File size: 3,238 Bytes
4c231da | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 | from pathlib import Path
import pytest
import torch
from atomtoken import Config, FSQ, StructureTokenizer, load_assembly, tokenizer_loss
from atomtoken.geometry import neighbors, signed_volumes
EXAMPLE = Path(__file__).parents[1] / "examples" / "peptide_ligand.json"
@pytest.fixture()
def assembly():
torch.manual_seed(19)
torch.set_num_threads(1)
return (*load_assembly(EXAMPLE), StructureTokenizer(Config.tiny()).eval())
def test_all_codes_round_trip():
fsq = FSQ(5)
with torch.no_grad():
fsq.project.weight.copy_(torch.eye(5))
fsq.project.bias.zero_()
ids = torch.arange(8192)
coords = fsq.from_ids(ids)
actual, codes, _ = fsq(torch.atanh(coords.clamp(-.999999, .999999)))
assert torch.equal(actual, ids)
torch.testing.assert_close(codes, coords)
with pytest.raises(ValueError):
fsq.from_ids(torch.tensor([8192]))
def test_rigid_motion_and_chirality(assembly):
chem, x, mask, model = assembly
rot, _ = torch.linalg.qr(torch.randn(3,3))
rot[:, 0] *= torch.linalg.det(rot).sign()
transformed = x@rot + torch.tensor([7., -3., 2.])
h, _ = model.encoder(chem,x,mask)
h2, _ = model.encoder(chem,transformed,mask)
torch.testing.assert_close(h,h2,rtol=2e-5,atol=3e-5)
ids = model.encode(chem,x,mask)
assert torch.equal(ids, model.encode(chem,transformed,mask))
assert ids.shape == (5,)
code = model.fsq.from_ids(ids)
v = model.decoder(chem,code,x/10,.4)
v2 = model.decoder(chem,code,transformed/10,.4)
torch.testing.assert_close(v2,v@rot,rtol=2e-4,atol=1e-5)
vol = signed_volumes(x,chem.triplets)
assert vol.abs().max()>.1
torch.testing.assert_close(vol, -signed_volumes(-x,chem.triplets))
def test_missing_atoms_and_graph_types(assembly):
chem, x, mask, model = assembly
# A stretched covalent link remains present, including when spatial caps are zero.
far = x.clone()
far[0] += 100
dst, src, rel, _ = neighbors(chem, far, mask, caps=(0,0))
assert len(dst)==2*len(chem.bonds) and ((dst==0)&(src==1)&(rel==0)).any()
dst, src, rel, _ = neighbors(chem, x, mask, caps=(1,1))
assert (rel==2).any() and (rel==1).any()
for kind in (1,2):
assert torch.bincount(dst[rel==kind], minlength=len(x)).max()<=1
mask[[3,11]] = False
orig = model.encode(chem,x,mask)
x[~mask] = torch.nan
assert torch.equal(orig,model.encode(chem,x,mask)) and orig[-1]==-1
loss,_=tokenizer_loss(model,chem,x,mask)
assert torch.isfinite(loss)
loss.backward()
for module in (model.encoder,model.fsq,model.decoder):
grads=[p.grad for p in module.parameters() if p.grad is not None]
assert all(g.isfinite().all() for g in grads) and sum(g.abs().sum() for g in grads)>0
def test_decoder_uses_codes_and_seed(assembly):
chem, x, mask, model = assembly
ids = model.encode(chem,x,mask)
sample = lambda i: model.decode(chem,i,steps=3,generator=torch.Generator().manual_seed(5))
a = sample(ids)
x.zero_()
assert torch.equal(a,sample(ids)) and a.shape == (12,3)
assert not torch.allclose(a,sample((ids+27)%8192))
assert a.isfinite().all()
torch.testing.assert_close(a.mean(0),torch.zeros(3),atol=1e-5,rtol=0)
|