search

featurecat / lizzie

Lizzie - Leela Zero Interface
GNU General Public License v3.0
970 stars 228 forks source link

Analytical anomaly with SGF with fixed stone placement #889

Closed hope366 closed 3 years ago

hope366 commented 3 years ago

I will explain how to reproduce this bug.

  1. Register multiple engines in lizzie in advance.For example, katago-eigen, katago-eigenavx2, katago-opencl, leelazero, etc. You may register multiple items with exactly the same contents.
  2. Prepare an SGF with a fixed stone arrangement instead of a normal SGF. Please use the following SGF this time. (; AP [MultiGo: 4.2.4] SZ [19] MULTIGOGM [1] AB [dd] [qd] AW [dp] [pp])
  3. After starting lizzie, load the above SGF and analyze it. You should probably see a number around 50%.
  4. Then switch to another engine. At this time, a value exceeding 90% should be displayed. (It should be around 50% originally)

Also, this bug may or may not occur due to subtle changes in the placement of the stones. For example, the SGF below did not have this bug. (; AP [MultiGo: 4.2.4] SZ [19] MULTIGOGM [1] AB [dd] [pd] AW [dp] [pp])

This bug occurred in all three types of lizzie: lizzie0.7.3, Lizzie-patched-0.7.4_210704b, Lizzie-improved-v2.7. The engines I used for verification are katago1.9.1eigen,katago1.9.1eigenavx2, katago1.9.1opencl, leelazero0.17. We were able to confirm this bug in all these engines.
We have only verified it in a Windows 10 environment.

hebaeba commented 3 years ago

After launching Lizzie, this bug seems to occur when loading a SGF with fixed stone placement and then switching engines. The same bug occurred when I first selected katago (GPU), loaded the SGF above, and then changed to katago-eigen. Thus, it seems to be independent of the engine type.