fix: use proper initilization for embedding layer
Browse files- modeling_lora.py +28 -11
modeling_lora.py
CHANGED
@@ -11,20 +11,37 @@ from torch.nn import Parameter
|
|
11 |
from .modeling_bert import BertModel, BertPreTrainedModel, JinaBertConfig
|
12 |
|
13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
class LoRAParametrization(nn.Module):
|
15 |
-
def __init__(self, fan_in, fan_out,
|
16 |
super().__init__()
|
17 |
# if weight is stored as (fan_out, fan_in), the memory layout of A & B follows (W + BA)x
|
18 |
# otherwise, it's x(W + AB). This allows us to tie the weights between linear layers and embeddings
|
|
|
19 |
self.swap = (lambda x: (x[1], x[0])) if fan_in_fan_out else (lambda x: x)
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
nn.
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
|
|
|
|
28 |
self.lora_alpha, self.rank = lora_alpha, rank
|
29 |
self.scaling = lora_alpha / rank
|
30 |
self.lora_dropout = nn.Dropout(p=lora_dropout_p) if lora_dropout_p > 0 else lambda x: x
|
@@ -55,14 +72,14 @@ class LoRAParametrization(nn.Module):
|
|
55 |
def from_linear(cls, layer, num_adaptions=1, rank=4, lora_dropout_p=0.0, lora_alpha=1):
|
56 |
fan_out, fan_in = layer.weight.shape
|
57 |
return cls(
|
58 |
-
fan_in, fan_out, num_adaptions=num_adaptions,
|
59 |
)
|
60 |
|
61 |
@classmethod
|
62 |
def from_embedding(cls, layer, num_adaptions=1, rank=4, lora_dropout_p=0.0, lora_alpha=1):
|
63 |
fan_in, fan_out = layer.weight.shape
|
64 |
return cls(
|
65 |
-
fan_in, fan_out, num_adaptions=num_adaptions,
|
66 |
)
|
67 |
|
68 |
@classmethod
|
|
|
11 |
from .modeling_bert import BertModel, BertPreTrainedModel, JinaBertConfig
|
12 |
|
13 |
|
14 |
+
def initialized_weights(shape, num_adaptions, init='kaiming'):
|
15 |
+
weight_data = []
|
16 |
+
for _ in range(num_adaptions):
|
17 |
+
new_adaption = torch.zeros(shape)
|
18 |
+
if init == 'kaiming':
|
19 |
+
nn.init.kaiming_uniform_(new_adaption, a=math.sqrt(5))
|
20 |
+
elif init == 'normal':
|
21 |
+
nn.init.normal_(new_adaption)
|
22 |
+
else:
|
23 |
+
raise NotImplementedError
|
24 |
+
weight_data.append(new_adaption)
|
25 |
+
return torch.stack(weight_data, dim=0)
|
26 |
+
|
27 |
+
|
28 |
class LoRAParametrization(nn.Module):
|
29 |
+
def __init__(self, fan_in, fan_out, layer_type='linear', num_adaptions=1, rank=4, lora_dropout_p=0.0, lora_alpha=1):
|
30 |
super().__init__()
|
31 |
# if weight is stored as (fan_out, fan_in), the memory layout of A & B follows (W + BA)x
|
32 |
# otherwise, it's x(W + AB). This allows us to tie the weights between linear layers and embeddings
|
33 |
+
fan_in_fan_out = (layer_type == 'embedding')
|
34 |
self.swap = (lambda x: (x[1], x[0])) if fan_in_fan_out else (lambda x: x)
|
35 |
+
|
36 |
+
if layer_type == 'linear':
|
37 |
+
self.lora_A = nn.Parameter(initialized_weights((rank, fan_in), num_adaptions, init='kaiming'))
|
38 |
+
self.lora_B = nn.Parameter(torch.zeros((num_adaptions, fan_out, rank)))
|
39 |
+
elif layer_type == 'embedding':
|
40 |
+
self.lora_A = nn.Parameter(torch.zeros((num_adaptions, fan_in, rank)))
|
41 |
+
self.lora_B = nn.Parameter(initialized_weights((rank, fan_out), num_adaptions=num_adaptions, init='normal'))
|
42 |
+
else:
|
43 |
+
raise NotImplementedError
|
44 |
+
|
45 |
self.lora_alpha, self.rank = lora_alpha, rank
|
46 |
self.scaling = lora_alpha / rank
|
47 |
self.lora_dropout = nn.Dropout(p=lora_dropout_p) if lora_dropout_p > 0 else lambda x: x
|
|
|
72 |
def from_linear(cls, layer, num_adaptions=1, rank=4, lora_dropout_p=0.0, lora_alpha=1):
|
73 |
fan_out, fan_in = layer.weight.shape
|
74 |
return cls(
|
75 |
+
fan_in, fan_out, num_adaptions=num_adaptions, layer_type='linear', rank=rank, lora_dropout_p=lora_dropout_p, lora_alpha=lora_alpha
|
76 |
)
|
77 |
|
78 |
@classmethod
|
79 |
def from_embedding(cls, layer, num_adaptions=1, rank=4, lora_dropout_p=0.0, lora_alpha=1):
|
80 |
fan_in, fan_out = layer.weight.shape
|
81 |
return cls(
|
82 |
+
fan_in, fan_out, num_adaptions=num_adaptions, layer_type='embedding', rank=rank, lora_dropout_p=lora_dropout_p, lora_alpha=lora_alpha
|
83 |
)
|
84 |
|
85 |
@classmethod
|