CelebA_multicov.py

from __future__ import print_function, division

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

import torch
import torch.nn as nn
import torch.nn.functional as F

import torchvision
from torchvision import models, transforms, datasets

import torch.optim as optim
from torch.utils.data import DataLoader, Subset

import PIL.Image as Image
from tqdm import tqdm
import os
import time

from sklearn.metrics import f1_score

from model.vgg16 import *
# from model.hybrid_CNN import Hybrid_Conv2d

experiment_name = 'hybrid_bs_32_lr_1e-5_multicov_latefuse40'
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')


# get dataset
def load_data(batch_size, use_subset=True):
    """
    return the train/val/test dataloader
    """
    
    transform = transforms.Compose([
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.5] * 3, std=[0.5] * 3)
    ])
    
    train_dataset = datasets.CelebA(root='./data',
                                    split='train',
                                    target_type='attr',
                                    transform=transform,
                                    download=False)
    val_dataset = datasets.CelebA(root='./data',
                                    split='valid',
                                    target_type='attr',
                                    transform=transform,
                                    download=False)
    test_dataset = datasets.CelebA(root='./data',
                                    split='test',
                                    target_type='attr',
                                    transform=transform,
                                    download=False)
    
    indices_train = list(range(700))
    indices_val = list(range(150))    
    indices_test = list(range(150))
    
    train_subset = Subset(train_dataset, indices_train)
    val_subset = Subset(train_dataset, indices_val)
    test_subset = Subset(test_dataset, indices_test)

    # data loader
    train_loader = DataLoader(dataset=train_subset if use_subset else train_dataset, 
                                batch_size=batch_size,
                                shuffle=True)
    val_loader = DataLoader(dataset=val_subset if use_subset else val_dataset,
                                batch_size=batch_size,
                                shuffle=False)
    test_loader = DataLoader(dataset=test_subset if use_subset else test_dataset,
                                batch_size=batch_size,
                                shuffle=False)
    
    return train_loader, val_loader, test_loader


def initialize_model(model, learning_rate, num_classes):
    """
    initialize the model (pretrained vgg16_bn)
    define loss function and optimizer and move data to gpu if available
    
    return:
        model, loss function(criterion), optimizer
    """
    
    num_ftrs = model.classifier[6].in_features
    model.classifier[6] = nn.Linear(num_ftrs, num_classes)
    model = model.to(device)
    
    # Define loss function and optimizer
    criterion = nn.CrossEntropyLoss()   # potential alternative: nn.BCELoss()
    optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
    return model, criterion, optimizer

def make_plots(step_hist, loss_hist, epoch=0):
    plt.plot(step_hist, loss_hist)
    plt.xlabel('train_iterations')
    plt.ylabel('Loss')
    # plt.title('epoch'+str(epoch+1))
    plt.title(experiment_name)
    plt.savefig(experiment_name)
    plt.clf()


def train(train_loader, model, criterion, optimizer, num_epochs):
    """
    Move data to GPU memory and train for specified number of epochs
    Also plot the loss function and save it in `Figures/`
    Trained model is saved as `cnn.ckpt`
    """
    for epoch in range(num_epochs): # repeat the entire training `num_epochs` times
        # for each training sample
        loss_hist = []
        step_hist = []
        for i, (images, labels) in tqdm(enumerate(train_loader)):
         
            zero_one_labels = (labels + 1) // 2       # map from {-1, 1} to {0, 1}
            
            label = zero_one_labels[:, 2]             # attractiveness label
            cov_attr_1 = zero_one_labels[:, 31]       # smiling   
            cov_attr_2 = zero_one_labels[:, 39]       # young
            cov_attr_3 = zero_one_labels[:, 19]       # high_cheeekbones
            
            cov_attrs = torch.stack((cov_attr_1, cov_attr_2, cov_attr_3)).T # (minibatch, num_cov) e.g. (32, 3)
            
            # move to gpu if available
            images = images.to(device)
            cov_attrs = cov_attrs.to(device)
            label = label.to(device)
            
            # forward pass
            if isinstance(model, VGG):
                outputs = model(images)               # baseline vgg
            
            else:
                outputs = model(images, cov_attrs)    # hybrid model takes covariate here
            
            loss = criterion(outputs, label) 
            
            # backward
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
            
            if (i+1) % 50 == 0:
                print('Epoch: [{}/{}], Step[{}/{}], Loss:{:.4f}' \
                        .format(epoch+1, num_epochs, i+1, len(train_loader), loss.item()))
                with open(experiment_name+'.txt', 'a') as f:
                    print('Epoch: [{}/{}], Step[{}/{}], Loss:{:.4f}' \
                        .format(epoch+1, num_epochs, i+1, len(train_loader), loss.item()), file=f)
                loss_hist.append(loss.item())
                step_hist.append(i+1)
        
        make_plots(step_hist, loss_hist, epoch)
        
    torch.save(model.state_dict(), experiment_name+'.ckpt')


def evaluate(val_loader, model):
    """
    Run the validation set on the trained model
    """
    # uncomment if you want to load from checkpoint
    # model_path = "{}.ckpt".format(experiment_name)
    # state_dict = torch.load(model_path)
    # model.load_state_dict(state_dict)
    
    model.eval() 
    with torch.no_grad():
        correct = 0
        total = 0
        y_true = []
        y_pred = []
        for images, labels in tqdm(val_loader):
            
            zero_one_labels = (labels + 1) // 2 # map from {-1, 1} to {0, 1}
            
            label = labels[:, 2]
            cov_attr_1 = zero_one_labels[:, 31]       # smiling   
            cov_attr_2 = zero_one_labels[:, 39]       # young
            cov_attr_3 = zero_one_labels[:, 19]       # high_cheeekbones
            
            cov_attrs = torch.stack((cov_attr_1, cov_attr_2, cov_attr_3)).T # (minibatch, num_cov) e.g. (32, 3)
            
            # move to device
            images = images.to(device)
            cov_attrs = cov_attrs.to(device)
            label = label.to(device)
            
            # forward pass
            if isinstance(model, VGG):
                outputs = model(images)               # baseline vgg
            
            else: 
                outputs = model(images, cov_attrs)    # hybrid model takes covariate here
                
            _, predicted = torch.max(outputs.data, dim=1)
            
            # accumulate stats
            y_true.append(label.cpu().numpy()) # in the one
            y_pred.append(predicted.cpu().numpy())
            total += label.size(0) # number of elements in the tensor
            correct += (label == predicted).sum().item()
        
        y_true = np.concatenate(y_true)
        y_pred = np.concatenate(y_pred)
        
        print('F1 Score: {}'.format(f1_score(y_true, y_pred, average='macro')))
        print('Validation accuracy: {}'.format(correct / total))
        with open(experiment_name+'.txt', 'a') as f:
            print('F1 Score: {}'.format(f1_score(y_true, y_pred, average='macro')), file=f)
            print('Validation accuracy: {}'.format(correct / total), file=f)
    
    
def main():
    # hyper parameters    
    num_epochs = 1
    num_classes = 2
    batch_size = 32
    learning_rate = 1e-5
    model_name = HybridVGG16_v40()
    # model_name = vgg16_bn(pretrained=True) # baseline model
    
    print("Loading data...")
    train_loader, val_loader, test_loader = load_data(batch_size, use_subset=True)
    
    print("Initializing model...")
    model, criterion, optimizer = initialize_model(model_name, learning_rate, num_classes)
   
    print("Start training... \n")
    train(train_loader, model, criterion, optimizer, num_epochs)
    
    print("Start evaluating... \n")
    evaluate(val_loader, model)    

if __name__ == "__main__":
    main()