|
| 1 | +# -*- coding: utf-8 -*- |
| 2 | + |
| 3 | +""" |
| 4 | +A pytorch implementation of DeepFM for rates prediction problem. |
| 5 | +""" |
| 6 | + |
| 7 | +import torch |
| 8 | +import torch.nn as nn |
| 9 | +import torch.nn.functional as F |
| 10 | +import torch.optim as optim |
| 11 | + |
| 12 | +from time import time |
| 13 | + |
| 14 | + |
| 15 | +class DeepFM(nn.Module): |
| 16 | + """ |
| 17 | + A DeepFM network with RMSE loss for rates prediction problem. |
| 18 | +
|
| 19 | + There are two parts in the architecture of this network: fm part for low |
| 20 | + order interactions of features and deep part for higher order. In this |
| 21 | + network, we use bachnorm and dropout technology for all hidden layers, |
| 22 | + and "Adam" method for optimazation. |
| 23 | +
|
| 24 | + You may find more details in this paper: |
| 25 | + DeepFM: A Factorization-Machine based Neural Network for CTR Prediction, |
| 26 | + Huifeng Guo, Ruiming Tang, Yunming Yey, Zhenguo Li, Xiuqiang He. |
| 27 | + """ |
| 28 | + |
| 29 | + def __init__(self, feature_sizes, embedding_size=4, |
| 30 | + hidden_dims=[32, 32], num_classes=10, dropout=[0.5, 0.5], |
| 31 | + use_cuda=True, verbose=False): |
| 32 | + """ |
| 33 | + Initialize a new network |
| 34 | +
|
| 35 | + Inputs: |
| 36 | + - feature_size: A list of integer giving the size of features for each field. |
| 37 | + - embedding_size: An integer giving size of feature embedding. |
| 38 | + - hidden_dims: A list of integer giving the size of each hidden layer. |
| 39 | + - num_classes: An integer giving the number of classes to predict. For example, |
| 40 | + someone may rate 1,2,3,4 or 5 stars to a film. |
| 41 | + - batch_size: An integer giving size of instances used in each interation. |
| 42 | + - use_cuda: Bool, Using cuda or not |
| 43 | + - verbose: Bool |
| 44 | + """ |
| 45 | + super().__init__() |
| 46 | + self.field_size = len(feature_sizes) |
| 47 | + self.feature_sizes = feature_sizes |
| 48 | + self.embedding_size = embedding_size |
| 49 | + self.hidden_dims = hidden_dims |
| 50 | + self.num_classes = num_classes |
| 51 | + self.dtype = torch.int32 |
| 52 | + self.bias = torch.nn.Parameter(torch.randn(1)) |
| 53 | + """ |
| 54 | + check if use cuda |
| 55 | + """ |
| 56 | + if use_cuda and torch.cuda.is_available(): |
| 57 | + self.device = torch.device('cuda') |
| 58 | + else: |
| 59 | + self.device = torch.device('cpu') |
| 60 | + """ |
| 61 | + init fm part |
| 62 | + """ |
| 63 | + self.fm_first_order_embeddings = nn.ModuleList( |
| 64 | + [nn.Embedding(feature_size, 1) for feature_size in self.feature_sizes]) |
| 65 | + self.fm_second_order_embeddings = nn.ModuleList( |
| 66 | + [nn.Embedding(feature_size, self.embedding_size) for feature_size in self.feature_sizes]) |
| 67 | + """ |
| 68 | + init deep part |
| 69 | + """ |
| 70 | + all_dims = [self.field_size * self.embedding_size] + \ |
| 71 | + self.hidden_dims + [self.num_classes] |
| 72 | + for i in range(1, len(hidden_dims) + 1): |
| 73 | + setattr(self, 'linear_'+str(i), |
| 74 | + nn.Linear(all_dims[i-1], all_dims[i])) |
| 75 | + # nn.init.kaiming_normal_(self.fc1.weight) |
| 76 | + setattr(self, 'batchNorm_' + str(i), |
| 77 | + nn.BatchNorm1d(all_dims[i])) |
| 78 | + setattr(self, 'dropout_'+str(i), |
| 79 | + nn.Dropout(dropout[i-1])) |
| 80 | + |
| 81 | + def forward(self, Xi, Xv): |
| 82 | + """ |
| 83 | + Forward process of network. |
| 84 | +
|
| 85 | + Inputs: |
| 86 | + - Xi: A tensor of input's index, shape of (N, D, 1) |
| 87 | + - Xv: A tensor of input's value, shape of (N, D, 1) |
| 88 | + """ |
| 89 | + """ |
| 90 | + fm part |
| 91 | + """ |
| 92 | + fm_first_order_emb_arr = [(torch.sum(emb(Xi[:, i, :]), 1).t() * \ |
| 93 | + Xv[:, i]).t() for i, emb in enumerate(self.fm_first_order_embeddings)] |
| 94 | + fm_first_order = torch.cat(fm_first_order_emb_arr, 1) |
| 95 | + # use 2xy = (x+y)^2 - x^2 - y^2 reduce calculation |
| 96 | + fm_second_order_emb_arr = [(torch.sum(emb(Xi[:, i, :]), 1).t() * \ |
| 97 | + Xv[:, i]).t() for i, emb in enumerate(self.fm_second_order_embeddings)] |
| 98 | + fm_sum_second_order_emb = sum(fm_second_order_emb_arr) |
| 99 | + fm_sum_second_order_emb_square = fm_sum_second_order_emb * \ |
| 100 | + fm_sum_second_order_emb # (x+y)^2 |
| 101 | + fm_second_order_emb_square = [ |
| 102 | + item*item for item in fm_second_order_emb_arr] |
| 103 | + fm_second_order_emb_square_sum = sum( |
| 104 | + fm_second_order_emb_square) # x^2+y^2 |
| 105 | + fm_second_order = (fm_sum_second_order_emb_square - |
| 106 | + fm_second_order_emb_square_sum) * 0.5 |
| 107 | + """ |
| 108 | + deep part |
| 109 | + """ |
| 110 | + deep_emb = torch.cat(fm_second_order_emb_arr, 1) |
| 111 | + deep_out = deep_emb |
| 112 | + for i in range(1, self.hidden_dims + 1): |
| 113 | + deep_out = getattr(self, 'linear_' + str(i))(deep_out) |
| 114 | + deep_out = getattr(self, 'batchNorm_' + str(i))(deep_out) |
| 115 | + deep_out = getattr(self, 'dropout_' + str(i))(deep_out) |
| 116 | + """ |
| 117 | + sum |
| 118 | + """ |
| 119 | + total_sum = torch.sum(fm_first_order, 1) + \ |
| 120 | + torch.sum(fm_second_order, 1) + torch.sum(deep_out, 1) + self.bias |
| 121 | + return total_sum |
| 122 | + |
| 123 | + def train(self, loader_train, loader_val, optimizer, epochs=1, verbose=False, print_every=100): |
| 124 | + """ |
| 125 | + Training a model. |
| 126 | +
|
| 127 | + Inputs: |
| 128 | + - loader_train: I |
| 129 | + - loader_val: . |
| 130 | + - optimizer: Abstraction of optimizer used in training process, e.g., "torch.optim.Adam()""torch.optim.SGD()". |
| 131 | + - epochs: Integer, number of epochs. |
| 132 | + - verbose: Bool, if print. |
| 133 | + - print_every: Integer, print after every number of iterations. |
| 134 | + """ |
| 135 | + """ |
| 136 | + load input data |
| 137 | + """ |
| 138 | + model = self.train().to(device=self.device) |
| 139 | + criterion = F.binary_cross_entropy_with_logits |
| 140 | + |
| 141 | + for _ in range(epochs): |
| 142 | + for t, (xi, xv, y) in enumerate(loader_train): |
| 143 | + model.train() |
| 144 | + xi = xi.to(device=self.device, dtype=self.dtype) |
| 145 | + xv = xv.to(device=self.device, dtype=self.dtype) |
| 146 | + y = y.to(device=self.device, dtype=torch.long) |
| 147 | + |
| 148 | + total = model(xi, xv) |
| 149 | + loss = criterion(total, y) |
| 150 | + optimizer.zero_grad() |
| 151 | + loss.backward() |
| 152 | + optimizer.step() |
| 153 | + |
| 154 | + if verbose and t % print_every == 0: |
| 155 | + print('Iteration %d, loss = %.4f' % (t, loss.item())) |
| 156 | + self.check_accuracy(loader_val, model) |
| 157 | + print() |
| 158 | + |
| 159 | + def check_accuracy(self, loader, model): |
| 160 | + if loader.dataset.train: |
| 161 | + print('Checking accuracy on validation set') |
| 162 | + else: |
| 163 | + print('Checking accuracy on test set') |
| 164 | + num_correct = 0 |
| 165 | + num_samples = 0 |
| 166 | + model.eval() # set model to evaluation mode |
| 167 | + with torch.no_grad(): |
| 168 | + for xi, xv, y in loader: |
| 169 | + xi = xi.to(device=self.device, dtype=self.dtype) # move to device, e.g. GPU |
| 170 | + xv = xv.to(device=self.device, dtype=self.dtype) |
| 171 | + y = y.to(device=self.device, dtype=torch.long) |
| 172 | + total = model(xi, xv) |
| 173 | + preds = (F.sigmoid(total) > 0.5) |
| 174 | + num_correct += (preds == y).sum() |
| 175 | + num_samples += preds.size(0) |
| 176 | + acc = float(num_correct) / num_samples |
| 177 | + print('Got %d / %d correct (%.2f%%)' % (num_correct, num_samples, 100 * acc)) |
| 178 | + |
| 179 | + |
| 180 | + |
| 181 | + |
| 182 | + |
0 commit comments