pytorch深度學(xué)習(xí)

The recommendation is a simple algorithm that works on the principle of data filtering. The algorithm finds a pattern between two users and recommends or provides additional relevant information to a user in choosing a product or services.

該建議是一種基于數(shù)據(jù)過濾原理的簡(jiǎn)單算法。 該算法在兩個(gè)用戶之間找到一種模式，并在選擇產(chǎn)品或服務(wù)時(shí)向用戶推薦或提供其他相關(guān)信息。

TL;DR Please follow this link to directly jump into full source code required to prepare the dataset and to train the model.

TL; DR請(qǐng)單擊此鏈接直接跳至準(zhǔn)備數(shù)據(jù)集和訓(xùn)練模型所需的完整源代碼。

協(xié)同過濾 (Collaborative Filtering)

Whenever we go to Amazon or any online store, we get recommendations stating that “Customers who brought this item also bought”. These recommendations based on your preference is obtained by an algorithm which predicts based on the previous buying made, this algorithm is called Collaborative Filtering. The main idea or concept in this algorithm is that network learns the similarity between two products based on buying history for example a person bought bread mostly buys milk, whenever we buy bread algorithm suggest us to buy milk.

每當(dāng)我們?nèi)嗰R遜或任何在線商店時(shí)，我們都會(huì)得到建議，說明“攜帶此商品的顧客也購(gòu)買了”。 這些根據(jù)您的偏好提供的建議是通過一種算法進(jìn)行的，該算法根據(jù)之前的購(gòu)買進(jìn)行預(yù)測(cè)，該算法稱為“ 協(xié)同過濾” 。 該算法的主要思想或概念是，網(wǎng)絡(luò)基于購(gòu)買歷史來學(xué)習(xí)兩種產(chǎn)品之間的相似性，例如，每當(dāng)我們購(gòu)買面包算法建議我們購(gòu)買牛奶時(shí)，一個(gè)人購(gòu)買的面包主要是購(gòu)買牛奶。

To implement this we are using a Movie dataset called MovieLens (https://grouplens.org/datasets/movielens/). This is an open-source dataset available in grouplens.org, The data set has 25000095 ratings and 1093360 tag applications across 62423 movies. These data are created by 162541 users between 09 January 1995 and 21 November 2019.

為了實(shí)現(xiàn)這一點(diǎn)，我們使用了一個(gè)名為MovieLens ( https://grouplens.org/datasets/movielens/ )的Movie數(shù)據(jù)集。 這是grouplens.org中可用的開源數(shù)據(jù)集。該數(shù)據(jù)集在62423個(gè)電影中具有25000095評(píng)級(jí)和1093360標(biāo)簽應(yīng)用程序。 這些數(shù)據(jù)由162541用戶在1995年1月9日至2019年11月21日期間創(chuàng)建。

Dataset has

數(shù)據(jù)集有

• User Ids
  用戶編號(hào)
• Movie Ids
  電影ID
• Rating data file Structure (ratings.csv)
  評(píng)級(jí)數(shù)據(jù)文件結(jié)構(gòu)(ratings.csv)

userId,movieId,rating,timestamp

• Tags Data File Structure (tags.csv)
  標(biāo)簽數(shù)據(jù)文件結(jié)構(gòu)(tags.csv)

userId,movieId,tag,timestamp

• Movies Data File Structure (movies.csv)
  電影數(shù)據(jù)文件結(jié)構(gòu)(movies.csv)

movieId,title,genres

We will be analyzing two major data frames rating.csv and movies.csv. Below shows the image of sample head() from each data frame.

我們將分析兩個(gè)主要的數(shù)據(jù)幀rating.csv和movies.csv。 下面顯示了每個(gè)數(shù)據(jù)幀中的head()樣本圖像。

A machine learning algorithm accepts only an array of numerical values, we just can't send the above dataset directly. There are a lot of embedding approaches most widely used approach is one-hot encoding, next is word2vec. One hot encoding, the columns with categorical data are numbered based on the number of categories like 0/1. We have more than 1000 category data, so we created a Neural network-based embedding of data.

機(jī)器學(xué)習(xí)算法僅接受數(shù)值數(shù)組，我們無法直接發(fā)送上述數(shù)據(jù)集。 嵌入方法有很多，最廣泛使用的方法是單熱編碼，其次是word2vec。 一種熱編碼，具有類別數(shù)據(jù)的列是根據(jù)類別的數(shù)量(例如0/1)進(jìn)行編號(hào)的。 我們有1000多個(gè)類別數(shù)據(jù)，因此我們創(chuàng)建了基于神經(jīng)網(wǎng)絡(luò)的數(shù)據(jù)嵌入。

網(wǎng)絡(luò) (Network)

With the PyTorch framework, we created an embedding network, which takes in the Number of users and Number of movies as input. The network takes the output of Movies embedding and User embeddings as inputs, which concatenate into a column (array). The network has 4 layers starting with dropout layers, then 3 fully connected layers with relu activation and a dropout. Drop out as added to randomize the network and increase the learning capability of the network. Finally, output layer with a sigmoid activation function.

使用PyTorch框架，我們創(chuàng)建了一個(gè)嵌入網(wǎng)絡(luò)，該網(wǎng)絡(luò)將用戶數(shù)和電影數(shù)作為輸入。 網(wǎng)絡(luò)將“電影”嵌入和“用戶”嵌入的輸出作為輸入，并連接到一列(數(shù)組)中。 該網(wǎng)絡(luò)具有4個(gè)層，其中第1個(gè)層是Dropout層，然后是3個(gè)完全連接的層，具有relu激活和一個(gè)dropout。 退出添加以使網(wǎng)絡(luò)隨機(jī)化并增加網(wǎng)絡(luò)的學(xué)習(xí)能力。 最后，輸出層具有S型激活功能。

Below snippet shows the network implementation with the PyTorch framework. (https://gist.github.com/9aec88bc33b50dbf9c6390bbeb42ba0b.git)

下面的代碼片段顯示了使用PyTorch框架的網(wǎng)絡(luò)實(shí)現(xiàn)。 ( https://gist.github.com/9aec88bc33b50dbf9c6390bbeb42ba0b.git )

class EmbeddingNet(nn.Module):"""Creates a dense network with embedding layers.Args:n_users: Number of unique users in the dataset.n_movies: Number of unique movies in the dataset.n_factors: Number of columns in the embeddings matrix.embedding_dropout: Dropout rate to apply right after embeddings layer.hidden:A single integer or a list of integers defining the number of units in hidden layer(s).dropouts: A single integer or a list of integers defining the dropout layers rates applyied right after each of hidden layers."""def __init__(self, n_users, n_movies,n_factors=50, embedding_dropout=0.02, hidden=10, dropouts=0.2):super().__init__()hidden = get_list(hidden)dropouts = get_list(dropouts)n_last = hidden[-1]def gen_layers(n_in):"""A generator that yields a sequence of hidden layers and their activations/dropouts.Note that the function captures `hidden` and `dropouts` values from the outer scope."""nonlocal hidden, dropoutsassert len(dropouts) <= len(hidden)for n_out, rate in zip_longest(hidden, dropouts):yield nn.Linear(n_in, n_out)yield nn.ReLU()if rate is not None and rate > 0.:yield nn.Dropout(rate)n_in = n_outself.u = nn.Embedding(n_users, n_factors)self.m = nn.Embedding(n_movies, n_factors)self.drop = nn.Dropout(embedding_dropout)self.hidden = nn.Sequential(*list(gen_layers(n_factors * 2)))self.fc = nn.Linear(n_last, 1)self._init()def forward(self, users, movies, minmax=None):features = torch.cat([self.u(users), self.m(movies)], dim=1)x = self.drop(features)x = self.hidden(x)out = torch.sigmoid(self.fc(x))if minmax is not None:min_rating, max_rating = minmaxout = out*(max_rating - min_rating + 1) + min_rating - 0.5return outdef _init(self):"""Setup embeddings and hidden layers with reasonable initial values."""def init(m):if type(m) == nn.Linear:torch.nn.init.xavier_uniform_(m.weight)m.bias.data.fill_(0.01)self.u.weight.data.uniform_(-0.05, 0.05)self.m.weight.data.uniform_(-0.05, 0.05)self.hidden.apply(init)init(self.fc)def get_list(n):if isinstance(n, (int, float)):return [n]elif hasattr(n, '__iter__'):return list(n)raise TypeError('layers configuraiton should be a single number or a list of numbers')

For example with a network with 100,200,300 as a number of hidden layers and dropouts. The output network with be as shown below.

例如，對(duì)于一個(gè)具有100,200,300的網(wǎng)絡(luò)，其中包含許多隱藏層和缺失。 輸出網(wǎng)絡(luò)如下所示。

EmbeddingNet(n, m, n_factors=150, hidden=[100, 200, 300], dropouts=[0.25, 0.5])

EmbeddingNet(n，m，n_factors = 150，隱藏= [100，200，300]，輟學(xué)= [0.25，0.5])

訓(xùn)練循環(huán) (Training Loop)

Below is the snippet of the training loop with Mean Squared Loss as quality measurement metrics and Adam function are used as an optimizer. Training parameters are chosen based on my previous experience, you can change it according to your own style.

下面是訓(xùn)練循環(huán)的代碼段，其中均方根損耗作為質(zhì)量測(cè)量指標(biāo)和Adam函數(shù)用作優(yōu)化器。 訓(xùn)練參數(shù)是根據(jù)我以前的經(jīng)驗(yàn)選擇的，您可以根據(jù)自己的風(fēng)格進(jìn)行更改。

# training loop parameters lr = 1e-3 wd = 1e-5 bs = 2000 n_epochs = 100 patience = 10 no_improvements = 0 best_loss = np.inf best_weights = None history = [] lr_history = []# use GPU if available identifier = 'cuda:0' if torch.cuda.is_available() else 'cpu' device = torch.device(identifier)# setting up network, optimizer and learning rate scheduler net.to(device) criterion = nn.MSELoss(reduction='sum') optimizer = optim.Adam(net.parameters(), lr=lr, weight_decay=wd) iterations_per_epoch = int(math.ceil(dataset_sizes['train'] // bs)) sched_func = cosine(t_max=iterations_per_epoch * 2, eta_min=lr/10) scheduler = CyclicLR(optimizer, sched_func)fmt = '[{epoch:03d}/{total:03d}] train: {train:.4f} - val: {val:.4f}'# start training for epoch in range(n_epochs):stats = {'epoch': epoch + 1, 'total': n_epochs}for phase in ('train', 'val'):training = phase == 'train'running_loss = 0.0n_batches = 0iterator = batches(*datasets[phase], shuffle=training, bs=bs)for batch in iterator:x_batch, y_batch = [b.to(device) for b in batch]optimizer.zero_grad()with torch.set_grad_enabled(training):outputs = net(x_batch[:, 1], x_batch[:, 0], minmax)loss = criterion(outputs, y_batch)if training:scheduler.step()loss.backward()optimizer.step()lr_history.extend(scheduler.get_lr())running_loss += loss.item()epoch_loss = running_loss / dataset_sizes[phase]stats[phase] = epoch_lossif phase == 'val':if epoch_loss < best_loss:print('loss improvement on epoch: %d' % (epoch + 1))best_loss = epoch_lossbest_weights = copy.deepcopy(net.state_dict())no_improvements = 0else:no_improvements += 1history.append(stats)print(fmt.format(**stats))if no_improvements >= patience:break

Please follow this link to see the full source code required to prepare the dataset and to train the model.

請(qǐng)點(diǎn)擊此鏈接查看準(zhǔn)備數(shù)據(jù)集和訓(xùn)練模型所需的完整源代碼。

結(jié)論 (Conclusion)

After 10–15mins of training the network resulted in an accuracy of 0.8853. The above graph shows the learning curve of the network during training and validation over 16 epochs. PyTorch is a powerful framework that has the potential to easily scaling it up large datasets. There is another data set available at Kaggle (https://www.kaggle.com/ibtesama/getting-started-with-a-movie-recommendation-system?) that would be great to try it.

經(jīng)過10-15分鐘的培訓(xùn)，網(wǎng)絡(luò)得出的精度為0.8853。 上圖顯示了在超過16個(gè)時(shí)期的訓(xùn)練和驗(yàn)證過程中網(wǎng)絡(luò)的學(xué)習(xí)曲線。 PyTorch是一個(gè)功能強(qiáng)大的框架，具有輕松擴(kuò)展大型數(shù)據(jù)集的潛力。 Kaggle上還有另一個(gè)可用的數(shù)據(jù)集( https://www.kaggle.com/ibtesama/getting-started-with-a-movie-recommendation-system？ )，可以嘗試一下。

Interested in learning and working with Python language, Machine Learning, Data science, even Robotics. Then probably you would be interested in my blog where I am talking about various programming topics and provide links to textbooks and guides I’ve found interesting.

對(duì)學(xué)習(xí)和使用Python語言，機(jī)器學(xué)習(xí)，數(shù)據(jù)科學(xué)甚至機(jī)器人技術(shù)感興趣。 然后，您可能會(huì)對(duì) 我 在 博客 中談?wù)摳鞣N編程主題并提供指向我發(fā)現(xiàn)很有趣的教科書和指南的鏈接感興趣。

Please check out the Instagram page for updates https://www.instagram.com/rudraalabs/. As always, feel free to send us any questions or feedback you might have.

請(qǐng)查看 Instagram 頁面以獲取更新 https://www.instagram.com/rudraalabs/ 。 與往常一樣，隨時(shí)向我們發(fā)送您可能有的任何問題或反饋。

Thank you for reading the blog.

感謝您閱讀博客。

翻譯自: https://medium.com/swlh/recommendation-system-implementation-with-deep-learning-and-pytorch-a03ee84a96f4

pytorch深度學(xué)習(xí)

總結(jié)

以上是生活随笔為你收集整理的pytorch深度学习_深度学习和PyTorch的推荐系统实施的全部?jī)?nèi)容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網(wǎng)站內(nèi)容還不錯(cuò)，歡迎將生活随笔推薦給好友。