aiuNN/src/aiunn/finetune.py

201 lines
7.4 KiB
Python

import torch
import pandas as pd
from albumentations import (
Compose, Resize, Normalize, RandomBrightnessContrast,
HorizontalFlip, VerticalFlip, Rotate, GaussianBlur
)
from albumentations.pytorch import ToTensorV2
from PIL import Image, ImageFile
import io
import base64
import numpy as np
from torch import nn
from torch.utils.data import random_split, DataLoader
from aiia import AIIA, AIIAConfig, AIIABase, AIIABaseShared, AIIAmoe, AIIAchunked, AIIArecursive
from torch.amp import autocast, GradScaler
from tqdm import tqdm
from torch.utils.checkpoint import checkpoint
class aiuNNDataset(torch.utils.data.Dataset):
def __init__(self, parquet_path):
self.df = pd.read_parquet(parquet_path, columns=['image_512', 'image_1024']).head(10000)
self.augmentation = Compose([
RandomBrightnessContrast(p=0.5),
HorizontalFlip(p=0.5),
VerticalFlip(p=0.5),
Rotate(limit=45, p=0.5),
GaussianBlur(blur_limit=(3, 7), p=0.5),
Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
ToTensorV2()
])
def __len__(self):
return len(self.df)
def load_image(self, image_data):
try:
if isinstance(image_data, str):
image_data = base64.b64decode(image_data)
if not isinstance(image_data, bytes):
raise ValueError("Invalid image data format")
image_stream = io.BytesIO(image_data)
ImageFile.LOAD_TRUNCATED_IMAGES = True
image = Image.open(image_stream).convert('RGB')
image_array = np.array(image)
return image_array
except Exception as e:
raise RuntimeError(f"Error loading image: {str(e)}")
finally:
if 'image_stream' in locals():
image_stream.close()
def __getitem__(self, idx):
row = self.df.iloc[idx]
low_res_image = self.load_image(row['image_512'])
high_res_image = self.load_image(row['image_1024'])
augmented_low = self.augmentation(image=low_res_image)
augmented_high = self.augmentation(image=high_res_image)
return {
'low_res': augmented_low['image'],
'high_res': augmented_high['image']
}
class Upscaler(nn.Module):
"""
Transforms the base model's final feature map using a transposed convolution.
The base model produces a feature map of size 512x512.
This layer upsamples by a factor of 2 (yielding 1024x1024) and maps the hidden features
to the output channels using a single ConvTranspose2d layer.
"""
def __init__(self, base_model: AIIABase):
super(Upscaler, self).__init__()
self.base_model = base_model
# Instead of adding separate upsampling and convolutional layers, we use a ConvTranspose2d layer.
self.last_transform = nn.ConvTranspose2d(
in_channels=base_model.config.hidden_size,
out_channels=base_model.config.num_channels,
kernel_size=base_model.config.kernel_size,
stride=2,
padding=1,
output_padding=1
)
def forward(self, x):
features = self.base_model(x)
return self.last_transform(features)
def finetune_model(model: nn.Module, datasets: list[str], batch_size=1, epochs=10, accumulation_steps=8, use_checkpoint=False):
# Load and concatenate datasets.
loaded_datasets = [aiuNNDataset(d) for d in datasets]
combined_dataset = torch.utils.data.ConcatDataset(loaded_datasets)
train_size = int(0.8 * len(combined_dataset))
val_size = len(combined_dataset) - train_size
train_dataset, val_dataset = random_split(combined_dataset, [train_size, val_size])
train_loader = DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4,
pin_memory=True,
persistent_workers=True
)
val_loader = DataLoader(
val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=4,
pin_memory=True,
persistent_workers=True
)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
if device.type == 'cuda':
current_device = torch.cuda.current_device()
torch.cuda.set_per_process_memory_fraction(0.95, device=current_device)
model = model.to(device)
criterion = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=model.base_model.config.learning_rate)
scaler = GradScaler()
best_val_loss = float('inf')
for epoch in range(epochs):
model.train()
train_loss = 0.0
optimizer.zero_grad()
for i, batch in enumerate(tqdm(train_loader, desc=f"Epoch {epoch+1}/Training"), start=1):
if torch.cuda.is_available():
torch.cuda.empty_cache()
low_res = batch['low_res'].to(device)
high_res = batch['high_res'].to(device)
with autocast(device_type="cuda"):
if use_checkpoint:
low_res = batch['low_res'].to(device).requires_grad_()
features = checkpoint(lambda x: model(x), low_res)
else:
features = model(low_res)
loss = criterion(features, high_res) / accumulation_steps
scaler.scale(loss).backward()
train_loss += loss.item() * accumulation_steps
if i % accumulation_steps == 0:
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
if (i % accumulation_steps) != 0:
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
avg_train_loss = train_loss / len(train_loader)
print(f"Epoch {epoch+1}, Training Loss: {avg_train_loss:.4f}")
model.eval()
val_loss = 0.0
with torch.no_grad():
for batch in tqdm(val_loader, desc="Validation"):
if torch.cuda.is_available():
torch.cuda.empty_cache()
low_res = batch['low_res'].to(device)
high_res = batch['high_res'].to(device)
with autocast(device_type="cuda"):
outputs = model(low_res)
loss = criterion(outputs, high_res)
val_loss += loss.item()
avg_val_loss = val_loss / len(val_loader)
print(f"Epoch {epoch+1}, Validation Loss: {avg_val_loss:.4f}")
if avg_val_loss < best_val_loss:
best_val_loss = avg_val_loss
model.base_model.save("best_model")
return model
def main():
BATCH_SIZE = 1
ACCUMULATION_STEPS = 8
USE_CHECKPOINT = False
# Load the base model using the provided configuration (e.g., hidden_size=512, num_channels=3, etc.)
base_model = AIIABase.load("/root/vision/AIIA/AIIA-base-512")
# Wrap the base model with our modified Upscaler that transforms its last layer.
model = Upscaler(base_model)
print("Modified model architecture with transformed final layer:")
print(base_model.config)
finetune_model(
model=model,
datasets=[
"/root/training_data/vision-dataset/image_upscaler.parquet",
"/root/training_data/vision-dataset/image_vec_upscaler.parquet"
],
batch_size=BATCH_SIZE,
epochs=10,
accumulation_steps=ACCUMULATION_STEPS,
use_checkpoint=USE_CHECKPOINT
)
if __name__ == '__main__':
main()