Delete ml_tinygpt.py
Browse files- ml_tinygpt.py +0 -66
ml_tinygpt.py
DELETED
|
@@ -1,66 +0,0 @@
|
|
| 1 |
-
import torch
|
| 2 |
-
import torch.nn.functional as F
|
| 3 |
-
from model import MiniGPT
|
| 4 |
-
from dataset import MiniBPETokenizr,SimpleTokenizr
|
| 5 |
-
import json
|
| 6 |
-
import os
|
| 7 |
-
|
| 8 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 9 |
-
|
| 10 |
-
# Load tokenizer
|
| 11 |
-
tokenizer = SimpleTokenizr()
|
| 12 |
-
tokenizer.load("./customchatbot-v1/trained-mini-gpt/tokenizer.json")
|
| 13 |
-
|
| 14 |
-
# Load model
|
| 15 |
-
model = MiniGPT(vocab_size=len(tokenizer))
|
| 16 |
-
model.load_state_dict(torch.load("./customchatbot-v1/trained-mini-gpt/mini-gpt.pth", map_location=device) if os.path.exists("./customchatbot-v1/trained-mini-gpt/mini-gpt.pth") else torch.load("./customchatbot-v1/trained-mini-gpt/checkpoint-mini-gpt.pth", map_location=device)["model_state_dict"] )
|
| 17 |
-
model.eval().to(device)
|
| 18 |
-
totalparams = sum(p.numel() for p in model.parameters())
|
| 19 |
-
print(f"Model total params: {totalparams:,}")
|
| 20 |
-
|
| 21 |
-
def sample_token(logits, temperature=1.0):
|
| 22 |
-
logits = logits / temperature
|
| 23 |
-
logits = torch.nan_to_num(logits, nan=-1e9)
|
| 24 |
-
probs = F.softmax(logits, dim=-1)
|
| 25 |
-
|
| 26 |
-
if torch.any(torch.isnan(probs)) or torch.any(probs < 0):
|
| 27 |
-
print("⚠️ Invalid probs detected. Using uniform fallback.")
|
| 28 |
-
probs = torch.ones_like(probs) / probs.size(-1)
|
| 29 |
-
|
| 30 |
-
return torch.multinomial(probs, num_samples=1).item()
|
| 31 |
-
|
| 32 |
-
def generate_reply(prompt, max_tokens=100):
|
| 33 |
-
tokens = tokenizer.encode(prompt)
|
| 34 |
-
if not tokens:
|
| 35 |
-
print("⚠️ Empty prompt after encoding.")
|
| 36 |
-
return
|
| 37 |
-
input_ids = torch.tensor(tokens, dtype=torch.long).unsqueeze(0).to(device)
|
| 38 |
-
generated = []
|
| 39 |
-
|
| 40 |
-
with torch.no_grad():
|
| 41 |
-
for _ in range(max_tokens):
|
| 42 |
-
logits = model(input_ids)
|
| 43 |
-
logits = logits[:, -1, :]
|
| 44 |
-
next_token = sample_token(logits)
|
| 45 |
-
generated.append(next_token)
|
| 46 |
-
|
| 47 |
-
next_str = tokenizer.itos.get(next_token, "")
|
| 48 |
-
encoded_text = tokenizer.encode(next_str)
|
| 49 |
-
decoded_text = tokenizer.decode(encoded_text)
|
| 50 |
-
print(decoded_text, end=" ", flush=True)
|
| 51 |
-
|
| 52 |
-
if next_str == "<END>":
|
| 53 |
-
break
|
| 54 |
-
|
| 55 |
-
input_ids = torch.cat([input_ids, torch.tensor([[next_token]]).to(device)], dim=1)
|
| 56 |
-
print()
|
| 57 |
-
|
| 58 |
-
# Chat loop
|
| 59 |
-
print("🧠 MiniGPT Chat (type 'exit' to quit')")
|
| 60 |
-
while True:
|
| 61 |
-
user_input = input("User: ")
|
| 62 |
-
if user_input.lower() == "exit":
|
| 63 |
-
break
|
| 64 |
-
prompt = f"^User: {user_input}\nMiniGPT:"
|
| 65 |
-
print("MiniGPT: ", end="", flush=True)
|
| 66 |
-
generate_reply(prompt)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|