Spaces:
Sleeping
Sleeping
yoonusajwardapiit
commited on
Commit
•
49b2bf5
1
Parent(s):
2047d88
Upload app.py
Browse files
app.py
CHANGED
@@ -1,33 +1,101 @@
|
|
1 |
-
|
2 |
import gradio as gr
|
3 |
import torch
|
4 |
import torch.nn as nn
|
5 |
|
6 |
-
# Define your custom model class
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
class BigramLanguageModel(nn.Module):
|
8 |
def __init__(self):
|
9 |
super().__init__()
|
10 |
-
# Example layers (adjust as needed for your model)
|
11 |
self.token_embedding_table = nn.Embedding(61, 64)
|
12 |
self.position_embedding_table = nn.Embedding(32, 64)
|
13 |
-
self.blocks = nn.Sequential(*[
|
14 |
self.ln_f = nn.LayerNorm(64)
|
15 |
self.lm_head = nn.Linear(64, 61)
|
16 |
|
17 |
-
def forward(self, idx):
|
18 |
-
|
19 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
|
21 |
-
def generate(self, idx, max_new_tokens
|
22 |
-
|
23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
-
# Load
|
26 |
def load_model():
|
27 |
model = BigramLanguageModel()
|
28 |
model_url = "https://huggingface.co/yoonusajwardapiit/triptuner/resolve/main/pytorch_model.bin"
|
29 |
model_weights = torch.hub.load_state_dict_from_url(model_url, map_location=torch.device('cpu'), weights_only=True)
|
30 |
-
model.load_state_dict(model_weights)
|
31 |
model.eval()
|
32 |
return model
|
33 |
|
|
|
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
import torch.nn as nn
|
4 |
|
5 |
+
# Define your custom model class with detailed layer structures
|
6 |
+
class Head(nn.Module):
|
7 |
+
def __init__(self, head_size):
|
8 |
+
super().__init__()
|
9 |
+
self.key = nn.Linear(64, head_size, bias=False)
|
10 |
+
self.query = nn.Linear(64, head_size, bias=False)
|
11 |
+
self.value = nn.Linear(64, head_size, bias=False)
|
12 |
+
self.register_buffer('tril', torch.tril(torch.ones(32, 32)))
|
13 |
+
self.dropout = nn.Dropout(0.1)
|
14 |
+
|
15 |
+
def forward(self, x):
|
16 |
+
B, T, C = x.shape
|
17 |
+
k = self.key(x)
|
18 |
+
q = self.query(x)
|
19 |
+
wei = q @ k.transpose(-2, -1) * C**-0.5
|
20 |
+
wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf'))
|
21 |
+
wei = nn.functional.softmax(wei, dim=-1)
|
22 |
+
wei = self.dropout(wei)
|
23 |
+
v = self.value(x)
|
24 |
+
return wei @ v
|
25 |
+
|
26 |
+
class MultiHeadAttention(nn.Module):
|
27 |
+
def __init__(self, num_heads, head_size):
|
28 |
+
super().__init__()
|
29 |
+
self.heads = nn.ModuleList([Head(head_size) for _ in range(num_heads)])
|
30 |
+
self.proj = nn.Linear(64, 64)
|
31 |
+
self.dropout = nn.Dropout(0.1)
|
32 |
+
|
33 |
+
def forward(self, x):
|
34 |
+
out = torch.cat([h(x) for h in self.heads], dim=-1)
|
35 |
+
return self.dropout(self.proj(out))
|
36 |
+
|
37 |
+
class FeedForward(nn.Module):
|
38 |
+
def __init__(self, n_embd):
|
39 |
+
super().__init__()
|
40 |
+
self.net = nn.Sequential(
|
41 |
+
nn.Linear(n_embd, 4 * n_embd),
|
42 |
+
nn.ReLU(),
|
43 |
+
nn.Linear(4 * n_embd, n_embd),
|
44 |
+
nn.Dropout(0.1),
|
45 |
+
)
|
46 |
+
|
47 |
+
def forward(self, x):
|
48 |
+
return self.net(x)
|
49 |
+
|
50 |
+
class Block(nn.Module):
|
51 |
+
def __init__(self, n_embd, n_head):
|
52 |
+
super().__init__()
|
53 |
+
head_size = n_embd // n_head
|
54 |
+
self.sa = MultiHeadAttention(n_head, head_size)
|
55 |
+
self.ffwd = FeedForward(n_embd)
|
56 |
+
self.ln1 = nn.LayerNorm(n_embd)
|
57 |
+
self.ln2 = nn.LayerNorm(n_embd)
|
58 |
+
|
59 |
+
def forward(self, x):
|
60 |
+
x = x + self.sa(self.ln1(x))
|
61 |
+
x = x + self.ffwd(self.ln2(x))
|
62 |
+
return x
|
63 |
+
|
64 |
class BigramLanguageModel(nn.Module):
|
65 |
def __init__(self):
|
66 |
super().__init__()
|
|
|
67 |
self.token_embedding_table = nn.Embedding(61, 64)
|
68 |
self.position_embedding_table = nn.Embedding(32, 64)
|
69 |
+
self.blocks = nn.Sequential(*[Block(64, n_head=4) for _ in range(4)])
|
70 |
self.ln_f = nn.LayerNorm(64)
|
71 |
self.lm_head = nn.Linear(64, 61)
|
72 |
|
73 |
+
def forward(self, idx, targets=None):
|
74 |
+
B, T = idx.shape
|
75 |
+
tok_emb = self.token_embedding_table(idx)
|
76 |
+
pos_emb = self.position_embedding_table(torch.arange(T, device=idx.device))
|
77 |
+
x = tok_emb + pos_emb
|
78 |
+
x = self.blocks(x)
|
79 |
+
x = self.ln_f(x)
|
80 |
+
logits = self.lm_head(x)
|
81 |
+
return logits, None
|
82 |
|
83 |
+
def generate(self, idx, max_new_tokens):
|
84 |
+
for _ in range(max_new_tokens):
|
85 |
+
idx_cond = idx[:, -32:]
|
86 |
+
logits, _ = self(idx_cond)
|
87 |
+
logits = logits[:, -1, :]
|
88 |
+
probs = nn.functional.softmax(logits, dim=-1)
|
89 |
+
idx_next = torch.multinomial(probs, num_samples=1)
|
90 |
+
idx = torch.cat((idx, idx_next), dim=1)
|
91 |
+
return idx
|
92 |
|
93 |
+
# Load the model with strict=False to handle missing or unexpected keys
|
94 |
def load_model():
|
95 |
model = BigramLanguageModel()
|
96 |
model_url = "https://huggingface.co/yoonusajwardapiit/triptuner/resolve/main/pytorch_model.bin"
|
97 |
model_weights = torch.hub.load_state_dict_from_url(model_url, map_location=torch.device('cpu'), weights_only=True)
|
98 |
+
model.load_state_dict(model_weights, strict=False)
|
99 |
model.eval()
|
100 |
return model
|
101 |
|