Skip to content

fix(dpmodel/pt/pd/jax): pass trainable to layer & support JAX trainable & support TF tensor fitting trainable #4793

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

Merged
merged 10 commits into from
Jul 8, 2025
Merged

fix(dpmodel/pt/pd/jax): pass trainable to layer & support JAX trainable & support TF tensor fitting trainable #4793

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
b9111c8
fix(dpmodel/pt/pd/jax): pass trainable to layer & support JAX trainable
njzjz Jun 10, 2025
7d7e043
bump the version of Layer data
njzjz Jun 10, 2025
77cc091
fix pd trainable
njzjz Jun 10, 2025
1c429d2
TF trainable
njzjz Jun 11, 2025
2c72228
handle list of trainable
njzjz Jun 11, 2025
36a23df
refactor(make_fitting_network): update trainable parameter to accept …
njzjz Jun 11, 2025
23a6bdd
trainable is Noen
njzjz Jun 12, 2025
bd3c17f
tf: dipole & polar trainable
njzjz Jun 13, 2025
d5840e6
Update network.py
njzjz Jun 18, 2025
94ce346
Update deepmd/dpmodel/utils/network.py
njzjz Jul 8, 2025
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
13 changes: 13 additions & 0 deletions deepmd/dpmodel/descriptor/dpa1.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -319,6 +319,7 @@ def __init__(
trainable_ln=trainable_ln,
ln_eps=ln_eps,
seed=child_seed(seed, 0),
trainable=trainable,
)
self.use_econf_tebd = use_econf_tebd
self.use_tebd_bias = use_tebd_bias
Expand All @@ -333,6 +334,7 @@ def __init__(
use_tebd_bias=use_tebd_bias,
type_map=type_map,
seed=child_seed(seed, 1),
trainable=trainable,
)
self.tebd_dim = tebd_dim
self.concat_output_tebd = concat_output_tebd
Expand Down Expand Up @@ -691,6 +693,7 @@ def __init__(
ln_eps: Optional[float] = 1e-5,
smooth: bool = True,
seed: Optional[Union[int, list[int]]] = None,
trainable: bool = True,
) -> None:
self.rcut = rcut
self.rcut_smth = rcut_smth
Expand Down Expand Up @@ -741,6 +744,7 @@ def __init__(
self.resnet_dt,
self.precision,
seed=child_seed(seed, 0),
trainable=trainable,
)
self.embeddings = embeddings
if self.tebd_input_mode in ["strip"]:
Expand All @@ -756,6 +760,7 @@ def __init__(
self.resnet_dt,
self.precision,
seed=child_seed(seed, 1),
trainable=trainable,
)
self.embeddings_strip = embeddings_strip
else:
Expand All @@ -774,6 +779,7 @@ def __init__(
smooth=self.smooth,
precision=self.precision,
seed=child_seed(seed, 2),
trainable=trainable,
)

wanted_shape = (self.ntypes, self.nnei, 4)
Expand Down Expand Up @@ -1186,6 +1192,7 @@ def __init__(
smooth: bool = True,
precision: str = DEFAULT_PRECISION,
seed: Optional[Union[int, list[int]]] = None,
trainable: bool = True,
) -> None:
"""Construct a neighbor-wise attention net."""
super().__init__()
Expand Down Expand Up @@ -1219,6 +1226,7 @@ def __init__(
smooth=smooth,
precision=precision,
seed=child_seed(seed, ii),
trainable=trainable,
)
for ii in range(layer_num)
]
Expand Down Expand Up @@ -1314,6 +1322,7 @@ def __init__(
smooth: bool = True,
precision: str = DEFAULT_PRECISION,
seed: Optional[Union[int, list[int]]] = None,
trainable: bool = True,
) -> None:
"""Construct a neighbor-wise attention layer."""
super().__init__()
Expand All @@ -1340,6 +1349,7 @@ def __init__(
smooth=smooth,
precision=precision,
seed=child_seed(seed, 0),
trainable=trainable,
)
self.attn_layer_norm = LayerNorm(
self.embed_dim,
Expand Down Expand Up @@ -1420,6 +1430,7 @@ def __init__(
smooth: bool = True,
precision: str = DEFAULT_PRECISION,
seed: Optional[Union[int, list[int]]] = None,
trainable: bool = True,
) -> None:
"""Construct a multi-head neighbor-wise attention net."""
super().__init__()
Expand Down Expand Up @@ -1449,6 +1460,7 @@ def __init__(
use_timestep=False,
precision=precision,
seed=child_seed(seed, 0),
trainable=trainable,
)
self.out_proj = NativeLayer(
hidden_dim,
Expand All @@ -1457,6 +1469,7 @@ def __init__(
use_timestep=False,
precision=precision,
seed=child_seed(seed, 1),
trainable=trainable,
)

def call(self, query, nei_mask, input_r=None, sw=None, attnw_shift=20.0):
Expand Down
6 changes: 6 additions & 0 deletions deepmd/dpmodel/descriptor/dpa2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -474,6 +474,7 @@ def init_subclass_params(sub_data, sub_class):
smooth=smooth,
type_one_side=self.repinit_args.type_one_side,
seed=child_seed(seed, 0),
trainable=trainable,
)
self.use_three_body = self.repinit_args.use_three_body
if self.use_three_body:
Expand All @@ -493,6 +494,7 @@ def init_subclass_params(sub_data, sub_class):
resnet_dt=self.repinit_args.resnet_dt,
smooth=smooth,
seed=child_seed(seed, 5),
trainable=trainable,
)
else:
self.repinit_three_body = None
Expand Down Expand Up @@ -533,6 +535,7 @@ def init_subclass_params(sub_data, sub_class):
g1_out_mlp=self.repformer_args.g1_out_mlp,
ln_eps=self.repformer_args.ln_eps,
seed=child_seed(seed, 1),
trainable=trainable,
)
self.rcsl_list = [
(self.repformers.get_rcut(), self.repformers.get_nsel()),
Expand Down Expand Up @@ -562,6 +565,7 @@ def init_subclass_params(sub_data, sub_class):
use_tebd_bias=use_tebd_bias,
type_map=type_map,
seed=child_seed(seed, 2),
trainable=trainable,
)
self.concat_output_tebd = concat_output_tebd
self.precision = precision
Expand All @@ -585,6 +589,7 @@ def init_subclass_params(sub_data, sub_class):
bias=False,
precision=precision,
seed=child_seed(seed, 3),
trainable=trainable,
)
self.tebd_transform = None
if self.add_tebd_to_repinit_out:
Expand All @@ -594,6 +599,7 @@ def init_subclass_params(sub_data, sub_class):
bias=False,
precision=precision,
seed=child_seed(seed, 4),
trainable=trainable,
)
assert self.repinit.rcut > self.repformers.rcut
assert self.repinit.sel[0] > self.repformers.sel[0]
Expand Down
2 changes: 2 additions & 0 deletions deepmd/dpmodel/descriptor/dpa3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -357,6 +357,7 @@ def init_subclass_params(sub_data, sub_class):
env_protection=env_protection,
precision=precision,
seed=child_seed(seed, 1),
trainable=trainable,
)

self.use_econf_tebd = use_econf_tebd
Expand All @@ -374,6 +375,7 @@ def init_subclass_params(sub_data, sub_class):
use_tebd_bias=use_tebd_bias,
type_map=type_map,
seed=child_seed(seed, 2),
trainable=trainable,
)
self.concat_output_tebd = concat_output_tebd
self.precision = precision
Expand Down
33 changes: 31 additions & 2 deletions deepmd/dpmodel/descriptor/repflows.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -167,6 +167,8 @@ class DescrptBlockRepflows(NativeOP, DescriptorBlock):
For example, when using paddings, there may be zero distances of neighbors, which may make division by zero error during environment matrix calculations without protection.
seed : int, optional
Random seed for parameter initialization.
trainable : bool, default: True
Whether the block is trainable
"""

def __init__(
Expand Down Expand Up @@ -205,6 +207,7 @@ def __init__(
sel_reduce_factor: float = 10.0,
use_loc_mapping: bool = True,
seed: Optional[Union[int, list[int]]] = None,
trainable: bool = True,
) -> None:
super().__init__()
self.e_rcut = float(e_rcut)
Expand Down Expand Up @@ -269,10 +272,19 @@ def __init__(
self.seed = seed

self.edge_embd = NativeLayer(
1, self.e_dim, precision=precision, seed=child_seed(seed, 0)
1,
self.e_dim,
precision=precision,
seed=child_seed(seed, 0),
trainable=trainable,
)
self.angle_embd = NativeLayer(
1, self.a_dim, precision=precision, bias=False, seed=child_seed(seed, 1)
1,
self.a_dim,
precision=precision,
bias=False,
seed=child_seed(seed, 1),
trainable=trainable,
)
layers = []
for ii in range(nlayers):
Expand Down Expand Up @@ -304,6 +316,7 @@ def __init__(
sel_reduce_factor=self.sel_reduce_factor,
smooth_edge_update=self.smooth_edge_update,
seed=child_seed(child_seed(seed, 1), ii),
trainable=trainable,
)
)
self.layers = layers
Expand Down Expand Up @@ -860,6 +873,7 @@ def __init__(
update_residual_init: str = "const",
precision: str = "float64",
seed: Optional[Union[int, list[int]]] = None,
trainable: bool = True,
) -> None:
super().__init__()
self.epsilon = 1e-4 # protection of 1./nnei
Expand Down Expand Up @@ -922,6 +936,7 @@ def __init__(
n_dim,
precision=precision,
seed=child_seed(seed, 0),
trainable=trainable,
)
if self.update_style == "res_residual":
self.n_residual.append(
Expand All @@ -931,6 +946,7 @@ def __init__(
self.update_residual_init,
precision=precision,
seed=child_seed(seed, 1),
trainable=trainable,
)
)

Expand All @@ -941,6 +957,7 @@ def __init__(
n_dim,
precision=precision,
seed=child_seed(seed, 2),
trainable=trainable,
)
if self.update_style == "res_residual":
self.n_residual.append(
Expand All @@ -950,6 +967,7 @@ def __init__(
self.update_residual_init,
precision=precision,
seed=child_seed(seed, 3),
trainable=trainable,
)
)

Expand All @@ -959,6 +977,7 @@ def __init__(
self.n_multi_edge_message * n_dim,
precision=precision,
seed=child_seed(seed, 4),
trainable=trainable,
)
if self.update_style == "res_residual":
for head_index in range(self.n_multi_edge_message):
Expand All @@ -969,6 +988,7 @@ def __init__(
self.update_residual_init,
precision=precision,
seed=child_seed(child_seed(seed, 5), head_index),
trainable=trainable,
)
)

Expand All @@ -978,6 +998,7 @@ def __init__(
e_dim,
precision=precision,
seed=child_seed(seed, 6),
trainable=trainable,
)
if self.update_style == "res_residual":
self.e_residual.append(
Expand All @@ -987,6 +1008,7 @@ def __init__(
self.update_residual_init,
precision=precision,
seed=child_seed(seed, 7),
trainable=trainable,
)
)

Expand Down Expand Up @@ -1015,13 +1037,15 @@ def __init__(
precision=precision,
bias=False,
seed=child_seed(seed, 8),
trainable=trainable,
)
self.a_compress_e_linear = NativeLayer(
self.e_dim,
self.e_a_compress_dim,
precision=precision,
bias=False,
seed=child_seed(seed, 9),
trainable=trainable,
)
else:
self.a_compress_n_linear = None
Expand All @@ -1033,12 +1057,14 @@ def __init__(
self.e_dim,
precision=precision,
seed=child_seed(seed, 10),
trainable=trainable,
)
self.edge_angle_linear2 = NativeLayer(
self.e_dim,
self.e_dim,
precision=precision,
seed=child_seed(seed, 11),
trainable=trainable,
)
if self.update_style == "res_residual":
self.e_residual.append(
Expand All @@ -1048,6 +1074,7 @@ def __init__(
self.update_residual_init,
precision=precision,
seed=child_seed(seed, 12),
trainable=trainable,
)
)

Expand All @@ -1057,6 +1084,7 @@ def __init__(
self.a_dim,
precision=precision,
seed=child_seed(seed, 13),
trainable=trainable,
)
if self.update_style == "res_residual":
self.a_residual.append(
Expand All @@ -1066,6 +1094,7 @@ def __init__(
self.update_residual_init,
precision=precision,
seed=child_seed(seed, 14),
trainable=trainable,
)
)
else:
Expand Down
Loading
Loading