_registry[self.model_type](checkpoint=model_path)
|
20 |
+
model.to(device=self.device)
|
21 |
+
self.model = model
|
22 |
+
|
23 |
+
return self.model
|
24 |
+
|
25 |
+
def load_image(self, image_path):
|
26 |
+
image = cv2.imread(image_path)
|
27 |
+
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
|
28 |
+
return image
|
29 |
+
|
30 |
+
def load_video(self, video_path):
|
31 |
+
cap = cv2.VideoCapture(video_path)
|
32 |
+
frame_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
|
33 |
+
frame_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
34 |
+
fourcc = cv2.VideoWriter_fourcc(*"XVID")
|
35 |
+
fps = int(cap.get(cv2.CAP_PROP_FPS))
|
36 |
+
out = cv2.VideoWriter("output.mp4", fourcc, fps, (frame_width, frame_height))
|
37 |
+
|
38 |
+
return cap, out
|
39 |
+
|
40 |
+
def predict(self, frame, model_type, points_per_side, points_per_batch):
|
41 |
+
model = self.load_model(model_type)
|
42 |
+
mask_generator = SamAutomaticMaskGenerator(
|
43 |
+
model, points_per_side=points_per_side, points_per_batch=points_per_batch
|
44 |
+
)
|
45 |
+
masks = mask_generator.generate(frame)
|
46 |
+
|
47 |
+
return frame, masks
|
48 |
+
|
49 |
+
def save_image(self, source, model_type, points_per_side, points_per_batch):
|
50 |
+
read_image = self.load_image(source)
|
51 |
+
image, anns = self.predict(read_image, model_type, points_per_side, points_per_batch)
|
52 |
+
if len(anns) == 0:
|
53 |
+
return
|
54 |
+
|
55 |
+
sorted_anns = sorted(anns, key=(lambda x: x["area"]), reverse=True)
|
56 |
+
mask_image = np.zeros((anns[0]["segmentation"].shape[0], anns[0]["segmentation"].shape[1], 3), dtype=np.uint8)
|
57 |
+
colors = np.random.randint(0, 255, size=(256, 3), dtype=np.uint8)
|
58 |
+
for i, ann in enumerate(sorted_anns):
|
59 |
+
m = ann["segmentation"]
|
60 |
+
img = np.ones((m.shape[0], m.shape[1], 3), dtype=np.uint8)
|
61 |
+
color = colors[i % 256]
|
62 |
+
for i in range(3):
|
63 |
+
img[:, :, 0] = color[0]
|
64 |
+
img[:, :, 1] = color[1]
|
65 |
+
img[:, :, 2] = color[2]
|
66 |
+
img = cv2.bitwise_and(img, img, mask=m.astype(np.uint8))
|
67 |
+
img = cv2.addWeighted(img, 0.35, np.zeros_like(img), 0.65, 0)
|
68 |
+
mask_image = cv2.add(mask_image, img)
|
69 |
+
|
70 |
+
combined_mask = cv2.add(image, mask_image)
|
71 |
+
cv2.imwrite("output.jpg", combined_mask)
|
72 |
+
|
73 |
+
return "output.jpg"
|
74 |
+
|
75 |
+
def save_video(self, source, model_type, points_per_side, points_per_batch):
|
76 |
+
cap, out = self.load_video()
|
77 |
+
colors = np.random.randint(0, 255, size=(256, 3), dtype=np.uint8)
|
78 |
+
|
79 |
+
while True:
|
80 |
+
ret, frame = cap.read()
|
81 |
+
if not ret:
|
82 |
+
break
|
83 |
+
|
84 |
+
image, anns = self.predict(frame)
|
85 |
+
if len(anns) == 0:
|
86 |
+
continue
|
87 |
+
|
88 |
+
sorted_anns = sorted(anns, key=(lambda x: x["area"]), reverse=True)
|
89 |
+
mask_image = np.zeros(
|
90 |
+
(anns[0]["segmentation"].shape[0], anns[0]["segmentation"].shape[1], 3), dtype=np.uint8
|
91 |
+
)
|
92 |
+
|
93 |
+
for i, ann in enumerate(sorted_anns):
|
94 |
+
if ann["area"] > 5000:
|
95 |
+
m = ann["segmentation"]
|
96 |
+
color = colors[i % 256] # Her nesne için farklı bir renk kullan
|
97 |
+
img = np.zeros((m.shape[0], m.shape[1], 3), dtype=np.uint8)
|
98 |
+
img[:, :, 0] = color[0]
|
99 |
+
img[:, :, 1] = color[1]
|
100 |
+
img[:, :, 2] = color[2]
|
101 |
+
img = cv2.bitwise_and(img, img, mask=m.astype(np.uint8))
|
102 |
+
img = cv2.addWeighted(img, 0.35, np.zeros_like(img), 0.65, 0)
|
103 |
+
mask_image = cv2.add(mask_image, img)
|
104 |
+
|
105 |
+
combined_mask = cv2.add(frame, mask_image)
|
106 |
+
out.write(combined_mask)
|
107 |
+
|
108 |
+
out.release()
|
109 |
+
cap.release()
|
110 |
+
cv2.destroyAllWindows()
|
111 |
+
|
112 |
+
return "output.mp4"
|
metaseg/modeling/__init__.py
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from metaseg.modeling.image_encoder import ImageEncoderViT
|
8 |
+
from metaseg.modeling.mask_decoder import MaskDecoder
|
9 |
+
from metaseg.modeling.prompt_encoder import PromptEncoder
|
10 |
+
from metaseg.modeling.sam import Sam
|
11 |
+
from metaseg.modeling.transformer import TwoWayTransformer
|
metaseg/modeling/__pycache__/__init__.cpython-310.pyc
ADDED
Binary file (484 Bytes). View file
|
|
metaseg/modeling/__pycache__/common.cpython-310.pyc
ADDED
Binary file (1.75 kB). View file
|
|
metaseg/modeling/__pycache__/image_encoder.cpython-310.pyc
ADDED
Binary file (12.6 kB). View file
|
|
metaseg/modeling/__pycache__/mask_decoder.cpython-310.pyc
ADDED
Binary file (5.46 kB). View file
|
|
metaseg/modeling/__pycache__/prompt_encoder.cpython-310.pyc
ADDED
Binary file (7.68 kB). View file
|
|
metaseg/modeling/__pycache__/sam.cpython-310.pyc
ADDED
Binary file (6.74 kB). View file
|
|
metaseg/modeling/__pycache__/transformer.cpython-310.pyc
ADDED
Binary file (6.59 kB). View file
|
|
metaseg/modeling/common.py
ADDED
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from typing import Type
|
8 |
+
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
|
12 |
+
|
13 |
+
class MLPBlock(nn.Module):
|
14 |
+
def __init__(
|
15 |
+
self,
|
16 |
+
embedding_dim: int,
|
17 |
+
mlp_dim: int,
|
18 |
+
act: Type[nn.Module] = nn.GELU,
|
19 |
+
) -> None:
|
20 |
+
super().__init__()
|
21 |
+
self.lin1 = nn.Linear(embedding_dim, mlp_dim)
|
22 |
+
self.lin2 = nn.Linear(mlp_dim, embedding_dim)
|
23 |
+
self.act = act()
|
24 |
+
|
25 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
26 |
+
return self.lin2(self.act(self.lin1(x)))
|
27 |
+
|
28 |
+
|
29 |
+
# From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa
|
30 |
+
# Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119 # noqa
|
31 |
+
class LayerNorm2d(nn.Module):
|
32 |
+
def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
|
33 |
+
super().__init__()
|
34 |
+
self.weight = nn.Parameter(torch.ones(num_channels))
|
35 |
+
self.bias = nn.Parameter(torch.zeros(num_channels))
|
36 |
+
self.eps = eps
|
37 |
+
|
38 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
39 |
+
u = x.mean(1, keepdim=True)
|
40 |
+
s = (x - u).pow(2).mean(1, keepdim=True)
|
41 |
+
x = (x - u) / torch.sqrt(s + self.eps)
|
42 |
+
x = self.weight[:, None, None] * x + self.bias[:, None, None]
|
43 |
+
return x
|
metaseg/modeling/image_encoder.py
ADDED
@@ -0,0 +1,389 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from typing import Optional, Tuple, Type
|
8 |
+
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
import torch.nn.functional as F
|
12 |
+
|
13 |
+
from metaseg.modeling.common import LayerNorm2d, MLPBlock
|
14 |
+
|
15 |
+
|
16 |
+
# This class and its supporting functions below lightly adapted from the ViTDet backbone available at: https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/vit.py # noqa
|
17 |
+
class ImageEncoderViT(nn.Module):
|
18 |
+
def __init__(
|
19 |
+
self,
|
20 |
+
img_size: int = 1024,
|
21 |
+
patch_size: int = 16,
|
22 |
+
in_chans: int = 3,
|
23 |
+
embed_dim: int = 768,
|
24 |
+
depth: int = 12,
|
25 |
+
num_heads: int = 12,
|
26 |
+
mlp_ratio: float = 4.0,
|
27 |
+
out_chans: int = 256,
|
28 |
+
qkv_bias: bool = True,
|
29 |
+
norm_layer: Type[nn.Module] = nn.LayerNorm,
|
30 |
+
act_layer: Type[nn.Module] = nn.GELU,
|
31 |
+
use_abs_pos: bool = True,
|
32 |
+
use_rel_pos: bool = False,
|
33 |
+
rel_pos_zero_init: bool = True,
|
34 |
+
window_size: int = 0,
|
35 |
+
global_attn_indexes: Tuple[int, ...] = (),
|
36 |
+
) -> None:
|
37 |
+
"""
|
38 |
+
Args:
|
39 |
+
img_size (int): Input image size.
|
40 |
+
patch_size (int): Patch size.
|
41 |
+
in_chans (int): Number of input image channels.
|
42 |
+
embed_dim (int): Patch embedding dimension.
|
43 |
+
depth (int): Depth of ViT.
|
44 |
+
num_heads (int): Number of attention heads in each ViT block.
|
45 |
+
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
|
46 |
+
qkv_bias (bool): If True, add a learnable bias to query, key, value.
|
47 |
+
norm_layer (nn.Module): Normalization layer.
|
48 |
+
act_layer (nn.Module): Activation layer.
|
49 |
+
use_abs_pos (bool): If True, use absolute positional embeddings.
|
50 |
+
use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
|
51 |
+
rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
|
52 |
+
window_size (int): Window size for window attention blocks.
|
53 |
+
global_attn_indexes (list): Indexes for blocks using global attention.
|
54 |
+
"""
|
55 |
+
super().__init__()
|
56 |
+
self.img_size = img_size
|
57 |
+
|
58 |
+
self.patch_embed = PatchEmbed(
|
59 |
+
kernel_size=(patch_size, patch_size),
|
60 |
+
stride=(patch_size, patch_size),
|
61 |
+
in_chans=in_chans,
|
62 |
+
embed_dim=embed_dim,
|
63 |
+
)
|
64 |
+
|
65 |
+
self.pos_embed: Optional[nn.Parameter] = None
|
66 |
+
if use_abs_pos:
|
67 |
+
# Initialize absolute positional embedding with pretrain image size.
|
68 |
+
self.pos_embed = nn.Parameter(torch.zeros(1, img_size // patch_size, img_size // patch_size, embed_dim))
|
69 |
+
|
70 |
+
self.blocks = nn.ModuleList()
|
71 |
+
for i in range(depth):
|
72 |
+
block = Block(
|
73 |
+
dim=embed_dim,
|
74 |
+
num_heads=num_heads,
|
75 |
+
mlp_ratio=mlp_ratio,
|
76 |
+
qkv_bias=qkv_bias,
|
77 |
+
norm_layer=norm_layer,
|
78 |
+
act_layer=act_layer,
|
79 |
+
use_rel_pos=use_rel_pos,
|
80 |
+
rel_pos_zero_init=rel_pos_zero_init,
|
81 |
+
window_size=window_size if i not in global_attn_indexes else 0,
|
82 |
+
input_size=(img_size // patch_size, img_size // patch_size),
|
83 |
+
)
|
84 |
+
self.blocks.append(block)
|
85 |
+
|
86 |
+
self.neck = nn.Sequential(
|
87 |
+
nn.Conv2d(
|
88 |
+
embed_dim,
|
89 |
+
out_chans,
|
90 |
+
kernel_size=1,
|
91 |
+
bias=False,
|
92 |
+
),
|
93 |
+
LayerNorm2d(out_chans),
|
94 |
+
nn.Conv2d(
|
95 |
+
out_chans,
|
96 |
+
out_chans,
|
97 |
+
kernel_size=3,
|
98 |
+
padding=1,
|
99 |
+
bias=False,
|
100 |
+
),
|
101 |
+
LayerNorm2d(out_chans),
|
102 |
+
)
|
103 |
+
|
104 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
105 |
+
x = self.patch_embed(x)
|
106 |
+
if self.pos_embed is not None:
|
107 |
+
x = x + self.pos_embed
|
108 |
+
|
109 |
+
for blk in self.blocks:
|
110 |
+
x = blk(x)
|
111 |
+
|
112 |
+
x = self.neck(x.permute(0, 3, 1, 2))
|
113 |
+
|
114 |
+
return x
|
115 |
+
|
116 |
+
|
117 |
+
class Block(nn.Module):
|
118 |
+
"""Transformer blocks with support of window attention and residual propagation blocks"""
|
119 |
+
|
120 |
+
def __init__(
|
121 |
+
self,
|
122 |
+
dim: int,
|
123 |
+
num_heads: int,
|
124 |
+
mlp_ratio: float = 4.0,
|
125 |
+
qkv_bias: bool = True,
|
126 |
+
norm_layer: Type[nn.Module] = nn.LayerNorm,
|
127 |
+
act_layer: Type[nn.Module] = nn.GELU,
|
128 |
+
use_rel_pos: bool = False,
|
129 |
+
rel_pos_zero_init: bool = True,
|
130 |
+
window_size: int = 0,
|
131 |
+
input_size: Optional[Tuple[int, int]] = None,
|
132 |
+
) -> None:
|
133 |
+
"""
|
134 |
+
Args:
|
135 |
+
dim (int): Number of input channels.
|
136 |
+
num_heads (int): Number of attention heads in each ViT block.
|
137 |
+
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
|
138 |
+
qkv_bias (bool): If True, add a learnable bias to query, key, value.
|
139 |
+
norm_layer (nn.Module): Normalization layer.
|
140 |
+
act_layer (nn.Module): Activation layer.
|
141 |
+
use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
|
142 |
+
rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
|
143 |
+
window_size (int): Window size for window attention blocks. If it equals 0, then
|
144 |
+
use global attention.
|
145 |
+
input_size (int or None): Input resolution for calculating the relative positional
|
146 |
+
parameter size.
|
147 |
+
"""
|
148 |
+
super().__init__()
|
149 |
+
self.norm1 = norm_layer(dim)
|
150 |
+
self.attn = Attention(
|
151 |
+
dim,
|
152 |
+
num_heads=num_heads,
|
153 |
+
qkv_bias=qkv_bias,
|
154 |
+
use_rel_pos=use_rel_pos,
|
155 |
+
rel_pos_zero_init=rel_pos_zero_init,
|
156 |
+
input_size=input_size if window_size == 0 else (window_size, window_size),
|
157 |
+
)
|
158 |
+
|
159 |
+
self.norm2 = norm_layer(dim)
|
160 |
+
self.mlp = MLPBlock(embedding_dim=dim, mlp_dim=int(dim * mlp_ratio), act=act_layer)
|
161 |
+
|
162 |
+
self.window_size = window_size
|
163 |
+
|
164 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
165 |
+
shortcut = x
|
166 |
+
x = self.norm1(x)
|
167 |
+
# Window partition
|
168 |
+
if self.window_size > 0:
|
169 |
+
H, W = x.shape[1], x.shape[2]
|
170 |
+
x, pad_hw = window_partition(x, self.window_size)
|
171 |
+
|
172 |
+
x = self.attn(x)
|
173 |
+
# Reverse window partition
|
174 |
+
if self.window_size > 0:
|
175 |
+
x = window_unpartition(x, self.window_size, pad_hw, (H, W))
|
176 |
+
|
177 |
+
x = shortcut + x
|
178 |
+
x = x + self.mlp(self.norm2(x))
|
179 |
+
|
180 |
+
return x
|
181 |
+
|
182 |
+
|
183 |
+
class Attention(nn.Module):
|
184 |
+
"""Multi-head Attention block with relative position embeddings."""
|
185 |
+
|
186 |
+
def __init__(
|
187 |
+
self,
|
188 |
+
dim: int,
|
189 |
+
num_heads: int = 8,
|
190 |
+
qkv_bias: bool = True,
|
191 |
+
use_rel_pos: bool = False,
|
192 |
+
rel_pos_zero_init: bool = True,
|
193 |
+
input_size: Optional[Tuple[int, int]] = None,
|
194 |
+
) -> None:
|
195 |
+
"""
|
196 |
+
Args:
|
197 |
+
dim (int): Number of input channels.
|
198 |
+
num_heads (int): Number of attention heads.
|
199 |
+
qkv_bias (bool: If True, add a learnable bias to query, key, value.
|
200 |
+
rel_pos (bool): If True, add relative positional embeddings to the attention map.
|
201 |
+
rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
|
202 |
+
input_size (int or None): Input resolution for calculating the relative positional
|
203 |
+
parameter size.
|
204 |
+
"""
|
205 |
+
super().__init__()
|
206 |
+
self.num_heads = num_heads
|
207 |
+
head_dim = dim // num_heads
|
208 |
+
self.scale = head_dim ** -0.5
|
209 |
+
|
210 |
+
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
211 |
+
self.proj = nn.Linear(dim, dim)
|
212 |
+
|
213 |
+
self.use_rel_pos = use_rel_pos
|
214 |
+
if self.use_rel_pos:
|
215 |
+
assert input_size is not None, "Input size must be provided if using relative positional encoding."
|
216 |
+
# initialize relative positional embeddings
|
217 |
+
self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
|
218 |
+
self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))
|
219 |
+
|
220 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
221 |
+
B, H, W, _ = x.shape
|
222 |
+
# qkv with shape (3, B, nHead, H * W, C)
|
223 |
+
qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
|
224 |
+
# q, k, v with shape (B * nHead, H * W, C)
|
225 |
+
q, k, v = qkv.reshape(3, B * self.num_heads, H * W, -1).unbind(0)
|
226 |
+
|
227 |
+
attn = (q * self.scale) @ k.transpose(-2, -1)
|
228 |
+
|
229 |
+
if self.use_rel_pos:
|
230 |
+
attn = add_decomposed_rel_pos(attn, q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W))
|
231 |
+
|
232 |
+
attn = attn.softmax(dim=-1)
|
233 |
+
x = (attn @ v).view(B, self.num_heads, H, W, -1).permute(0, 2, 3, 1, 4).reshape(B, H, W, -1)
|
234 |
+
x = self.proj(x)
|
235 |
+
|
236 |
+
return x
|
237 |
+
|
238 |
+
|
239 |
+
def window_partition(x: torch.Tensor, window_size: int) -> Tuple[torch.Tensor, Tuple[int, int]]:
|
240 |
+
"""
|
241 |
+
Partition into non-overlapping windows with padding if needed.
|
242 |
+
Args:
|
243 |
+
x (tensor): input tokens with [B, H, W, C].
|
244 |
+
window_size (int): window size.
|
245 |
+
|
246 |
+
Returns:
|
247 |
+
windows: windows after partition with [B * num_windows, window_size, window_size, C].
|
248 |
+
(Hp, Wp): padded height and width before partition
|
249 |
+
"""
|
250 |
+
B, H, W, C = x.shape
|
251 |
+
|
252 |
+
pad_h = (window_size - H % window_size) % window_size
|
253 |
+
pad_w = (window_size - W % window_size) % window_size
|
254 |
+
if pad_h > 0 or pad_w > 0:
|
255 |
+
x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
|
256 |
+
Hp, Wp = H + pad_h, W + pad_w
|
257 |
+
|
258 |
+
x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
|
259 |
+
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
|
260 |
+
return windows, (Hp, Wp)
|
261 |
+
|
262 |
+
|
263 |
+
def window_unpartition(
|
264 |
+
windows: torch.Tensor, window_size: int, pad_hw: Tuple[int, int], hw: Tuple[int, int]
|
265 |
+
) -> torch.Tensor:
|
266 |
+
"""
|
267 |
+
Window unpartition into original sequences and removing padding.
|
268 |
+
Args:
|
269 |
+
x (tensor): input tokens with [B * num_windows, window_size, window_size, C].
|
270 |
+
window_size (int): window size.
|
271 |
+
pad_hw (Tuple): padded height and width (Hp, Wp).
|
272 |
+
hw (Tuple): original height and width (H, W) before padding.
|
273 |
+
|
274 |
+
Returns:
|
275 |
+
x: unpartitioned sequences with [B, H, W, C].
|
276 |
+
"""
|
277 |
+
Hp, Wp = pad_hw
|
278 |
+
H, W = hw
|
279 |
+
B = windows.shape[0] // (Hp * Wp // window_size // window_size)
|
280 |
+
x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
|
281 |
+
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
|
282 |
+
|
283 |
+
if Hp > H or Wp > W:
|
284 |
+
x = x[:, :H, :W, :].contiguous()
|
285 |
+
return x
|
286 |
+
|
287 |
+
|
288 |
+
def get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:
|
289 |
+
"""
|
290 |
+
Get relative positional embeddings according to the relative positions of
|
291 |
+
query and key sizes.
|
292 |
+
Args:
|
293 |
+
q_size (int): size of query q.
|
294 |
+
k_size (int): size of key k.
|
295 |
+
rel_pos (Tensor): relative position embeddings (L, C).
|
296 |
+
|
297 |
+
Returns:
|
298 |
+
Extracted positional embeddings according to relative positions.
|
299 |
+
"""
|
300 |
+
max_rel_dist = int(2 * max(q_size, k_size) - 1)
|
301 |
+
# Interpolate rel pos if needed.
|
302 |
+
if rel_pos.shape[0] != max_rel_dist:
|
303 |
+
# Interpolate rel pos.
|
304 |
+
rel_pos_resized = F.interpolate(
|
305 |
+
rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
|
306 |
+
size=max_rel_dist,
|
307 |
+
mode="linear",
|
308 |
+
)
|
309 |
+
rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
|
310 |
+
else:
|
311 |
+
rel_pos_resized = rel_pos
|
312 |
+
|
313 |
+
# Scale the coords with short length if shapes for q and k are different.
|
314 |
+
q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
|
315 |
+
k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
|
316 |
+
relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
|
317 |
+
|
318 |
+
return rel_pos_resized[relative_coords.long()]
|
319 |
+
|
320 |
+
|
321 |
+
def add_decomposed_rel_pos(
|
322 |
+
attn: torch.Tensor,
|
323 |
+
q: torch.Tensor,
|
324 |
+
rel_pos_h: torch.Tensor,
|
325 |
+
rel_pos_w: torch.Tensor,
|
326 |
+
q_size: Tuple[int, int],
|
327 |
+
k_size: Tuple[int, int],
|
328 |
+
) -> torch.Tensor:
|
329 |
+
"""
|
330 |
+
Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.
|
331 |
+
https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py # noqa B950
|
332 |
+
Args:
|
333 |
+
attn (Tensor): attention map.
|
334 |
+
q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).
|
335 |
+
rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.
|
336 |
+
rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.
|
337 |
+
q_size (Tuple): spatial sequence size of query q with (q_h, q_w).
|
338 |
+
k_size (Tuple): spatial sequence size of key k with (k_h, k_w).
|
339 |
+
|
340 |
+
Returns:
|
341 |
+
attn (Tensor): attention map with added relative positional embeddings.
|
342 |
+
"""
|
343 |
+
q_h, q_w = q_size
|
344 |
+
k_h, k_w = k_size
|
345 |
+
Rh = get_rel_pos(q_h, k_h, rel_pos_h)
|
346 |
+
Rw = get_rel_pos(q_w, k_w, rel_pos_w)
|
347 |
+
|
348 |
+
B, _, dim = q.shape
|
349 |
+
r_q = q.reshape(B, q_h, q_w, dim)
|
350 |
+
rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh)
|
351 |
+
rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw)
|
352 |
+
|
353 |
+
attn = (attn.view(B, q_h, q_w, k_h, k_w) + rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]).view(
|
354 |
+
B, q_h * q_w, k_h * k_w
|
355 |
+
)
|
356 |
+
|
357 |
+
return attn
|
358 |
+
|
359 |
+
|
360 |
+
class PatchEmbed(nn.Module):
|
361 |
+
"""
|
362 |
+
Image to Patch Embedding.
|
363 |
+
"""
|
364 |
+
|
365 |
+
def __init__(
|
366 |
+
self,
|
367 |
+
kernel_size: Tuple[int, int] = (16, 16),
|
368 |
+
stride: Tuple[int, int] = (16, 16),
|
369 |
+
padding: Tuple[int, int] = (0, 0),
|
370 |
+
in_chans: int = 3,
|
371 |
+
embed_dim: int = 768,
|
372 |
+
) -> None:
|
373 |
+
"""
|
374 |
+
Args:
|
375 |
+
kernel_size (Tuple): kernel size of the projection layer.
|
376 |
+
stride (Tuple): stride of the projection layer.
|
377 |
+
padding (Tuple): padding size of the projection layer.
|
378 |
+
in_chans (int): Number of input image channels.
|
379 |
+
embed_dim (int): embed_dim (int): Patch embedding dimension.
|
380 |
+
"""
|
381 |
+
super().__init__()
|
382 |
+
|
383 |
+
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding)
|
384 |
+
|
385 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
386 |
+
x = self.proj(x)
|
387 |
+
# B C H W -> B H W C
|
388 |
+
x = x.permute(0, 2, 3, 1)
|
389 |
+
return x
|
metaseg/modeling/mask_decoder.py
ADDED
@@ -0,0 +1,169 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from typing import List, Tuple, Type
|
8 |
+
|
9 |
+
import torch
|
10 |
+
from torch import nn
|
11 |
+
from torch.nn import functional as F
|
12 |
+
|
13 |
+
from metaseg.modeling.common import LayerNorm2d
|
14 |
+
|
15 |
+
|
16 |
+
class MaskDecoder(nn.Module):
|
17 |
+
def __init__(
|
18 |
+
self,
|
19 |
+
*,
|
20 |
+
transformer_dim: int,
|
21 |
+
transformer: nn.Module,
|
22 |
+
num_multimask_outputs: int = 3,
|
23 |
+
activation: Type[nn.Module] = nn.GELU,
|
24 |
+
iou_head_depth: int = 3,
|
25 |
+
iou_head_hidden_dim: int = 256,
|
26 |
+
) -> None:
|
27 |
+
"""
|
28 |
+
Predicts masks given an image and prompt embeddings, using a
|
29 |
+
tranformer architecture.
|
30 |
+
|
31 |
+
Arguments:
|
32 |
+
transformer_dim (int): the channel dimension of the transformer
|
33 |
+
transformer (nn.Module): the transformer used to predict masks
|
34 |
+
num_multimask_outputs (int): the number of masks to predict
|
35 |
+
when disambiguating masks
|
36 |
+
activation (nn.Module): the type of activation to use when
|
37 |
+
upscaling masks
|
38 |
+
iou_head_depth (int): the depth of the MLP used to predict
|
39 |
+
mask quality
|
40 |
+
iou_head_hidden_dim (int): the hidden dimension of the MLP
|
41 |
+
used to predict mask quality
|
42 |
+
"""
|
43 |
+
super().__init__()
|
44 |
+
self.transformer_dim = transformer_dim
|
45 |
+
self.transformer = transformer
|
46 |
+
|
47 |
+
self.num_multimask_outputs = num_multimask_outputs
|
48 |
+
|
49 |
+
self.iou_token = nn.Embedding(1, transformer_dim)
|
50 |
+
self.num_mask_tokens = num_multimask_outputs + 1
|
51 |
+
self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)
|
52 |
+
|
53 |
+
self.output_upscaling = nn.Sequential(
|
54 |
+
nn.ConvTranspose2d(transformer_dim, transformer_dim // 4, kernel_size=2, stride=2),
|
55 |
+
LayerNorm2d(transformer_dim // 4),
|
56 |
+
activation(),
|
57 |
+
nn.ConvTranspose2d(transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2),
|
58 |
+
activation(),
|
59 |
+
)
|
60 |
+
self.output_hypernetworks_mlps = nn.ModuleList(
|
61 |
+
[MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3) for i in range(self.num_mask_tokens)]
|
62 |
+
)
|
63 |
+
|
64 |
+
self.iou_prediction_head = MLP(transformer_dim, iou_head_hidden_dim, self.num_mask_tokens, iou_head_depth)
|
65 |
+
|
66 |
+
def forward(
|
67 |
+
self,
|
68 |
+
image_embeddings: torch.Tensor,
|
69 |
+
image_pe: torch.Tensor,
|
70 |
+
sparse_prompt_embeddings: torch.Tensor,
|
71 |
+
dense_prompt_embeddings: torch.Tensor,
|
72 |
+
multimask_output: bool,
|
73 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
74 |
+
"""
|
75 |
+
Predict masks given image and prompt embeddings.
|
76 |
+
|
77 |
+
Arguments:
|
78 |
+
image_embeddings (torch.Tensor): the embeddings from the image encoder
|
79 |
+
image_pe (torch.Tensor): positional encoding with the shape of image_embeddings
|
80 |
+
sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes
|
81 |
+
dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs
|
82 |
+
multimask_output (bool): Whether to return multiple masks or a single
|
83 |
+
mask.
|
84 |
+
|
85 |
+
Returns:
|
86 |
+
torch.Tensor: batched predicted masks
|
87 |
+
torch.Tensor: batched predictions of mask quality
|
88 |
+
"""
|
89 |
+
masks, iou_pred = self.predict_masks(
|
90 |
+
image_embeddings=image_embeddings,
|
91 |
+
image_pe=image_pe,
|
92 |
+
sparse_prompt_embeddings=sparse_prompt_embeddings,
|
93 |
+
dense_prompt_embeddings=dense_prompt_embeddings,
|
94 |
+
)
|
95 |
+
|
96 |
+
# Select the correct mask or masks for outptu
|
97 |
+
if multimask_output:
|
98 |
+
mask_slice = slice(1, None)
|
99 |
+
else:
|
100 |
+
mask_slice = slice(0, 1)
|
101 |
+
masks = masks[:, mask_slice, :, :]
|
102 |
+
iou_pred = iou_pred[:, mask_slice]
|
103 |
+
|
104 |
+
# Prepare output
|
105 |
+
return masks, iou_pred
|
106 |
+
|
107 |
+
def predict_masks(
|
108 |
+
self,
|
109 |
+
image_embeddings: torch.Tensor,
|
110 |
+
image_pe: torch.Tensor,
|
111 |
+
sparse_prompt_embeddings: torch.Tensor,
|
112 |
+
dense_prompt_embeddings: torch.Tensor,
|
113 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
114 |
+
"""Predicts masks. See 'forward' for more details."""
|
115 |
+
# Concatenate output tokens
|
116 |
+
output_tokens = torch.cat([self.iou_token.weight, self.mask_tokens.weight], dim=0)
|
117 |
+
output_tokens = output_tokens.unsqueeze(0).expand(sparse_prompt_embeddings.size(0), -1, -1)
|
118 |
+
tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)
|
119 |
+
|
120 |
+
# Expand per-image data in batch direction to be per-mask
|
121 |
+
src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
|
122 |
+
src = src + dense_prompt_embeddings
|
123 |
+
pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
|
124 |
+
b, c, h, w = src.shape
|
125 |
+
|
126 |
+
# Run the transformer
|
127 |
+
hs, src = self.transformer(src, pos_src, tokens)
|
128 |
+
iou_token_out = hs[:, 0, :]
|
129 |
+
mask_tokens_out = hs[:, 1 : (1 + self.num_mask_tokens), :]
|
130 |
+
|
131 |
+
# Upscale mask embeddings and predict masks using the mask tokens
|
132 |
+
src = src.transpose(1, 2).view(b, c, h, w)
|
133 |
+
upscaled_embedding = self.output_upscaling(src)
|
134 |
+
hyper_in_list: List[torch.Tensor] = []
|
135 |
+
for i in range(self.num_mask_tokens):
|
136 |
+
hyper_in_list.append(self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :]))
|
137 |
+
hyper_in = torch.stack(hyper_in_list, dim=1)
|
138 |
+
b, c, h, w = upscaled_embedding.shape
|
139 |
+
masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)
|
140 |
+
|
141 |
+
# Generate mask quality predictions
|
142 |
+
iou_pred = self.iou_prediction_head(iou_token_out)
|
143 |
+
|
144 |
+
return masks, iou_pred
|
145 |
+
|
146 |
+
|
147 |
+
# Lightly adapted from
|
148 |
+
# https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa
|
149 |
+
class MLP(nn.Module):
|
150 |
+
def __init__(
|
151 |
+
self,
|
152 |
+
input_dim: int,
|
153 |
+
hidden_dim: int,
|
154 |
+
output_dim: int,
|
155 |
+
num_layers: int,
|
156 |
+
sigmoid_output: bool = False,
|
157 |
+
) -> None:
|
158 |
+
super().__init__()
|
159 |
+
self.num_layers = num_layers
|
160 |
+
h = [hidden_dim] * (num_layers - 1)
|
161 |
+
self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
|
162 |
+
self.sigmoid_output = sigmoid_output
|
163 |
+
|
164 |
+
def forward(self, x):
|
165 |
+
for i, layer in enumerate(self.layers):
|
166 |
+
x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
|
167 |
+
if self.sigmoid_output:
|
168 |
+
x = F.sigmoid(x)
|
169 |
+
return x
|
metaseg/modeling/prompt_encoder.py
ADDED
@@ -0,0 +1,212 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from typing import Any, Optional, Tuple, Type
|
8 |
+
|
9 |
+
import numpy as np
|
10 |
+
import torch
|
11 |
+
from torch import nn
|
12 |
+
|
13 |
+
from metaseg.modeling.common import LayerNorm2d
|
14 |
+
|
15 |
+
|
16 |
+
class PromptEncoder(nn.Module):
|
17 |
+
def __init__(
|
18 |
+
self,
|
19 |
+
embed_dim: int,
|
20 |
+
image_embedding_size: Tuple[int, int],
|
21 |
+
input_image_size: Tuple[int, int],
|
22 |
+
mask_in_chans: int,
|
23 |
+
activation: Type[nn.Module] = nn.GELU,
|
24 |
+
) -> None:
|
25 |
+
"""
|
26 |
+
Encodes prompts for input to SAM's mask decoder.
|
27 |
+
|
28 |
+
Arguments:
|
29 |
+
embed_dim (int): The prompts' embedding dimension
|
30 |
+
image_embedding_size (tuple(int, int)): The spatial size of the
|
31 |
+
image embedding, as (H, W).
|
32 |
+
input_image_size (int): The padded size of the image as input
|
33 |
+
to the image encoder, as (H, W).
|
34 |
+
mask_in_chans (int): The number of hidden channels used for
|
35 |
+
encoding input masks.
|
36 |
+
activation (nn.Module): The activation to use when encoding
|
37 |
+
input masks.
|
38 |
+
"""
|
39 |
+
super().__init__()
|
40 |
+
self.embed_dim = embed_dim
|
41 |
+
self.input_image_size = input_image_size
|
42 |
+
self.image_embedding_size = image_embedding_size
|
43 |
+
self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)
|
44 |
+
|
45 |
+
self.num_point_embeddings: int = 4 # pos/neg point + 2 box corners
|
46 |
+
point_embeddings = [nn.Embedding(1, embed_dim) for i in range(self.num_point_embeddings)]
|
47 |
+
self.point_embeddings = nn.ModuleList(point_embeddings)
|
48 |
+
self.not_a_point_embed = nn.Embedding(1, embed_dim)
|
49 |
+
|
50 |
+
self.mask_input_size = (4 * image_embedding_size[0], 4 * image_embedding_size[1])
|
51 |
+
self.mask_downscaling = nn.Sequential(
|
52 |
+
nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),
|
53 |
+
LayerNorm2d(mask_in_chans // 4),
|
54 |
+
activation(),
|
55 |
+
nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),
|
56 |
+
LayerNorm2d(mask_in_chans),
|
57 |
+
activation(),
|
58 |
+
nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),
|
59 |
+
)
|
60 |
+
self.no_mask_embed = nn.Embedding(1, embed_dim)
|
61 |
+
|
62 |
+
def get_dense_pe(self) -> torch.Tensor:
|
63 |
+
"""
|
64 |
+
Returns the positional encoding used to encode point prompts,
|
65 |
+
applied to a dense set of points the shape of the image encoding.
|
66 |
+
|
67 |
+
Returns:
|
68 |
+
torch.Tensor: Positional encoding with shape
|
69 |
+
1x(embed_dim)x(embedding_h)x(embedding_w)
|
70 |
+
"""
|
71 |
+
return self.pe_layer(self.image_embedding_size).unsqueeze(0)
|
72 |
+
|
73 |
+
def _embed_points(
|
74 |
+
self,
|
75 |
+
points: torch.Tensor,
|
76 |
+
labels: torch.Tensor,
|
77 |
+
pad: bool,
|
78 |
+
) -> torch.Tensor:
|
79 |
+
"""Embeds point prompts."""
|
80 |
+
points = points + 0.5 # Shift to center of pixel
|
81 |
+
if pad:
|
82 |
+
padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)
|
83 |
+
padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)
|
84 |
+
points = torch.cat([points, padding_point], dim=1)
|
85 |
+
labels = torch.cat([labels, padding_label], dim=1)
|
86 |
+
point_embedding = self.pe_layer.forward_with_coords(points, self.input_image_size)
|
87 |
+
point_embedding[labels == -1] = 0.0
|
88 |
+
point_embedding[labels == -1] += self.not_a_point_embed.weight
|
89 |
+
point_embedding[labels == 0] += self.point_embeddings[0].weight
|
90 |
+
point_embedding[labels == 1] += self.point_embeddings[1].weight
|
91 |
+
return point_embedding
|
92 |
+
|
93 |
+
def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
|
94 |
+
"""Embeds box prompts."""
|
95 |
+
boxes = boxes + 0.5 # Shift to center of pixel
|
96 |
+
coords = boxes.reshape(-1, 2, 2)
|
97 |
+
corner_embedding = self.pe_layer.forward_with_coords(coords, self.input_image_size)
|
98 |
+
corner_embedding[:, 0, :] += self.point_embeddings[2].weight
|
99 |
+
corner_embedding[:, 1, :] += self.point_embeddings[3].weight
|
100 |
+
return corner_embedding
|
101 |
+
|
102 |
+
def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:
|
103 |
+
"""Embeds mask inputs."""
|
104 |
+
mask_embedding = self.mask_downscaling(masks)
|
105 |
+
return mask_embedding
|
106 |
+
|
107 |
+
def _get_batch_size(
|
108 |
+
self,
|
109 |
+
points: Optional[Tuple[torch.Tensor, torch.Tensor]],
|
110 |
+
boxes: Optional[torch.Tensor],
|
111 |
+
masks: Optional[torch.Tensor],
|
112 |
+
) -> int:
|
113 |
+
"""
|
114 |
+
Gets the batch size of the output given the batch size of the input prompts.
|
115 |
+
"""
|
116 |
+
if points is not None:
|
117 |
+
return points[0].shape[0]
|
118 |
+
elif boxes is not None:
|
119 |
+
return boxes.shape[0]
|
120 |
+
elif masks is not None:
|
121 |
+
return masks.shape[0]
|
122 |
+
else:
|
123 |
+
return 1
|
124 |
+
|
125 |
+
def _get_device(self) -> torch.device:
|
126 |
+
return self.point_embeddings[0].weight.device
|
127 |
+
|
128 |
+
def forward(
|
129 |
+
self,
|
130 |
+
points: Optional[Tuple[torch.Tensor, torch.Tensor]],
|
131 |
+
boxes: Optional[torch.Tensor],
|
132 |
+
masks: Optional[torch.Tensor],
|
133 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
134 |
+
"""
|
135 |
+
Embeds different types of prompts, returning both sparse and dense
|
136 |
+
embeddings.
|
137 |
+
|
138 |
+
Arguments:
|
139 |
+
points (tuple(torch.Tensor, torch.Tensor) or none): point coordinates
|
140 |
+
and labels to embed.
|
141 |
+
boxes (torch.Tensor or none): boxes to embed
|
142 |
+
masks (torch.Tensor or none): masks to embed
|
143 |
+
|
144 |
+
Returns:
|
145 |
+
torch.Tensor: sparse embeddings for the points and boxes, with shape
|
146 |
+
BxNx(embed_dim), where N is determined by the number of input points
|
147 |
+
and boxes.
|
148 |
+
torch.Tensor: dense embeddings for the masks, in the shape
|
149 |
+
Bx(embed_dim)x(embed_H)x(embed_W)
|
150 |
+
"""
|
151 |
+
bs = self._get_batch_size(points, boxes, masks)
|
152 |
+
sparse_embeddings = torch.empty((bs, 0, self.embed_dim), device=self._get_device())
|
153 |
+
if points is not None:
|
154 |
+
coords, labels = points
|
155 |
+
point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))
|
156 |
+
sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)
|
157 |
+
if boxes is not None:
|
158 |
+
box_embeddings = self._embed_boxes(boxes)
|
159 |
+
sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)
|
160 |
+
|
161 |
+
if masks is not None:
|
162 |
+
dense_embeddings = self._embed_masks(masks)
|
163 |
+
else:
|
164 |
+
dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(
|
165 |
+
bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]
|
166 |
+
)
|
167 |
+
|
168 |
+
return sparse_embeddings, dense_embeddings
|
169 |
+
|
170 |
+
|
171 |
+
class PositionEmbeddingRandom(nn.Module):
|
172 |
+
"""
|
173 |
+
Positional encoding using random spatial frequencies.
|
174 |
+
"""
|
175 |
+
|
176 |
+
def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
|
177 |
+
super().__init__()
|
178 |
+
if scale is None or scale <= 0.0:
|
179 |
+
scale = 1.0
|
180 |
+
self.register_buffer(
|
181 |
+
"positional_encoding_gaussian_matrix",
|
182 |
+
scale * torch.randn((2, num_pos_feats)),
|
183 |
+
)
|
184 |
+
|
185 |
+
def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
|
186 |
+
"""Positionally encode points that are normalized to [0,1]."""
|
187 |
+
# assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape
|
188 |
+
coords = 2 * coords - 1
|
189 |
+
coords = coords @ self.positional_encoding_gaussian_matrix
|
190 |
+
coords = 2 * np.pi * coords
|
191 |
+
# outputs d_1 x ... x d_n x C shape
|
192 |
+
return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
|
193 |
+
|
194 |
+
def forward(self, size: Tuple[int, int]) -> torch.Tensor:
|
195 |
+
"""Generate positional encoding for a grid of the specified size."""
|
196 |
+
h, w = size
|
197 |
+
device: Any = self.positional_encoding_gaussian_matrix.device
|
198 |
+
grid = torch.ones((h, w), device=device, dtype=torch.float32)
|
199 |
+
y_embed = grid.cumsum(dim=0) - 0.5
|
200 |
+
x_embed = grid.cumsum(dim=1) - 0.5
|
201 |
+
y_embed = y_embed / h
|
202 |
+
x_embed = x_embed / w
|
203 |
+
|
204 |
+
pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
|
205 |
+
return pe.permute(2, 0, 1) # C x H x W
|
206 |
+
|
207 |
+
def forward_with_coords(self, coords_input: torch.Tensor, image_size: Tuple[int, int]) -> torch.Tensor:
|
208 |
+
"""Positionally encode points that are not normalized to [0,1]."""
|
209 |
+
coords = coords_input.clone()
|
210 |
+
coords[:, :, 0] = coords[:, :, 0] / image_size[1]
|
211 |
+
coords[:, :, 1] = coords[:, :, 1] / image_size[0]
|
212 |
+
return self._pe_encoding(coords.to(torch.float)) # B x N x C
|
metaseg/modeling/sam.py
ADDED
@@ -0,0 +1,174 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from typing import Any, Dict, List, Tuple
|
8 |
+
|
9 |
+
import torch
|
10 |
+
from torch import nn
|
11 |
+
from torch.nn import functional as F
|
12 |
+
|
13 |
+
from metaseg.modeling.image_encoder import ImageEncoderViT
|
14 |
+
from metaseg.modeling.mask_decoder import MaskDecoder
|
15 |
+
from metaseg.modeling.prompt_encoder import PromptEncoder
|
16 |
+
|
17 |
+
|
18 |
+
class Sam(nn.Module):
|
19 |
+
mask_threshold: float = 0.0
|
20 |
+
image_format: str = "RGB"
|
21 |
+
|
22 |
+
def __init__(
|
23 |
+
self,
|
24 |
+
image_encoder: ImageEncoderViT,
|
25 |
+
prompt_encoder: PromptEncoder,
|
26 |
+
mask_decoder: MaskDecoder,
|
27 |
+
pixel_mean: List[float] = [123.675, 116.28, 103.53],
|
28 |
+
pixel_std: List[float] = [58.395, 57.12, 57.375],
|
29 |
+
) -> None:
|
30 |
+
"""
|
31 |
+
SAM predicts object masks from an image and input prompts.
|
32 |
+
|
33 |
+
Arguments:
|
34 |
+
image_encoder (ImageEncoderViT): The backbone used to encode the
|
35 |
+
image into image embeddings that allow for efficient mask prediction.
|
36 |
+
prompt_encoder (PromptEncoder): Encodes various types of input prompts.
|
37 |
+
mask_decoder (MaskDecoder): Predicts masks from the image embeddings
|
38 |
+
and encoded prompts.
|
39 |
+
pixel_mean (list(float)): Mean values for normalizing pixels in the input image.
|
40 |
+
pixel_std (list(float)): Std values for normalizing pixels in the input image.
|
41 |
+
"""
|
42 |
+
super().__init__()
|
43 |
+
self.image_encoder = image_encoder
|
44 |
+
self.prompt_encoder = prompt_encoder
|
45 |
+
self.mask_decoder = mask_decoder
|
46 |
+
self.register_buffer("pixel_mean", torch.Tensor(pixel_mean).view(-1, 1, 1), False)
|
47 |
+
self.register_buffer("pixel_std", torch.Tensor(pixel_std).view(-1, 1, 1), False)
|
48 |
+
|
49 |
+
@property
|
50 |
+
def device(self) -> Any:
|
51 |
+
return self.pixel_mean.device
|
52 |
+
|
53 |
+
@torch.no_grad()
|
54 |
+
def forward(
|
55 |
+
self,
|
56 |
+
batched_input: List[Dict[str, Any]],
|
57 |
+
multimask_output: bool,
|
58 |
+
) -> List[Dict[str, torch.Tensor]]:
|
59 |
+
"""
|
60 |
+
Predicts masks end-to-end from provided images and prompts.
|
61 |
+
If prompts are not known in advance, using SamPredictor is
|
62 |
+
recommended over calling the model directly.
|
63 |
+
|
64 |
+
Arguments:
|
65 |
+
batched_input (list(dict)): A list over input images, each a
|
66 |
+
dictionary with the following keys. A prompt key can be
|
67 |
+
excluded if it is not present.
|
68 |
+
'image': The image as a torch tensor in 3xHxW format,
|
69 |
+
already transformed for input to the model.
|
70 |
+
'original_size': (tuple(int, int)) The original size of
|
71 |
+
the image before transformation, as (H, W).
|
72 |
+
'point_coords': (torch.Tensor) Batched point prompts for
|
73 |
+
this image, with shape BxNx2. Already transformed to the
|
74 |
+
input frame of the model.
|
75 |
+
'point_labels': (torch.Tensor) Batched labels for point prompts,
|
76 |
+
with shape BxN.
|
77 |
+
'boxes': (torch.Tensor) Batched box inputs, with shape Bx4.
|
78 |
+
Already transformed to the input frame of the model.
|
79 |
+
'mask_inputs': (torch.Tensor) Batched mask inputs to the model,
|
80 |
+
in the form Bx1xHxW.
|
81 |
+
multimask_output (bool): Whether the model should predict multiple
|
82 |
+
disambiguating masks, or return a single mask.
|
83 |
+
|
84 |
+
Returns:
|
85 |
+
(list(dict)): A list over input images, where each element is
|
86 |
+
as dictionary with the following keys.
|
87 |
+
'masks': (torch.Tensor) Batched binary mask predictions,
|
88 |
+
with shape BxCxHxW, where B is the number of input promts,
|
89 |
+
C is determiend by multimask_output, and (H, W) is the
|
90 |
+
original size of the image.
|
91 |
+
'iou_predictions': (torch.Tensor) The model's predictions
|
92 |
+
of mask quality, in shape BxC.
|
93 |
+
'low_res_logits': (torch.Tensor) Low resolution logits with
|
94 |
+
shape BxCxHxW, where H=W=256. Can be passed as mask input
|
95 |
+
to subsequent iterations of prediction.
|
96 |
+
"""
|
97 |
+
input_images = torch.stack([self.preprocess(x["image"]) for x in batched_input], dim=0)
|
98 |
+
image_embeddings = self.image_encoder(input_images)
|
99 |
+
|
100 |
+
outputs = []
|
101 |
+
for image_record, curr_embedding in zip(batched_input, image_embeddings):
|
102 |
+
if "point_coords" in image_record:
|
103 |
+
points = (image_record["point_coords"], image_record["point_labels"])
|
104 |
+
else:
|
105 |
+
points = None
|
106 |
+
sparse_embeddings, dense_embeddings = self.prompt_encoder(
|
107 |
+
points=points,
|
108 |
+
boxes=image_record.get("boxes", None),
|
109 |
+
masks=image_record.get("mask_inputs", None),
|
110 |
+
)
|
111 |
+
low_res_masks, iou_predictions = self.mask_decoder(
|
112 |
+
image_embeddings=curr_embedding.unsqueeze(0),
|
113 |
+
image_pe=self.prompt_encoder.get_dense_pe(),
|
114 |
+
sparse_prompt_embeddings=sparse_embeddings,
|
115 |
+
dense_prompt_embeddings=dense_embeddings,
|
116 |
+
multimask_output=multimask_output,
|
117 |
+
)
|
118 |
+
masks = self.postprocess_masks(
|
119 |
+
low_res_masks,
|
120 |
+
input_size=image_record["image"].shape[-2:],
|
121 |
+
original_size=image_record["original_size"],
|
122 |
+
)
|
123 |
+
masks = masks > self.mask_threshold
|
124 |
+
outputs.append(
|
125 |
+
{
|
126 |
+
"masks": masks,
|
127 |
+
"iou_predictions": iou_predictions,
|
128 |
+
"low_res_logits": low_res_masks,
|
129 |
+
}
|
130 |
+
)
|
131 |
+
return outputs
|
132 |
+
|
133 |
+
def postprocess_masks(
|
134 |
+
self,
|
135 |
+
masks: torch.Tensor,
|
136 |
+
input_size: Tuple[int, ...],
|
137 |
+
original_size: Tuple[int, ...],
|
138 |
+
) -> torch.Tensor:
|
139 |
+
"""
|
140 |
+
Remove padding and upscale masks to the original image size.
|
141 |
+
|
142 |
+
Arguments:
|
143 |
+
masks (torch.Tensor): Batched masks from the mask_decoder,
|
144 |
+
in BxCxHxW format.
|
145 |
+
input_size (tuple(int, int)): The size of the image input to the
|
146 |
+
model, in (H, W) format. Used to remove padding.
|
147 |
+
original_size (tuple(int, int)): The original size of the image
|
148 |
+
before resizing for input to the model, in (H, W) format.
|
149 |
+
|
150 |
+
Returns:
|
151 |
+
(torch.Tensor): Batched masks in BxCxHxW format, where (H, W)
|
152 |
+
is given by original_size.
|
153 |
+
"""
|
154 |
+
masks = F.interpolate(
|
155 |
+
masks,
|
156 |
+
(self.image_encoder.img_size, self.image_encoder.img_size),
|
157 |
+
mode="bilinear",
|
158 |
+
align_corners=False,
|
159 |
+
)
|
160 |
+
masks = masks[..., : input_size[0], : input_size[1]]
|
161 |
+
masks = F.interpolate(masks, original_size, mode="bilinear", align_corners=False)
|
162 |
+
return masks
|
163 |
+
|
164 |
+
def preprocess(self, x: torch.Tensor) -> torch.Tensor:
|
165 |
+
"""Normalize pixel values and pad to a square input."""
|
166 |
+
# Normalize colors
|
167 |
+
x = (x - self.pixel_mean) / self.pixel_std
|
168 |
+
|
169 |
+
# Pad
|
170 |
+
h, w = x.shape[-2:]
|
171 |
+
padh = self.image_encoder.img_size - h
|
172 |
+
padw = self.image_encoder.img_size - w
|
173 |
+
x = F.pad(x, (0, padw, 0, padh))
|
174 |
+
return x
|
metaseg/modeling/transformer.py
ADDED
@@ -0,0 +1,232 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
import math
|
8 |
+
from typing import Tuple, Type
|
9 |
+
|
10 |
+
import torch
|
11 |
+
from torch import Tensor, nn
|
12 |
+
|
13 |
+
from metaseg.modeling.common import MLPBlock
|
14 |
+
|
15 |
+
|
16 |
+
class TwoWayTransformer(nn.Module):
|
17 |
+
def __init__(
|
18 |
+
self,
|
19 |
+
depth: int,
|
20 |
+
embedding_dim: int,
|
21 |
+
num_heads: int,
|
22 |
+
mlp_dim: int,
|
23 |
+
activation: Type[nn.Module] = nn.ReLU,
|
24 |
+
attention_downsample_rate: int = 2,
|
25 |
+
) -> None:
|
26 |
+
"""
|
27 |
+
A transformer decoder that attends to an input image using
|
28 |
+
queries whose positional embedding is supplied.
|
29 |
+
|
30 |
+
Args:
|
31 |
+
depth (int): number of layers in the transformer
|
32 |
+
embedding_dim (int): the channel dimension for the input embeddings
|
33 |
+
num_heads (int): the number of heads for multihead attention. Must
|
34 |
+
divide embedding_dim
|
35 |
+
mlp_dim (int): the channel dimension internal to the MLP block
|
36 |
+
activation (nn.Module): the activation to use in the MLP block
|
37 |
+
"""
|
38 |
+
super().__init__()
|
39 |
+
self.depth = depth
|
40 |
+
self.embedding_dim = embedding_dim
|
41 |
+
self.num_heads = num_heads
|
42 |
+
self.mlp_dim = mlp_dim
|
43 |
+
self.layers = nn.ModuleList()
|
44 |
+
|
45 |
+
for i in range(depth):
|
46 |
+
self.layers.append(
|
47 |
+
TwoWayAttentionBlock(
|
48 |
+
embedding_dim=embedding_dim,
|
49 |
+
num_heads=num_heads,
|
50 |
+
mlp_dim=mlp_dim,
|
51 |
+
activation=activation,
|
52 |
+
attention_downsample_rate=attention_downsample_rate,
|
53 |
+
skip_first_layer_pe=(i == 0),
|
54 |
+
)
|
55 |
+
)
|
56 |
+
|
57 |
+
self.final_attn_token_to_image = Attention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate)
|
58 |
+
self.norm_final_attn = nn.LayerNorm(embedding_dim)
|
59 |
+
|
60 |
+
def forward(
|
61 |
+
self,
|
62 |
+
image_embedding: Tensor,
|
63 |
+
image_pe: Tensor,
|
64 |
+
point_embedding: Tensor,
|
65 |
+
) -> Tuple[Tensor, Tensor]:
|
66 |
+
"""
|
67 |
+
Args:
|
68 |
+
image_embedding (torch.Tensor): image to attend to. Should be shape
|
69 |
+
B x embedding_dim x h x w for any h and w.
|
70 |
+
image_pe (torch.Tensor): the positional encoding to add to the image. Must
|
71 |
+
have the same shape as image_embedding.
|
72 |
+
point_embedding (torch.Tensor): the embedding to add to the query points.
|
73 |
+
Must have shape B x N_points x embedding_dim for any N_points.
|
74 |
+
|
75 |
+
Returns:
|
76 |
+
torch.Tensor: the processed point_embedding
|
77 |
+
torch.Tensor: the processed image_embedding
|
78 |
+
"""
|
79 |
+
# BxCxHxW -> BxHWxC == B x N_image_tokens x C
|
80 |
+
bs, c, h, w = image_embedding.shape
|
81 |
+
image_embedding = image_embedding.flatten(2).permute(0, 2, 1)
|
82 |
+
image_pe = image_pe.flatten(2).permute(0, 2, 1)
|
83 |
+
|
84 |
+
# Prepare queries
|
85 |
+
queries = point_embedding
|
86 |
+
keys = image_embedding
|
87 |
+
|
88 |
+
# Apply transformer blocks and final layernorm
|
89 |
+
for layer in self.layers:
|
90 |
+
queries, keys = layer(
|
91 |
+
queries=queries,
|
92 |
+
keys=keys,
|
93 |
+
query_pe=point_embedding,
|
94 |
+
key_pe=image_pe,
|
95 |
+
)
|
96 |
+
|
97 |
+
# Apply the final attenion layer from the points to the image
|
98 |
+
q = queries + point_embedding
|
99 |
+
k = keys + image_pe
|
100 |
+
attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)
|
101 |
+
queries = queries + attn_out
|
102 |
+
queries = self.norm_final_attn(queries)
|
103 |
+
|
104 |
+
return queries, keys
|
105 |
+
|
106 |
+
|
107 |
+
class TwoWayAttentionBlock(nn.Module):
|
108 |
+
def __init__(
|
109 |
+
self,
|
110 |
+
embedding_dim: int,
|
111 |
+
num_heads: int,
|
112 |
+
mlp_dim: int = 2048,
|
113 |
+
activation: Type[nn.Module] = nn.ReLU,
|
114 |
+
attention_downsample_rate: int = 2,
|
115 |
+
skip_first_layer_pe: bool = False,
|
116 |
+
) -> None:
|
117 |
+
"""
|
118 |
+
A transformer block with four layers: (1) self-attention of sparse
|
119 |
+
inputs, (2) cross attention of sparse inputs to dense inputs, (3) mlp
|
120 |
+
block on sparse inputs, and (4) cross attention of dense inputs to sparse
|
121 |
+
inputs.
|
122 |
+
|
123 |
+
Arguments:
|
124 |
+
embedding_dim (int): the channel dimension of the embeddings
|
125 |
+
num_heads (int): the number of heads in the attention layers
|
126 |
+
mlp_dim (int): the hidden dimension of the mlp block
|
127 |
+
activation (nn.Module): the activation of the mlp block
|
128 |
+
skip_first_layer_pe (bool): skip the PE on the first layer
|
129 |
+
"""
|
130 |
+
super().__init__()
|
131 |
+
self.self_attn = Attention(embedding_dim, num_heads)
|
132 |
+
self.norm1 = nn.LayerNorm(embedding_dim)
|
133 |
+
|
134 |
+
self.cross_attn_token_to_image = Attention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate)
|
135 |
+
self.norm2 = nn.LayerNorm(embedding_dim)
|
136 |
+
|
137 |
+
self.mlp = MLPBlock(embedding_dim, mlp_dim, activation)
|
138 |
+
self.norm3 = nn.LayerNorm(embedding_dim)
|
139 |
+
|
140 |
+
self.norm4 = nn.LayerNorm(embedding_dim)
|
141 |
+
self.cross_attn_image_to_token = Attention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate)
|
142 |
+
|
143 |
+
self.skip_first_layer_pe = skip_first_layer_pe
|
144 |
+
|
145 |
+
def forward(self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor) -> Tuple[Tensor, Tensor]:
|
146 |
+
# Self attention block
|
147 |
+
if self.skip_first_layer_pe:
|
148 |
+
queries = self.self_attn(q=queries, k=queries, v=queries)
|
149 |
+
else:
|
150 |
+
q = queries + query_pe
|
151 |
+
attn_out = self.self_attn(q=q, k=q, v=queries)
|
152 |
+
queries = queries + attn_out
|
153 |
+
queries = self.norm1(queries)
|
154 |
+
|
155 |
+
# Cross attention block, tokens attending to image embedding
|
156 |
+
q = queries + query_pe
|
157 |
+
k = keys + key_pe
|
158 |
+
attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)
|
159 |
+
queries = queries + attn_out
|
160 |
+
queries = self.norm2(queries)
|
161 |
+
|
162 |
+
# MLP block
|
163 |
+
mlp_out = self.mlp(queries)
|
164 |
+
queries = queries + mlp_out
|
165 |
+
queries = self.norm3(queries)
|
166 |
+
|
167 |
+
# Cross attention block, image embedding attending to tokens
|
168 |
+
q = queries + query_pe
|
169 |
+
k = keys + key_pe
|
170 |
+
attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)
|
171 |
+
keys = keys + attn_out
|
172 |
+
keys = self.norm4(keys)
|
173 |
+
|
174 |
+
return queries, keys
|
175 |
+
|
176 |
+
|
177 |
+
class Attention(nn.Module):
|
178 |
+
"""
|
179 |
+
An attention layer that allows for downscaling the size of the embedding
|
180 |
+
after projection to queries, keys, and values.
|
181 |
+
"""
|
182 |
+
|
183 |
+
def __init__(
|
184 |
+
self,
|
185 |
+
embedding_dim: int,
|
186 |
+
num_heads: int,
|
187 |
+
downsample_rate: int = 1,
|
188 |
+
) -> None:
|
189 |
+
super().__init__()
|
190 |
+
self.embedding_dim = embedding_dim
|
191 |
+
self.internal_dim = embedding_dim // downsample_rate
|
192 |
+
self.num_heads = num_heads
|
193 |
+
assert self.internal_dim % num_heads == 0, "num_heads must divide embedding_dim."
|
194 |
+
|
195 |
+
self.q_proj = nn.Linear(embedding_dim, self.internal_dim)
|
196 |
+
self.k_proj = nn.Linear(embedding_dim, self.internal_dim)
|
197 |
+
self.v_proj = nn.Linear(embedding_dim, self.internal_dim)
|
198 |
+
self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
|
199 |
+
|
200 |
+
def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:
|
201 |
+
b, n, c = x.shape
|
202 |
+
x = x.reshape(b, n, num_heads, c // num_heads)
|
203 |
+
return x.transpose(1, 2) # B x N_heads x N_tokens x C_per_head
|
204 |
+
|
205 |
+
def _recombine_heads(self, x: Tensor) -> Tensor:
|
206 |
+
b, n_heads, n_tokens, c_per_head = x.shape
|
207 |
+
x = x.transpose(1, 2)
|
208 |
+
return x.reshape(b, n_tokens, n_heads * c_per_head) # B x N_tokens x C
|
209 |
+
|
210 |
+
def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:
|
211 |
+
# Input projections
|
212 |
+
q = self.q_proj(q)
|
213 |
+
k = self.k_proj(k)
|
214 |
+
v = self.v_proj(v)
|
215 |
+
|
216 |
+
# Separate into heads
|
217 |
+
q = self._separate_heads(q, self.num_heads)
|
218 |
+
k = self._separate_heads(k, self.num_heads)
|
219 |
+
v = self._separate_heads(v, self.num_heads)
|
220 |
+
|
221 |
+
# Attention
|
222 |
+
_, _, _, c_per_head = q.shape
|
223 |
+
attn = q @ k.permute(0, 1, 3, 2) # B x N_heads x N_tokens x N_tokens
|
224 |
+
attn = attn / math.sqrt(c_per_head)
|
225 |
+
attn = torch.softmax(attn, dim=-1)
|
226 |
+
|
227 |
+
# Get output
|
228 |
+
out = attn @ v
|
229 |
+
out = self._recombine_heads(out)
|
230 |
+
out = self.out_proj(out)
|
231 |
+
|
232 |
+
return out
|
metaseg/predictor.py
ADDED
@@ -0,0 +1,264 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from typing import Optional, Tuple
|
8 |
+
|
9 |
+
import numpy as np
|
10 |
+
import torch
|
11 |
+
|
12 |
+
from metaseg.modeling import Sam
|
13 |
+
from metaseg.utils.transforms import ResizeLongestSide
|
14 |
+
|
15 |
+
|
16 |
+
class SamPredictor:
|
17 |
+
def __init__(
|
18 |
+
self,
|
19 |
+
sam_model: Sam,
|
20 |
+
) -> None:
|
21 |
+
"""
|
22 |
+
Uses SAM to calculate the image embedding for an image, and then
|
23 |
+
allow repeated, efficient mask prediction given prompts.
|
24 |
+
|
25 |
+
Arguments:
|
26 |
+
sam_model (Sam): The model to use for mask prediction.
|
27 |
+
"""
|
28 |
+
super().__init__()
|
29 |
+
self.model = sam_model
|
30 |
+
self.transform = ResizeLongestSide(sam_model.image_encoder.img_size)
|
31 |
+
self.reset_image()
|
32 |
+
|
33 |
+
def set_image(
|
34 |
+
self,
|
35 |
+
image: np.ndarray,
|
36 |
+
image_format: str = "RGB",
|
37 |
+
) -> None:
|
38 |
+
"""
|
39 |
+
Calculates the image embeddings for the provided image, allowing
|
40 |
+
masks to be predicted with the 'predict' method.
|
41 |
+
|
42 |
+
Arguments:
|
43 |
+
image (np.ndarray): The image for calculating masks. Expects an
|
44 |
+
image in HWC uint8 format, with pixel values in [0, 255].
|
45 |
+
image_format (str): The color format of the image, in ['RGB', 'BGR'].
|
46 |
+
"""
|
47 |
+
assert image_format in [
|
48 |
+
"RGB",
|
49 |
+
"BGR",
|
50 |
+
], f"image_format must be in ['RGB', 'BGR'], is {image_format}."
|
51 |
+
if image_format != self.model.image_format:
|
52 |
+
image = image[..., ::-1]
|
53 |
+
|
54 |
+
# Transform the image to the form expected by the model
|
55 |
+
input_image = self.transform.apply_image(image)
|
56 |
+
input_image_torch = torch.as_tensor(input_image, device=self.device)
|
57 |
+
input_image_torch = input_image_torch.permute(2, 0, 1).contiguous()[None, :, :, :]
|
58 |
+
|
59 |
+
self.set_torch_image(input_image_torch, image.shape[:2])
|
60 |
+
|
61 |
+
@torch.no_grad()
|
62 |
+
def set_torch_image(
|
63 |
+
self,
|
64 |
+
transformed_image: torch.Tensor,
|
65 |
+
original_image_size: Tuple[int, ...],
|
66 |
+
) -> None:
|
67 |
+
"""
|
68 |
+
Calculates the image embeddings for the provided image, allowing
|
69 |
+
masks to be predicted with the 'predict' method. Expects the input
|
70 |
+
image to be already transformed to the format expected by the model.
|
71 |
+
|
72 |
+
Arguments:
|
73 |
+
transformed_image (torch.Tensor): The input image, with shape
|
74 |
+
1x3xHxW, which has been transformed with ResizeLongestSide.
|
75 |
+
original_image_size (tuple(int, int)): The size of the image
|
76 |
+
before transformation, in (H, W) format.
|
77 |
+
"""
|
78 |
+
assert (
|
79 |
+
len(transformed_image.shape) == 4
|
80 |
+
and transformed_image.shape[1] == 3
|
81 |
+
and max(*transformed_image.shape[2:]) == self.model.image_encoder.img_size
|
82 |
+
), f"set_torch_image input must be BCHW with long side {self.model.image_encoder.img_size}."
|
83 |
+
self.reset_image()
|
84 |
+
|
85 |
+
self.original_size = original_image_size
|
86 |
+
self.input_size = tuple(transformed_image.shape[-2:])
|
87 |
+
input_image = self.model.preprocess(transformed_image)
|
88 |
+
self.features = self.model.image_encoder(input_image)
|
89 |
+
self.is_image_set = True
|
90 |
+
|
91 |
+
def predict(
|
92 |
+
self,
|
93 |
+
point_coords: Optional[np.ndarray] = None,
|
94 |
+
point_labels: Optional[np.ndarray] = None,
|
95 |
+
box: Optional[np.ndarray] = None,
|
96 |
+
mask_input: Optional[np.ndarray] = None,
|
97 |
+
multimask_output: bool = True,
|
98 |
+
return_logits: bool = False,
|
99 |
+
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
100 |
+
"""
|
101 |
+
Predict masks for the given input prompts, using the currently set image.
|
102 |
+
|
103 |
+
Arguments:
|
104 |
+
point_coords (np.ndarray or None): A Nx2 array of point prompts to the
|
105 |
+
model. Each point is in (X,Y) in pixels.
|
106 |
+
point_labels (np.ndarray or None): A length N array of labels for the
|
107 |
+
point prompts. 1 indicates a foreground point and 0 indicates a
|
108 |
+
background point.
|
109 |
+
box (np.ndarray or None): A length 4 array given a box prompt to the
|
110 |
+
model, in XYXY format.
|
111 |
+
mask_input (np.ndarray): A low resolution mask input to the model, typically
|
112 |
+
coming from a previous prediction iteration. Has form 1xHxW, where
|
113 |
+
for SAM, H=W=256.
|
114 |
+
multimask_output (bool): If true, the model will return three masks.
|
115 |
+
For ambiguous input prompts (such as a single click), this will often
|
116 |
+
produce better masks than a single prediction. If only a single
|
117 |
+
mask is needed, the model's predicted quality score can be used
|
118 |
+
to select the best mask. For non-ambiguous prompts, such as multiple
|
119 |
+
input prompts, multimask_output=False can give better results.
|
120 |
+
return_logits (bool): If true, returns un-thresholded masks logits
|
121 |
+
instead of a binary mask.
|
122 |
+
|
123 |
+
Returns:
|
124 |
+
(np.ndarray): The output masks in CxHxW format, where C is the
|
125 |
+
number of masks, and (H, W) is the original image size.
|
126 |
+
(np.ndarray): An array of length C containing the model's
|
127 |
+
predictions for the quality of each mask.
|
128 |
+
(np.ndarray): An array of shape CxHxW, where C is the number
|
129 |
+
of masks and H=W=256. These low resolution logits can be passed to
|
130 |
+
a subsequent iteration as mask input.
|
131 |
+
"""
|
132 |
+
if not self.is_image_set:
|
133 |
+
raise RuntimeError("An image must be set with .set_image(...) before mask prediction.")
|
134 |
+
|
135 |
+
# Transform input prompts
|
136 |
+
coords_torch, labels_torch, box_torch, mask_input_torch = None, None, None, None
|
137 |
+
if point_coords is not None:
|
138 |
+
assert point_labels is not None, "point_labels must be supplied if point_coords is supplied."
|
139 |
+
point_coords = self.transform.apply_coords(point_coords, self.original_size)
|
140 |
+
coords_torch = torch.as_tensor(point_coords, dtype=torch.float, device=self.device)
|
141 |
+
labels_torch = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
|
142 |
+
coords_torch, labels_torch = coords_torch[None, :, :], labels_torch[None, :]
|
143 |
+
if box is not None:
|
144 |
+
box = self.transform.apply_boxes(box, self.original_size)
|
145 |
+
box_torch = torch.as_tensor(box, dtype=torch.float, device=self.device)
|
146 |
+
box_torch = box_torch[None, :]
|
147 |
+
if mask_input is not None:
|
148 |
+
mask_input_torch = torch.as_tensor(mask_input, dtype=torch.float, device=self.device)
|
149 |
+
mask_input_torch = mask_input_torch[None, :, :, :]
|
150 |
+
|
151 |
+
masks, iou_predictions, low_res_masks = self.predict_torch(
|
152 |
+
coords_torch,
|
153 |
+
labels_torch,
|
154 |
+
box_torch,
|
155 |
+
mask_input_torch,
|
156 |
+
multimask_output,
|
157 |
+
return_logits=return_logits,
|
158 |
+
)
|
159 |
+
|
160 |
+
masks = masks[0].detach().cpu().numpy()
|
161 |
+
iou_predictions = iou_predictions[0].detach().cpu().numpy()
|
162 |
+
low_res_masks = low_res_masks[0].detach().cpu().numpy()
|
163 |
+
return masks, iou_predictions, low_res_masks
|
164 |
+
|
165 |
+
@torch.no_grad()
|
166 |
+
def predict_torch(
|
167 |
+
self,
|
168 |
+
point_coords: Optional[torch.Tensor],
|
169 |
+
point_labels: Optional[torch.Tensor],
|
170 |
+
boxes: Optional[torch.Tensor] = None,
|
171 |
+
mask_input: Optional[torch.Tensor] = None,
|
172 |
+
multimask_output: bool = True,
|
173 |
+
return_logits: bool = False,
|
174 |
+
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
175 |
+
"""
|
176 |
+
Predict masks for the given input prompts, using the currently set image.
|
177 |
+
Input prompts are batched torch tensors and are expected to already be
|
178 |
+
transformed to the input frame using ResizeLongestSide.
|
179 |
+
|
180 |
+
Arguments:
|
181 |
+
point_coords (torch.Tensor or None): A BxNx2 array of point prompts to the
|
182 |
+
model. Each point is in (X,Y) in pixels.
|
183 |
+
point_labels (torch.Tensor or None): A BxN array of labels for the
|
184 |
+
point prompts. 1 indicates a foreground point and 0 indicates a
|
185 |
+
background point.
|
186 |
+
box (np.ndarray or None): A Bx4 array given a box prompt to the
|
187 |
+
model, in XYXY format.
|
188 |
+
mask_input (np.ndarray): A low resolution mask input to the model, typically
|
189 |
+
coming from a previous prediction iteration. Has form Bx1xHxW, where
|
190 |
+
for SAM, H=W=256. Masks returned by a previous iteration of the
|
191 |
+
predict method do not need further transformation.
|
192 |
+
multimask_output (bool): If true, the model will return three masks.
|
193 |
+
For ambiguous input prompts (such as a single click), this will often
|
194 |
+
produce better masks than a single prediction. If only a single
|
195 |
+
mask is needed, the model's predicted quality score can be used
|
196 |
+
to select the best mask. For non-ambiguous prompts, such as multiple
|
197 |
+
input prompts, multimask_output=False can give better results.
|
198 |
+
return_logits (bool): If true, returns un-thresholded masks logits
|
199 |
+
instead of a binary mask.
|
200 |
+
|
201 |
+
Returns:
|
202 |
+
(torch.Tensor): The output masks in BxCxHxW format, where C is the
|
203 |
+
number of masks, and (H, W) is the original image size.
|
204 |
+
(torch.Tensor): An array of shape BxC containing the model's
|
205 |
+
predictions for the quality of each mask.
|
206 |
+
(torch.Tensor): An array of shape BxCxHxW, where C is the number
|
207 |
+
of masks and H=W=256. These low res logits can be passed to
|
208 |
+
a subsequent iteration as mask input.
|
209 |
+
"""
|
210 |
+
if not self.is_image_set:
|
211 |
+
raise RuntimeError("An image must be set with .set_image(...) before mask prediction.")
|
212 |
+
|
213 |
+
if point_coords is not None:
|
214 |
+
points = (point_coords, point_labels)
|
215 |
+
else:
|
216 |
+
points = None
|
217 |
+
|
218 |
+
# Embed prompts
|
219 |
+
sparse_embeddings, dense_embeddings = self.model.prompt_encoder(
|
220 |
+
points=points,
|
221 |
+
boxes=boxes,
|
222 |
+
masks=mask_input,
|
223 |
+
)
|
224 |
+
|
225 |
+
# Predict masks
|
226 |
+
low_res_masks, iou_predictions = self.model.mask_decoder(
|
227 |
+
image_embeddings=self.features,
|
228 |
+
image_pe=self.model.prompt_encoder.get_dense_pe(),
|
229 |
+
sparse_prompt_embeddings=sparse_embeddings,
|
230 |
+
dense_prompt_embeddings=dense_embeddings,
|
231 |
+
multimask_output=multimask_output,
|
232 |
+
)
|
233 |
+
|
234 |
+
# Upscale the masks to the original image resolution
|
235 |
+
masks = self.model.postprocess_masks(low_res_masks, self.input_size, self.original_size)
|
236 |
+
|
237 |
+
if not return_logits:
|
238 |
+
masks = masks > self.model.mask_threshold
|
239 |
+
|
240 |
+
return masks, iou_predictions, low_res_masks
|
241 |
+
|
242 |
+
def get_image_embedding(self) -> torch.Tensor:
|
243 |
+
"""
|
244 |
+
Returns the image embeddings for the currently set image, with
|
245 |
+
shape 1xCxHxW, where C is the embedding dimension and (H,W) are
|
246 |
+
the embedding spatial dimension of SAM (typically C=256, H=W=64).
|
247 |
+
"""
|
248 |
+
if not self.is_image_set:
|
249 |
+
raise RuntimeError("An image must be set with .set_image(...) to generate an embedding.")
|
250 |
+
assert self.features is not None, "Features must exist if an image has been set."
|
251 |
+
return self.features
|
252 |
+
|
253 |
+
@property
|
254 |
+
def device(self) -> torch.device:
|
255 |
+
return self.model.device
|
256 |
+
|
257 |
+
def reset_image(self) -> None:
|
258 |
+
"""Resets the currently set image."""
|
259 |
+
self.is_image_set = False
|
260 |
+
self.features = None
|
261 |
+
self.orig_h = None
|
262 |
+
self.orig_w = None
|
263 |
+
self.input_h = None
|
264 |
+
self.input_w = None
|
metaseg/utils/__init__.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
metaseg/utils/__pycache__/__init__.cpython-310.pyc
ADDED
Binary file (159 Bytes). View file
|
|
metaseg/utils/__pycache__/amg.cpython-310.pyc
ADDED
Binary file (12 kB). View file
|
|
metaseg/utils/__pycache__/file.cpython-310.pyc
ADDED
Binary file (1.16 kB). View file
|
|
metaseg/utils/__pycache__/transforms.cpython-310.pyc
ADDED
Binary file (3.9 kB). View file
|
|
metaseg/utils/amg.py
ADDED
@@ -0,0 +1,330 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
import math
|
8 |
+
from copy import deepcopy
|
9 |
+
from itertools import product
|
10 |
+
from typing import Any, Dict, Generator, ItemsView, List, Tuple
|
11 |
+
|
12 |
+
import numpy as np
|
13 |
+
import torch
|
14 |
+
|
15 |
+
|
16 |
+
class MaskData:
|
17 |
+
"""
|
18 |
+
A structure for storing masks and their related data in batched format.
|
19 |
+
Implements basic filtering and concatenation.
|
20 |
+
"""
|
21 |
+
|
22 |
+
def __init__(self, **kwargs) -> None:
|
23 |
+
for v in kwargs.values():
|
24 |
+
assert isinstance(
|
25 |
+
v, (list, np.ndarray, torch.Tensor)
|
26 |
+
), "MaskData only supports list, numpy arrays, and torch tensors."
|
27 |
+
self._stats = dict(**kwargs)
|
28 |
+
|
29 |
+
def __setitem__(self, key: str, item: Any) -> None:
|
30 |
+
assert isinstance(
|
31 |
+
item, (list, np.ndarray, torch.Tensor)
|
32 |
+
), "MaskData only supports list, numpy arrays, and torch tensors."
|
33 |
+
self._stats[key] = item
|
34 |
+
|
35 |
+
def __delitem__(self, key: str) -> None:
|
36 |
+
del self._stats[key]
|
37 |
+
|
38 |
+
def __getitem__(self, key: str) -> Any:
|
39 |
+
return self._stats[key]
|
40 |
+
|
41 |
+
def items(self) -> ItemsView[str, Any]:
|
42 |
+
return self._stats.items()
|
43 |
+
|
44 |
+
def filter(self, keep: torch.Tensor) -> None:
|
45 |
+
for k, v in self._stats.items():
|
46 |
+
if v is None:
|
47 |
+
self._stats[k] = None
|
48 |
+
elif isinstance(v, torch.Tensor):
|
49 |
+
self._stats[k] = v[torch.as_tensor(keep, device=v.device)]
|
50 |
+
elif isinstance(v, np.ndarray):
|
51 |
+
self._stats[k] = v[keep.detach().cpu().numpy()]
|
52 |
+
elif isinstance(v, list) and keep.dtype == torch.bool:
|
53 |
+
self._stats[k] = [a for i, a in enumerate(v) if keep[i]]
|
54 |
+
elif isinstance(v, list):
|
55 |
+
self._stats[k] = [v[i] for i in keep]
|
56 |
+
else:
|
57 |
+
raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
|
58 |
+
|
59 |
+
def cat(self, new_stats: "MaskData") -> None:
|
60 |
+
for k, v in new_stats.items():
|
61 |
+
if k not in self._stats or self._stats[k] is None:
|
62 |
+
self._stats[k] = deepcopy(v)
|
63 |
+
elif isinstance(v, torch.Tensor):
|
64 |
+
self._stats[k] = torch.cat([self._stats[k], v], dim=0)
|
65 |
+
elif isinstance(v, np.ndarray):
|
66 |
+
self._stats[k] = np.concatenate([self._stats[k], v], axis=0)
|
67 |
+
elif isinstance(v, list):
|
68 |
+
self._stats[k] = self._stats[k] + deepcopy(v)
|
69 |
+
else:
|
70 |
+
raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
|
71 |
+
|
72 |
+
def to_numpy(self) -> None:
|
73 |
+
for k, v in self._stats.items():
|
74 |
+
if isinstance(v, torch.Tensor):
|
75 |
+
self._stats[k] = v.detach().cpu().numpy()
|
76 |
+
|
77 |
+
|
78 |
+
def is_box_near_crop_edge(
|
79 |
+
boxes: torch.Tensor, crop_box: List[int], orig_box: List[int], atol: float = 20.0
|
80 |
+
) -> torch.Tensor:
|
81 |
+
"""Filter masks at the edge of a crop, but not at the edge of the original image."""
|
82 |
+
crop_box_torch = torch.as_tensor(crop_box, dtype=torch.float, device=boxes.device)
|
83 |
+
orig_box_torch = torch.as_tensor(orig_box, dtype=torch.float, device=boxes.device)
|
84 |
+
boxes = uncrop_boxes_xyxy(boxes, crop_box).float()
|
85 |
+
near_crop_edge = torch.isclose(boxes, crop_box_torch[None, :], atol=atol, rtol=0)
|
86 |
+
near_image_edge = torch.isclose(boxes, orig_box_torch[None, :], atol=atol, rtol=0)
|
87 |
+
near_crop_edge = torch.logical_and(near_crop_edge, ~near_image_edge)
|
88 |
+
return torch.any(near_crop_edge, dim=1)
|
89 |
+
|
90 |
+
|
91 |
+
def box_xyxy_to_xywh(box_xyxy: torch.Tensor) -> torch.Tensor:
|
92 |
+
box_xywh = deepcopy(box_xyxy)
|
93 |
+
box_xywh[2] = box_xywh[2] - box_xywh[0]
|
94 |
+
box_xywh[3] = box_xywh[3] - box_xywh[1]
|
95 |
+
return box_xywh
|
96 |
+
|
97 |
+
|
98 |
+
def batch_iterator(batch_size: int, *args) -> Generator[List[Any], None, None]:
|
99 |
+
assert len(args) > 0 and all(
|
100 |
+
len(a) == len(args[0]) for a in args
|
101 |
+
), "Batched iteration must have inputs of all the same size."
|
102 |
+
n_batches = len(args[0]) // batch_size + int(len(args[0]) % batch_size != 0)
|
103 |
+
for b in range(n_batches):
|
104 |
+
yield [arg[b * batch_size : (b + 1) * batch_size] for arg in args]
|
105 |
+
|
106 |
+
|
107 |
+
def mask_to_rle_pytorch(tensor: torch.Tensor) -> List[Dict[str, Any]]:
|
108 |
+
"""
|
109 |
+
Encodes masks to an uncompressed RLE, in the format expected by
|
110 |
+
pycoco tools.
|
111 |
+
"""
|
112 |
+
# Put in fortran order and flatten h,w
|
113 |
+
b, h, w = tensor.shape
|
114 |
+
tensor = tensor.permute(0, 2, 1).flatten(1)
|
115 |
+
|
116 |
+
# Compute change indices
|
117 |
+
diff = tensor[:, 1:] ^ tensor[:, :-1]
|
118 |
+
change_indices = diff.nonzero()
|
119 |
+
|
120 |
+
# Encode run length
|
121 |
+
out = []
|
122 |
+
for i in range(b):
|
123 |
+
cur_idxs = change_indices[change_indices[:, 0] == i, 1]
|
124 |
+
cur_idxs = torch.cat(
|
125 |
+
[
|
126 |
+
torch.tensor([0], dtype=cur_idxs.dtype, device=cur_idxs.device),
|
127 |
+
cur_idxs + 1,
|
128 |
+
torch.tensor([h * w], dtype=cur_idxs.dtype, device=cur_idxs.device),
|
129 |
+
]
|
130 |
+
)
|
131 |
+
btw_idxs = cur_idxs[1:] - cur_idxs[:-1]
|
132 |
+
counts = [] if tensor[i, 0] == 0 else [0]
|
133 |
+
counts.extend(btw_idxs.detach().cpu().tolist())
|
134 |
+
out.append({"size": [h, w], "counts": counts})
|
135 |
+
return out
|
136 |
+
|
137 |
+
|
138 |
+
def rle_to_mask(rle: Dict[str, Any]) -> np.ndarray:
|
139 |
+
"""Compute a binary mask from an uncompressed RLE."""
|
140 |
+
h, w = rle["size"]
|
141 |
+
mask = np.empty(h * w, dtype=bool)
|
142 |
+
idx = 0
|
143 |
+
parity = False
|
144 |
+
for count in rle["counts"]:
|
145 |
+
mask[idx : idx + count] = parity
|
146 |
+
idx += count
|
147 |
+
parity ^= True
|
148 |
+
mask = mask.reshape(w, h)
|
149 |
+
return mask.transpose() # Put in C order
|
150 |
+
|
151 |
+
|
152 |
+
def area_from_rle(rle: Dict[str, Any]) -> int:
|
153 |
+
return sum(rle["counts"][1::2])
|
154 |
+
|
155 |
+
|
156 |
+
def calculate_stability_score(masks: torch.Tensor, mask_threshold: float, threshold_offset: float) -> torch.Tensor:
|
157 |
+
"""
|
158 |
+
Computes the stability score for a batch of masks. The stability
|
159 |
+
score is the IoU between the binary masks obtained by thresholding
|
160 |
+
the predicted mask logits at high and low values.
|
161 |
+
"""
|
162 |
+
# One mask is always contained inside the other.
|
163 |
+
# Save memory by preventing unnecesary cast to torch.int64
|
164 |
+
intersections = (masks > (mask_threshold + threshold_offset)).sum(-1, dtype=torch.int16).sum(-1, dtype=torch.int32)
|
165 |
+
unions = (masks > (mask_threshold - threshold_offset)).sum(-1, dtype=torch.int16).sum(-1, dtype=torch.int32)
|
166 |
+
return intersections / unions
|
167 |
+
|
168 |
+
|
169 |
+
def build_point_grid(n_per_side: int) -> np.ndarray:
|
170 |
+
"""Generates a 2D grid of points evenly spaced in [0,1]x[0,1]."""
|
171 |
+
offset = 1 / (2 * n_per_side)
|
172 |
+
points_one_side = np.linspace(offset, 1 - offset, n_per_side)
|
173 |
+
points_x = np.tile(points_one_side[None, :], (n_per_side, 1))
|
174 |
+
points_y = np.tile(points_one_side[:, None], (1, n_per_side))
|
175 |
+
points = np.stack([points_x, points_y], axis=-1).reshape(-1, 2)
|
176 |
+
return points
|
177 |
+
|
178 |
+
|
179 |
+
def build_all_layer_point_grids(n_per_side: int, n_layers: int, scale_per_layer: int) -> List[np.ndarray]:
|
180 |
+
"""Generates point grids for all crop layers."""
|
181 |
+
points_by_layer = []
|
182 |
+
for i in range(n_layers + 1):
|
183 |
+
n_points = int(n_per_side / (scale_per_layer ** i))
|
184 |
+
points_by_layer.append(build_point_grid(n_points))
|
185 |
+
return points_by_layer
|
186 |
+
|
187 |
+
|
188 |
+
def generate_crop_boxes(
|
189 |
+
im_size: Tuple[int, ...], n_layers: int, overlap_ratio: float
|
190 |
+
) -> Tuple[List[List[int]], List[int]]:
|
191 |
+
"""
|
192 |
+
Generates a list of crop boxes of different sizes. Each layer
|
193 |
+
has (2**i)**2 boxes for the ith layer.
|
194 |
+
"""
|
195 |
+
crop_boxes, layer_idxs = [], []
|
196 |
+
im_h, im_w = im_size
|
197 |
+
short_side = min(im_h, im_w)
|
198 |
+
|
199 |
+
# Original image
|
200 |
+
crop_boxes.append([0, 0, im_w, im_h])
|
201 |
+
layer_idxs.append(0)
|
202 |
+
|
203 |
+
def crop_len(orig_len, n_crops, overlap):
|
204 |
+
return int(math.ceil((overlap * (n_crops - 1) + orig_len) / n_crops))
|
205 |
+
|
206 |
+
for i_layer in range(n_layers):
|
207 |
+
n_crops_per_side = 2 ** (i_layer + 1)
|
208 |
+
overlap = int(overlap_ratio * short_side * (2 / n_crops_per_side))
|
209 |
+
|
210 |
+
crop_w = crop_len(im_w, n_crops_per_side, overlap)
|
211 |
+
crop_h = crop_len(im_h, n_crops_per_side, overlap)
|
212 |
+
|
213 |
+
crop_box_x0 = [int((crop_w - overlap) * i) for i in range(n_crops_per_side)]
|
214 |
+
crop_box_y0 = [int((crop_h - overlap) * i) for i in range(n_crops_per_side)]
|
215 |
+
|
216 |
+
# Crops in XYWH format
|
217 |
+
for x0, y0 in product(crop_box_x0, crop_box_y0):
|
218 |
+
box = [x0, y0, min(x0 + crop_w, im_w), min(y0 + crop_h, im_h)]
|
219 |
+
crop_boxes.append(box)
|
220 |
+
layer_idxs.append(i_layer + 1)
|
221 |
+
|
222 |
+
return crop_boxes, layer_idxs
|
223 |
+
|
224 |
+
|
225 |
+
def uncrop_boxes_xyxy(boxes: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
|
226 |
+
x0, y0, _, _ = crop_box
|
227 |
+
offset = torch.tensor([[x0, y0, x0, y0]], device=boxes.device)
|
228 |
+
# Check if boxes has a channel dimension
|
229 |
+
if len(boxes.shape) == 3:
|
230 |
+
offset = offset.unsqueeze(1)
|
231 |
+
return boxes + offset
|
232 |
+
|
233 |
+
|
234 |
+
def uncrop_points(points: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
|
235 |
+
x0, y0, _, _ = crop_box
|
236 |
+
offset = torch.tensor([[x0, y0]], device=points.device)
|
237 |
+
# Check if points has a channel dimension
|
238 |
+
if len(points.shape) == 3:
|
239 |
+
offset = offset.unsqueeze(1)
|
240 |
+
return points + offset
|
241 |
+
|
242 |
+
|
243 |
+
def uncrop_masks(masks: torch.Tensor, crop_box: List[int], orig_h: int, orig_w: int) -> torch.Tensor:
|
244 |
+
x0, y0, x1, y1 = crop_box
|
245 |
+
if x0 == 0 and y0 == 0 and x1 == orig_w and y1 == orig_h:
|
246 |
+
return masks
|
247 |
+
# Coordinate transform masks
|
248 |
+
pad_x, pad_y = orig_w - (x1 - x0), orig_h - (y1 - y0)
|
249 |
+
pad = (x0, pad_x - x0, y0, pad_y - y0)
|
250 |
+
return torch.nn.functional.pad(masks, pad, value=0)
|
251 |
+
|
252 |
+
|
253 |
+
def remove_small_regions(mask: np.ndarray, area_thresh: float, mode: str) -> Tuple[np.ndarray, bool]:
|
254 |
+
"""
|
255 |
+
Removes small disconnected regions and holes in a mask. Returns the
|
256 |
+
mask and an indicator of if the mask has been modified.
|
257 |
+
"""
|
258 |
+
import cv2 # type: ignore
|
259 |
+
|
260 |
+
assert mode in ["holes", "islands"]
|
261 |
+
correct_holes = mode == "holes"
|
262 |
+
working_mask = (correct_holes ^ mask).astype(np.uint8)
|
263 |
+
n_labels, regions, stats, _ = cv2.connectedComponentsWithStats(working_mask, 8)
|
264 |
+
sizes = stats[:, -1][1:] # Row 0 is background label
|
265 |
+
small_regions = [i + 1 for i, s in enumerate(sizes) if s < area_thresh]
|
266 |
+
if len(small_regions) == 0:
|
267 |
+
return mask, False
|
268 |
+
fill_labels = [0] + small_regions
|
269 |
+
if not correct_holes:
|
270 |
+
fill_labels = [i for i in range(n_labels) if i not in fill_labels]
|
271 |
+
# If every region is below threshold, keep largest
|
272 |
+
if len(fill_labels) == 0:
|
273 |
+
fill_labels = [int(np.argmax(sizes)) + 1]
|
274 |
+
mask = np.isin(regions, fill_labels)
|
275 |
+
return mask, True
|
276 |
+
|
277 |
+
|
278 |
+
def coco_encode_rle(uncompressed_rle: Dict[str, Any]) -> Dict[str, Any]:
|
279 |
+
from pycocotools import mask as mask_utils # type: ignore
|
280 |
+
|
281 |
+
h, w = uncompressed_rle["size"]
|
282 |
+
rle = mask_utils.frPyObjects(uncompressed_rle, h, w)
|
283 |
+
rle["counts"] = rle["counts"].decode("utf-8") # Necessary to serialize with json
|
284 |
+
return rle
|
285 |
+
|
286 |
+
|
287 |
+
def batched_mask_to_box(masks: torch.Tensor) -> torch.Tensor:
|
288 |
+
"""
|
289 |
+
Calculates boxes in XYXY format around masks. Return [0,0,0,0] for
|
290 |
+
an empty mask. For input shape C1xC2x...xHxW, the output shape is C1xC2x...x4.
|
291 |
+
"""
|
292 |
+
# torch.max below raises an error on empty inputs, just skip in this case
|
293 |
+
if torch.numel(masks) == 0:
|
294 |
+
return torch.zeros(*masks.shape[:-2], 4, device=masks.device)
|
295 |
+
|
296 |
+
# Normalize shape to CxHxW
|
297 |
+
shape = masks.shape
|
298 |
+
h, w = shape[-2:]
|
299 |
+
if len(shape) > 2:
|
300 |
+
masks = masks.flatten(0, -3)
|
301 |
+
else:
|
302 |
+
masks = masks.unsqueeze(0)
|
303 |
+
|
304 |
+
# Get top and bottom edges
|
305 |
+
in_height, _ = torch.max(masks, dim=-1)
|
306 |
+
in_height_coords = in_height * torch.arange(h, device=in_height.device)[None, :]
|
307 |
+
bottom_edges, _ = torch.max(in_height_coords, dim=-1)
|
308 |
+
in_height_coords = in_height_coords + h * (~in_height)
|
309 |
+
top_edges, _ = torch.min(in_height_coords, dim=-1)
|
310 |
+
|
311 |
+
# Get left and right edges
|
312 |
+
in_width, _ = torch.max(masks, dim=-2)
|
313 |
+
in_width_coords = in_width * torch.arange(w, device=in_width.device)[None, :]
|
314 |
+
right_edges, _ = torch.max(in_width_coords, dim=-1)
|
315 |
+
in_width_coords = in_width_coords + w * (~in_width)
|
316 |
+
left_edges, _ = torch.min(in_width_coords, dim=-1)
|
317 |
+
|
318 |
+
# If the mask is empty the right edge will be to the left of the left edge.
|
319 |
+
# Replace these boxes with [0, 0, 0, 0]
|
320 |
+
empty_filter = (right_edges < left_edges) | (bottom_edges < top_edges)
|
321 |
+
out = torch.stack([left_edges, top_edges, right_edges, bottom_edges], dim=-1)
|
322 |
+
out = out * (~empty_filter).unsqueeze(-1)
|
323 |
+
|
324 |
+
# Return to original shape
|
325 |
+
if len(shape) > 2:
|
326 |
+
out = out.reshape(*shape[:-2], 4)
|
327 |
+
else:
|
328 |
+
out = out[0]
|
329 |
+
|
330 |
+
return out
|
metaseg/utils/file.py
ADDED
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import urllib.request
|
3 |
+
|
4 |
+
|
5 |
+
def download_model(model_type):
|
6 |
+
"""
|
7 |
+
model_type: str, A string representing the model type. It can be 'vit_h', 'vit_l', or 'vit_b'.
|
8 |
+
"""
|
9 |
+
|
10 |
+
# A dictionary containing model types as keys and their respective URLs as values
|
11 |
+
model_urls = {
|
12 |
+
"vit_h": "https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth",
|
13 |
+
"vit_l": "https://dl.fbaipublicfiles.com/segment_anything/sam_vit_l_0b3195.pth",
|
14 |
+
"vit_b": "https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth",
|
15 |
+
}
|
16 |
+
|
17 |
+
# Check if the model file already exists and model_type is in model_urls
|
18 |
+
filename = f"{model_type}.pth"
|
19 |
+
if not os.path.exists(filename) and model_type in model_urls:
|
20 |
+
url = model_urls[model_type]
|
21 |
+
print(f"Downloading {model_type} model from {url}...")
|
22 |
+
urllib.request.urlretrieve(url, filename)
|
23 |
+
print(f"{model_type} model has been successfully downloaded and saved as '{filename}'.")
|
24 |
+
elif os.path.exists(filename):
|
25 |
+
print(f"{model_type} model already exists as '{filename}'. Skipping download.")
|
26 |
+
else:
|
27 |
+
raise ValueError("Invalid model type. It should be 'vit_h', 'vit_l', or 'vit_b'.")
|
28 |
+
|
29 |
+
return filename
|
30 |
+
|
31 |
+
|
32 |
+
download_model("vit_b")
|
metaseg/utils/onnx.py
ADDED
@@ -0,0 +1,138 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from typing import Tuple
|
8 |
+
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
from torch.nn import functional as F
|
12 |
+
|
13 |
+
from metaseg.modeling import Sam
|
14 |
+
from metaseg.utils.amg import calculate_stability_score
|
15 |
+
|
16 |
+
|
17 |
+
class SamOnnxModel(nn.Module):
|
18 |
+
"""
|
19 |
+
This model should not be called directly, but is used in ONNX export.
|
20 |
+
It combines the prompt encoder, mask decoder, and mask postprocessing of Sam,
|
21 |
+
with some functions modified to enable model tracing. Also supports extra
|
22 |
+
options controlling what information. See the ONNX export script for details.
|
23 |
+
"""
|
24 |
+
|
25 |
+
def __init__(
|
26 |
+
self,
|
27 |
+
model: Sam,
|
28 |
+
return_single_mask: bool,
|
29 |
+
use_stability_score: bool = False,
|
30 |
+
return_extra_metrics: bool = False,
|
31 |
+
) -> None:
|
32 |
+
super().__init__()
|
33 |
+
self.mask_decoder = model.mask_decoder
|
34 |
+
self.model = model
|
35 |
+
self.img_size = model.image_encoder.img_size
|
36 |
+
self.return_single_mask = return_single_mask
|
37 |
+
self.use_stability_score = use_stability_score
|
38 |
+
self.stability_score_offset = 1.0
|
39 |
+
self.return_extra_metrics = return_extra_metrics
|
40 |
+
|
41 |
+
@staticmethod
|
42 |
+
def resize_longest_image_size(input_image_size: torch.Tensor, longest_side: int) -> torch.Tensor:
|
43 |
+
input_image_size = input_image_size.to(torch.float32)
|
44 |
+
scale = longest_side / torch.max(input_image_size)
|
45 |
+
transformed_size = scale * input_image_size
|
46 |
+
transformed_size = torch.floor(transformed_size + 0.5).to(torch.int64)
|
47 |
+
return transformed_size
|
48 |
+
|
49 |
+
def _embed_points(self, point_coords: torch.Tensor, point_labels: torch.Tensor) -> torch.Tensor:
|
50 |
+
point_coords = point_coords + 0.5
|
51 |
+
point_coords = point_coords / self.img_size
|
52 |
+
point_embedding = self.model.prompt_encoder.pe_layer._pe_encoding(point_coords)
|
53 |
+
point_labels = point_labels.unsqueeze(-1).expand_as(point_embedding)
|
54 |
+
|
55 |
+
point_embedding = point_embedding * (point_labels != -1)
|
56 |
+
point_embedding = point_embedding + self.model.prompt_encoder.not_a_point_embed.weight * (point_labels == -1)
|
57 |
+
|
58 |
+
for i in range(self.model.prompt_encoder.num_point_embeddings):
|
59 |
+
point_embedding = point_embedding + self.model.prompt_encoder.point_embeddings[i].weight * (
|
60 |
+
point_labels == i
|
61 |
+
)
|
62 |
+
|
63 |
+
return point_embedding
|
64 |
+
|
65 |
+
def _embed_masks(self, input_mask: torch.Tensor, has_mask_input: torch.Tensor) -> torch.Tensor:
|
66 |
+
mask_embedding = has_mask_input * self.model.prompt_encoder.mask_downscaling(input_mask)
|
67 |
+
mask_embedding = mask_embedding + (1 - has_mask_input) * self.model.prompt_encoder.no_mask_embed.weight.reshape(
|
68 |
+
1, -1, 1, 1
|
69 |
+
)
|
70 |
+
return mask_embedding
|
71 |
+
|
72 |
+
def mask_postprocessing(self, masks: torch.Tensor, orig_im_size: torch.Tensor) -> torch.Tensor:
|
73 |
+
masks = F.interpolate(
|
74 |
+
masks,
|
75 |
+
size=(self.img_size, self.img_size),
|
76 |
+
mode="bilinear",
|
77 |
+
align_corners=False,
|
78 |
+
)
|
79 |
+
|
80 |
+
prepadded_size = self.resize_longest_image_size(orig_im_size, self.img_size)
|
81 |
+
masks = masks[..., : int(prepadded_size[0]), : int(prepadded_size[1])]
|
82 |
+
|
83 |
+
orig_im_size = orig_im_size.to(torch.int64)
|
84 |
+
h, w = orig_im_size[0], orig_im_size[1]
|
85 |
+
masks = F.interpolate(masks, size=(h, w), mode="bilinear", align_corners=False)
|
86 |
+
return masks
|
87 |
+
|
88 |
+
def select_masks(
|
89 |
+
self, masks: torch.Tensor, iou_preds: torch.Tensor, num_points: int
|
90 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
91 |
+
# Determine if we should return the multiclick mask or not from the number of points.
|
92 |
+
# The reweighting is used to avoid control flow.
|
93 |
+
score_reweight = torch.tensor([[1000] + [0] * (self.model.mask_decoder.num_mask_tokens - 1)]).to(
|
94 |
+
iou_preds.device
|
95 |
+
)
|
96 |
+
score = iou_preds + (num_points - 2.5) * score_reweight
|
97 |
+
best_idx = torch.argmax(score, dim=1)
|
98 |
+
masks = masks[torch.arange(masks.shape[0]), best_idx, :, :].unsqueeze(1)
|
99 |
+
iou_preds = iou_preds[torch.arange(masks.shape[0]), best_idx].unsqueeze(1)
|
100 |
+
|
101 |
+
return masks, iou_preds
|
102 |
+
|
103 |
+
@torch.no_grad()
|
104 |
+
def forward(
|
105 |
+
self,
|
106 |
+
image_embeddings: torch.Tensor,
|
107 |
+
point_coords: torch.Tensor,
|
108 |
+
point_labels: torch.Tensor,
|
109 |
+
mask_input: torch.Tensor,
|
110 |
+
has_mask_input: torch.Tensor,
|
111 |
+
orig_im_size: torch.Tensor,
|
112 |
+
):
|
113 |
+
sparse_embedding = self._embed_points(point_coords, point_labels)
|
114 |
+
dense_embedding = self._embed_masks(mask_input, has_mask_input)
|
115 |
+
|
116 |
+
masks, scores = self.model.mask_decoder.predict_masks(
|
117 |
+
image_embeddings=image_embeddings,
|
118 |
+
image_pe=self.model.prompt_encoder.get_dense_pe(),
|
119 |
+
sparse_prompt_embeddings=sparse_embedding,
|
120 |
+
dense_prompt_embeddings=dense_embedding,
|
121 |
+
)
|
122 |
+
|
123 |
+
if self.use_stability_score:
|
124 |
+
scores = calculate_stability_score(masks, self.model.mask_threshold, self.stability_score_offset)
|
125 |
+
|
126 |
+
if self.return_single_mask:
|
127 |
+
masks, scores = self.select_masks(masks, scores, point_coords.shape[1])
|
128 |
+
|
129 |
+
upscaled_masks = self.mask_postprocessing(masks, orig_im_size)
|
130 |
+
|
131 |
+
if self.return_extra_metrics:
|
132 |
+
stability_scores = calculate_stability_score(
|
133 |
+
upscaled_masks, self.model.mask_threshold, self.stability_score_offset
|
134 |
+
)
|
135 |
+
areas = (upscaled_masks > self.model.mask_threshold).sum(-1).sum(-1)
|
136 |
+
return upscaled_masks, scores, stability_scores, areas, masks
|
137 |
+
|
138 |
+
return upscaled_masks, scores, masks
|
metaseg/utils/transforms.py
ADDED
@@ -0,0 +1,92 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from copy import deepcopy
|
8 |
+
from typing import Tuple
|
9 |
+
|
10 |
+
import numpy as np
|
11 |
+
import torch
|
12 |
+
from torch.nn import functional as F
|
13 |
+
from torchvision.transforms.functional import resize, to_pil_image # type: ignore
|
14 |
+
|
15 |
+
|
16 |
+
class ResizeLongestSide:
|
17 |
+
"""
|
18 |
+
Resizes images to longest side 'target_length', as well as provides
|
19 |
+
methods for resizing coordinates and boxes. Provides methods for
|
20 |
+
transforming both numpy array and batched torch tensors.
|
21 |
+
"""
|
22 |
+
|
23 |
+
def __init__(self, target_length: int) -> None:
|
24 |
+
self.target_length = target_length
|
25 |
+
|
26 |
+
def apply_image(self, image: np.ndarray) -> np.ndarray:
|
27 |
+
"""
|
28 |
+
Expects a numpy array with shape HxWxC in uint8 format.
|
29 |
+
"""
|
30 |
+
target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)
|
31 |
+
return np.array(resize(to_pil_image(image), target_size))
|
32 |
+
|
33 |
+
def apply_coords(self, coords: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
|
34 |
+
"""
|
35 |
+
Expects a numpy array of length 2 in the final dimension. Requires the
|
36 |
+
original image size in (H, W) format.
|
37 |
+
"""
|
38 |
+
old_h, old_w = original_size
|
39 |
+
new_h, new_w = self.get_preprocess_shape(original_size[0], original_size[1], self.target_length)
|
40 |
+
coords = deepcopy(coords).astype(float)
|
41 |
+
coords[..., 0] = coords[..., 0] * (new_w / old_w)
|
42 |
+
coords[..., 1] = coords[..., 1] * (new_h / old_h)
|
43 |
+
return coords
|
44 |
+
|
45 |
+
def apply_boxes(self, boxes: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
|
46 |
+
"""
|
47 |
+
Expects a numpy array shape Bx4. Requires the original image size
|
48 |
+
in (H, W) format.
|
49 |
+
"""
|
50 |
+
boxes = self.apply_coords(boxes.reshape(-1, 2, 2), original_size)
|
51 |
+
return boxes.reshape(-1, 4)
|
52 |
+
|
53 |
+
def apply_image_torch(self, image: torch.Tensor) -> torch.Tensor:
|
54 |
+
"""
|
55 |
+
Expects batched images with shape BxCxHxW and float format. This
|
56 |
+
transformation may not exactly match apply_image. apply_image is
|
57 |
+
the transformation expected by the model.
|
58 |
+
"""
|
59 |
+
# Expects an image in BCHW format. May not exactly match apply_image.
|
60 |
+
target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)
|
61 |
+
return F.interpolate(image, target_size, mode="bilinear", align_corners=False, antialias=True)
|
62 |
+
|
63 |
+
def apply_coords_torch(self, coords: torch.Tensor, original_size: Tuple[int, ...]) -> torch.Tensor:
|
64 |
+
"""
|
65 |
+
Expects a torch tensor with length 2 in the last dimension. Requires the
|
66 |
+
original image size in (H, W) format.
|
67 |
+
"""
|
68 |
+
old_h, old_w = original_size
|
69 |
+
new_h, new_w = self.get_preprocess_shape(original_size[0], original_size[1], self.target_length)
|
70 |
+
coords = deepcopy(coords).to(torch.float)
|
71 |
+
coords[..., 0] = coords[..., 0] * (new_w / old_w)
|
72 |
+
coords[..., 1] = coords[..., 1] * (new_h / old_h)
|
73 |
+
return coords
|
74 |
+
|
75 |
+
def apply_boxes_torch(self, boxes: torch.Tensor, original_size: Tuple[int, ...]) -> torch.Tensor:
|
76 |
+
"""
|
77 |
+
Expects a torch tensor with shape Bx4. Requires the original image
|
78 |
+
size in (H, W) format.
|
79 |
+
"""
|
80 |
+
boxes = self.apply_coords_torch(boxes.reshape(-1, 2, 2), original_size)
|
81 |
+
return boxes.reshape(-1, 4)
|
82 |
+
|
83 |
+
@staticmethod
|
84 |
+
def get_preprocess_shape(oldh: int, oldw: int, long_side_length: int) -> Tuple[int, int]:
|
85 |
+
"""
|
86 |
+
Compute the output size given input size and target long side length.
|
87 |
+
"""
|
88 |
+
scale = long_side_length * 1.0 / max(oldh, oldw)
|
89 |
+
newh, neww = oldh * scale, oldw * scale
|
90 |
+
neww = int(neww + 0.5)
|
91 |
+
newh = int(newh + 0.5)
|
92 |
+
return (newh, neww)
|
pyproject.toml
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
[tool.black]
|
2 |
+
line-length = 120
|
3 |
+
|
4 |
+
[tool.isort]
|
5 |
+
line_length = 120
|
6 |
+
profile = "black"
|
requirements.txt
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
onnxruntime
|
2 |
+
pycocotools
|
3 |
+
torch>=1.7
|
4 |
+
torchvision>=0.8
|
5 |
+
|
6 |
+
# code formatting
|
7 |
+
black==21.7b0
|
8 |
+
flake8==3.9.2
|
9 |
+
isort==5.9.2
|
10 |
+
click==8.0.4
|
scripts/amg.py
ADDED
@@ -0,0 +1,233 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
import argparse
|
8 |
+
import json
|
9 |
+
import os
|
10 |
+
from typing import Any, Dict, List
|
11 |
+
|
12 |
+
import cv2 # type: ignore
|
13 |
+
|
14 |
+
from metaseg import SamAutomaticMaskGenerator, sam_model_registry
|
15 |
+
|
16 |
+
parser = argparse.ArgumentParser(
|
17 |
+
description=(
|
18 |
+
"Runs automatic mask generation on an input image or directory of images, "
|
19 |
+
"and outputs masks as either PNGs or COCO-style RLEs. Requires open-cv, "
|
20 |
+
"as well as pycocotools if saving in RLE format."
|
21 |
+
)
|
22 |
+
)
|
23 |
+
|
24 |
+
parser.add_argument(
|
25 |
+
"--input",
|
26 |
+
type=str,
|
27 |
+
required=True,
|
28 |
+
help="Path to either a single input image or folder of images.",
|
29 |
+
)
|
30 |
+
|
31 |
+
parser.add_argument(
|
32 |
+
"--output",
|
33 |
+
type=str,
|
34 |
+
required=True,
|
35 |
+
help=(
|
36 |
+
"Path to the directory where masks will be output. Output will be either a folder "
|
37 |
+
"of PNGs per image or a single json with COCO-style masks."
|
38 |
+
),
|
39 |
+
)
|
40 |
+
|
41 |
+
parser.add_argument(
|
42 |
+
"--model-type",
|
43 |
+
type=str,
|
44 |
+
default="default",
|
45 |
+
help="The type of model to load, in ['default', 'vit_l', 'vit_b']",
|
46 |
+
)
|
47 |
+
|
48 |
+
parser.add_argument(
|
49 |
+
"--checkpoint",
|
50 |
+
type=str,
|
51 |
+
required=True,
|
52 |
+
help="The path to the SAM checkpoint to use for mask generation.",
|
53 |
+
)
|
54 |
+
|
55 |
+
parser.add_argument("--device", type=str, default="cuda", help="The device to run generation on.")
|
56 |
+
|
57 |
+
parser.add_argument(
|
58 |
+
"--convert-to-rle",
|
59 |
+
action="store_true",
|
60 |
+
help=("Save masks as COCO RLEs in a single json instead of as a folder of PNGs. " "Requires pycocotools."),
|
61 |
+
)
|
62 |
+
|
63 |
+
amg_settings = parser.add_argument_group("AMG Settings")
|
64 |
+
|
65 |
+
amg_settings.add_argument(
|
66 |
+
"--points-per-side",
|
67 |
+
type=int,
|
68 |
+
default=None,
|
69 |
+
help="Generate masks by sampling a grid over the image with this many points to a side.",
|
70 |
+
)
|
71 |
+
|
72 |
+
amg_settings.add_argument(
|
73 |
+
"--points-per-batch",
|
74 |
+
type=int,
|
75 |
+
default=None,
|
76 |
+
help="How many input points to process simultaneously in one batch.",
|
77 |
+
)
|
78 |
+
|
79 |
+
amg_settings.add_argument(
|
80 |
+
"--pred-iou-thresh",
|
81 |
+
type=float,
|
82 |
+
default=None,
|
83 |
+
help="Exclude masks with a predicted score from the model that is lower than this threshold.",
|
84 |
+
)
|
85 |
+
|
86 |
+
amg_settings.add_argument(
|
87 |
+
"--stability-score-thresh",
|
88 |
+
type=float,
|
89 |
+
default=None,
|
90 |
+
help="Exclude masks with a stability score lower than this threshold.",
|
91 |
+
)
|
92 |
+
|
93 |
+
amg_settings.add_argument(
|
94 |
+
"--stability-score-offset",
|
95 |
+
type=float,
|
96 |
+
default=None,
|
97 |
+
help="Larger values perturb the mask more when measuring stability score.",
|
98 |
+
)
|
99 |
+
|
100 |
+
amg_settings.add_argument(
|
101 |
+
"--box-nms-thresh",
|
102 |
+
type=float,
|
103 |
+
default=None,
|
104 |
+
help="The overlap threshold for excluding a duplicate mask.",
|
105 |
+
)
|
106 |
+
|
107 |
+
amg_settings.add_argument(
|
108 |
+
"--crop-n-layers",
|
109 |
+
type=int,
|
110 |
+
default=None,
|
111 |
+
help=(
|
112 |
+
"If >0, mask generation is run on smaller crops of the image to generate more masks. "
|
113 |
+
"The value sets how many different scales to crop at."
|
114 |
+
),
|
115 |
+
)
|
116 |
+
|
117 |
+
amg_settings.add_argument(
|
118 |
+
"--crop-nms-thresh",
|
119 |
+
type=float,
|
120 |
+
default=None,
|
121 |
+
help="The overlap threshold for excluding duplicate masks across different crops.",
|
122 |
+
)
|
123 |
+
|
124 |
+
amg_settings.add_argument(
|
125 |
+
"--crop-overlap-ratio",
|
126 |
+
type=int,
|
127 |
+
default=None,
|
128 |
+
help="Larger numbers mean image crops will overlap more.",
|
129 |
+
)
|
130 |
+
|
131 |
+
amg_settings.add_argument(
|
132 |
+
"--crop-n-points-downscale-factor",
|
133 |
+
type=int,
|
134 |
+
default=None,
|
135 |
+
help="The number of points-per-side in each layer of crop is reduced by this factor.",
|
136 |
+
)
|
137 |
+
|
138 |
+
amg_settings.add_argument(
|
139 |
+
"--min-mask-region-area",
|
140 |
+
type=int,
|
141 |
+
default=None,
|
142 |
+
help=(
|
143 |
+
"Disconnected mask regions or holes with area smaller than this value "
|
144 |
+
"in pixels are removed by postprocessing."
|
145 |
+
),
|
146 |
+
)
|
147 |
+
|
148 |
+
|
149 |
+
def write_masks_to_folder(masks: List[Dict[str, Any]], path: str) -> None:
|
150 |
+
header = "id,area,bbox_x0,bbox_y0,bbox_w,bbox_h,point_input_x,point_input_y,predicted_iou,stability_score,crop_box_x0,crop_box_y0,crop_box_w,crop_box_h" # noqa
|
151 |
+
metadata = [header]
|
152 |
+
for i, mask_data in enumerate(masks):
|
153 |
+
mask = mask_data["segmentation"]
|
154 |
+
filename = f"{i}.png"
|
155 |
+
cv2.imwrite(os.path.join(path, filename), mask * 255)
|
156 |
+
mask_metadata = [
|
157 |
+
str(i),
|
158 |
+
str(mask_data["area"]),
|
159 |
+
*[str(x) for x in mask_data["bbox"]],
|
160 |
+
*[str(x) for x in mask_data["point_coords"][0]],
|
161 |
+
str(mask_data["predicted_iou"]),
|
162 |
+
str(mask_data["stability_score"]),
|
163 |
+
*[str(x) for x in mask_data["crop_box"]],
|
164 |
+
]
|
165 |
+
row = ",".join(mask_metadata)
|
166 |
+
metadata.append(row)
|
167 |
+
metadata_path = os.path.join(path, "metadata.csv")
|
168 |
+
with open(metadata_path, "w") as f:
|
169 |
+
f.write("\n".join(metadata))
|
170 |
+
|
171 |
+
return
|
172 |
+
|
173 |
+
|
174 |
+
def get_amg_kwargs(args):
|
175 |
+
amg_kwargs = {
|
176 |
+
"points_per_side": args.points_per_side,
|
177 |
+
"points_per_batch": args.points_per_batch,
|
178 |
+
"pred_iou_thresh": args.pred_iou_thresh,
|
179 |
+
"stability_score_thresh": args.stability_score_thresh,
|
180 |
+
"stability_score_offset": args.stability_score_offset,
|
181 |
+
"box_nms_thresh": args.box_nms_thresh,
|
182 |
+
"crop_n_layers": args.crop_n_layers,
|
183 |
+
"crop_nms_thresh": args.crop_nms_thresh,
|
184 |
+
"crop_overlap_ratio": args.crop_overlap_ratio,
|
185 |
+
"crop_n_points_downscale_factor": args.crop_n_points_downscale_factor,
|
186 |
+
"min_mask_region_area": args.min_mask_region_area,
|
187 |
+
}
|
188 |
+
amg_kwargs = {k: v for k, v in amg_kwargs.items() if v is not None}
|
189 |
+
return amg_kwargs
|
190 |
+
|
191 |
+
|
192 |
+
def main(args: argparse.Namespace) -> None:
|
193 |
+
print("Loading model...")
|
194 |
+
sam = sam_model_registry[args.model_type](checkpoint=args.checkpoint)
|
195 |
+
_ = sam.to(device=args.device)
|
196 |
+
output_mode = "coco_rle" if args.convert_to_rle else "binary_mask"
|
197 |
+
amg_kwargs = get_amg_kwargs(args)
|
198 |
+
generator = SamAutomaticMaskGenerator(sam, output_mode=output_mode, **amg_kwargs)
|
199 |
+
|
200 |
+
if not os.path.isdir(args.input):
|
201 |
+
targets = [args.input]
|
202 |
+
else:
|
203 |
+
targets = [f for f in os.listdir(args.input) if not os.path.isdir(os.path.join(args.input, f))]
|
204 |
+
targets = [os.path.join(args.input, f) for f in targets]
|
205 |
+
|
206 |
+
os.makedirs(args.output, exist_ok=True)
|
207 |
+
|
208 |
+
for t in targets:
|
209 |
+
print(f"Processing '{t}'...")
|
210 |
+
image = cv2.imread(t)
|
211 |
+
if image is None:
|
212 |
+
print(f"Could not load '{t}' as an image, skipping...")
|
213 |
+
continue
|
214 |
+
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
|
215 |
+
|
216 |
+
masks = generator.generate(image)
|
217 |
+
|
218 |
+
base = os.path.basename(t)
|
219 |
+
base = os.path.splitext(base)[0]
|
220 |
+
save_base = os.path.join(args.output, base)
|
221 |
+
if output_mode == "binary_mask":
|
222 |
+
os.makedirs(save_base, exist_ok=False)
|
223 |
+
write_masks_to_folder(masks, save_base)
|
224 |
+
else:
|
225 |
+
save_file = save_base + ".json"
|
226 |
+
with open(save_file, "w") as f:
|
227 |
+
json.dump(masks, f)
|
228 |
+
print("Done!")
|
229 |
+
|
230 |
+
|
231 |
+
if __name__ == "__main__":
|
232 |
+
args = parser.parse_args()
|
233 |
+
main(args)
|
scripts/code_format.sh
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
black . --config pyproject.toml
|
2 |
+
isort .
|
scripts/export_onnx_model.py
ADDED
@@ -0,0 +1,198 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
import argparse
|
8 |
+
import warnings
|
9 |
+
|
10 |
+
import torch
|
11 |
+
|
12 |
+
from metaseg import build_sam, build_sam_vit_b, build_sam_vit_l
|
13 |
+
from metaseg.utils.onnx import SamOnnxModel
|
14 |
+
|
15 |
+
try:
|
16 |
+
import onnxruntime # type: ignore
|
17 |
+
|
18 |
+
onnxruntime_exists = True
|
19 |
+
except ImportError:
|
20 |
+
onnxruntime_exists = False
|
21 |
+
|
22 |
+
parser = argparse.ArgumentParser(description="Export the SAM prompt encoder and mask decoder to an ONNX model.")
|
23 |
+
|
24 |
+
parser.add_argument("--checkpoint", type=str, required=True, help="The path to the SAM model checkpoint.")
|
25 |
+
|
26 |
+
parser.add_argument("--output", type=str, required=True, help="The filename to save the ONNX model to.")
|
27 |
+
|
28 |
+
parser.add_argument(
|
29 |
+
"--model-type",
|
30 |
+
type=str,
|
31 |
+
default="default",
|
32 |
+
help="In ['default', 'vit_b', 'vit_l']. Which type of SAM model to export.",
|
33 |
+
)
|
34 |
+
|
35 |
+
parser.add_argument(
|
36 |
+
"--return-single-mask",
|
37 |
+
action="store_true",
|
38 |
+
help=(
|
39 |
+
"If true, the exported ONNX model will only return the best mask, "
|
40 |
+
"instead of returning multiple masks. For high resolution images "
|
41 |
+
"this can improve runtime when upscaling masks is expensive."
|
42 |
+
),
|
43 |
+
)
|
44 |
+
|
45 |
+
parser.add_argument(
|
46 |
+
"--opset",
|
47 |
+
type=int,
|
48 |
+
default=17,
|
49 |
+
help="The ONNX opset version to use. Must be >=11",
|
50 |
+
)
|
51 |
+
|
52 |
+
parser.add_argument(
|
53 |
+
"--quantize-out",
|
54 |
+
type=str,
|
55 |
+
default=None,
|
56 |
+
help=(
|
57 |
+
"If set, will quantize the model and save it with this name. "
|
58 |
+
"Quantization is performed with quantize_dynamic from onnxruntime.quantization.quantize."
|
59 |
+
),
|
60 |
+
)
|
61 |
+
|
62 |
+
parser.add_argument(
|
63 |
+
"--gelu-approximate",
|
64 |
+
action="store_true",
|
65 |
+
help=(
|
66 |
+
"Replace GELU operations with approximations using tanh. Useful "
|
67 |
+
"for some runtimes that have slow or unimplemented erf ops, used in GELU."
|
68 |
+
),
|
69 |
+
)
|
70 |
+
|
71 |
+
parser.add_argument(
|
72 |
+
"--use-stability-score",
|
73 |
+
action="store_true",
|
74 |
+
help=(
|
75 |
+
"Replaces the model's predicted mask quality score with the stability "
|
76 |
+
"score calculated on the low resolution masks using an offset of 1.0. "
|
77 |
+
),
|
78 |
+
)
|
79 |
+
|
80 |
+
parser.add_argument(
|
81 |
+
"--return-extra-metrics",
|
82 |
+
action="store_true",
|
83 |
+
help=(
|
84 |
+
"The model will return five results: (masks, scores, stability_scores, "
|
85 |
+
"areas, low_res_logits) instead of the usual three. This can be "
|
86 |
+
"significantly slower for high resolution outputs."
|
87 |
+
),
|
88 |
+
)
|
89 |
+
|
90 |
+
|
91 |
+
def run_export(
|
92 |
+
model_type: str,
|
93 |
+
checkpoint: str,
|
94 |
+
output: str,
|
95 |
+
opset: int,
|
96 |
+
return_single_mask: bool,
|
97 |
+
gelu_approximate: bool = False,
|
98 |
+
use_stability_score: bool = False,
|
99 |
+
return_extra_metrics=False,
|
100 |
+
):
|
101 |
+
print("Loading model...")
|
102 |
+
if model_type == "vit_b":
|
103 |
+
sam = build_sam_vit_b(checkpoint)
|
104 |
+
elif model_type == "vit_l":
|
105 |
+
sam = build_sam_vit_l(checkpoint)
|
106 |
+
else:
|
107 |
+
sam = build_sam(checkpoint)
|
108 |
+
|
109 |
+
onnx_model = SamOnnxModel(
|
110 |
+
model=sam,
|
111 |
+
return_single_mask=return_single_mask,
|
112 |
+
use_stability_score=use_stability_score,
|
113 |
+
return_extra_metrics=return_extra_metrics,
|
114 |
+
)
|
115 |
+
|
116 |
+
if gelu_approximate:
|
117 |
+
for n, m in onnx_model.named_modules():
|
118 |
+
if isinstance(m, torch.nn.GELU):
|
119 |
+
m.approximate = "tanh"
|
120 |
+
|
121 |
+
dynamic_axes = {
|
122 |
+
"point_coords": {1: "num_points"},
|
123 |
+
"point_labels": {1: "num_points"},
|
124 |
+
}
|
125 |
+
|
126 |
+
embed_dim = sam.prompt_encoder.embed_dim
|
127 |
+
embed_size = sam.prompt_encoder.image_embedding_size
|
128 |
+
mask_input_size = [4 * x for x in embed_size]
|
129 |
+
dummy_inputs = {
|
130 |
+
"image_embeddings": torch.randn(1, embed_dim, *embed_size, dtype=torch.float),
|
131 |
+
"point_coords": torch.randint(low=0, high=1024, size=(1, 5, 2), dtype=torch.float),
|
132 |
+
"point_labels": torch.randint(low=0, high=4, size=(1, 5), dtype=torch.float),
|
133 |
+
"mask_input": torch.randn(1, 1, *mask_input_size, dtype=torch.float),
|
134 |
+
"has_mask_input": torch.tensor([1], dtype=torch.float),
|
135 |
+
"orig_im_size": torch.tensor([1500, 2250], dtype=torch.float),
|
136 |
+
}
|
137 |
+
|
138 |
+
_ = onnx_model(**dummy_inputs)
|
139 |
+
|
140 |
+
output_names = ["masks", "iou_predictions", "low_res_masks"]
|
141 |
+
|
142 |
+
with warnings.catch_warnings():
|
143 |
+
warnings.filterwarnings("ignore", category=torch.jit.TracerWarning)
|
144 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
145 |
+
with open(output, "wb") as f:
|
146 |
+
print(f"Exporing onnx model to {output}...")
|
147 |
+
torch.onnx.export(
|
148 |
+
onnx_model,
|
149 |
+
tuple(dummy_inputs.values()),
|
150 |
+
f,
|
151 |
+
export_params=True,
|
152 |
+
verbose=False,
|
153 |
+
opset_version=opset,
|
154 |
+
do_constant_folding=True,
|
155 |
+
input_names=list(dummy_inputs.keys()),
|
156 |
+
output_names=output_names,
|
157 |
+
dynamic_axes=dynamic_axes,
|
158 |
+
)
|
159 |
+
|
160 |
+
if onnxruntime_exists:
|
161 |
+
ort_inputs = {k: to_numpy(v) for k, v in dummy_inputs.items()}
|
162 |
+
ort_session = onnxruntime.InferenceSession(output)
|
163 |
+
_ = ort_session.run(None, ort_inputs)
|
164 |
+
print("Model has successfully been run with ONNXRuntime.")
|
165 |
+
|
166 |
+
|
167 |
+
def to_numpy(tensor):
|
168 |
+
return tensor.cpu().numpy()
|
169 |
+
|
170 |
+
|
171 |
+
if __name__ == "__main__":
|
172 |
+
args = parser.parse_args()
|
173 |
+
run_export(
|
174 |
+
model_type=args.model_type,
|
175 |
+
checkpoint=args.checkpoint,
|
176 |
+
output=args.output,
|
177 |
+
opset=args.opset,
|
178 |
+
return_single_mask=args.return_single_mask,
|
179 |
+
gelu_approximate=args.gelu_approximate,
|
180 |
+
use_stability_score=args.use_stability_score,
|
181 |
+
return_extra_metrics=args.return_extra_metrics,
|
182 |
+
)
|
183 |
+
|
184 |
+
if args.quantize_out is not None:
|
185 |
+
assert onnxruntime_exists, "onnxruntime is required to quantize the model."
|
186 |
+
from onnxruntime.quantization import QuantType # type: ignore
|
187 |
+
from onnxruntime.quantization.quantize import quantize_dynamic # type: ignore
|
188 |
+
|
189 |
+
print(f"Quantizing model and writing to {args.quantize_out}...")
|
190 |
+
quantize_dynamic(
|
191 |
+
model_input=args.output,
|
192 |
+
model_output=args.quantize_out,
|
193 |
+
optimize_model=True,
|
194 |
+
per_channel=False,
|
195 |
+
reduce_range=False,
|
196 |
+
weight_type=QuantType.QUInt8,
|
197 |
+
)
|
198 |
+
print("Done!")
|
scripts/package.sh
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
python setup.py sdist
|
2 |
+
twine upload dist/*
|
setup.cfg
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
[isort]
|
2 |
+
line_length=100
|
3 |
+
multi_line_output=3
|
4 |
+
include_trailing_comma=True
|
5 |
+
known_standard_library=numpy,setuptools
|
6 |
+
skip_glob=*/__init__.py
|
7 |
+
known_myself=segment_anything
|
8 |
+
known_third_party=matplotlib,cv2,torch,torchvision,pycocotools,onnx,black,isort
|
9 |
+
no_lines_before=STDLIB,THIRDPARTY
|
10 |
+
sections=FUTURE,STDLIB,THIRDPARTY,MYSELF,FIRSTPARTY,LOCALFOLDER
|
11 |
+
default_section=FIRSTPARTY
|
setup.py
ADDED
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
import io
|
8 |
+
import os
|
9 |
+
import re
|
10 |
+
|
11 |
+
from setuptools import find_packages, setup
|
12 |
+
|
13 |
+
|
14 |
+
def get_long_description():
|
15 |
+
base_dir = os.path.abspath(os.path.dirname(__file__))
|
16 |
+
with io.open(os.path.join(base_dir, "README.md"), encoding="utf-8") as f:
|
17 |
+
return f.read()
|
18 |
+
|
19 |
+
|
20 |
+
def get_requirements():
|
21 |
+
with open("requirements.txt") as f:
|
22 |
+
return f.read().splitlines()
|
23 |
+
|
24 |
+
|
25 |
+
def get_version():
|
26 |
+
current_dir = os.path.abspath(os.path.dirname(__file__))
|
27 |
+
version_file = os.path.join(current_dir, "metaseg", "__init__.py")
|
28 |
+
with io.open(version_file, encoding="utf-8") as f:
|
29 |
+
return re.search(r'^__version__ = [\'"]([^\'"]*)[\'"]', f.read(), re.M).group(1)
|
30 |
+
|
31 |
+
|
32 |
+
_ALL_REQUIREMENTS = ["matplotlib", "pycocotools", "opencv-python", "onnx", "onnxruntime"]
|
33 |
+
|
34 |
+
_DEV_REQUIREMENTS = [
|
35 |
+
"black==23.*",
|
36 |
+
"isort==5.12.0",
|
37 |
+
"flake8",
|
38 |
+
"mypy",
|
39 |
+
]
|
40 |
+
|
41 |
+
extras = {
|
42 |
+
"all": _ALL_REQUIREMENTS,
|
43 |
+
"dev": _DEV_REQUIREMENTS,
|
44 |
+
}
|
45 |
+
|
46 |
+
setup(
|
47 |
+
name="metaseg",
|
48 |
+
license="Apache-2.0",
|
49 |
+
author="kadirnar",
|
50 |
+
long_description=get_long_description(),
|
51 |
+
long_description_content_type="text/markdown",
|
52 |
+
url="https://github.com/kadirnar/segment-anything-pip",
|
53 |
+
version=get_version(),
|
54 |
+
install_requires=get_requirements(),
|
55 |
+
packages=find_packages(exclude=("notebook")),
|
56 |
+
extras_require=extras,
|
57 |
+
python_requires=">=3.8",
|
58 |
+
)
|