Create deep_analysis_output.txt

deep_analysis_output.txt

@@ -0,0 +1,252 @@
+================================================================================
+FLOW MATCH RELAY — FULL ANALYSIS TOOLKIT
+  Device: cuda
+================================================================================
+Loading weights: 100%
+ 169/169 [00:00<00:00, 4115.48it/s, Materializing param=unet.time_emb.3.weight]
+  Params: 6,746,403 (relay: 76,384, 1.1%)
+  Relay modules: 2
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+TEST 1: Relay Diagnostics — Drift, Gates, Anchor Geometry
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+  unet.mid_block1.relay:
+    Patches: 16, Anchors/patch: 16, Patch dim: 16
+    Drift (rad):  mean=0.096787  std=0.031425  min=0.025191  max=0.204950
+    Drift (deg):  mean=5.55°  max=11.74°
+    Gates:        mean=0.0594  std=0.0017  min=0.0560  max=0.0619
+    Patch 0: anchor_cos mean=0.0443 max=0.6491 min=-0.5109
+    Patch 1: anchor_cos mean=0.0329 max=0.6991 min=-0.6696
+    Patch 2: anchor_cos mean=-0.0137 max=0.5337 min=-0.6256
+    Patch 3: anchor_cos mean=0.0025 max=0.5802 min=-0.5198
+    Near 0.29154: 0.0% of anchors within ±0.05
+    Per-patch mean drift:
+      Patch  0: 0.083389 rad (4.78°)
+      Patch  1: 0.105304 rad (6.03°)
+      Patch  2: 0.084447 rad (4.84°)
+      Patch  3: 0.115432 rad (6.61°)
+      Patch  4: 0.103072 rad (5.91°)
+      Patch  5: 0.053514 rad (3.07°)
+      Patch  6: 0.116086 rad (6.65°)
+      Patch  7: 0.074704 rad (4.28°)
+      Patch  8: 0.119158 rad (6.83°)
+      Patch  9: 0.134659 rad (7.72°)
+      Patch 10: 0.112943 rad (6.47°)
+      Patch 11: 0.076762 rad (4.40°)
+      Patch 12: 0.114448 rad (6.56°)
+      Patch 13: 0.109745 rad (6.29°)
+      Patch 14: 0.056048 rad (3.21°)
+      Patch 15: 0.088878 rad (5.09°)
+
+  unet.mid_block2.relay:
+    Patches: 16, Anchors/patch: 16, Patch dim: 16
+    Drift (rad):  mean=0.116423  std=0.031319  min=0.045233  max=0.246318
+    Drift (deg):  mean=6.67°  max=14.11°
+    Gates:        mean=0.0658  std=0.0019  min=0.0627  max=0.0702
+    Patch 0: anchor_cos mean=-0.0104 max=0.5409 min=-0.5740
+    Patch 1: anchor_cos mean=0.0241 max=0.5879 min=-0.5389
+    Patch 2: anchor_cos mean=0.0198 max=0.6510 min=-0.5917
+    Patch 3: anchor_cos mean=0.0060 max=0.5366 min=-0.7307
+    Near 0.29154: 0.4% of anchors within ±0.05
+    Per-patch mean drift:
+      Patch  0: 0.119493 rad (6.85°)
+      Patch  1: 0.139048 rad (7.97°)
+      Patch  2: 0.096759 rad (5.54°)
+      Patch  3: 0.159487 rad (9.14°)
+      Patch  4: 0.123855 rad (7.10°)
+      Patch  5: 0.094995 rad (5.44°)
+      Patch  6: 0.129639 rad (7.43°)
+      Patch  7: 0.077257 rad (4.43°)
+      Patch  8: 0.127187 rad (7.29°)
+      Patch  9: 0.114844 rad (6.58°)
+      Patch 10: 0.118358 rad (6.78°)
+      Patch 11: 0.097721 rad (5.60°)
+      Patch 12: 0.128776 rad (7.38°)
+      Patch 13: 0.127976 rad (7.33°)
+      Patch 14: 0.086075 rad (4.93°)
+      Patch 15: 0.121296 rad (6.95°)
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+TEST 2: Bottleneck Feature Geometry — CV at the relay point
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+  CV of bottleneck features at different timesteps:
+       t                                   module       CV    eff_d     norm
+    0.00                    unet.mid_block1.relay   0.5503     34.5    51.61
+    0.00                          unet.mid_block1   0.5316     34.5    51.61
+    0.00                            unet.mid_attn   0.4850     38.8    51.91
+    0.00                    unet.mid_block2.relay   0.5030     40.8    63.07
+    0.00                          unet.mid_block2   0.4623     40.8    63.07
+    0.25                    unet.mid_block1.relay   0.4621     43.1    52.02
+    0.25                          unet.mid_block1   0.4761     43.1    52.02
+    0.25                            unet.mid_attn   0.4429     47.2    52.35
+    0.25                    unet.mid_block2.relay   0.4353     48.7    63.47
+    0.25                          unet.mid_block2   0.4243     48.7    63.47
+    0.50                    unet.mid_block1.relay   0.4567     41.3    51.88
+    0.50                          unet.mid_block1   0.4775     41.3    51.88
+    0.50                            unet.mid_attn   0.4195     45.0    52.22
+    0.50                    unet.mid_block2.relay   0.4303     46.3    63.28
+    0.50                          unet.mid_block2   0.4266     46.3    63.28
+    0.75                    unet.mid_block1.relay   0.4972     35.2    51.85
+    0.75                          unet.mid_block1   0.5265     35.2    51.85
+    0.75                            unet.mid_attn   0.4821     39.5    52.07
+    0.75                    unet.mid_block2.relay   0.4743     40.7    63.12
+    0.75                          unet.mid_block2   0.4939     40.7    63.12
+    1.00                    unet.mid_block1.relay   0.5709     24.0    58.44
+    1.00                          unet.mid_block1   0.5983     24.0    58.44
+    1.00                            unet.mid_attn   0.5804     24.8    58.74
+    1.00                    unet.mid_block2.relay   0.6099     25.0    70.53
+    1.00                          unet.mid_block2   0.6065     25.0    70.53
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+TEST 3: Per-Class Anchor Utilization
+  Which anchors activate for each class?
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+  Nearest anchor distribution per class (Patch 0):
+       class     0     1     2     3     4     5     6     7     8     9    10    11    12    13    14    15
+       plane   0%    0%    0%    0%    0%    0%    0%   22%█   0%    0%    0%    0%    0%    0%   77%█   0% 
+        auto   0%    0%    0%    0%    0%    0%    0%   20%█   0%    0%    0%    0%    0%    0%   80%█   0% 
+        bird   0%    0%    1%    0%    0%    0%    0%   42%█   0%    0%    0%    0%    0%    0%   57%█   0% 
+         cat   0%    0%    0%    0%    0%    0%    0%   40%█   0%    0%    0%    0%    0%    0%   60%█   0% 
+        deer   0%    0%    1%    0%    0%    0%    0%   58%█   0%    0%    0%    0%    0%    0%   41%█   0% 
+         dog   0%    0%    0%    0%    0%    0%    0%   39%█   0%    0%    0%    0%    0%    0%   61%█   0% 
+        frog   0%    0%    0%    0%    0%    0%    0%   58%█   0%    0%    0%    0%    0%    0%   42%█   0% 
+       horse   0%    0%    0%    0%    0%    0%    0%   57%█   0%    0%    0%    0%    0%    0%   43%█   0% 
+        ship   0%    0%    0%    0%    0%    0%    0%    2%    0%    0%    0%    0%    0%    0%   98%█   0% 
+       truck   0%    0%    0%    0%    0%    0%    0%   54%█   0%    0%    0%    0%    0%    0%   46%█   0% 
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+TEST 4: Gate Dynamics — do relay gates respond to timestep?
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  Note: gates are learned parameters, not t-dependent.
+  Measuring relay output magnitude at different t instead.
+
+       t   relay_Δ_norm    relay_Δ_cos   input_norm  output_norm
+    0.00       116.3409     0.01136649       321.42       423.60
+    0.10       116.6277     0.01213849       323.16       424.56
+    0.25       116.7582     0.01237500       329.56       430.23
+    0.50       117.6399     0.01344323       331.23       431.26
+    0.75       118.0994     0.01383245       334.28       433.97
+    0.90       118.1683     0.01333457       350.58       449.37
+    1.00       119.9183     0.01262599       384.39       482.87
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+TEST 5: Generation Quality — Per-Class Diversity
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+       class  intra_cos  intra_std       CV     norm
+       plane     0.8748     0.0821   0.6407    33.17
+        auto     0.8073     0.0672   0.3739    27.07
+        bird     0.8779     0.0694   0.6898    27.62
+         cat     0.7880     0.0993   0.5233    26.13
+        deer     0.8695     0.0630   0.5286    26.79
+         dog     0.8332     0.0703   0.4365    27.46
+        frog     0.8332     0.0711   0.5007    25.48
+       horse     0.8313     0.0810   0.4687    29.01
+        ship     0.8920     0.0519   0.5069    32.44
+       truck     0.8311     0.0585   0.3973    29.58
+
+  ✓ Saved per-class grids to analysis/
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+TEST 6: Velocity Field — how does v_pred behave across t?
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+       t     v_norm      v_std   v·target    v_cos_t
+    0.05      53.05     0.9593     0.8501     0.3426
+    0.10      55.90     1.0099     0.8958     0.2447
+    0.25      58.23     1.0513     0.9352     0.1561
+    0.50      58.96     1.0641     0.9486     0.1258
+    0.75      58.85     1.0615     0.9411     0.1444
+    0.90      57.49     1.0360     0.9244     0.1876
+    0.95      56.50     1.0187     0.9126     0.2168
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+TEST 7: Ablation — Relay ON vs OFF during generation
+  Disable relay gates, measure generation difference
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  Relay ON  — mean pixel: 0.4528  std: 0.2227
+  Relay OFF — mean pixel: 0.4524  std: 0.2232
+  Pixel diff:    0.002449
+  Cosine sim:    0.999977
+  Max pixel Δ:   0.074433
+  ✓ Saved analysis/relay_ablation.png (top=ON, bottom=OFF)
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+TEST 8: Anchor Constellation Structure
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+  unet.mid_block1.relay:
+    Home↔Current cos: mean=0.994831  min=0.979071
+    Patch 0 anchor spread: mean_cos=0.0443  max_cos=0.6491  min_cos=-0.5109
+    Patch 1 anchor spread: mean_cos=0.0329  max_cos=0.6991  min_cos=-0.6696
+    Patch 2 anchor spread: mean_cos=-0.0137  max_cos=0.5337  min_cos=-0.6256
+    Patch 3 anchor spread: mean_cos=0.0025  max_cos=0.5802  min_cos=-0.5198
+    Patch 0 anchor eff_dim: 11.3 / 16
+    Patch 1 anchor eff_dim: 11.9 / 16
+    Patch 2 anchor eff_dim: 11.7 / 16
+    Patch 3 anchor eff_dim: 11.5 / 16
+
+  unet.mid_block2.relay:
+    Home↔Current cos: mean=0.992746  min=0.969817
+    Patch 0 anchor spread: mean_cos=-0.0104  max_cos=0.5409  min_cos=-0.5740
+    Patch 1 anchor spread: mean_cos=0.0241  max_cos=0.5879  min_cos=-0.5389
+    Patch 2 anchor spread: mean_cos=0.0198  max_cos=0.6510  min_cos=-0.5917
+    Patch 3 anchor spread: mean_cos=0.0060  max_cos=0.5366  min_cos=-0.7307
+    Patch 0 anchor eff_dim: 11.8 / 16
+    Patch 1 anchor eff_dim: 11.9 / 16
+    Patch 2 anchor eff_dim: 12.0 / 16
+    Patch 3 anchor eff_dim: 11.4 / 16
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+TEST 9: Sampling Trajectory — CV through ODE steps
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+    step      t     x_norm      x_std   CV_pixel
+       0   1.00      54.32     0.9801     0.0110
+       1   0.98      53.21     0.9602     0.0111
+       5   0.90      48.88     0.8819     0.0123
+      10   0.80      43.61     0.7869     0.0112
+      20   0.60      33.95     0.6123     0.0241
+      30   0.40      26.37     0.4764     0.0707
+      40   0.20      22.88     0.4184     0.1773
+      49   0.02      24.85     0.4587     0.2497
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+TEST 10: Inter-Class vs Intra-Class Separation
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  Inter-class cosine similarity matrix:
+            plan  auto  bird   cat  deer   dog  frog  hors  ship  truc
+     plane   1.0  0.99  0.99  0.98  0.98  0.97  0.98  0.99  1.00  0.98
+      auto  0.99   1.0  0.99  0.98  0.98  0.97  0.98  0.99  0.99  0.99
+      bird  0.99  0.99   1.0  0.99  0.99  0.99  0.99  1.00  0.99  0.98
+       cat  0.98  0.98  0.99   1.0  0.99  1.00  0.99  0.99  0.97  0.96
+      deer  0.98  0.98  0.99  0.99   1.0  0.99  1.00  0.99  0.97  0.96
+       dog  0.97  0.97  0.99  1.00  0.99   1.0  0.99  0.98  0.96  0.95
+      frog  0.98  0.98  0.99  0.99  1.00  0.99   1.0  0.99  0.97  0.97
+     horse  0.99  0.99  1.00  0.99  0.99  0.98  0.99   1.0  0.98  0.98
+      ship  1.00  0.99  0.99  0.97  0.97  0.96  0.97  0.98   1.0  0.99
+     truck  0.98  0.99  0.98  0.96  0.96  0.95  0.97  0.98  0.99   1.0
+
+  Intra-class cos: 0.8438 ± 0.0318
+  Inter-class cos: 0.8305 ± 0.0230
+  Separation ratio: 1.02×
+
+================================================================================
+ANALYSIS COMPLETE
+================================================================================
+
+  Files saved to analysis/:
+    - class_*.png:           per-class generated samples
+    - all_classes.png:       4 samples per class, 10 columns
+    - relay_ablation.png:    relay ON (top) vs OFF (bottom)
+
+  Key metrics to look for:
+    1. Anchor drift → did any converge near 0.29154?
+    2. Gate values → did they learn to open from init (0.047)?
+    3. Per-class anchor utilization → class-specific routing?
+    4. Relay ablation → does turning off the relay change generation?
+    5. Intra/inter-class ratio → > 1.0 means classes are separable
+    6. Velocity cosine → higher = better flow matching
+    7. CV through ODE → how does geometry evolve during generation?
+
+================================================================================