def InvariantPointAttentionTest(args, config, global_config):
print('InvariantPointAttentionTest')
feat, params, res = load_data(args, 'InvariantPointAttention')
conf = config.model.heads.structure_module
attn_single = InvariantPointAttention(
conf,
global_config,
num_feat_1d=feat['inputs_1d'].shape[-1],
num_feat_2d=feat['inputs_2d'].shape[-1])
attn_single.load_weights_from_af2(params,
rel_path='invariant_point_attention')
attn_batch = InvariantPointAttentionB(
conf,
global_config,
num_feat_1d=feat['inputs_1d'].shape[-1],
num_feat_2d=feat['inputs_2d'].shape[-1])
attn_batch.load_weights_from_af2(params,
rel_path='invariant_point_attention')
print('inputs1d:', feat['inputs_1d'].size())
print('inputs2d:', feat['inputs_2d'].size())
print('activations:', feat['activations'].size())
print('mask:', feat['mask'].size())
batch_size = 8
inputs_1d_batch = feat['inputs_1d'][None, ...].repeat(batch_size, 1, 1)
inputs_2d_batch = feat['inputs_2d'][None, ...].repeat(batch_size, 1, 1, 1)
activations_batch = feat['activations'][None, ...].repeat(batch_size, 1, 1)
mask_batch = feat['mask'][None, ...].repeat(batch_size, 1, 1)
qa_single = QuatAffine.from_tensor(
feat['activations'].to(dtype=torch.float32))
qa_batch = QuatAffine.from_tensor(
activations_batch.to(dtype=torch.float32))
res_single = attn_single(inputs_1d=feat['inputs_1d'],
inputs_2d=feat['inputs_2d'],
mask=feat['mask'],
affine=qa_single)
res_batch = attn_batch(inputs_1d=inputs_1d_batch,
inputs_2d=inputs_2d_batch,
mask=mask_batch,
affine=qa_batch)
print(check_recursive(res_single, res))
print(check_recursive(res_batch[0, ...], res))
for i in range(batch_size):
err = torch.sum(torch.abs(res_batch[i, ...] - res_single))
print(i, err.item())
assert err < 1e-2
def pre_compose_test(args, name):
(tensor, update), res = load_data(args, name)
qa = QuatAffine.from_tensor(tensor)
this_res = qa.pre_compose(update).to_tensor()
err, max_err, mean_err = check_recursive(res, this_res)
print(f'Max error = {max_err}, mean error = {mean_err} total error = {err}')
print(f'Success = {(max_err < 1e-4) and (mean_err < 1e-5)}')
def invert_point_test(args, name):
(tensor, point), res = load_data(args, name)
qa = QuatAffine.from_tensor(tensor)
this_res = qa.invert_point(point)
err, max_err, mean_err = check_recursive(res, this_res)
print(f'Max error = {max_err}, mean error = {mean_err} total error = {err}')
print(f'Success = {(max_err < 1e-4) and (mean_err < 1e-5)}')
def test_torsion_angles_to_frames(args):
print('test_torsion_angles_to_frames')
(activations, aatype, torsion_angles_sin_cos), res = load_data(args, 'test_torsion_angles_to_frames')
rigs = QuatAffine.from_tensor(activations).to_rigids()
this_res = torsion_angles_to_frames(aatype=aatype, backb_to_global=rigs, torsion_angles_sin_cos=torsion_angles_sin_cos)
this_res = affine.rigids_to_tensor_flat12(this_res)
print(check_recursive(this_res, res))
def forward(self, activations:torch.Tensor, sequence_mask:torch.Tensor, update_affine:bool, initial_act:torch.Tensor,
is_training:bool=False, static_feat_2d:torch.Tensor=None, aatype:torch.Tensor=None):
affine = QuatAffine.from_tensor(activations['affine'])
act = activations['act']
affine.cast_to(dtype=act.dtype) #Casting to float16
attn = self.attention_module(inputs_1d=act, inputs_2d=static_feat_2d, mask=sequence_mask, affine=affine)
act = act + attn
act = self.attention_layer_norm(act)
input_act = act
for i in range(self.config.num_layer_in_transition):
act = self.transition[i](act)
if i < self.config.num_layer_in_transition - 1:
act = self.relu(act)
act = act + input_act
act = self.transition_layer_norm(act)
if update_affine:
affine_update = self.affine_update(act)
#Not sure if it helps
affine.cast_to(dtype=torch.float32) #Casting to float32 to avoid multiplying rotations in float16
affine = affine.pre_compose(affine_update.to(dtype=torch.float32))
affine.cast_to(dtype=act.dtype) #Casting back to float16
sc = self.side_chain(affine.scale_translation(self.config.position_scale), [act, initial_act], aatype)
#Casting final affines to float32
outputs = {'affine': affine.to_tensor().to(dtype=torch.float32), 'sc': sc}
new_activations = {'act': act, 'affine': affine.apply_rotation_tensor_fn(torch.detach).to_tensor()}
return new_activations, outputs
def init_test(args, name):
args, res = load_data(args, name)
qa = QuatAffine.from_tensor(*args)
this_res = qa.quaternion, qa.translation, qa.rotation
for a,b in zip(res, this_res):
err, max_err, mean_err = check_recursive(a,b)
print(f'Max error = {max_err}, mean error = {mean_err} total error = {err}')
print(f'Success = {(max_err < 1e-4) and (mean_err < 1e-5)}')
def to_tensor_test(args, name):
(tensor, ), res = load_data(args, name)
qa = QuatAffine.from_tensor(tensor)
qa = qa.apply_rotation_tensor_fn(lambda t: t+1.0)
this_res = qa.to_tensor()
err, max_err, mean_err = check_recursive(res,this_res)
print(f'Max error = {max_err}, mean error = {mean_err} total error = {err}')
print(f'Success = {(max_err < 1e-4) and (mean_err < 1e-5)}')
def test_frames_and_literature_positions_to_atom14_pos(args):
print('test_frames_and_literature_positions_to_atom14_pos')
(activations, aatype, torsion_angles_sin_cos), res = load_data(args, 'test_frames_and_literature_positions_to_atom14_pos')
rigs = QuatAffine.from_tensor(activations).to_rigids()
all_frames = torsion_angles_to_frames(aatype=aatype, backb_to_global=rigs, torsion_angles_sin_cos=torsion_angles_sin_cos)
this_res = frames_and_literature_positions_to_atom14_pos(aatype=aatype, all_frames_to_global=all_frames)
this_res = affine.vecs_to_tensor(this_res)
print(check_recursive(this_res, res))
def apply_rot_func_test(args, name):
(tensor, ), res = load_data(args, name)
qa = QuatAffine.from_tensor(tensor)
qa = qa.apply_rotation_tensor_fn(lambda t: t+1.0)
this_res = qa.quaternion, qa.translation, qa.rotation
for a,b in zip(res, this_res):
err, max_err, mean_err = check_recursive(a,b)
print(f'Max error = {max_err}, mean error = {mean_err} total error = {err}')
print(f'Success = {(max_err < 1e-4) and (mean_err < 1e-5)}')
def InvariantPointAttentionTest(args, config, global_config):
print('InvariantPointAttentionTest')
feat, params, res = load_data(args, 'InvariantPointAttention')
conf = config.model.heads.structure_module
attn = InvariantPointAttention( conf, global_config,
num_res=feat['inputs_1d'].shape[-2],
num_feat_1d=feat['inputs_1d'].shape[-1],
num_feat_2d=feat['inputs_2d'].shape[-1])
attn.load_weights_from_af2(params, rel_path='invariant_point_attention')
qa = QuatAffine.from_tensor(feat['activations'].to(dtype=torch.float32))
this_res = attn(inputs_1d = feat['inputs_1d'], inputs_2d = feat['inputs_2d'], mask=feat['mask'], affine=qa)
print(check_recursive(this_res, res))
def loss(self, value: Dict[str, torch.Tensor],
batch: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
predicted_affine = QuatAffine.from_tensor(
value['structure_module']['final_affines'])
true_affine = QuatAffine.from_tensor(batch['backbone_affine_tensor'])
mask = batch['backbone_affine_mask']
square_mask = mask[:, None] * mask[None, :]
num_bins = self.config.num_bins
breaks = value['predicted_aligned_error']['breaks']
logits = value['predicted_aligned_error']['logits']
def _local_frame_points(affine):
points = [x.unsqueeze(dim=-2) for x in affine.translation]
return affine.invert_point(points, extra_dims=1)
error_dist2_xyz = [
torch.square(x - y)
for x, y in zip(_local_frame_points(predicted_affine),
_local_frame_points(true_affine))
]
error_dist2 = sum(error_dist2_xyz)
error_dist2.detach()
sq_breaks = torch.square(breaks)
true_bins = torch.sum(
(error_dist2[..., None] > sq_breaks).to(dtype=torch.int32), dim=-1)
errors = self.loss_function(logits, true_bins)
loss = torch.sum(errors * square_mask, dim=(-2, -1)) / (
1e-8 + torch.sum(square_mask, dim=(-2, -1)))
if self.config.filter_by_resolution:
loss *= ((batch['resolution'] >= self.config.min_resolution) &
(batch['resolution'] < self.config.max_resolution)).to(
dtype=torch.float32)
return {'loss': loss}
def MultiRigidSidechainTest(args, config, global_config):
print('MultiRigidSidechainTest')
feat, params, res = load_data(args, 'MultiRigidSidechain')
conf = config.model.heads.structure_module.sidechain
for key in params.keys():
print(key)
for param in params[key].keys():
print('\t' + param)
attn = MultiRigidSidechain( conf, global_config,
repr_dim=feat['representations_list'][0].shape[-1],
num_repr=len(feat['representations_list'])
)
attn.load_weights_from_af2(params, rel_path='rigid_sidechain')
qa = QuatAffine.from_tensor(feat['activations'].to(dtype=torch.float32))
this_res = attn(affine = qa, representations_list = feat['representations_list'], aatype=feat['aatype'])
this_res['atom_pos'] = affine.vecs_to_tensor(this_res['atom_pos'])
this_res['frames'] = affine.rigids_to_tensor_flat12(this_res['frames'])
print(check_recursive(this_res, res))
def backbone_loss(self, ret:Dict[str, torch.Tensor], value:Dict[str, torch.Tensor], batch:Dict[str, torch.Tensor]) -> None: affine_trajectory = QuatAffine.from_tensor(value['traj']) rigid_trajectory = affine_trajectory.to_rigids() # gt_trajectory = QuatAffine.from_tensor(batch['backbone_affine_tensor']) # gt_rigid = gt_trajectory.to_rigids() gt_rigid = rigids_from_tensor_flat12(batch['backbone_affine_tensor']) backbone_mask = batch['backbone_affine_mask'] fape_loss_fn = functools.partial(protein.frame_aligned_point_error, l1_clamp_distance=self.config.fape.clamp_distance, length_scale=self.config.fape.loss_unit_distance) fape_loss = [] for i in range(value['traj'].size(0)): pred = rigids_apply(lambda x: x[i,...], rigid_trajectory) fape_loss.append(fape_loss_fn(pred, gt_rigid, backbone_mask, pred.trans, gt_rigid.trans, backbone_mask)) fape_loss = torch.stack(fape_loss, dim=0) # print(gt_rigid.trans) # print(gt_rigid.trans) if 'use_clamped_fape' in batch: use_clamped_fape = torch.Tensor(batch['use_clamped_fape'], dtype=torch.float32) unclamped_fape_loss_fn = functools.partial( protein.frame_aligned_point_error, l1_clamp_distance=None, length_scale=self.config.fape.loss_unit_distance) fape_loss_unclamped = [] for i in range(value['traj'].size(0)): pred = rigids_apply(lambda x: x[i,...], rigid_trajectory) fape_loss_unclamped = unclamped_fape_loss_fn( pred, gt_rigid, backbone_mask, pred.trans, gt_rigid.trans, backbone_mask) fape_loss_unclamped = torch.stack(fape_loss_unclamped, dim=0) fape_loss = fape_loss*use_clamped_fape + fape_loss_unclamped*(1.0-use_clamped_fape) ret['fape'] = fape_loss[-1] ret['loss'] += torch.mean(fape_loss)