Spaces:
Sleeping
Sleeping
Commit
·
3821e91
1
Parent(s):
83a075b
First commit
Browse files- app.py +58 -0
- input.txt +0 -0
- model.py +149 -0
- nanogpt.pth +3 -0
- requirements.txt +81 -0
app.py
ADDED
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import gradio as gr
|
3 |
+
from model import BigramLanguageModel
|
4 |
+
|
5 |
+
cuda = torch.cuda.is_available()
|
6 |
+
device = 'cuda' if cuda else 'cpu'
|
7 |
+
|
8 |
+
model = BigramLanguageModel()
|
9 |
+
model.load_state_dict(torch.load("model.pth", map_location=torch.device(device)), strict=False)
|
10 |
+
|
11 |
+
# read text file
|
12 |
+
with open('input.txt', 'r', encoding='utf-8') as f:
|
13 |
+
text = f.read()
|
14 |
+
|
15 |
+
# collect all the unique characters that occur in this text
|
16 |
+
chars = sorted(list(set(text)))
|
17 |
+
vocab_size = len(chars)
|
18 |
+
|
19 |
+
# create a maaping from charaters that occur in this text
|
20 |
+
stoi = { ch:i for i,ch in enumerate(chars) }
|
21 |
+
itos = { i:ch for i,ch in enumerate(chars) }
|
22 |
+
encode = lambda s: [stoi[c] for c in s] # encoder: take a string, output a list of integers
|
23 |
+
decode = lambda l: ''.join([itos[i] for i in l]) # decoder: take a list of integers, output a
|
24 |
+
|
25 |
+
def inference(input_text, max_new_tokens=500):
|
26 |
+
context = torch.tensor(encode(input_text), dtype=torch.long, device=device)
|
27 |
+
|
28 |
+
output_text = decode(model.generate(context, max_new_tokens=max_new_tokens)[0].tolist())
|
29 |
+
|
30 |
+
return output_text
|
31 |
+
|
32 |
+
title = "NanoGPT trained on Shakespeare Plays dataset"
|
33 |
+
description = "A simple Gradio interface to generate text from gpt model trained on Shakespeare Plays"
|
34 |
+
examples = [["Shape", 500],
|
35 |
+
["Answer", 500],
|
36 |
+
["Ideology", 500],
|
37 |
+
["Absorb", 500],
|
38 |
+
["Triangle", 500],
|
39 |
+
["Listen", 500],
|
40 |
+
["Census", 500],
|
41 |
+
["Balance", 500],
|
42 |
+
["Representative", 500],
|
43 |
+
["Cinema", 500],
|
44 |
+
]
|
45 |
+
demo = gr.Interface(
|
46 |
+
inference,
|
47 |
+
inputs = [
|
48 |
+
gr.Textbox(label="Enter any word", type="text"),
|
49 |
+
gr.Slider(minimum=100, maximum=10000, step=100, value=500, label="Max character to generate")
|
50 |
+
],
|
51 |
+
outputs = [
|
52 |
+
gr.Textbox(label="Output", type="text")
|
53 |
+
],
|
54 |
+
title = title,
|
55 |
+
description = description,
|
56 |
+
examples = examples,
|
57 |
+
)
|
58 |
+
demo.launch()
|
input.txt
ADDED
The diff for this file is too large to render.
See raw diff
|
|
model.py
ADDED
@@ -0,0 +1,149 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
|
5 |
+
n_embd = 384
|
6 |
+
block_size = 256
|
7 |
+
dropout = 0.1
|
8 |
+
n_head = 8
|
9 |
+
n_layer = 6
|
10 |
+
|
11 |
+
# read text file
|
12 |
+
with open('input.txt', 'r', encoding='utf-8') as f:
|
13 |
+
text = f.read()
|
14 |
+
|
15 |
+
# collect all the unique characters that occur in this text
|
16 |
+
chars = sorted(list(set(text)))
|
17 |
+
vocab_size = len(chars)
|
18 |
+
|
19 |
+
cuda = torch.cuda.is_available()
|
20 |
+
device = 'cuda' if cuda else 'cpu'
|
21 |
+
|
22 |
+
class Head(nn.Module):
|
23 |
+
""" one head of self-attention"""
|
24 |
+
|
25 |
+
def __init__(self, head_size):
|
26 |
+
super().__init__()
|
27 |
+
self.key = nn.Linear(n_embd, head_size, bias=False)
|
28 |
+
self.query = nn.Linear(n_embd, head_size, bias=False)
|
29 |
+
self.value = nn.Linear(n_embd, head_size, bias=False)
|
30 |
+
self.register_buffer('trill', torch.tril(torch.ones(block_size, block_size)))
|
31 |
+
self.dropout = nn.Dropout(dropout)
|
32 |
+
|
33 |
+
def forward(self, x):
|
34 |
+
B,T,C = x.shape
|
35 |
+
k = self.key(x)
|
36 |
+
q = self.query(x)
|
37 |
+
# compute attention scores ("affinities")
|
38 |
+
wei = q @ k.transpose(-2, -1) * C**-0.5
|
39 |
+
wei = wei.masked_fill(self.trill[:T, :T] == 0, float('-inf'))
|
40 |
+
wei = F.softmax(wei, dim=-1)
|
41 |
+
wei = self.dropout(wei)
|
42 |
+
# perform the weighted aggregation of the values
|
43 |
+
v = self.value(x)
|
44 |
+
out = wei @ v
|
45 |
+
return out
|
46 |
+
|
47 |
+
class MultiHeadAttention(nn.Module):
|
48 |
+
""" multiple heads of self-attention in parallel """
|
49 |
+
|
50 |
+
def __init__(self, num_heads, head_size):
|
51 |
+
super().__init__()
|
52 |
+
self.heads = nn.ModuleList([Head(head_size) for _ in range(num_heads)])
|
53 |
+
self.proj = nn.Linear(n_embd, n_embd)
|
54 |
+
self.dropout = nn.Dropout(dropout)
|
55 |
+
|
56 |
+
def forward(self, x):
|
57 |
+
out = torch.cat([h(x) for h in self.heads], dim=-1)
|
58 |
+
out = self.dropout(self.proj(out))
|
59 |
+
return out
|
60 |
+
|
61 |
+
class FeedForward(nn.Module):
|
62 |
+
""" a simple linear layer followed by a non-linearity """
|
63 |
+
|
64 |
+
def __init__(self, n_embd):
|
65 |
+
super().__init__()
|
66 |
+
self.net = nn.Sequential(
|
67 |
+
nn.Linear(n_embd, 4 * n_embd),
|
68 |
+
nn.ReLU(),
|
69 |
+
nn.Linear(4 * n_embd, n_embd),
|
70 |
+
nn.Dropout(dropout)
|
71 |
+
)
|
72 |
+
|
73 |
+
def forward(self, x):
|
74 |
+
return self.net(x)
|
75 |
+
|
76 |
+
class Block(nn.Module):
|
77 |
+
""" Transformer block: communication followed by computation"""
|
78 |
+
|
79 |
+
def __init__(self, n_embd, n_head) -> None:
|
80 |
+
|
81 |
+
super().__init__()
|
82 |
+
head_size = n_embd // n_head
|
83 |
+
self.sa = MultiHeadAttention(n_head, head_size)
|
84 |
+
self.ffwd = FeedForward(n_embd)
|
85 |
+
self.ln1 = nn.LayerNorm(n_embd)
|
86 |
+
self.ln2 = nn.LayerNorm(n_embd)
|
87 |
+
|
88 |
+
def forward(self, x):
|
89 |
+
x = x + self.sa(self.ln1(x))
|
90 |
+
x = x + self.ffwd(self.ln2(x))
|
91 |
+
return (x)
|
92 |
+
|
93 |
+
|
94 |
+
class BigramLanguageModel(nn.Module):
|
95 |
+
|
96 |
+
def __init__(self):
|
97 |
+
super().__init__()
|
98 |
+
# each token directly reads off the logits for the next token from a lookup table
|
99 |
+
self.token_embedding_table = nn.Embedding(vocab_size, n_embd)
|
100 |
+
self.position_embedding_table = nn.Embedding(block_size, n_embd)
|
101 |
+
# self.sa_head = MultiHeadAttention(4, n_embd//4)
|
102 |
+
# self.ffwd = FeedForward(n_embd)
|
103 |
+
# self.blocks = nn.Sequential(
|
104 |
+
# Block(n_embd, n_head=4),
|
105 |
+
# Block(n_embd, n_head=4),
|
106 |
+
# Block(n_embd, n_head=4),
|
107 |
+
# nn.LayerNorm(n_embd)
|
108 |
+
# )
|
109 |
+
self.blocks = nn.Sequential(*[Block(n_embd, n_head) for _ in range (n_layer)])
|
110 |
+
self.ln_f = nn.LayerNorm(n_embd)
|
111 |
+
self.lm_head = nn.Linear(n_embd, vocab_size)
|
112 |
+
|
113 |
+
def forward(self, idx, targets=None):
|
114 |
+
B, T = idx.shape
|
115 |
+
# idx and targets are both (B,T) tensor of integers
|
116 |
+
tok_emb = self.token_embedding_table(idx) # (B,T,C)
|
117 |
+
pos_emb = self.position_embedding_table(torch.arange(T, device=device))
|
118 |
+
x = tok_emb + pos_emb
|
119 |
+
# x = self.sa_head(x)
|
120 |
+
# x = self.ffwd(x)
|
121 |
+
x = self.blocks(x)
|
122 |
+
x = self.ln_f(x)
|
123 |
+
logits = self.lm_head(x)
|
124 |
+
|
125 |
+
if targets is None:
|
126 |
+
loss = None
|
127 |
+
else:
|
128 |
+
B, T, C = logits.shape
|
129 |
+
logits = logits.view(B*T, C)
|
130 |
+
targets = targets.view(B*T)
|
131 |
+
loss = F.cross_entropy(logits, targets)
|
132 |
+
|
133 |
+
return logits, loss
|
134 |
+
|
135 |
+
def generate(self, idx, max_new_tokens):
|
136 |
+
# idx is (B, T) array of indices in the current context
|
137 |
+
for _ in range(max_new_tokens):
|
138 |
+
idx_cond = idx[:, -block_size:]
|
139 |
+
# get the predictions
|
140 |
+
logits, loss = self(idx_cond)
|
141 |
+
# focus only on the last time step
|
142 |
+
logits = logits[:, -1, :] # becomes (B, C)
|
143 |
+
# apply softmax to get probabilities
|
144 |
+
probs = F.softmax(logits, dim=-1) # (B, C)
|
145 |
+
# sample from the distribution
|
146 |
+
idx_next = torch.multinomial(probs, num_samples=1) # (B, 1)
|
147 |
+
# append sampled index to the running sequence
|
148 |
+
idx = torch.cat((idx, idx_next), dim=1) # (B, T+1)
|
149 |
+
return idx
|
nanogpt.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:cf87f20f3e71ff9fbd3781c4f4969cbeebe0f73ec85838b7edac1c2e7f86dd29
|
3 |
+
size 55835502
|
requirements.txt
ADDED
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
aiofiles==23.2.1
|
2 |
+
altair==5.3.0
|
3 |
+
annotated-types==0.7.0
|
4 |
+
anyio==4.4.0
|
5 |
+
attrs==23.2.0
|
6 |
+
certifi==2024.6.2
|
7 |
+
charset-normalizer==3.3.2
|
8 |
+
click==8.1.7
|
9 |
+
colorama==0.4.6
|
10 |
+
contourpy==1.2.1
|
11 |
+
cycler==0.12.1
|
12 |
+
dnspython==2.6.1
|
13 |
+
email_validator==2.1.1
|
14 |
+
fastapi==0.111.0
|
15 |
+
fastapi-cli==0.0.4
|
16 |
+
ffmpy==0.3.2
|
17 |
+
filelock==3.13.1
|
18 |
+
fonttools==4.53.0
|
19 |
+
fsspec==2024.2.0
|
20 |
+
gradio==4.36.1
|
21 |
+
gradio_client==1.0.1
|
22 |
+
h11==0.14.0
|
23 |
+
httpcore==1.0.5
|
24 |
+
httptools==0.6.1
|
25 |
+
httpx==0.27.0
|
26 |
+
huggingface-hub==0.23.4
|
27 |
+
idna==3.7
|
28 |
+
importlib_resources==6.4.0
|
29 |
+
intel-openmp==2021.4.0
|
30 |
+
Jinja2==3.1.3
|
31 |
+
jsonschema==4.22.0
|
32 |
+
jsonschema-specifications==2023.12.1
|
33 |
+
kiwisolver==1.4.5
|
34 |
+
markdown-it-py==3.0.0
|
35 |
+
MarkupSafe==2.1.5
|
36 |
+
matplotlib==3.9.0
|
37 |
+
mdurl==0.1.2
|
38 |
+
mkl==2021.4.0
|
39 |
+
mpmath==1.3.0
|
40 |
+
networkx==3.2.1
|
41 |
+
numpy==1.26.3
|
42 |
+
orjson==3.10.5
|
43 |
+
packaging==24.1
|
44 |
+
pandas==2.2.2
|
45 |
+
pillow==10.2.0
|
46 |
+
pydantic==2.7.4
|
47 |
+
pydantic_core==2.18.4
|
48 |
+
pydub==0.25.1
|
49 |
+
Pygments==2.18.0
|
50 |
+
pyparsing==3.1.2
|
51 |
+
python-dateutil==2.9.0.post0
|
52 |
+
python-dotenv==1.0.1
|
53 |
+
python-multipart==0.0.9
|
54 |
+
pytz==2024.1
|
55 |
+
PyYAML==6.0.1
|
56 |
+
referencing==0.35.1
|
57 |
+
requests==2.32.3
|
58 |
+
rich==13.7.1
|
59 |
+
rpds-py==0.18.1
|
60 |
+
ruff==0.4.9
|
61 |
+
semantic-version==2.10.0
|
62 |
+
shellingham==1.5.4
|
63 |
+
six==1.16.0
|
64 |
+
sniffio==1.3.1
|
65 |
+
starlette==0.37.2
|
66 |
+
sympy==1.12
|
67 |
+
tbb==2021.11.0
|
68 |
+
tomlkit==0.12.0
|
69 |
+
toolz==0.12.1
|
70 |
+
torch==2.3.1
|
71 |
+
torchaudio==2.3.1
|
72 |
+
torchvision==0.18.1
|
73 |
+
tqdm==4.66.4
|
74 |
+
typer==0.12.3
|
75 |
+
typing_extensions==4.9.0
|
76 |
+
tzdata==2024.1
|
77 |
+
ujson==5.10.0
|
78 |
+
urllib3==2.2.1
|
79 |
+
uvicorn==0.30.1
|
80 |
+
watchfiles==0.22.0
|
81 |
+
websockets==11.0.3
|