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)