Create train_model.py
Browse files- train_model.py +47 -0
train_model.py
ADDED
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Install required packages first:
|
2 |
+
# pip install torch transformers safetensors
|
3 |
+
|
4 |
+
import torch
|
5 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
6 |
+
|
7 |
+
# -----------------------------
|
8 |
+
# 1️⃣ Load the trained model
|
9 |
+
# -----------------------------
|
10 |
+
model_path = "./mini_gpt_safetensor" # folder where model was saved
|
11 |
+
|
12 |
+
print("📥 Loading model and tokenizer...")
|
13 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
14 |
+
tokenizer.pad_token = tokenizer.eos_token # GPT models don't have pad_token
|
15 |
+
|
16 |
+
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto") # load model
|
17 |
+
|
18 |
+
# -----------------------------
|
19 |
+
# 2️⃣ Generate text
|
20 |
+
# -----------------------------
|
21 |
+
def generate_text(prompt, max_length=50):
|
22 |
+
# Tokenize prompt
|
23 |
+
input_ids = tokenizer(prompt, return_tensors="pt").input_ids
|
24 |
+
input_ids = input_ids.to(model.device)
|
25 |
+
|
26 |
+
# Generate text
|
27 |
+
output_ids = model.generate(
|
28 |
+
input_ids,
|
29 |
+
max_length=max_length,
|
30 |
+
do_sample=True, # for randomness
|
31 |
+
top_k=50, # sample from top 50 tokens
|
32 |
+
top_p=0.95, # nucleus sampling
|
33 |
+
temperature=0.7,
|
34 |
+
num_return_sequences=1
|
35 |
+
)
|
36 |
+
|
37 |
+
# Decode output
|
38 |
+
output_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
|
39 |
+
return output_text
|
40 |
+
|
41 |
+
# -----------------------------
|
42 |
+
# 3️⃣ Test generation
|
43 |
+
# -----------------------------
|
44 |
+
prompt = "Hello, I am training a mini GPT model"
|
45 |
+
generated_text = generate_text(prompt, max_length=50)
|
46 |
+
print("\n📝 Generated text:")
|
47 |
+
print(generated_text)
|