Skip to content

Remove inference multiplicity #36

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Remove inference multiplicity #36

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 0 additions & 3 deletions configs/experiment/test.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,3 @@ model:
tau: 0.01
log_timesteps: True
w_cutoff: 0.99

processor:
inference_multiplicity: 1
86 changes: 53 additions & 33 deletions src/simplefold/inference.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -183,7 +183,6 @@ def initialize_others(args, device):
scale=16.0,
ref_scale=5.0,
multiplicity=1,
inference_multiplicity=args.nsample_per_protein,
backend=args.backend,
)

Expand All @@ -210,44 +209,65 @@ def generate_structure(
model, plddt_latent_module, plddt_out_module, device
):
# run inference for target protein
if args.backend == "torch":
noise = torch.randn_like(batch['coords']).to(device)
elif args.backend == "mlx":
noise = mx.random.normal(batch['coords'].shape)
out_dict = sampler.sample(model, flow, noise, batch)

if args.plddt:
coord_samples = []
pad_mask = batch["atom_pad_mask"]
if args.plddt and plddt_latent_module is not None and plddt_out_module is not None:
compute_plddt = True
plddt_samples = []
else:
compute_plddt = False
plddt_samples = None
plddts = None
for _ in range(args.nsample_per_protein):
if args.backend == "torch":
t = torch.ones(batch['coords'].shape[0], device=device)
# use unscaled coords to extract latent for pLDDT prediction
out_feat = plddt_latent_module(
out_dict["denoised_coords"].detach(), t, batch)
plddt_out_dict = plddt_out_module(
out_feat["latent"].detach(),
batch,
)
noise = torch.randn_like(batch["coords"]).to(device)
elif args.backend == "mlx":
t = mx.ones(batch['coords'].shape[0])
# use unscaled coords to extract latent for pLDDT prediction
out_feat = plddt_latent_module(
out_dict["denoised_coords"], t, batch)
plddt_out_dict = plddt_out_module(
out_feat["latent"],
batch,
)
# scale pLDDT to [0, 100]
plddts = plddt_out_dict["plddt"] * 100.0
else:
plddts = None
noise = mx.random.normal(batch["coords"].shape)
out_dict = sampler.sample(model, flow, noise, batch)

if compute_plddt:
if args.backend == "torch":
t = torch.ones(batch['coords'].shape[0], device=device)
out_feat = plddt_latent_module(
out_dict["denoised_coords"].detach(), t, batch
)
plddt_out_dict = plddt_out_module(
out_feat["latent"].detach(),
batch,
)
elif args.backend == "mlx":
t = mx.ones(batch['coords'].shape[0])
out_feat = plddt_latent_module(
out_dict["denoised_coords"], t, batch
)
plddt_out_dict = plddt_out_module(
out_feat["latent"],
batch,
)
# scale pLDDT to [0, 100]
plddt_samples.append(plddt_out_dict["plddt"] * 100.0)

out_dict = processor.postprocess(out_dict, batch)
if args.backend == "torch":
coord_samples.append(out_dict["denoised_coords"].detach())
else:
coord_samples.append(out_dict["denoised_coords"])

out_dict = processor.postprocess(out_dict, batch)
# sampled_coord = out_dict['denoised_coords'].detach()
if args.backend == "torch":
sampled_coord = out_dict['denoised_coords'].detach()
sampled_coord = torch.cat(coord_samples, dim=0)
pad_mask = pad_mask.detach().repeat_interleave(
args.nsample_per_protein, dim=0
)
if compute_plddt:
plddts = torch.cat(plddt_samples, dim=0).detach()
else:
sampled_coord = out_dict['denoised_coords']
sampled_coord = mx.concatenate(coord_samples, axis=0)
pad_mask = mx.concatenate(
[pad_mask] * args.nsample_per_protein, axis=0
)
if compute_plddt:
plddts = mx.concatenate(plddt_samples, axis=0)

pad_mask = batch['atom_pad_mask']
return sampled_coord, pad_mask, plddts


Expand Down
6 changes: 2 additions & 4 deletions src/simplefold/processor/protein_processor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,15 +26,13 @@ def __init__(
scale=16.0,
ref_scale=5.0,
multiplicity=1,
inference_multiplicity=1,
backend="torch",
):
self.device = device
self.scale = scale
self.ref_scale = ref_scale
# if multiplicity > 1, effective batch size is multiplicity * batch_size
self.multiplicity = multiplicity
self.inference_multiplicity = inference_multiplicity
self.backend = backend
if self.backend == "mlx":
self.center_random_fn = mlx_center_random
Expand Down Expand Up @@ -69,7 +67,7 @@ def process_esm(

esmaa = af2_idx_to_esm_idx(aatype, mask, af2_to_esm)

multiplicity = self.multiplicity if not inference else self.inference_multiplicity
multiplicity = self.multiplicity if not inference else 1

esm_s_, _ = compute_language_model_representations(
esmaa, esm_model, esm_dict, backend=self.backend
Expand Down Expand Up @@ -167,7 +165,7 @@ def preprocess_inference(self, batch, esm_model=None, esm_dict=None, af2_to_esm=
batch_size, -1)
batch['mol_index'] = mol_index

batch = self.batch_to_device(batch, multiplicity=self.inference_multiplicity)
batch = self.batch_to_device(batch)

if esm_model is not None and batch.get('esm_s', None) is None:
print("Processing ESM features for inference...")
Expand Down
85 changes: 52 additions & 33 deletions src/simplefold/wrapper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -270,7 +270,6 @@ def initialize_others(self):
scale=16.0,
ref_scale=5.0,
multiplicity=1,
inference_multiplicity=self.nsample_per_protein,
backend=self.backend,
)

Expand Down Expand Up @@ -320,49 +319,69 @@ def process_input(self, aa_seq):

def run_inference(self, batch, model, plddt_model, device):
# run inference for target protein
if self.backend == "torch":
noise = torch.randn_like(batch["coords"]).to(device)
elif self.backend == "mlx":
noise = mx.random.normal(batch["coords"].shape)
out_dict = self.sampler.sample(model, self.flow, noise, batch)

plddt_out_module = plddt_model["plddt_out_module"]
plddt_latent_module = plddt_model["plddt_latent_module"]

coord_samples = []
if plddt_latent_module is None or plddt_out_module is None:
plddts = None
compute_plddt = False
plddt_samples = None
else:
compute_plddt = True
plddt_samples = []
pad_mask = batch["atom_pad_mask"]
plddts = None

for _ in range(self.nsample_per_protein):
if self.backend == "torch":
t = torch.ones(batch["coords"].shape[0], device=device)
# use unscaled coords to extract latent for pLDDT prediction
out_feat = plddt_latent_module(
out_dict["denoised_coords"].detach(), t, batch
)
plddt_out_dict = plddt_out_module(
out_feat["latent"].detach(),
batch,
)
noise = torch.randn_like(batch["coords"]).to(device)
elif self.backend == "mlx":
t = mx.ones(batch["coords"].shape[0])
# use unscaled coords to extract latent for pLDDT prediction
out_feat = plddt_latent_module(out_dict["denoised_coords"], t, batch)
plddt_out_dict = plddt_out_module(
out_feat["latent"],
batch,
)
# scale pLDDT to [0, 100]
plddts = plddt_out_dict["plddt"] * 100.0

out_dict = self.processor.postprocess(out_dict, batch)
# sampled_coord = out_dict['denoised_coords'].detach()
noise = mx.random.normal(batch["coords"].shape)
out_dict = self.sampler.sample(model, self.flow, noise, batch)

if compute_plddt:
if self.backend == "torch":
t = torch.ones(batch["coords"].shape[0], device=device)
out_feat = plddt_latent_module(
out_dict["denoised_coords"].detach(), t, batch
)
plddt_out_dict = plddt_out_module(
out_feat["latent"].detach(),
batch,
)
elif self.backend == "mlx":
t = mx.ones(batch["coords"].shape[0])
out_feat = plddt_latent_module(out_dict["denoised_coords"], t, batch)
plddt_out_dict = plddt_out_module(
out_feat["latent"],
batch,
)
# scale pLDDT to [0, 100]
plddt_samples.append(plddt_out_dict["plddt"] * 100.0)

out_dict = self.processor.postprocess(out_dict, batch)
if self.backend == "torch":
coord_samples.append(out_dict["denoised_coords"].detach())
else:
coord_samples.append(out_dict["denoised_coords"])

if self.backend == "torch":
sampled_coord = out_dict["denoised_coords"].detach()
sampled_coord = torch.cat(coord_samples, dim=0)
pad_mask = pad_mask.detach().repeat_interleave(
self.nsample_per_protein, dim=0
)
if compute_plddt:
plddts = torch.cat(plddt_samples, dim=0).detach()
else:
sampled_coord = out_dict["denoised_coords"]
sampled_coord = mx.concatenate(coord_samples, axis=0)
pad_mask = mx.concatenate(
[pad_mask] * self.nsample_per_protein, axis=0
)
if compute_plddt:
plddts = mx.concatenate(plddt_samples, axis=0)

return {
"sampled_coord": sampled_coord,
"pad_mask": batch["atom_pad_mask"],
"pad_mask": pad_mask,
"plddts": plddts,
}

Expand Down