Upload 3 files
Browse files- architecture.py +159 -0
- requirements.txt +7 -0
- tokenizer.py +51 -0
architecture.py
ADDED
|
@@ -0,0 +1,159 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import math
|
| 2 |
+
|
| 3 |
+
import torch
|
| 4 |
+
|
| 5 |
+
|
| 6 |
+
class PositionalEncoding(torch.nn.Module):
|
| 7 |
+
"""
|
| 8 |
+
https://pytorch.org/tutorials/beginner/transformer_tutorial.html
|
| 9 |
+
"""
|
| 10 |
+
|
| 11 |
+
def __init__(self, d_model: int, max_len: int = 512):
|
| 12 |
+
super().__init__()
|
| 13 |
+
|
| 14 |
+
position = torch.arange(max_len).unsqueeze(1)
|
| 15 |
+
div_term = torch.exp(
|
| 16 |
+
torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model)
|
| 17 |
+
)
|
| 18 |
+
|
| 19 |
+
pe = torch.zeros(max_len, d_model)
|
| 20 |
+
pe[:, : d_model // 2] = torch.sin(position * div_term)
|
| 21 |
+
pe[:, d_model // 2 :] = torch.cos(position * div_term)
|
| 22 |
+
|
| 23 |
+
self.register_buffer("pe", pe)
|
| 24 |
+
|
| 25 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 26 |
+
x = x + self.pe[: x.size(0)]
|
| 27 |
+
return x
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
class MultiheadSelfAttention(torch.nn.Module):
|
| 31 |
+
def __init__(self, embed_dim: int, num_heads: int = 8):
|
| 32 |
+
super().__init__()
|
| 33 |
+
|
| 34 |
+
self.embed_dim = embed_dim
|
| 35 |
+
self.num_heads = num_heads
|
| 36 |
+
|
| 37 |
+
self.query = torch.nn.Linear(
|
| 38 |
+
in_features=embed_dim,
|
| 39 |
+
out_features=embed_dim,
|
| 40 |
+
)
|
| 41 |
+
self.key = torch.nn.Linear(
|
| 42 |
+
in_features=embed_dim,
|
| 43 |
+
out_features=embed_dim,
|
| 44 |
+
)
|
| 45 |
+
self.value = torch.nn.Linear(
|
| 46 |
+
in_features=embed_dim,
|
| 47 |
+
out_features=embed_dim,
|
| 48 |
+
)
|
| 49 |
+
|
| 50 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 51 |
+
q = self.query(x).view(x.shape[0], self.num_heads, -1).transpose(0, 1)
|
| 52 |
+
k = self.key(x).view(x.shape[0], self.num_heads, -1).permute(1, 2, 0)
|
| 53 |
+
v = self.value(x).view(x.shape[0], self.num_heads, -1).transpose(0, 1)
|
| 54 |
+
qk = torch.softmax(
|
| 55 |
+
torch.matmul(q, k) / (self.embed_dim / self.num_heads) ** 0.5,
|
| 56 |
+
dim=-1,
|
| 57 |
+
)
|
| 58 |
+
qkv = torch.matmul(qk, v).transpose(0, 1).reshape(x.shape[0], -1)
|
| 59 |
+
return qkv
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
class Block(torch.nn.Module):
|
| 63 |
+
def __init__(self, d_model: int, num_heads: int = 8, eps: float = 1e-6):
|
| 64 |
+
super().__init__()
|
| 65 |
+
|
| 66 |
+
self.ln1 = torch.nn.LayerNorm(normalized_shape=d_model, eps=eps)
|
| 67 |
+
self.attn = MultiheadSelfAttention(embed_dim=d_model, num_heads=num_heads)
|
| 68 |
+
self.ln2 = torch.nn.LayerNorm(normalized_shape=d_model, eps=eps)
|
| 69 |
+
self.linear1 = torch.nn.Linear(in_features=d_model, out_features=d_model * 4)
|
| 70 |
+
self.linear2 = torch.nn.Linear(in_features=d_model * 4, out_features=d_model)
|
| 71 |
+
|
| 72 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 73 |
+
ln1 = self.ln1(x)
|
| 74 |
+
attn = self.attn(ln1)
|
| 75 |
+
ln2 = self.ln2(x + attn)
|
| 76 |
+
mlp = self.linear2(torch.relu(self.linear1(ln2)))
|
| 77 |
+
return mlp + x + attn
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
class Head(torch.nn.Module):
|
| 81 |
+
def __init__(
|
| 82 |
+
self,
|
| 83 |
+
d_model: int,
|
| 84 |
+
eps: float = 1e-6,
|
| 85 |
+
):
|
| 86 |
+
super().__init__()
|
| 87 |
+
|
| 88 |
+
self.d_model = d_model
|
| 89 |
+
self.eps = eps
|
| 90 |
+
|
| 91 |
+
self.ln = torch.nn.LayerNorm(normalized_shape=d_model, eps=eps)
|
| 92 |
+
self.linear1 = torch.nn.Linear(in_features=d_model, out_features=d_model)
|
| 93 |
+
self.linear2 = torch.nn.Linear(in_features=d_model, out_features=d_model)
|
| 94 |
+
self.tanh_layer = torch.nn.Linear(in_features=d_model * 2, out_features=d_model)
|
| 95 |
+
|
| 96 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 97 |
+
ln = self.ln(x)
|
| 98 |
+
mlp = torch.exp(self.linear2(torch.nn.functional.elu(self.linear1(ln))))
|
| 99 |
+
res = torch.cat(
|
| 100 |
+
[
|
| 101 |
+
ln.sum(dim=0) / ln.shape[0],
|
| 102 |
+
(mlp * ln).sum(dim=0) / mlp.sum(dim=0),
|
| 103 |
+
]
|
| 104 |
+
)
|
| 105 |
+
res = torch.tanh(self.tanh_layer(res))
|
| 106 |
+
res /= (res**2).sum() ** 0.5
|
| 107 |
+
res /= (res**2).sum() ** 0.5
|
| 108 |
+
return res
|
| 109 |
+
|
| 110 |
+
|
| 111 |
+
class MUSE(torch.nn.Module):
|
| 112 |
+
def __init__(
|
| 113 |
+
self,
|
| 114 |
+
num_embeddings: int,
|
| 115 |
+
embedding_dim: int,
|
| 116 |
+
d_model: int,
|
| 117 |
+
num_heads: int,
|
| 118 |
+
eps: float = 1e-6,
|
| 119 |
+
):
|
| 120 |
+
super().__init__()
|
| 121 |
+
|
| 122 |
+
self.num_embeddings = num_embeddings
|
| 123 |
+
self.embedding_dim = embedding_dim
|
| 124 |
+
self.d_model = d_model
|
| 125 |
+
self.num_heads = num_heads
|
| 126 |
+
self.eps = eps
|
| 127 |
+
|
| 128 |
+
self.embedding = torch.nn.Embedding(
|
| 129 |
+
num_embeddings=num_embeddings,
|
| 130 |
+
embedding_dim=embedding_dim,
|
| 131 |
+
)
|
| 132 |
+
self.linear = torch.nn.Linear(
|
| 133 |
+
in_features=embedding_dim,
|
| 134 |
+
out_features=d_model,
|
| 135 |
+
)
|
| 136 |
+
self.pe = PositionalEncoding(
|
| 137 |
+
d_model=d_model,
|
| 138 |
+
max_len=512, # TODO: remove hardcode
|
| 139 |
+
)
|
| 140 |
+
self.block0 = Block(d_model=d_model)
|
| 141 |
+
self.block1 = Block(d_model=d_model)
|
| 142 |
+
self.block2 = Block(d_model=d_model)
|
| 143 |
+
self.block3 = Block(d_model=d_model)
|
| 144 |
+
self.block4 = Block(d_model=d_model)
|
| 145 |
+
self.block5 = Block(d_model=d_model)
|
| 146 |
+
self.head = Head(d_model=d_model)
|
| 147 |
+
|
| 148 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 149 |
+
x = self.embedding(x)
|
| 150 |
+
x = self.linear(x)
|
| 151 |
+
x = self.pe(x)
|
| 152 |
+
x = self.block0(x)
|
| 153 |
+
x = self.block1(x)
|
| 154 |
+
x = self.block2(x)
|
| 155 |
+
x = self.block3(x)
|
| 156 |
+
x = self.block4(x)
|
| 157 |
+
x = self.block5(x)
|
| 158 |
+
x = self.head(x)
|
| 159 |
+
return x
|
requirements.txt
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
onnx==1.16.0
|
| 2 |
+
onnxruntime==1.18.0
|
| 3 |
+
onnxruntime_extensions==0.10.1
|
| 4 |
+
tensorflow==2.16.1
|
| 5 |
+
tensorflow-hub==0.16.1
|
| 6 |
+
tensorflow-text==2.16.1
|
| 7 |
+
torch==2.3.0
|
tokenizer.py
ADDED
|
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from tensorflow.core.protobuf.saved_model_pb2 import SavedModel
|
| 3 |
+
from tensorflow.python.saved_model.loader_impl import parse_saved_model
|
| 4 |
+
from tensorflow_text.python.ops.sentencepiece_tokenizer import SentencepieceTokenizer
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
def _get_tokenizer_from_saved_model(saved_model: SavedModel) -> SentencepieceTokenizer:
|
| 8 |
+
"""
|
| 9 |
+
Get tokenizer from tf SavedModel.
|
| 10 |
+
:param SavedModel saved_model: tf SavedModel.
|
| 11 |
+
:return: tokenizer.
|
| 12 |
+
:rtype: SentencepieceTokenizer
|
| 13 |
+
"""
|
| 14 |
+
|
| 15 |
+
# extract functions that contain SentencePiece somewhere in there
|
| 16 |
+
functions_with_sp = [
|
| 17 |
+
f
|
| 18 |
+
for f in saved_model.meta_graphs[0].graph_def.library.function
|
| 19 |
+
if "tokenizer" in str(f).lower()
|
| 20 |
+
]
|
| 21 |
+
|
| 22 |
+
assert (
|
| 23 |
+
len(functions_with_sp) == 1
|
| 24 |
+
), f"len(functions_with_sp) = {len(functions_with_sp)}"
|
| 25 |
+
|
| 26 |
+
# find SentencePieceOp (contains the model) in the found function
|
| 27 |
+
nodes_with_sp = [
|
| 28 |
+
n for n in functions_with_sp[0].node_def if n.op == "SentencepieceOp"
|
| 29 |
+
]
|
| 30 |
+
|
| 31 |
+
assert len(nodes_with_sp) == 1, f"len(nodes_with_sp) = {len(nodes_with_sp)}"
|
| 32 |
+
|
| 33 |
+
# we can pretty much save the model into a file since it does not change
|
| 34 |
+
model = nodes_with_sp[0].attr["model"].s
|
| 35 |
+
|
| 36 |
+
# instantiate the model
|
| 37 |
+
tokenizer = SentencepieceTokenizer(model)
|
| 38 |
+
|
| 39 |
+
return tokenizer
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
def get_tokenizer(model_path: str) -> SentencepieceTokenizer:
|
| 43 |
+
tokenizer = _get_tokenizer_from_saved_model(parse_saved_model(model_path))
|
| 44 |
+
return tokenizer
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
def tokenize(
|
| 48 |
+
sentence: str, # TODO: add batch processing
|
| 49 |
+
tokenizer: SentencepieceTokenizer,
|
| 50 |
+
) -> torch.Tensor:
|
| 51 |
+
return torch.LongTensor([1] + tokenizer.tokenize([sentence]).to_list()[0] + [2])
|