242 lines
7.2 KiB
Markdown
242 lines
7.2 KiB
Markdown
---
|
|
library_name: transformers
|
|
tags: []
|
|
---
|
|
|
|
# SUMMARY
|
|
|
|
Just a model using to learn Fine Tuning of 'gpt2-medium'
|
|
- on a self made datasets
|
|
- on a self made special tokens
|
|
- on a multiple fine tuned with ~15K dataset (in progress mode)
|
|
|
|
If interested in how I got to this point and how I created the datasets you can visit:
|
|
[Crafting GPT2 for Personalized AI-Preparing Data the Long Way](https://medium.com/@deeokay/the-soul-in-the-machine-crafting-gpt2-for-personalized-ai-9d38be3f635f)
|
|
<!-- Provide a quick summary of what the model is/does. -->
|
|
|
|
|
|
|
|
# FINE TUNED - BASE MODEL
|
|
I would consider this [GPT2-medium-custom-v1.0](https://huggingface.co/Deeokay/GPT2-medium-custom-v1.0) a the base model to start my Fine Tuning 2.0 on specific Datasets.
|
|
- Previous models of this: gpt-special-tokens-medium(1~4) are consider beta check-points to this
|
|
|
|
This model is available to test on Ollama [Deeokay/mediumgpt2](https://ollama.com/deeokay/mediumgpt2) it is not perfect and I am still working out some stuff, but I am quite proud that I was able to make it this far.
|
|
Please note, the acutal GGUF file is also included in this repository if you would like to create your own versions (templates etc.)
|
|
|
|
|
|
## DECLARING NEW SPECIAL TOKENS
|
|
|
|
```python
|
|
special_tokens_dict = {
|
|
'eos_token': '<|STOP|>',
|
|
'bos_token': '<|STOP|>',
|
|
'pad_token': '<|PAD|>',
|
|
'additional_special_tokens': ['<|BEGIN_QUERY|>', '<|BEGIN_QUERY|>',
|
|
'<|BEGIN_ANALYSIS|>', '<|END_ANALYSIS|>',
|
|
'<|BEGIN_RESPONSE|>', '<|END_RESPONSE|>',
|
|
'<|BEGIN_SENTIMENT|>', '<|END_SENTIMENT|>',
|
|
'<|BEGIN_CLASSIFICATION|>', '<|END_CLASSIFICATION|>',]
|
|
}
|
|
|
|
tokenizer.add_special_tokens(special_tokens_dict)
|
|
model.resize_token_embeddings(len(tokenizer))
|
|
|
|
tokenizer.eos_token_id = tokenizer.convert_tokens_to_ids('<|STOP|>')
|
|
tokenizer.bos_token_id = tokenizer.convert_tokens_to_ids('<|STOP|>')
|
|
tokenizer.pad_token_id = tokenizer.convert_tokens_to_ids('<|PAD|>')
|
|
```
|
|
|
|
The order of tokens is as follows:
|
|
|
|
```python
|
|
def combine_text(user_prompt, analysis, sentiment, new_response, classification):
|
|
user_q = f"<|STOP|><|BEGIN_QUERY|>{user_prompt}<|END_QUERY|>"
|
|
analysis = f"<|BEGIN_ANALYSIS|>{analysis}<|END_ANALYSIS|>"
|
|
new_response = f"<|BEGIN_RESPONSE|>{new_response}<|END_RESPONSE|>"
|
|
classification = f"<|BEGIN_CLASSIFICATION|>{classification}<|END_CLASSIFICATION|>"
|
|
sentiment = f"<|BEGIN_SENTIMENT|>Sentiment: {sentiment}<|END_SENTIMENT|><|STOP|>"
|
|
return user_q + analysis + new_response + classification + sentiment
|
|
```
|
|
|
|
## INFERANCING
|
|
|
|
I am currently testing two ways, if anyone knows a better one, please let me know!
|
|
|
|
```python
|
|
import torch
|
|
from transformers import GPT2LMHeadModel, GPT2Tokenizer
|
|
|
|
models_folder = "Deeokay/gpt2-medium-custom-v1.0"
|
|
|
|
model = GPT2LMHeadModel.from_pretrained(models_folder)
|
|
tokenizer = GPT2Tokenizer.from_pretrained(models_folder)
|
|
|
|
# Device configuration <<change as needed>>
|
|
device = torch.device("cpu")
|
|
model.to(device)
|
|
|
|
```
|
|
|
|
### OPTION 1 INFERFENCE
|
|
|
|
```python
|
|
import time
|
|
|
|
class Stopwatch:
|
|
def __init__(self):
|
|
self.start_time = None
|
|
self.end_time = None
|
|
|
|
def start(self):
|
|
self.start_time = time.time()
|
|
|
|
def stop(self):
|
|
self.end_time = time.time()
|
|
|
|
def elapsed_time(self):
|
|
if self.start_time is None:
|
|
return "Stopwatch hasn't been started"
|
|
if self.end_time is None:
|
|
return "Stopwatch hasn't been stopped"
|
|
return self.end_time - self.start_time
|
|
|
|
stopwatch1 = Stopwatch()
|
|
|
|
def generate_response(input_text, max_length=250):
|
|
|
|
stopwatch1.start()
|
|
|
|
# Prepare the input
|
|
# input_text = f"<|BEGIN_QUERY|>{input_text}<|END_QUERY|><|BEGIN_ANALYSIS|>{input_text}<|END_ANALYSIS|><|BEGIN_RESPONSE|>"
|
|
input_text = f"<|BEGIN_QUERY|>{input_text}<|END_QUERY|><|BEGIN_ANALYSIS|>"
|
|
|
|
input_ids = tokenizer.encode(input_text, return_tensors="pt").to(device)
|
|
|
|
# Create attention mask
|
|
attention_mask = torch.ones_like(input_ids).to(device)
|
|
|
|
# Generate
|
|
output = model.generate(
|
|
input_ids,
|
|
max_new_tokens=max_length,
|
|
num_return_sequences=1,
|
|
no_repeat_ngram_size=2,
|
|
attention_mask=attention_mask,
|
|
pad_token_id=tokenizer.eos_token_id,
|
|
eos_token_id=tokenizer.convert_tokens_to_ids('<|STOP|>'),
|
|
)
|
|
|
|
stopwatch1.stop()
|
|
|
|
return tokenizer.decode(output[0], skip_special_tokens=False)
|
|
```
|
|
|
|
### OPTION 2 INFERNCE
|
|
|
|
```python
|
|
import time
|
|
|
|
class Stopwatch:
|
|
def __init__(self):
|
|
self.start_time = None
|
|
self.end_time = None
|
|
|
|
def start(self):
|
|
self.start_time = time.time()
|
|
|
|
def stop(self):
|
|
self.end_time = time.time()
|
|
|
|
def elapsed_time(self):
|
|
if self.start_time is None:
|
|
return "Stopwatch hasn't been started"
|
|
if self.end_time is None:
|
|
return "Stopwatch hasn't been stopped"
|
|
return self.end_time - self.start_time
|
|
|
|
stopwatch2 = Stopwatch()
|
|
|
|
def generate_response2(input_text, max_length=250):
|
|
|
|
stopwatch2.start()
|
|
|
|
# Prepare the input
|
|
# input_text = f"<|BEGIN_QUERY|>{input_text}<|END_QUERY|><|BEGIN_ANALYSIS|>{input_text}<|END_ANALYSIS|><|BEGIN_RESPONSE|>"
|
|
input_text = f"<|BEGIN_QUERY|>{input_text}<|END_QUERY|><|BEGIN_ANALYSIS|>"
|
|
input_ids = tokenizer.encode(input_text, return_tensors="pt").to(device)
|
|
|
|
# Create attention mask
|
|
attention_mask = torch.ones_like(input_ids).to(device)
|
|
|
|
# # 2ND OPTION FOR : Generate
|
|
output = model.generate(
|
|
input_ids,
|
|
max_new_tokens=max_length,
|
|
attention_mask=attention_mask,
|
|
do_sample=True,
|
|
temperature=0.4, # this can be played around
|
|
top_k=60, # this can be played around
|
|
no_repeat_ngram_size=2,
|
|
pad_token_id=tokenizer.pad_token_id,
|
|
eos_token_id=tokenizer.eos_token_id,
|
|
)
|
|
|
|
stopwatch2.stop()
|
|
return tokenizer.decode(output[0], skip_special_tokens=False)
|
|
```
|
|
### DECODING ANSWER
|
|
|
|
When I need just the response
|
|
|
|
```python
|
|
def decode(text):
|
|
full_text = text
|
|
|
|
# Extract the response part
|
|
start_token = "<|BEGIN_RESPONSE|>"
|
|
end_token = "<|END_RESPONSE|>"
|
|
start_idx = full_text.find(start_token)
|
|
end_idx = full_text.find(end_token)
|
|
|
|
if start_idx != -1 and end_idx != -1:
|
|
response = full_text[start_idx + len(start_token):end_idx].strip()
|
|
else:
|
|
response = full_text.strip()
|
|
|
|
return response
|
|
```
|
|
|
|
### MY SETUP
|
|
|
|
I use the stopwatch to time the responses and I use both inference to see the difference
|
|
|
|
```python
|
|
input_text = "Who is Steve Jobs and what was contribution?"
|
|
response1_full = generate_response(input_text)
|
|
#response1 = decode(response1_full)
|
|
print(f"Input: {input_text}")
|
|
print("=======================================")
|
|
print(f"Response1: {response1_full}")
|
|
elapsed1 = stopwatch1.elapsed_time()
|
|
print(f"Process took {elapsed1:.4f} seconds")
|
|
print("=======================================")
|
|
response2_full = generate_response2(input_text)
|
|
#response2 = decode(response2_full)
|
|
print(f"Response2: {response2_full}")
|
|
elapsed2 = stopwatch2.elapsed_time()
|
|
print(f"Process took {elapsed2:.4f} seconds")
|
|
print("=======================================")
|
|
```
|
|
|
|
|
|
### Out-of-Scope Use
|
|
|
|
Well everything that has a factual data.. trust at your own risk!
|
|
|
|
Never tested on mathamatical knowledge.
|
|
|
|
I quite enjoy how the response feels closer to what I had in mind..
|
|
|
|
|
|
|