WeCanopy / classification /classification.py
Luecke's picture
classification model
7b4e127
# -*- coding: utf-8 -*-
"""classification.ipynb
Automatically generated by Colab.
Original file is located at
https://colab.research.google.com/drive/1JuZNV3fqC5XQ0L-jhIyVRbIDPfWWGkVI
"""
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, models, transforms
from torch.utils.data import DataLoader
from torch.utils.data import DataLoader, random_split
import os
import matplotlib.pyplot as plt
import random
from PIL import Image
import numpy as np
import pandas as pd
# Define the data directories
data_dir = 'drive/MyDrive/Ai_Hackathon_2024/plant_data/data_for_training'
augmented_data_dir = 'drive/MyDrive/Ai_Hackathon_2024/plant_data/augmented_data'
# Define the desired number of images per class
N = 50
# Define the augmentation transforms
augmentation_transforms = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(30),
transforms.RandomAffine(degrees=0, translate=(0.1, 0.1)),
transforms.RandomResizedCrop(size=(224, 224), scale=(0.8, 1.0)),
transforms.Pad(padding=10, padding_mode='reflect'), # Add padding with reflection
transforms.ToTensor(),
])
# Load the dataset
print('loading dataset...')
dataset = datasets.ImageFolder(data_dir)
class_names = dataset.classes
print('loaded dataset.')
# Function to save augmented images
def save_image(img, path, idx):
img.save(os.path.join(path, f'{idx}.png'))
# Augment the dataset
if not os.path.exists(augmented_data_dir):
os.makedirs(augmented_data_dir)
print('starting augmentation process...')
for class_idx in range(len(dataset.classes)):
print(f"class_idx = {class_idx}")
class_dir = os.path.join(augmented_data_dir, dataset.classes[class_idx])
if not os.path.exists(class_dir):
os.makedirs(class_dir)
class_images = [img_path for img_path, label in dataset.samples if label == class_idx]
current_count = 0
# Save original images first
for img_path in class_images:
img = Image.open(img_path)
save_image(img, class_dir, current_count)
current_count += 1
# If there are fewer than N images, augment the dataset
while current_count < N:
img_path = random.choice(class_images)
img = Image.open(img_path)
img = augmentation_transforms(img)
img = transforms.ToPILImage()(img) # Convert back to PIL Image
save_image(img, class_dir, current_count)
current_count += 1
print('Data augmentation completed.')
# Define the data directory
data_dir = augmented_data_dir #'drive/MyDrive/Ai_Hackathon_2024/plant_data/data_for_training'
# Set the random seed for reproducibility
seed = 42
torch.manual_seed(seed)
# Define transforms
data_transforms = transforms.Compose([
transforms.Resize((224, 224)),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(30),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# Create the dataset
full_dataset = datasets.ImageFolder(data_dir, transform=data_transforms)
# Define the train-validation split ratio
train_size = int(0.8 * len(full_dataset))
val_size = len(full_dataset) - train_size
# Split the dataset
train_dataset, val_dataset = random_split(full_dataset, [train_size, val_size], generator=torch.Generator().manual_seed(seed))
# Create data loaders
train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=32, shuffle=False)
# Load the pre-trained ResNet50 model
resnet50 = models.resnet50(weights='ResNet50_Weights.DEFAULT')
# Freeze the parameters of the pre-trained model
for param in resnet50.parameters():
param.requires_grad = False
# Remove the final fully connected layer
num_ftrs = resnet50.fc.in_features
resnet50.fc = nn.Identity() # Replace the final layer with an identity function to get the feature vectors
# Define a custom neural network with one hidden layer and an output layer
class CustomNet(nn.Module):
def __init__(self, num_ftrs, num_classes):
super(CustomNet, self).__init__()
self.resnet50 = resnet50
self.hidden = nn.Linear(num_ftrs, 512)
self.relu = nn.ReLU()
self.output = nn.Linear(512, num_classes)
def forward(self, x):
x = self.resnet50(x) # Extract features using the pre-trained model
x = self.hidden(x) # Pass through the hidden layer
x = self.relu(x) # Apply ReLU activation
x = self.output(x) # Output layer
return x
# Instantiate the custom network
num_classes = len(full_dataset.classes)
model = CustomNet(num_ftrs, num_classes)
# Move the model to the appropriate device
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = model.to(device)
# Define criterion and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)
def train_model(model, dataloaders, criterion, optimizer, num_epochs=10):
best_model_wts = model.state_dict()
best_acc = 0.0
train_losses = []
val_losses = []
for epoch in range(num_epochs):
print(f'Epoch {epoch}/{num_epochs - 1}')
print('-' * 10)
# Each epoch has a training and validation phase
for phase in ['train', 'val']:
if phase == 'train':
model.train()
else:
model.eval()
running_loss = 0.0
running_corrects = 0
for inputs, labels in dataloaders[phase]:
inputs, labels = inputs.to(device), labels.to(device)
# Zero the parameter gradients
optimizer.zero_grad()
# Forward
with torch.set_grad_enabled(phase == 'train'):
outputs = model(inputs)
_, preds = torch.max(outputs, 1)
loss = criterion(outputs, labels)
# Backward + optimize only if in training phase
if phase == 'train':
loss.backward()
optimizer.step()
running_loss += loss.item() * inputs.size(0)
running_corrects += torch.sum(preds == labels.data)
epoch_loss = running_loss / len(dataloaders[phase].dataset)
epoch_acc = running_corrects.double() / len(dataloaders[phase].dataset)
print(f'{phase} Loss: {epoch_loss:.4f} Acc: {epoch_acc:.4f}')
if phase == 'train':
train_losses.append(epoch_loss)
else:
val_losses.append(epoch_loss)
# Deep copy the model
if phase == 'val' and epoch_acc > best_acc:
best_acc = epoch_acc
best_model_wts = model.state_dict()
print('Best val Acc: {:4f}'.format(best_acc))
# Load best model weights
model.load_state_dict(best_model_wts)
# Plot the training and validation loss
plt.figure(figsize=(10, 5))
plt.plot(train_losses, label='Training Loss')
plt.plot(val_losses, label='Validation Loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend()
plt.show()
return model
# Create a dictionary to hold the dataloaders
dataloaders = {'train': train_loader, 'val': val_loader}
# Train and evaluate the model
model = train_model(model, dataloaders, criterion, optimizer, num_epochs=10)
# Save the model
torch.save(model.state_dict(), 'drive/MyDrive/Ai_Hackathon_2024/plant_data/fine_tuned_plant_classifier.pth')
# Function to evaluate the model
def evaluate_model(model, dataloader):
model.eval()
correct = 0
total = 0
all_preds = []
all_labels = []
with torch.no_grad():
for inputs, labels in dataloader:
inputs, labels = inputs.to(device), labels.to(device)
outputs = model(inputs)
_, preds = torch.max(outputs, 1)
all_preds.extend(preds.cpu().numpy())
all_labels.extend(labels.cpu().numpy())
correct += (preds == labels).sum().item()
total += labels.size(0)
accuracy = correct / total
return accuracy, all_preds, all_labels
# Evaluate the model
dataloader = DataLoader(full_dataset, batch_size=32, shuffle=True)
accuracy, all_preds, all_labels = evaluate_model(model, dataloader)
# Calculate the number of correct and incorrect predictions
correct_preds = sum(np.array(all_preds) == np.array(all_labels))
incorrect_preds = len(all_labels) - correct_preds
print(f'Total images: {len(all_labels)}')
print(f'Correct predictions: {correct_preds}')
print(f'Incorrect predictions: {incorrect_preds}')
print(f'Accuracy: {accuracy:.4f}')
##-----------------------------------------------------------##
real_dataset = datasets.ImageFolder('drive/MyDrive/Ai_Hackathon_2024/plant_data/data_for_training', transform=data_transforms)
# Evaluate the model
dataloader = DataLoader(real_dataset, batch_size=32, shuffle=True)
accuracy, all_preds, all_labels = evaluate_model(model, dataloader)
# Calculate the number of correct and incorrect predictions
correct_preds = sum(np.array(all_preds) == np.array(all_labels))
incorrect_preds = len(all_labels) - correct_preds
print('-'*10)
print(f'Total images: {len(all_labels)}')
print(f'Correct predictions: {correct_preds}')
print(f'Incorrect predictions: {incorrect_preds}')
print(f'Accuracy: {accuracy:.4f}')
# Function to load and preprocess the image
def process_image(image_path):
data_transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
image = Image.open(image_path).convert('RGB')
image = data_transform(image)# data_transforms(image) # <-- data transforms uses all the random cropping as well
image = image.unsqueeze(0) # Add batch dimension
return image
#----------------------------INFERENCE PART----------------------------
# Function to predict the class of a single image
def predict_single_image(image_path, model):
# Load the image and preprocess it
image = process_image(image_path)
# Load the model
model.eval()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = model.to(device)
# Pass the image through the model
with torch.no_grad():
image = image.to(device)
outputs = model(image)
probabilities = torch.nn.functional.softmax(outputs[0], dim=0)
# Return the class names and their probabilities as a Pandas Series
return pd.Series(probabilities.cpu().numpy(), index=class_names).sort_values(ascending=False)
def classify(img_path):
# Path to the single image
image_path = img_path
# Initialize your custom model
model = CustomNet(num_ftrs, num_classes)
# Load the trained model weights
model.load_state_dict(torch.load('./fine_tuned:plant_classifier.pth'))
# Predict the class probabilities
class_probabilities = predict_single_image(image_path, model)
return class_probabilities
#----------------------------INFERENCE PART----------------------------
## script to automatically include larger drone images
import os
import shutil
from PIL import Image
# Define the paths
source_dir = 'path/to/source_images' # The directory with new images
target_base_dir = 'path/to/training_images' # The base directory containing original class folders
new_base_dir = 'path/to/training_images_2' # The base directory for the new substructure
# Extract the class folders
class_folders = [d for d in os.listdir(target_base_dir) if os.path.isdir(os.path.join(target_base_dir, d))]
# Function to extract ID from a filename
def extract_id(filename):
return filename.split('_')[0] # Assumes ID is the first part of the filename separated by '_'
# Function to crop the middle section of an image
def crop_middle_section(image):
width, height = image.size
new_width = width // 3
new_height = height // 3
left = (width - new_width) // 2
top = (height - new_height) // 2
right = left + new_width
bottom = top + new_height
return image.crop((left, top, right, bottom))
# Create the new base directory if it does not exist
os.makedirs(new_base_dir, exist_ok=True)
# Create a dictionary to map IDs to their respective class folders
id_to_class_folder = {}
for class_folder in class_folders:
class_folder_path = os.path.join(target_base_dir, class_folder)
for filename in os.listdir(class_folder_path):
if os.path.isfile(os.path.join(class_folder_path, filename)):
file_id = extract_id(filename)
id_to_class_folder[file_id] = class_folder
# Copy and manipulate the matching images
for filename in os.listdir(source_dir):
if os.path.isfile(os.path.join(source_dir, filename)):
file_id = extract_id(filename)
if file_id in id_to_class_folder:
target_class_folder = id_to_class_folder[file_id]
new_class_folder_path = os.path.join(new_base_dir, target_class_folder)
os.makedirs(new_class_folder_path, exist_ok=True) # Create the class folder if it doesn't exist
target_path = os.path.join(new_class_folder_path, filename)
# Open and manipulate the image
image_path = os.path.join(source_dir, filename)
with Image.open(image_path) as img:
cropped_img = crop_middle_section(img)
cropped_img.save(target_path)
print(f'Copied and cropped {filename} to {new_class_folder_path}')
print('Image processing and copying completed.')