375 lines
13 KiB
Python
375 lines
13 KiB
Python
# -*- coding: utf-8 -*-
|
||
"""Week7_Exercise.ipynb
|
||
|
||
Automatically generated by Colab.
|
||
|
||
Original file is located at
|
||
https://colab.research.google.com/drive/1wxcBNWbsDDC_SwXnQZP2dmo7ddOxkJmU
|
||
"""
|
||
|
||
#my pip installtions (some of them were not used in the project but i initally planned to use them (and we're left here so that I after the project i revisit the notebook and update when i have time and more sources.))
|
||
!pip install -q --upgrade pip
|
||
!pip install -q transformers accelerate peft bitsandbytes trl sentencepiece safetensors
|
||
!pip install -q wandb
|
||
!pip install -q git+https://github.com/huggingface/peft.git@main
|
||
!pip install datasets==3.0.1
|
||
!pip install evaluate -q
|
||
!pip install --upgrade scikit-learn
|
||
|
||
#All imports
|
||
import os, random, json, re
|
||
import pandas as pd
|
||
import numpy as np
|
||
import torch
|
||
import matplotlib.pyplot as plt
|
||
from tqdm import tqdm
|
||
from sklearn.model_selection import train_test_split
|
||
from datasets import load_dataset
|
||
from IPython.display import Markdown as md
|
||
from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score
|
||
from transformers import (
|
||
AutoTokenizer,
|
||
AutoModelForCausalLM,
|
||
TrainingArguments,
|
||
Trainer,
|
||
DataCollatorForLanguageModeling
|
||
)
|
||
|
||
#I tried a lot of models and every time my colab ran out of RAM, i did a lot of tweakings including replacing models with smaller ones
|
||
#I also used try and error to come up with samples size, eval size etc that would fit in my limited T4 Ram (had started from sample size of 15k down to 200)
|
||
MODEL_NAME = "facebook/opt-125m"
|
||
SAMPLE_SIZE = 200
|
||
EVAL_SIZE = 50
|
||
MAX_LENGTH = 128
|
||
RANDOM_SEED = 42
|
||
|
||
#Seeting LoRa hyper parameters
|
||
LORA_R = 4
|
||
LORA_ALPHA = 8
|
||
LORA_DROPOUT = 0.05
|
||
#Target modules to apply LoRA. I kept these to just "q_proj" and "v_proj" to lower memory usage on a T4 GPU.
|
||
TARGET_MODULES = ["q_proj", "v_proj"]
|
||
|
||
#to make sure thes expriment is reproducible
|
||
random.seed(RANDOM_SEED)
|
||
np.random.seed(RANDOM_SEED)
|
||
torch.manual_seed(RANDOM_SEED)
|
||
|
||
#Hf data
|
||
DATASET_NAME = "McAuley-Lab/Amazon-Reviews-2023"
|
||
SUBSET = "raw_meta_Appliances"
|
||
|
||
#loading the data
|
||
dataset = load_dataset(DATASET_NAME, SUBSET, split="full")
|
||
df = dataset.to_pandas()
|
||
|
||
from datasets import Dataset, DatasetDict
|
||
|
||
# #this took forever to run making me update it
|
||
|
||
# Split into train/eval
|
||
split = dataset.train_test_split(test_size=0.2, seed=42)
|
||
train_dataset = split["train"]
|
||
eval_dataset = split["test"]
|
||
|
||
# Reduce dataset sizes for quick experimentation
|
||
MAX_TRAIN_SAMPLES = 2000 # or 2000 if you want it even faster
|
||
MAX_EVAL_SAMPLES = 500
|
||
|
||
train_dataset = train_dataset.shuffle(seed=42).select(range(min(MAX_TRAIN_SAMPLES, len(train_dataset))))
|
||
eval_dataset = eval_dataset.shuffle(seed=42).select(range(min(MAX_EVAL_SAMPLES, len(eval_dataset))))
|
||
|
||
|
||
|
||
# Wrap into a DatasetDict for Trainer compatibility
|
||
dataset = DatasetDict({"train": train_dataset, "eval": eval_dataset})
|
||
|
||
# Prepare columns for your preprocessing
|
||
# Rename relevant columns to match what preprocess_function expects
|
||
dataset = dataset.rename_columns({
|
||
"title": "input",
|
||
"price": "output"
|
||
})
|
||
|
||
# Add a fixed instruction since your dataset doesn’t have one
|
||
def add_instruction(example):
|
||
example["instruction"] = "Estimate the fair market price of this product in USD. Return only a single number."
|
||
return example
|
||
|
||
dataset = dataset.map(add_instruction)
|
||
|
||
print(dataset)
|
||
print(dataset["train"][0])
|
||
|
||
# somecleaning on prices
|
||
df["price_clean"] = pd.to_numeric(df["price"], errors="coerce")
|
||
|
||
#Print the data showing the price and price cleaned to see they are actual not all 0
|
||
print(df_clean[["title", "price", "price_clean"]].head(10))
|
||
print(f"\nNumber of valid price entries: {df_clean['price_clean'].notna().sum()}")
|
||
|
||
# Bringing the text fields togeter
|
||
def combine_text(row):
|
||
title = row["title"] or ""
|
||
features = " ".join(row["features"]) if isinstance(row["features"], list) else ""
|
||
description = " ".join(row["description"]) if isinstance(row["description"], list) else ""
|
||
return f"TITLE: {title}\n\nFEATURES: {features}\n\nDESCRIPTION: {description}"
|
||
|
||
df["text"] = df.apply(combine_text, axis=1)
|
||
|
||
df_clean = df.dropna(subset=["price_clean"]).reset_index(drop=True)
|
||
|
||
# trying to downsamble for RAM purposes-hoping the punshment to results wasn't much
|
||
if len(df_clean) > SAMPLE_SIZE + EVAL_SIZE:
|
||
df_sample = df_clean.sample(SAMPLE_SIZE + EVAL_SIZE, random_state=RANDOM_SEED).reset_index(drop=True)
|
||
else:
|
||
df_sample = df_clean.copy()
|
||
|
||
train_df, eval_df = train_test_split(df_sample, test_size=EVAL_SIZE, random_state=RANDOM_SEED)
|
||
|
||
# FHow to format the examples
|
||
def make_example(row):
|
||
instruction = "Estimate the fair market price of this product in USD. Return only a single number."
|
||
input_text = row["text"]
|
||
output = f"{float(row['price_clean']):.2f}"
|
||
return {"instruction": instruction, "input": input_text, "output": output}
|
||
|
||
train_examples = [make_example(r) for _, r in train_df.iterrows()]
|
||
eval_examples = [make_example(r) for _, r in eval_df.iterrows()]
|
||
|
||
# Saving into JSONL
|
||
with open("pricing_train.jsonl", "w") as f:
|
||
for ex in train_examples:
|
||
f.write(json.dumps(ex) + "\n")
|
||
|
||
with open("pricing_eval.jsonl", "w") as f:
|
||
for ex in eval_examples:
|
||
f.write(json.dumps(ex) + "\n")
|
||
|
||
#Check the price exists in the Saved JSON aboved
|
||
with open("pricing_train.jsonl") as f:
|
||
lines = [json.loads(line) for line in f]
|
||
|
||
print("Sample outputs from training data:")
|
||
for ex in lines[:5]:
|
||
print(ex["output"])
|
||
|
||
#A good formating for the llm
|
||
def format_for_model(ex):
|
||
return f"### Instruction:\n{ex['instruction']}\n\n### Input:\n{ex['input']}\n\n### Response:\n{ex['output']}"
|
||
|
||
#seeing the examples
|
||
print("Example formatted prompts (3):")
|
||
|
||
#iterating over the egs
|
||
for ex in train_examples[:3]:
|
||
print(format_for_model(ex))
|
||
print("-"*80)
|
||
|
||
#tokenization now
|
||
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
||
tokenizer.pad_token = tokenizer.eos_token
|
||
|
||
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, device_map="auto")
|
||
|
||
#check ifthe model successful
|
||
print(f"{MODEL_NAME} succceeded")
|
||
|
||
# Sample random evaluation
|
||
sample_eval = random.sample(eval_examples, 10)
|
||
|
||
baseline_preds, baseline_truths = [], []
|
||
|
||
#iteration over the evals
|
||
for ex in sample_eval:
|
||
prompt = f"### Instruction:\n{ex['instruction']}\n\n### Input:\n{ex['input']}\n\n### Response:"
|
||
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
||
output = model.generate(**inputs, max_new_tokens=50, temperature=0.2, do_sample=False)
|
||
text_output = tokenizer.decode(output[0], skip_special_tokens=True)
|
||
|
||
# Extract numeric prediction from model output
|
||
match = re.search(r"\$?(\d+(\.\d+)?)", text_output)
|
||
pred_price = float(match.group(1)) if match else None
|
||
true_price = float(ex["output"])
|
||
|
||
if pred_price is not None:
|
||
baseline_preds.append(pred_price)
|
||
baseline_truths.append(true_price)
|
||
|
||
print(f"Predicted: {pred_price}, True: {true_price}")
|
||
|
||
# Manual computation of metrics
|
||
if baseline_preds:
|
||
mae = mean_absolute_error(baseline_truths, baseline_preds)
|
||
mse = mean_squared_error(baseline_truths, baseline_preds)
|
||
rmse = mse ** 0.5 # take square root manually
|
||
print(f"\nBaseline MAE: ${mae:.2f}")
|
||
print(f"Baseline RMSE: ${rmse:.2f}")
|
||
|
||
#inspectthe data a little
|
||
print(dataset)
|
||
print(dataset["train"].column_names)
|
||
print(dataset["train"][0]) # show one sample
|
||
|
||
def preprocess_function(examples):
|
||
prompts = [
|
||
f"### Instruction:\n{instr}\n\n### Input:\n{inp}\n\n### Response:\n{out}"
|
||
for instr, inp, out in zip(examples["instruction"], examples["input"], examples["output"])
|
||
]
|
||
return tokenizer(prompts, truncation=True, padding="max_length", max_length=MAX_LENGTH)
|
||
|
||
tokenized_datasets = dataset.map(preprocess_function, batched=True)
|
||
|
||
data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
|
||
|
||
#updated for faster exp
|
||
training_args = TrainingArguments(
|
||
output_dir="./price-predictor-checkpoints",
|
||
num_train_epochs=1, # ⬅change from 2 to 1
|
||
per_device_train_batch_size=1,
|
||
gradient_accumulation_steps=2,
|
||
learning_rate=2e-4,
|
||
fp16=True,
|
||
save_total_limit=1,
|
||
logging_steps=10,
|
||
report_to="none",
|
||
)
|
||
|
||
#our trainer
|
||
trainer = Trainer(
|
||
model=model,
|
||
args=training_args,
|
||
train_dataset=tokenized_datasets["train"],
|
||
eval_dataset=tokenized_datasets["eval"],
|
||
tokenizer=tokenizer,
|
||
data_collator=data_collator
|
||
)
|
||
|
||
#outcomes
|
||
train_result = trainer.train()
|
||
trainer.save_model("./price-predictor-finetuned")
|
||
|
||
# Loading fine-tuned model
|
||
model = AutoModelForCausalLM.from_pretrained("./price-predictor-finetuned", device_map="auto")
|
||
tokenizer = AutoTokenizer.from_pretrained("./price-predictor-finetuned")
|
||
|
||
# Inspect one example from your fine-tuning eval dataset before using it
|
||
print("Inspecting one evaluation example (should have instruction, input, output):")
|
||
with open("pricing_eval.jsonl") as f:
|
||
sample_eval = [json.loads(line) for line in f][:3] # just a few samples
|
||
for ex in sample_eval:
|
||
print(json.dumps(ex, indent=2))
|
||
|
||
eval_dataset_small = load_dataset("json", data_files="pricing_eval.jsonl")["train"]
|
||
eval_dataset_small = eval_dataset_small.shuffle(seed=42).select(range(min(50, len(eval_dataset_small))))
|
||
|
||
for ex in eval_dataset_small.select(range(5)):
|
||
print("Output price:", ex["output"])
|
||
|
||
#iteration Over the tqdm
|
||
for ex in tqdm(eval_dataset_small, desc="Evaluating"):
|
||
prompt = f"### Instruction:\nEstimate the fair market price of this product in USD. Return only a single number.\n\n### Input:\n{ex['input']}\n\n### Response:"
|
||
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
||
with torch.no_grad():
|
||
output = model.generate(**inputs, max_new_tokens=20)
|
||
text = tokenizer.decode(output[0], skip_special_tokens=True)
|
||
|
||
# Extract numeric prediction
|
||
numbers = re.findall(r"[-+]?\d*\.\d+|\d+", text)
|
||
pred = float(numbers[-1]) if numbers else np.nan
|
||
|
||
pred_prices.append(pred)
|
||
true_prices.append(float(ex["output"]))
|
||
|
||
# --- Fix length mismatch and mask NaNs ---
|
||
import numpy as np
|
||
|
||
pred_prices = np.array(pred_prices, dtype=float)
|
||
true_prices = np.array(true_prices, dtype=float)
|
||
|
||
# Ensure both arrays are same length
|
||
min_len = min(len(pred_prices), len(true_prices))
|
||
pred_prices = pred_prices[:min_len]
|
||
true_prices = true_prices[:min_len]
|
||
|
||
# Filter out NaNs or nonsensical large predictions
|
||
mask = (~np.isnan(pred_prices)) & (pred_prices < 10000) # exclude any predictions above $10k
|
||
pred_prices = pred_prices[mask]
|
||
true_prices = true_prices[mask]
|
||
|
||
print("Arrays aligned:")
|
||
print("Preds:", len(pred_prices), "Truths:", len(true_prices))
|
||
|
||
# Filter out invalid predictions
|
||
mask = ~np.isnan(pred_prices)
|
||
|
||
pred_prices = np.array(pred_prices)[mask]
|
||
|
||
true_prices = np.array(true_prices)[mask]
|
||
|
||
# Compute metrics manually again
|
||
mae = mean_absolute_error(true_prices, pred_prices)
|
||
|
||
mse = mean_squared_error(true_prices, pred_prices)
|
||
|
||
rmse = np.sqrt(mse)
|
||
|
||
r2 = r2_score(true_prices, pred_prices)
|
||
|
||
print(f"Fine-Tuned Evaluation:\nMAE: ${mae:.2f}, RMSE: ${rmse:.2f}, R²: {r2:.4f}")
|
||
|
||
#see what was predicted
|
||
plt.figure(figsize=(6,6))
|
||
plt.scatter(true_prices, pred_prices, alpha=0.5)
|
||
plt.plot([0, max(true_prices)], [0, max(true_prices)], 'r--', label="Perfect Prediction")
|
||
plt.xlabel("Actual Price (USD)")
|
||
plt.ylabel("Predicted Price (USD)")
|
||
plt.title("Predicted vs Actual Prices")
|
||
plt.legend()
|
||
plt.grid(True)
|
||
plt.show()
|
||
|
||
#Zoom
|
||
plt.figure(figsize=(6,6))
|
||
plt.scatter(true_prices, pred_prices, alpha=0.6)
|
||
plt.plot([0, max(true_prices)], [0, max(true_prices)], 'r--', label="Perfect Prediction")
|
||
plt.xlabel("Actual Price (USD)")
|
||
plt.ylabel("Predicted Price (USD)")
|
||
plt.title("Predicted vs Actual Prices (Zoomed In)")
|
||
plt.ylim(0, 600) # Zoom y-axis
|
||
plt.legend()
|
||
plt.grid(True)
|
||
plt.show()
|
||
|
||
#check the distribution
|
||
errors = np.abs(pred_prices - true_prices)
|
||
plt.figure(figsize=(8,4))
|
||
plt.hist(errors, bins=30, edgecolor='k', alpha=0.7)
|
||
plt.title("Distribution of Absolute Errors")
|
||
plt.xlabel("Absolute Error ($)")
|
||
plt.ylabel("Frequency")
|
||
plt.show()
|
||
|
||
print(f"Average Error: ${np.mean(errors):.2f}, Median Error: ${np.median(errors):.2f}")
|
||
|
||
# Load the base model
|
||
model_name = "facebook/opt-125m"
|
||
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
|
||
|
||
# Sample 5 examples from your eval_df
|
||
examples = eval_df.sample(5, random_state=42)
|
||
|
||
for i, row in examples.iterrows():
|
||
prompt = f"### Instruction:\nEstimate the fair market price of this product in USD. Return only a single number.\n\n### Input:\n{row['text']}\n\n### Response:"
|
||
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
||
|
||
with torch.no_grad():
|
||
outputs = model.generate(**inputs, max_new_tokens=20)
|
||
|
||
prediction_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
||
|
||
print(f"\n--- Example {i} ---")
|
||
print("Prompt:\n", prompt[:200], "...")
|
||
print("Model output:\n", prediction_text)
|
||
print("Actual price:", row["price_clean"]) |