Fix bug - stepsize from windowed adaptation is reset to 1 when term_buffer == 0

[mitzimorris]

Submission Checklist

• Run unit tests: ./runTests.py src/test/unit
• Run cpplint: make cpplint
• Declare copyright holder and open-source license: see below

Summary

The underlying bug occurs when stepsize_adaptation.restart is called immediately before stepsize_adaptation.complete_adaptation; the former resets value x_bar_ to 0; the latter returns exp(x_bar_) which is 1. Added a check to complete_adaptation which doesn't update the stepsize when x_bar_ is 0.

I found this bug after refactoring the behavior of the windowed_adaptation object in order to make sure that the adaptation schedule was being computed correctly. The code uses names phase1, phase2, and phase3 to track the adaptation schedule. The windowed_adaptation object exposes a counter, cur_iter_ which is incremented by the adapter objects, adapt_diag_e_nuts et al.

Unit tests on the windowed_adaptation check that the schedule is computed properly.

Unit tests on the hmc_nuts_diag_e_adapt_sampler check that the adapted stepsize matches the stepsize used during sampling.

Intended Effect

If sampler config sets 'term_buffer=0', preserve the evolved stepsize from previous adaptation phases.

How to Verify

Unit tests

Side Effects

N/A

Documentation

N/A

Copyright and Licensing

Please list the copyright holder for the work you are submitting (this will be you or your assignee, such as a university or company):

Columbia University

By submitting this pull request, the copyright holder is agreeing to license the submitted work under the following licenses:

• Code: BSD 3-clause (https://opensource.org/licenses/BSD-3-Clause)
• Documentation: CC-BY 4.0 (https://creativecommons.org/licenses/by/4.0/)

[stan-buildbot]

---

Name	Old Result	New Result	Ratio	Performance change( 1 - new / old )
arma/arma.stan	0.32	0.3	1.07	6.21% faster
low_dim_corr_gauss/low_dim_corr_gauss.stan	0.01	0.01	1.08	7.05% faster
gp_regr/gen_gp_data.stan	0.03	0.02	1.05	4.49% faster
gp_regr/gp_regr.stan	0.1	0.09	1.06	6.07% faster
sir/sir.stan	70.63	66.35	1.06	6.06% faster
irt_2pl/irt_2pl.stan	4.23	3.98	1.06	6.04% faster
eight_schools/eight_schools.stan	0.06	0.05	1.07	6.27% faster
pkpd/sim_one_comp_mm_elim_abs.stan	0.28	0.24	1.16	13.83% faster
pkpd/one_comp_mm_elim_abs.stan	19.96	18.62	1.07	6.71% faster
garch/garch.stan	0.43	0.4	1.07	6.28% faster
low_dim_gauss_mix/low_dim_gauss_mix.stan	2.79	2.61	1.07	6.63% faster
arK/arK.stan	1.84	1.71	1.08	7.06% faster
gp_pois_regr/gp_pois_regr.stan	2.87	2.74	1.05	4.61% faster
low_dim_gauss_mix_collapse/low_dim_gauss_mix_collapse.stan	8.8	8.39	1.05	4.6% faster
performance.compilation	185.51	178.43	1.04	3.82% faster
Mean result: 1.0687889508974118

---

Jenkins Console Log
Blue Ocean
Commit hash: 3dcd12419d102d4023893031ac04c0b52dd79879

---

Machine information

No LSB modules are available. Distributor ID: Ubuntu Description: Ubuntu 20.04.3 LTS Release: 20.04 Codename: focal

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
Address sizes: 46 bits physical, 48 bits virtual
CPU(s): 80
On-line CPU(s) list: 0-79
Thread(s) per core: 2
Core(s) per socket: 20
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 85
Model name: Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
Stepping: 4
CPU MHz: 2400.000
CPU max MHz: 3700.0000
CPU min MHz: 1000.0000
BogoMIPS: 4800.00
Virtualization: VT-x
L1d cache: 1.3 MiB
L1i cache: 1.3 MiB
L2 cache: 40 MiB
L3 cache: 55 MiB
NUMA node0 CPU(s): 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78
NUMA node1 CPU(s): 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63,65,67,69,71,73,75,77,79
Vulnerability Gather data sampling: Mitigation; Microcode
Vulnerability Itlb multihit: KVM: Mitigation: VMX disabled
Vulnerability L1tf: Mitigation; PTE Inversion; VMX conditional cache flushes, SMT vulnerable
Vulnerability Mds: Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Meltdown: Mitigation; PTI
Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Mitigation; IBRS
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; IBRS; IBPB conditional; STIBP conditional; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Mitigation; Clear CPU buffers; SMT vulnerable
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single pti intel_ppin ssbd mba ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req pku ospke md_clear flush_l1d arch_capabilities

G++:
g++ (Ubuntu 9.4.0-1ubuntu1~20.04) 9.4.0
Copyright (C) 2019 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Clang:
clang version 10.0.0-4ubuntu1
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

[WardBrian]

I think this is incorrect. The condition added here means that the stepsize adaptation is never restarted, even as the metric changes, which means the dual averaging is over the entire warmup phase, not just the most recent window.

I think the correct solution will need to modify learn_variance and learn_covariance such that they return false when the next window is size 0

[WardBrian]

I think the correct condition is to update the current return true in the if (end_adaptation_window()) branch of the variance/covariance functions with return adapt_term_buffer_ > 0 || adapt_window_counter_ < adapt_next_window_;

At the end of metric adaptation, adapt_window_counter_ = adapt_next_window_ + 1. If the term buffer is 0, we're done, otherwise we're doing the final stepsize adaptation. But this is subtle, and will require some delicate testing, I think

[mitzimorris]

no, I think it's simpler than that - there's an off by one error in the logic of end_adaptation_window.

stan/src/stan/mcmc/windowed_adaptation.hpp

Lines 86 to 100 in 6d0d863

	bool end_adaptation_window() {
	return (adapt_window_counter_ == adapt_next_window_)
	&& (adapt_window_counter_ != num_warmup_);
	}
	void compute_next_window() {
	if (adapt_next_window_ == num_warmup_ - adapt_term_buffer_ - 1)
	return;
	adapt_window_size_ *= 2;
	adapt_next_window_ = adapt_window_counter_ + adapt_window_size_;
	if (adapt_next_window_ == num_warmup_ - adapt_term_buffer_ - 1)
	return;

[mitzimorris]

I've added a bunch of tests to check that this logic doesn't affect the overall number of warmup iterations and that the stepsize is not reset to 1 at the final iteration.

[WardBrian]

For the tests, I think it would be better to check that the third column isn't ==1 for the first sampling iteration, rather than relying on the string outputs.

I think there's also still an issue here, possibly a new one? I'm testing 3 configurations which should all have identical metric adaptation (at least, if my intuition here is correct), but the term_buffer=0 case still doesn't. They do appear to have identical draws up until the 1000th iteration, as we would expect

	num_warmup=1050 adapt term_buffer=50	num_warmup=1001 adapt term_buffer=1	num_warmup=1000 adapt term_buffer=0
adapted metric	0.463521	0.463521	0.411266

(this is for the bernoulli model with seed 1234).

I think the issue is that the adaptation_window() might have the same off-by-one error, meaning one extra iteration is included? Though I think that would mean there was an iteration of warmup where nothing is happening, so maybe the off-by-one isn't in the check but rather in setting the value to num_warmup_ - adapt_term_buffer_ - 1 in the first place?

[mitzimorris]

@WardBrian : Happy Birthday!!!

[stan-buildbot]

---

Name	Old Result	New Result	Ratio	Performance change( 1 - new / old )
arma/arma.stan	0.33	0.3	1.08	7.41% faster
low_dim_corr_gauss/low_dim_corr_gauss.stan	0.01	0.01	1.1	9.21% faster
gp_regr/gen_gp_data.stan	0.03	0.02	1.05	4.91% faster
gp_regr/gp_regr.stan	0.09	0.09	1.03	2.55% faster
sir/sir.stan	69.12	66.14	1.05	4.31% faster
irt_2pl/irt_2pl.stan	4.19	3.95	1.06	5.6% faster
eight_schools/eight_schools.stan	0.06	0.05	1.04	3.83% faster
pkpd/sim_one_comp_mm_elim_abs.stan	0.25	0.24	1.05	5.07% faster
pkpd/one_comp_mm_elim_abs.stan	19.9	18.82	1.06	5.44% faster
garch/garch.stan	0.45	0.4	1.13	11.23% faster
low_dim_gauss_mix/low_dim_gauss_mix.stan	2.76	2.64	1.04	4.16% faster
arK/arK.stan	1.81	1.73	1.04	4.16% faster
gp_pois_regr/gp_pois_regr.stan	2.85	2.72	1.05	4.7% faster
low_dim_gauss_mix_collapse/low_dim_gauss_mix_collapse.stan	8.91	8.37	1.06	6.03% faster
performance.compilation	180.47	179.41	1.01	0.59% faster
Mean result: 1.0564845412726898

---

Jenkins Console Log
Blue Ocean
Commit hash: 3dcd12419d102d4023893031ac04c0b52dd79879

---

Machine information

No LSB modules are available. Distributor ID: Ubuntu Description: Ubuntu 20.04.3 LTS Release: 20.04 Codename: focal

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
Address sizes: 46 bits physical, 48 bits virtual
CPU(s): 80
On-line CPU(s) list: 0-79
Thread(s) per core: 2
Core(s) per socket: 20
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 85
Model name: Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
Stepping: 4
CPU MHz: 2400.000
CPU max MHz: 3700.0000
CPU min MHz: 1000.0000
BogoMIPS: 4800.00
Virtualization: VT-x
L1d cache: 1.3 MiB
L1i cache: 1.3 MiB
L2 cache: 40 MiB
L3 cache: 55 MiB
NUMA node0 CPU(s): 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78
NUMA node1 CPU(s): 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63,65,67,69,71,73,75,77,79
Vulnerability Gather data sampling: Mitigation; Microcode
Vulnerability Itlb multihit: KVM: Mitigation: VMX disabled
Vulnerability L1tf: Mitigation; PTE Inversion; VMX conditional cache flushes, SMT vulnerable
Vulnerability Mds: Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Meltdown: Mitigation; PTI
Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Mitigation; IBRS
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; IBRS; IBPB conditional; STIBP conditional; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Mitigation; Clear CPU buffers; SMT vulnerable
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single pti intel_ppin ssbd mba ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req pku ospke md_clear flush_l1d arch_capabilities

G++:
g++ (Ubuntu 9.4.0-1ubuntu1~20.04) 9.4.0
Copyright (C) 2019 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Clang:
clang version 10.0.0-4ubuntu1
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

[mitzimorris]

I think there's also still an issue here, possibly a new one? I'm testing 3 configurations which should all have identical metric adaptation (at least, if my intuition here is correct), but the term_buffer=0 case still doesn't. They do appear to have identical draws up until the 1000th iteration, as we would expect

agreed - there is definitely a problem here. working on it.

[mitzimorris]

discretion is the better part of valor. refactoring not worth it. will submit fix for bug and unit tests in a new PR.