AbstractPhil commited on
Commit
ed7f648
Β·
verified Β·
1 Parent(s): 90fe452

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +70 -93
README.md CHANGED
@@ -5,21 +5,6 @@ license: apache-2.0
5
  First barrage tests and compilation tests are finished. The orbital is yielding a true representation of omega.
6
 
7
  The speed and accuracy is both representationally superior and compiles safe to full-graph.
8
-
9
- ```
10
- Flow fullgraph Raw Compiled
11
- ────────────────────── ──────────── ────────── ────────────
12
- OrbitalFlow OK 20.71ms 732us
13
- AlignmentFlow OK 9.11ms 9.02ms
14
-
15
- Flow Peak MB
16
- ────────────────────── ──────────
17
- OrbitalFlow 10.5
18
- AlignmentFlow 60.0
19
- Full ensemble 203.6
20
-
21
- ```
22
-
23
  ```
24
  geolip_core.linalg: available
25
  geolip.linalg backend:
@@ -40,23 +25,23 @@ geolip.linalg backend:
40
 
41
  Flow Params Shape Time Conf Res norm
42
  ────────────────────── ──────── ────────────── ────────── ──────── ──────────
43
- QuaternionFlow 70,725 (16, 64, 128) 264us 0.476 1.318
44
- QuaternionLiteFlow 87,109 (16, 64, 128) 190us 0.465 1.279
45
- VelocityFlow 70,210 (16, 64, 128) 170us 0.533 0.056
46
- MagnitudeFlow 73,293 (16, 64, 128) 20.58ms 0.481 0.753
47
- OrbitalFlow 25,445 (16, 64, 128) 20.70ms 0.551 3.464
48
- AlignmentFlow 37,186 (16, 64, 128) 43.67ms 0.480 0.131
49
 
50
  ========================================================================
51
  2. LINALG INTEGRATION
52
  ========================================================================
53
 
54
  Testing eigh dispatch in MagnitudeFlow and OrbitalFlow...
55
- magnitude finite=True conf=0.477
56
- orbital finite=True conf=0.501
57
 
58
- Gram eigenspectrum: shape=(16, 12) range=[0.0001, 6.6260]
59
- Eigenvector orth err: 8.39e+01
60
 
61
  ========================================================================
62
  3. MULTI-SCALE
@@ -65,39 +50,39 @@ geolip.linalg backend:
65
  OrbitalFlow across scales:
66
  Config B n k d Time OK
67
  ────────── ──── ───── ───── ───── ────────── ────
68
- tiny 4 16 8 64 12.05ms OK
69
- small 16 64 32 128 20.64ms OK
70
- medium 32 128 64 256 20.71ms OK
71
- large 64 256 128 256 20.75ms OK
72
- wide 8 512 256 512 20.74ms OK
73
 
74
  ========================================================================
75
  4. ENSEMBLE FUSION
76
  ========================================================================
77
 
78
- weighted: time=87.28ms norm=0.586 diversity=0.829
79
- quaternion conf=0.476Β±0.002 res=1.322
80
- quat_lite conf=0.465Β±0.002 res=1.285
81
- velocity conf=0.533Β±0.003 res=0.056
82
- magnitude conf=0.481Β±0.004 res=0.757
83
- orbital conf=0.554Β±0.002 res=1.644
84
- alignment conf=0.480Β±0.003 res=0.131
85
 
86
- gated: time=87.36ms norm=0.598 diversity=0.829
87
- quaternion conf=0.476Β±0.002 res=1.322
88
- quat_lite conf=0.465Β±0.002 res=1.285
89
- velocity conf=0.533Β±0.003 res=0.056
90
- magnitude conf=0.481Β±0.004 res=0.757
91
- orbital conf=0.554Β±0.002 res=1.644
92
- alignment conf=0.480Β±0.003 res=0.131
93
 
94
- residual: time=87.35ms norm=0.586 diversity=0.829
95
- quaternion conf=0.476Β±0.002 res=1.322
96
- quat_lite conf=0.465Β±0.002 res=1.285
97
- velocity conf=0.533Β±0.003 res=0.056
98
- magnitude conf=0.481Β±0.004 res=0.757
99
- orbital conf=0.554Β±0.002 res=1.644
100
- alignment conf=0.480Β±0.003 res=0.131
101
 
102
  ========================================================================
103
  5. GRADIENT HEALTH
@@ -106,23 +91,23 @@ geolip.linalg backend:
106
  Flow Loss Grad norm Status
107
  ────────────────── ────────── ──────────── ────────
108
  quaternion mse 8.90e-04 OK
109
- quat_lite mse 7.05e-04 OK
110
- velocity mse 2.78e-04 OK
111
- magnitude mse 7.92e-04 OK
112
- orbital mse 3.45e-03 OK
113
- alignment mse 1.27e-03 OK
114
- quaternion cosine 1.43e-01 OK
115
- quat_lite cosine 1.17e-01 OK
116
- velocity cosine 3.37e-02 OK
117
- magnitude cosine 1.98e-01 OK
118
- orbital cosine 2.82e-01 OK
119
- alignment cosine 9.83e-02 OK
120
- quaternion norm 1.09e-01 OK
121
- quat_lite norm 8.79e-02 OK
122
- velocity norm 2.06e-02 OK
123
- magnitude norm 3.99e-01 OK
124
- orbital norm 3.97e+00 OK
125
- alignment norm 1.01e-01 OK
126
 
127
  ========================================================================
128
  6. ABLATION (100 training steps, rotation target)
@@ -130,20 +115,18 @@ geolip.linalg backend:
130
 
131
  Configuration MSE Params
132
  ─────────────────────────────────── ────────── ──────────
133
- QuaternionFlow 0.0010 280,709
134
  QuaternionLiteFlow 0.0005 346,245
135
  VelocityFlow 0.0061 279,682
136
  MagnitudeFlow 0.0081 285,837
137
- OrbitalFlow 0.0456 91,813
138
- AlignmentFlow 0.0066 148,098
139
- Quat + Orbital 0.0024 372,524
140
  Velocity + Magnitude 0.0062 565,521
141
- Orbital + Alignment 0.0188 239,913
142
- Velocity + Orbital 0.1576 371,497
143
- Full (weighted) 0.0046 1,432,390
144
- /tmp/ipykernel_16934/1749400227.py:389: UserWarning: torch.linalg.svd: During SVD computation with the selected cusolver driver, batches 0, 1, 2, 3, 4, and other 27 batches failed to converge. A more accurate method will be used to compute the SVD as a fallback. Check doc at https://pytorch.org/docs/stable/generated/torch.linalg.svd.html (Triggered internally at /pytorch/aten/src/ATen/native/cuda/linalg/BatchLinearAlgebraLib.cpp:701.)
145
- U, _, Vh = torch.linalg.svd(C) # full dΓ—d SVD β€” not through geolip.linalg.svd
146
- Full (residual) FAILED: linalg.svd: (Batch element 0):
147
 
148
  ========================================================================
149
  7. COMPILE COMPATIBILITY
@@ -151,18 +134,12 @@ geolip.linalg backend:
151
 
152
  Flow fullgraph Raw Compiled
153
  ────────────────────── ──────────── ────────── ────────────
154
- QuaternionFlow OK 269us 191us
155
- QuaternionLiteFlow OK 191us 172us
156
- VelocityFlow OK 176us 161us
157
- /usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:321: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
158
- warnings.warn(
159
- /usr/local/lib/python3.12/dist-packages/torch/_inductor/lowering.py:1904: FutureWarning: `torch._prims_common.check` is deprecated and will be removed in the future. Please use `torch._check*` functions instead.
160
- check(
161
- MagnitudeFlow OK 20.68ms 733us
162
- /usr/local/lib/python3.12/dist-packages/torch/_inductor/lowering.py:1904: FutureWarning: `torch._prims_common.check` is deprecated and will be removed in the future. Please use `torch._check*` functions instead.
163
- check(
164
- OrbitalFlow OK 20.71ms 732us
165
- AlignmentFlow OK 9.11ms 9.02ms
166
 
167
  ========================================================================
168
  8. MEMORY (B=32, n=128, k=64, d=256)
@@ -174,9 +151,9 @@ geolip.linalg backend:
174
  QuaternionLiteFlow 28.3
175
  VelocityFlow 48.0
176
  MagnitudeFlow 24.2
177
- OrbitalFlow 10.5
178
- AlignmentFlow 60.0
179
- Full ensemble 203.6
180
 
181
  ========================================================================
182
  Done.
 
5
  First barrage tests and compilation tests are finished. The orbital is yielding a true representation of omega.
6
 
7
  The speed and accuracy is both representationally superior and compiles safe to full-graph.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8
  ```
9
  geolip_core.linalg: available
10
  geolip.linalg backend:
 
25
 
26
  Flow Params Shape Time Conf Res norm
27
  ────────────────────── ──────── ────────────── ────────── ──────── ──────────
28
+ QuaternionFlow 70,725 (16, 64, 128) 267us 0.489 1.355
29
+ QuaternionLiteFlow 87,109 (16, 64, 128) 194us 0.471 1.282
30
+ VelocityFlow 70,210 (16, 64, 128) 175us 0.495 0.058
31
+ MagnitudeFlow 73,293 (16, 64, 128) 315us 0.526 0.601
32
+ OrbitalFlow 25,445 (16, 64, 128) 378us 0.524 1.650
33
+ AlignmentFlow 8,922 (16, 64, 128) 347us 0.485 2.905
34
 
35
  ========================================================================
36
  2. LINALG INTEGRATION
37
  ========================================================================
38
 
39
  Testing eigh dispatch in MagnitudeFlow and OrbitalFlow...
40
+ magnitude finite=True conf=0.469
41
+ orbital finite=True conf=0.462
42
 
43
+ Gram eigenspectrum: shape=(16, 12) range=[0.0002, 7.5887]
44
+ Eigenvector orth err: 6.86e+01
45
 
46
  ========================================================================
47
  3. MULTI-SCALE
 
50
  OrbitalFlow across scales:
51
  Config B n k d Time OK
52
  ────────── ──── ───── ───── ───── ────────── ────
53
+ tiny 4 16 8 64 328us OK
54
+ small 16 64 32 128 377us OK
55
+ medium 32 128 64 256 379us OK
56
+ large 64 256 128 256 430us OK
57
+ wide 8 512 256 512 426us OK
58
 
59
  ========================================================================
60
  4. ENSEMBLE FUSION
61
  ========================================================================
62
 
63
+ weighted: time=2.02ms norm=0.805 diversity=0.905
64
+ quaternion conf=0.489Β±0.002 res=1.355
65
+ quat_lite conf=0.471Β±0.002 res=1.285
66
+ velocity conf=0.495Β±0.003 res=0.058
67
+ magnitude conf=0.526Β±0.004 res=0.608
68
+ orbital conf=0.524Β±0.002 res=1.647
69
+ alignment conf=0.485Β±0.014 res=3.110
70
 
71
+ gated: time=2.02ms norm=0.663 diversity=0.905
72
+ quaternion conf=0.489Β±0.002 res=1.355
73
+ quat_lite conf=0.471Β±0.002 res=1.285
74
+ velocity conf=0.495Β±0.003 res=0.058
75
+ magnitude conf=0.526Β±0.004 res=0.608
76
+ orbital conf=0.524Β±0.002 res=1.647
77
+ alignment conf=0.485Β±0.014 res=3.110
78
 
79
+ residual: time=2.03ms norm=0.805 diversity=0.905
80
+ quaternion conf=0.489Β±0.002 res=1.355
81
+ quat_lite conf=0.471Β±0.002 res=1.285
82
+ velocity conf=0.495Β±0.003 res=0.058
83
+ magnitude conf=0.526Β±0.004 res=0.608
84
+ orbital conf=0.524Β±0.002 res=1.647
85
+ alignment conf=0.485Β±0.014 res=3.110
86
 
87
  ========================================================================
88
  5. GRADIENT HEALTH
 
91
  Flow Loss Grad norm Status
92
  ────────────────── ────────── ──────────── ────────
93
  quaternion mse 8.90e-04 OK
94
+ quat_lite mse 6.98e-04 OK
95
+ velocity mse 3.14e-04 OK
96
+ magnitude mse 8.74e-04 OK
97
+ orbital mse 1.53e-03 OK
98
+ alignment mse 1.55e-02 OK
99
+ quaternion cosine 9.37e-02 OK
100
+ quat_lite cosine 7.39e-02 OK
101
+ velocity cosine 2.66e-02 OK
102
+ magnitude cosine 8.45e-02 OK
103
+ orbital cosine 1.91e-01 OK
104
+ alignment cosine 1.94e+00 OK
105
+ quaternion norm 8.57e-02 OK
106
+ quat_lite norm 6.37e-02 OK
107
+ velocity norm 7.90e-03 OK
108
+ magnitude norm 6.39e-02 OK
109
+ orbital norm 1.45e-01 OK
110
+ alignment norm 2.95e+00 OK
111
 
112
  ========================================================================
113
  6. ABLATION (100 training steps, rotation target)
 
115
 
116
  Configuration MSE Params
117
  ─────────────────────────────────── ────────── ──────────
118
+ QuaternionFlow 0.0011 280,709
119
  QuaternionLiteFlow 0.0005 346,245
120
  VelocityFlow 0.0061 279,682
121
  MagnitudeFlow 0.0081 285,837
122
+ OrbitalFlow 0.0031 91,813
123
+ AlignmentFlow 0.0081 26,010
124
+ Quat + Orbital 0.0009 372,524
125
  Velocity + Magnitude 0.0062 565,521
126
+ Orbital + Alignment 0.0026 117,825
127
+ Velocity + Orbital 0.0021 371,497
128
+ Full (weighted) 0.0007 1,310,302
129
+ Full (residual) 0.0007 1,310,302
 
 
130
 
131
  ========================================================================
132
  7. COMPILE COMPATIBILITY
 
134
 
135
  Flow fullgraph Raw Compiled
136
  ────────────────────── ──────────── ────────── ────────────
137
+ QuaternionFlow OK 274us 200us
138
+ QuaternionLiteFlow OK 199us 181us
139
+ VelocityFlow OK 181us 168us
140
+ MagnitudeFlow OK 328us 351us
141
+ OrbitalFlow OK 376us 347us
142
+ AlignmentFlow OK 348us 362us
 
 
 
 
 
 
143
 
144
  ========================================================================
145
  8. MEMORY (B=32, n=128, k=64, d=256)
 
151
  QuaternionLiteFlow 28.3
152
  VelocityFlow 48.0
153
  MagnitudeFlow 24.2
154
+ OrbitalFlow 10.9
155
+ AlignmentFlow 15.0
156
+ Full ensemble 178.7
157
 
158
  ========================================================================
159
  Done.