scripts/model.py

"""
Attribution: https://github.com/AIPI540/AIPI540-Deep-Learning-Applications/

Jon Reifschneider
Brinnae Bent 

"""

import os
import pandas as pd
import time
import torch
import numpy as np
import pandas as pd
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import TensorDataset
from sklearn.model_selection import train_test_split


class NNColabFiltering(nn.Module):

    def __init__(self, n_playlists, n_artists, embedding_dim_users, embedding_dim_items, n_activations, rating_range):
        super().__init__()
        self.user_embeddings = nn.Embedding(num_embeddings=n_playlists,embedding_dim=embedding_dim_users)
        self.item_embeddings = nn.Embedding(num_embeddings=n_artists,embedding_dim=embedding_dim_items)
        self.fc1 = nn.Linear(embedding_dim_users+embedding_dim_items,n_activations)
        self.fc2 = nn.Linear(n_activations,1)
        self.rating_range = rating_range

    def forward(self, X):
        # Get embeddings for minibatch
        embedded_users = self.user_embeddings(X[:,0])
        embedded_items = self.item_embeddings(X[:,1])
        # Concatenate user and item embeddings
        embeddings = torch.cat([embedded_users,embedded_items],dim=1)
        # Pass embeddings through network
        preds = self.fc1(embeddings)
        preds = F.relu(preds)
        preds = self.fc2(preds)
        # Scale predicted ratings to target-range [low,high]
        preds = torch.sigmoid(preds) * (self.rating_range[1]-self.rating_range[0]) + self.rating_range[0]
        return preds

def prep_dataloaders(X_train,y_train,X_val,y_val,batch_size):
    '''
    Loads the prefetched data from the output dir

    Inputs:
        X_train: training data features
        y_train: training data target
        X_val: validation data features
        y_val: validation data targets
        batch_size: the batch size to use

    Returns:
        trainloader: training dataloader
        valloader: validation dataloader
    '''
    # Convert training and test data to TensorDatasets
    trainset = TensorDataset(torch.from_numpy(np.array(X_train)).long(),
                            torch.from_numpy(np.array(y_train)).float())
    valset = TensorDataset(torch.from_numpy(np.array(X_val)).long(),
                            torch.from_numpy(np.array(y_val)).float())

    # Create Dataloaders for our training and test data to allow us to iterate over minibatches
    trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True)
    valloader = torch.utils.data.DataLoader(valset, batch_size=batch_size, shuffle=False)

    return trainloader, valloader

def train_model(model, criterion, optimizer, dataloaders, device, num_epochs=5, scheduler=None):
    '''
    Loads the prefetched data from the output dir

    Inputs:
        model: the model to train
        criterion: the criterion to use to train
        optimizer: the optimizer to use to train
        dataloaders: the dict of dataloaders to user in the training and validation 
        device: the torch defined cpu/gpu 
        num_epochs: number of epochs to use for training
        scheduler: the scheduler to use to train for training 

    Returns:
        costpaths: the loss for each epoch for validation and training
    '''
    model = model.to(device) 
    since = time.time()

    costpaths = {'train':[],'val':[]}

    for epoch in range(num_epochs):
        print('Epoch {}/{}'.format(epoch, num_epochs - 1))
        print('-' * 10)

        for phase in ['train', 'val']:
            if phase == 'train':
                model.train()  
            else:
                model.eval()  

            running_loss = 0.0

            index = 0
            for (inputs,labels) in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)

                optimizer.zero_grad()

                with torch.set_grad_enabled(phase == 'train'):
                    outputs = model.forward(inputs).view(-1)
                    loss = criterion(outputs, labels)

                    if phase == 'train':
                        loss.backward()
                        optimizer.step()

                running_loss += np.sqrt(loss.item()) * labels.size(0)
                print(f'\r{running_loss} {index} {(index / len(dataloaders[phase]))*100:.2f}%', end='')
                index +=1

            if (phase == 'train') and (scheduler is not None):
                scheduler.step()

            epoch_loss = running_loss / len(dataloaders[phase].dataset)
            costpaths[phase].append(epoch_loss)
            print('\n{} loss: {:.4f}'.format(phase, epoch_loss))

    time_elapsed = time.time() - since
    print('Training complete in {:.0f}m {:.0f}s'.format(
        time_elapsed // 60, time_elapsed % 60))

    return costpaths


if __name__ == '__main__':
    artists = pd.read_csv(os.getcwd() + '/data/processed/playlists.csv')
    X = artists.loc[:,['playlist_id','artist_album_id',]]
    y = artists.loc[:,'song_percent']

    # Split our data into training and test sets
    X_train, X_val, y_train, y_val = train_test_split(X,y,random_state=0, test_size=0.2)
    batchsize = 64
    trainloader,valloader = prep_dataloaders(X_train,y_train,X_val,y_val,batchsize)

    dataloaders = {'train':trainloader, 'val':valloader}
    n_users = X.loc[:,'playlist_id'].max()+1
    n_items = X.loc[:,'artist_album_id'].max()+1
    model = NNColabFiltering(n_users,n_items,embedding_dim_users=50, embedding_dim_items=50, n_activations = 100,rating_range=[0.,1.])
    criterion = nn.MSELoss()
    lr=0.001
    n_epochs=10
    wd=1e-3
    optimizer = optim.Adam(model.parameters(), lr=lr, weight_decay=wd)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    cost_paths = train_model(model,criterion,optimizer,dataloaders, device,n_epochs, scheduler=None)

    # Save the entire model
    torch.save(model, os.getcwd() + '/models/recommender.pt')