Bind GINO-Greif trilinear adapter, grouped support and CPU cost gates
This commit is contained in:
1 parent
367da00d71
commit
8810276beb
9 files changed
+2229
-15
No files matched your search
@@ -0,0 +1,109 @@
|
||||
"""CPU geometry, pinned spectral implementation and production guard replay."""
|
||||
|
||||
import ast
|
||||
from pathlib import Path
|
||||
import unittest
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
|
||||
import test_track_b_followup_preparation as followup
|
||||
import track_b_gino_greif_prepare as prep
|
||||
from track_b_gino_greif_adapter import COLUMNS, GRID, GreifGINO, deposit, gather, stencil
|
||||
from track_b_gino_shapenet_entry import read, sha
|
||||
|
||||
|
||||
class GreifAdapter(unittest.TestCase):
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
torch.set_num_threads(1)
|
||||
cls.candidate = read(prep.REVIEW_ROOT/'preparation-binding.json')
|
||||
if cls.candidate != prep.binding():
|
||||
raise ValueError('Candidate differs from the exact on-disk binding')
|
||||
|
||||
def test_all_19_columns_contribute_and_mass_is_preserved(self):
|
||||
ids, weights = stencil([[0.,0.,0.],[.12,.29,.35],[.2,.41,.46],[1.,1.,1.]], [1,4,8,21], 8)
|
||||
mass = torch.tensor([.1,1.,2.,3.], dtype=torch.float32)
|
||||
x = torch.arange(76, dtype=torch.float32).reshape(4,19) / 11
|
||||
grid, volume = deposit(x, mass, ids, weights, 8)
|
||||
self.assertEqual(grid.shape, (20,8,8,8))
|
||||
self.assertTrue(torch.allclose(volume.sum(), torch.tensor(1.)))
|
||||
expected = (mass/mass.sum()) @ x
|
||||
self.assertTrue(torch.allclose(grid[:19].reshape(19,-1) @ volume, expected))
|
||||
self.assertTrue(torch.equal(grid[-1].reshape(-1).bool(), volume>0))
|
||||
self.assertFalse(bool(torch.any(grid[:19].reshape(19,-1)[:, volume==0])))
|
||||
# At the outer corner distinct from others, all channels round trip exactly.
|
||||
self.assertTrue(torch.allclose(gather(grid[:19], ids, weights)[0], x[0]))
|
||||
|
||||
def test_exact_gather_on_affine_field_and_isotropic_translation(self):
|
||||
xyz = np.array([[0.,0.,0.],[.23,.6,.4],[1.,.7,.8]])
|
||||
ids = [23,25,58]
|
||||
corners, weights = stencil(xyz, ids, 8)
|
||||
c2, w2 = stencil(xyz + np.array([2.,-4.,3.]), ids, 8)
|
||||
self.assertTrue(torch.equal(c2,corners))
|
||||
self.assertTrue(torch.allclose(weights,w2,atol=1e-7))
|
||||
axes = torch.meshgrid(*(torch.arange(8,dtype=torch.float32) for _ in range(3)),indexing='ij')
|
||||
full = (3*axes[0] + 2*axes[1] - axes[2] + 10)[None]
|
||||
lattice = (xyz - (xyz.min(0)+xyz.max(0))/2)*5 + 3.5
|
||||
expected = torch.from_numpy(lattice @ np.array([3,2,-1]) + 10).float()
|
||||
self.assertTrue(torch.allclose(gather(full,corners,weights)[:,0],expected))
|
||||
with self.assertRaises(ValueError):
|
||||
stencil(xyz, [23,58,25], 8)
|
||||
|
||||
def test_pinned_model_full_64_cpu_forward_and_small_backward(self):
|
||||
xyz = [[0.,0.,0.],[.12,.29,.35],[.2,.41,.46],[1.,1.,1.]]
|
||||
raw = [2,4,8,12]
|
||||
x = torch.arange(76, dtype=torch.float32).reshape(4,19) / 100
|
||||
mass = torch.ones(4)
|
||||
with torch.no_grad():
|
||||
v, w = stencil(xyz, raw)
|
||||
core = GreifGINO().eval()
|
||||
y = core(x,mass,v,w)
|
||||
self.assertEqual(y.shape, (4,))
|
||||
self.assertTrue(torch.equal(y,torch.zeros(4)))
|
||||
self.assertEqual(core.size, GRID)
|
||||
v, w = stencil(xyz, raw, 8)
|
||||
core = GreifGINO(size=8,modes=(4,4,4))
|
||||
with torch.no_grad():
|
||||
core.readout[-1].weight.fill_(.01)
|
||||
core(x,mass,v,w).sum().backward()
|
||||
self.assertTrue(any(p.grad is not None and p.grad.abs().sum() > 0 for p in core.fno.parameters()))
|
||||
|
||||
def test_source_columns_and_h2_paired_domain(self):
|
||||
pilot = prep.read(prep.PILOT)
|
||||
source = Path(__file__).resolve().parents[1] / 'tools/gre24_features_v2.py'
|
||||
layout = next(ast.literal_eval(s.value) for s in ast.parse(source.read_text()).body if
|
||||
isinstance(s, ast.Assign) and any(isinstance(t, ast.Name) and t.id == 'COLUMNS' for t in s.targets))
|
||||
self.assertEqual(layout, COLUMNS)
|
||||
self.assertEqual(self.candidate['adapter_candidate']['columns'], list(layout))
|
||||
self.assertEqual([self.candidate['folds'][k]['counts']['test'] for k in ('0','1')],[121,91])
|
||||
self.assertEqual(self.candidate['h2_reference_support']['status'],'REFERENCE_MANIFESTS_ALIGNED')
|
||||
self.assertEqual(self.candidate['epochs'], pilot['full_epochs'])
|
||||
self.assertFalse(set(self.candidate['folds']['0']['parts']['test']) &
|
||||
set(self.candidate['folds']['1']['parts']['test']))
|
||||
|
||||
def test_real_production_guard_idle_and_busy_with_frozen_limits(self):
|
||||
for busy in (False, True):
|
||||
followup.FollowupPreparation().replay_guard('GINO-Greif', self.candidate, busy)
|
||||
|
||||
def test_release_gate_requires_both_full_folds_and_respects_phase_cap(self):
|
||||
c = self.candidate
|
||||
rows = {f:dict(fold=int(f),seed=0,epoch=1,learning_cases=d['counts']['learning'],
|
||||
learning_nodes=d['learning_nodes'],heldout_target_reads=0,
|
||||
epoch_seconds=50.+5*int(f),booked_gpu_seconds=100.) for f,d in c['folds'].items()}
|
||||
k = dict(other_unsettled=0.,approved_greif_cap=80000.,remaining_overhead=500.)
|
||||
result = prep.first_epoch_gate(c, rows, **k)
|
||||
self.assertTrue(result['release_epoch2'])
|
||||
self.assertEqual(result['charged_first_epochs_and_setup_gpu_seconds'],200.)
|
||||
self.assertEqual(result['remaining_train_gpu_seconds'],399*105.)
|
||||
self.assertFalse(prep.first_epoch_gate(c, {'0':rows['0']}, **k)['release_epoch2'])
|
||||
remainder=c['projection']['upper_remaining_before_other_bookings_gpu_seconds']
|
||||
blocked=prep.first_epoch_gate(c, rows, **dict(k,other_unsettled=remainder-10000))
|
||||
self.assertEqual(blocked['status'],'STOP_PROJECTED_CAP')
|
||||
self.assertFalse(blocked['release_epoch2'])
|
||||
with self.assertRaises(ValueError):
|
||||
prep.first_epoch_gate(c, dict(rows, **{'1':dict(rows['1'],learning_cases=373)}), **k)
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
unittest.main()
|
||||
Reference in new issue
Block a user