-
Notifications
You must be signed in to change notification settings - Fork 181
/
finetuning_utils.py
192 lines (181 loc) · 7.32 KB
/
finetuning_utils.py
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
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
import torch
from mace.tools.utils import AtomicNumberTable
def load_foundations_elements(
model: torch.nn.Module,
model_foundations: torch.nn.Module,
table: AtomicNumberTable,
load_readout=False,
use_shift=True,
use_scale=True,
max_L=2,
):
"""
Load the foundations of a model into a model for fine-tuning.
"""
assert model_foundations.r_max == model.r_max
z_table = AtomicNumberTable([int(z) for z in model_foundations.atomic_numbers])
model_heads = model.heads
new_z_table = table
num_species_foundations = len(z_table.zs)
num_channels_foundation = (
model_foundations.node_embedding.linear.weight.shape[0]
// num_species_foundations
)
indices_weights = [z_table.z_to_index(z) for z in new_z_table.zs]
num_radial = model.radial_embedding.out_dim
num_species = len(indices_weights)
max_ell = model.spherical_harmonics._lmax # pylint: disable=protected-access
model.node_embedding.linear.weight = torch.nn.Parameter(
model_foundations.node_embedding.linear.weight.view(
num_species_foundations, -1
)[indices_weights, :]
.flatten()
.clone()
/ (num_species_foundations / num_species) ** 0.5
)
if model.radial_embedding.bessel_fn.__class__.__name__ == "BesselBasis":
model.radial_embedding.bessel_fn.bessel_weights = torch.nn.Parameter(
model_foundations.radial_embedding.bessel_fn.bessel_weights.clone()
)
for i in range(int(model.num_interactions)):
model.interactions[i].linear_up.weight = torch.nn.Parameter(
model_foundations.interactions[i].linear_up.weight.clone()
)
model.interactions[i].avg_num_neighbors = model_foundations.interactions[
i
].avg_num_neighbors
for j in range(4): # Assuming 4 layers in conv_tp_weights,
layer_name = f"layer{j}"
if j == 0:
getattr(model.interactions[i].conv_tp_weights, layer_name).weight = (
torch.nn.Parameter(
getattr(
model_foundations.interactions[i].conv_tp_weights,
layer_name,
)
.weight[:num_radial, :]
.clone()
)
)
else:
getattr(model.interactions[i].conv_tp_weights, layer_name).weight = (
torch.nn.Parameter(
getattr(
model_foundations.interactions[i].conv_tp_weights,
layer_name,
).weight.clone()
)
)
model.interactions[i].linear.weight = torch.nn.Parameter(
model_foundations.interactions[i].linear.weight.clone()
)
if (
model.interactions[i].__class__.__name__
== "RealAgnosticResidualInteractionBlock"
):
model.interactions[i].skip_tp.weight = torch.nn.Parameter(
model_foundations.interactions[i]
.skip_tp.weight.reshape(
num_channels_foundation,
num_species_foundations,
num_channels_foundation,
)[:, indices_weights, :]
.flatten()
.clone()
/ (num_species_foundations / num_species) ** 0.5
)
else:
model.interactions[i].skip_tp.weight = torch.nn.Parameter(
model_foundations.interactions[i]
.skip_tp.weight.reshape(
num_channels_foundation,
(max_ell + 1),
num_species_foundations,
num_channels_foundation,
)[:, :, indices_weights, :]
.flatten()
.clone()
/ (num_species_foundations / num_species) ** 0.5
)
# Transferring products
for i in range(2): # Assuming 2 products modules
max_range = max_L + 1 if i == 0 else 1
for j in range(max_range): # Assuming 3 contractions in symmetric_contractions
model.products[i].symmetric_contractions.contractions[j].weights_max = (
torch.nn.Parameter(
model_foundations.products[i]
.symmetric_contractions.contractions[j]
.weights_max[indices_weights, :, :]
.clone()
)
)
for k in range(2): # Assuming 2 weights in each contraction
model.products[i].symmetric_contractions.contractions[j].weights[k] = (
torch.nn.Parameter(
model_foundations.products[i]
.symmetric_contractions.contractions[j]
.weights[k][indices_weights, :, :]
.clone()
)
)
model.products[i].linear.weight = torch.nn.Parameter(
model_foundations.products[i].linear.weight.clone()
)
if load_readout:
# Transferring readouts
model_readouts_zero_linear_weight = model.readouts[0].linear.weight.clone()
model_readouts_zero_linear_weight = (
model_foundations.readouts[0]
.linear.weight.view(num_channels_foundation, -1)
.repeat(1, len(model_heads))
.flatten()
.clone()
)
model.readouts[0].linear.weight = torch.nn.Parameter(
model_readouts_zero_linear_weight
)
shape_input_1 = (
model_foundations.readouts[1].linear_1.__dict__["irreps_out"].num_irreps
)
shape_output_1 = model.readouts[1].linear_1.__dict__["irreps_out"].num_irreps
model_readouts_one_linear_1_weight = model.readouts[1].linear_1.weight.clone()
model_readouts_one_linear_1_weight = (
model_foundations.readouts[1]
.linear_1.weight.view(num_channels_foundation, -1)
.repeat(1, len(model_heads))
.flatten()
.clone()
)
model.readouts[1].linear_1.weight = torch.nn.Parameter(
model_readouts_one_linear_1_weight
)
model_readouts_one_linear_2_weight = model.readouts[1].linear_2.weight.clone()
model_readouts_one_linear_2_weight = model_foundations.readouts[
1
].linear_2.weight.view(shape_input_1, -1).repeat(
len(model_heads), len(model_heads)
).flatten().clone() / (
((shape_input_1) / (shape_output_1)) ** 0.5
)
model.readouts[1].linear_2.weight = torch.nn.Parameter(
model_readouts_one_linear_2_weight
)
if model_foundations.scale_shift is not None:
if use_scale:
model.scale_shift.scale = model_foundations.scale_shift.scale.repeat(
len(model_heads)
).clone()
if use_shift:
model.scale_shift.shift = model_foundations.scale_shift.shift.repeat(
len(model_heads)
).clone()
return model
def load_foundations(
model,
model_foundations,
):
for name, param in model_foundations.named_parameters():
if name in model.state_dict().keys():
if "readouts" not in name:
model.state_dict()[name].copy_(param)
return model