train.py

import numpy as np
import os
import sys
import cv2
import torch
from torch.utils.data import DataLoader
from tqdm import tqdm
import torchvision
from torchvision import transforms
import torch.nn as nn
import torch.optim as optim
from torch.utils.data.sampler import SubsetRandomSampler


sys.path.append('/Users/suyashharlalka/Documents/workspace/gabit/acne_classification/code')

from dataset import AcneDataset
from utils import save_model, get_transforms_to_apply,get_model_architecture, get_training_params, get_criterion
import config
from base import TrainingObjective, ModelBackbone
import json



data_dir = config.DATASET_PATH
image_names = os.listdir(data_dir)
model_training = config.MODEL_TRAINING
isLimited = config.IS_LIMITED
batch_size = config.BATCH_SIZE
shuffle = config.SHUFFLE
num_workers = config.NUM_WORKERS

dataset = AcneDataset(data_dir, limit=isLimited)

validation_split = 0.2
dataset_size = len(dataset)
indices = list(range(dataset_size))
split = int(np.floor(validation_split * dataset_size))
if shuffle:
    np.random.seed(42)
    np.random.shuffle(indices)

train_indices, test_indices = indices[split:], indices[:split]
train_sampler = SubsetRandomSampler(train_indices)
test_sampler = SubsetRandomSampler(test_indices)

transform = get_transforms_to_apply()
dataset.transform = transform

train_dataloader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, sampler=train_sampler)
test_dataloader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, sampler=test_sampler)

num_classes = config.NUM_CLASSES
model = get_model_architecture()
training_params = get_training_params(model)
criterion = get_criterion()

optimizer = optim.Adam(training_params, lr=config.BASE_LR)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=config.LR_DECAY_STEP_SIZE, gamma=config.LR_DECAY_GAMMA)

device = torch.device('mps') if torch.backends.mps.is_available() else torch.device('cpu')
model.to(device)

config_json = {}
config_json['DATASET_PATH'] = config.DATASET_PATH
config_json['MODEL_DIR'] = config.MODEL_DIR
config_json['MODEL_OBJECTIVE'] = config.MODEL_OBJECTIVE.name
config_json['MODEL_BACKBONE'] = config.MODEL_BACKBONE.name
config_json['MODEL_INPUT_SIZE'] = config.MODEL_INPUT_SIZE
config_json['TRANSFORMS_TO_APPLY'] = config.TRANSFORMS_TO_APPLY
config_json['NUM_CLASSES'] = config.NUM_CLASSES
config_json['LAST_N_LAYERS_TO_TRAIN'] = config.LAST_N_LAYERS_TO_TRAIN
config_json['EPOCHS'] = config.EPOCHS
config_json['MODEL_TRAINING'] = config.MODEL_TRAINING
config_json['IS_LIMITED'] = config.IS_LIMITED
config_json['BATCH_SIZE'] = config.BATCH_SIZE
config_json['SHUFFLE'] = config.SHUFFLE
config_json['NUM_WORKERS'] = config.NUM_WORKERS
config_json['BASE_LR'] = config.BASE_LR
config_json['LR_DECAY_STEP_SIZE'] = config.LR_DECAY_STEP_SIZE
config_json['LR_DECAY_GAMMA'] = config.LR_DECAY_GAMMA


if model_training:
    num_epochs = config.EPOCHS
    for epoch in range(num_epochs):
        model.train()
        runningLoss = 0.0
        for i, (images, labels) in enumerate(tqdm(train_dataloader, desc="Processing", unit="batch")):
            images = images.to(device)
            labels = labels.to(device)
            outputs = model(images)
            loss = criterion(outputs.squeeze(), labels)
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
            runningLoss += loss.item()
        
        # scheduler.step()
        print(f'Epoch [{epoch+1}/{num_epochs}], Loss: {runningLoss/len(train_dataloader):.4f}')
        
        if 'LOSS' not in config_json:
            config_json['LOSS'] = []
        config_json['LOSS'].append(runningLoss/len(train_dataloader))

        if 'EPOCHS_DONE' not in config_json:
            config_json['EPOCHS_DONE'] = 0
        config_json['EPOCHS_DONE'] = epoch + 1

        if epoch == 0 :
            model_dir = save_model(model, config_json)
        else:
            model_dir = save_model(model, config_json, model_dir)
        
        if 'TRAINED_MODEL_DIR' not in config_json:
            config_json['TRAINED_MODEL_DIR'] = model_dir
        
        config_save_path = os.path.join(model_dir, 'config.json')
        with open(config_save_path, 'w') as f:
            json.dump(config_json, f)


# config_path = os.path.join(model_dir, 'config.json')   
config_path = '/Users/suyash.harlalka/Desktop/personal/acne_classification/model/model_1/config.json' 
with open(config_path, 'r') as f:
    config_loaded = json.load(f)

from sklearn.metrics import confusion_matrix

model_trained_path = os.path.join(config_loaded['TRAINED_MODEL_DIR'], 'model.pth')
model.load_state_dict(torch.load(model_trained_path))
model.eval()
with torch.no_grad():
    correct = 0
    total = 0
    mainLabel = []
    predictedLabel = []
    for images, labels in tqdm(test_dataloader, desc="Processing", unit="batch"):
        images = images.to(device)
        labels = labels.to(device)
        outputs = model(images)
        predicted = torch.round(outputs.data)
        predicted = predicted.squeeze(1)
        total += labels.size(0)
        correct += (predicted == labels).sum().item()

        mainLabel.extend(labels.cpu().numpy())
        predictedLabel.extend(predicted.cpu().numpy())
        
    print(f'Accuracy of the network on the {total} test images: {100 * correct / total}%')
    
    cft = confusion_matrix(mainLabel, predictedLabel, labels=[0, 1, 2, 3], normalize='true')
    print(cft)

    correct = 0
    total = 0
    mainLabel = []
    predictedLabel = []
    for images, labels in tqdm(train_dataloader, desc="Processing", unit="batch"):
        images = images.to(device)
        labels = labels.to(device)
        outputs = model(images)
        predicted = torch.round(outputs.data)
        predicted = predicted.squeeze(1)
        total += labels.size(0)
        correct += (predicted == labels).sum().item()

        mainLabel.extend(labels.cpu().numpy())
        predictedLabel.extend(predicted.cpu().numpy())
    
    print(f'Accuracy of the network on the {total} train images: {100 * correct / total}%')
    cft = confusion_matrix(mainLabel, predictedLabel, labels=[0, 1, 2, 3], normalize='true')
    print(cft)