# Copyright 2018 The Cornac Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
import numpy as np
from tqdm.auto import tqdm
from ..recommender import Recommender
from ..recommender import ANNMixin, MEASURE_DOT
from ...exception import CornacException
from ...utils import fast_dot
from ...utils import get_rng
from ...utils.init_utils import xavier_uniform
[docs]
class AMR(Recommender, ANNMixin):
"""Adversarial Training Towards Robust Multimedia Recommender System.
Parameters
----------
k: int, optional, default: 10
The dimension of the gamma latent factors.
k2: int, optional, default: 10
The dimension of the theta latent factors.
n_epochs: int, optional, default: 20
Maximum number of epochs for SGD.
batch_size: int, optional, default: 100
The batch size for SGD.
learning_rate: float, optional, default: 0.001
The learning rate for SGD.
lambda_w: float, optional, default: 0.01
The regularization hyper-parameter for latent factor weights.
lambda_b: float, optional, default: 0.01
The regularization hyper-parameter for biases.
lambda_e: float, optional, default: 0.0
The regularization hyper-parameter for embedding matrix E and beta prime vector.
lambda_adv: float, optional, default: 1.0
The regularization hyper-parameter in Eq. (8) and (10) for the adversarial sample loss.
use_gpu: boolean, optional, default: True
Whether or not to use GPU to speed up training.
trainable: boolean, optional, default: True
When False, the model is not trained and Cornac assumes that the model already \
pre-trained (U and V are not None).
verbose: boolean, optional, default: True
When True, running logs are displayed.
init_params: dictionary, optional, default: None
Initial parameters, e.g., init_params = {'Bi': beta_item, 'Gu': gamma_user,
'Gi': gamma_item, 'Tu': theta_user, 'E': emb_matrix, 'Bp': beta_prime}
seed: int, optional, default: None
Random seed for weight initialization.
References
----------
* Tang, J., Du, X., He, X., Yuan, F., Tian, Q., and Chua, T. (2020). Adversarial Training Towards Robust Multimedia Recommender System.
"""
def __init__(
self,
name="AMR",
k=10,
k2=10,
n_epochs=50,
batch_size=100,
learning_rate=0.005,
lambda_w=0.01,
lambda_b=0.01,
lambda_e=0.0,
lambda_adv=1.0,
use_gpu=False,
trainable=True,
verbose=True,
init_params=None,
seed=None,
):
super().__init__(name=name, trainable=trainable, verbose=verbose)
self.k = k
self.k2 = k2
self.n_epochs = n_epochs
self.batch_size = batch_size
self.learning_rate = learning_rate
self.lambda_w = lambda_w
self.lambda_b = lambda_b
self.lambda_e = lambda_e
self.lambda_adv = lambda_adv
self.use_gpu = use_gpu
self.seed = seed
# Init params if provided
self.init_params = {} if init_params is None else init_params
self.gamma_user = self.init_params.get("Gu", None)
self.gamma_item = self.init_params.get("Gi", None)
self.emb_matrix = self.init_params.get("E", None)
def _init(self, n_users, n_items, features):
rng = get_rng(self.seed)
if self.gamma_user is None:
self.gamma_user = xavier_uniform((n_users, self.k), rng)
if self.gamma_item is None:
self.gamma_item = xavier_uniform((n_items, self.k), rng)
if self.emb_matrix is None:
self.emb_matrix = xavier_uniform((features.shape[1], self.k), rng)
# pre-computed for faster evaluation
self.theta_item = np.matmul(features, self.emb_matrix)
[docs]
def fit(self, train_set, val_set=None):
"""Fit the model to observations.
Parameters
----------
train_set: :obj:`cornac.data.Dataset`, required
User-Item preference data as well as additional modalities.
val_set: :obj:`cornac.data.Dataset`, optional, default: None
User-Item preference data for model selection purposes (e.g., early stopping).
Returns
-------
self : object
"""
Recommender.fit(self, train_set, val_set)
if train_set.item_image is None:
raise CornacException("item_image modality is required but None.")
# Item visual feature from CNN
train_features = train_set.item_image.features[: self.total_items]
train_features = train_features.astype(np.float32)
self._init(
n_users=self.total_users, n_items=self.total_items, features=train_features
)
if self.trainable:
self._fit_torch(train_set, train_features)
return self
def _fit_torch(self, train_set, train_features):
import torch
def _l2_loss(*tensors):
l2_loss = 0
for tensor in tensors:
l2_loss += tensor.pow(2).sum()
return l2_loss / 2
def _inner(a, b):
return (a * b).sum(dim=1)
dtype = torch.float
device = (
torch.device("cuda:0")
if (self.use_gpu and torch.cuda.is_available())
else torch.device("cpu")
)
# set requireds_grad=True to get the adversarial gradient
# if F is not put into the optimization list of parameters
# it won't be updated
F = torch.tensor(train_features, device=device, dtype=dtype, requires_grad=True)
# Learned parameters
Gu = torch.tensor(
self.gamma_user, device=device, dtype=dtype, requires_grad=True
)
Gi = torch.tensor(
self.gamma_item, device=device, dtype=dtype, requires_grad=True
)
E = torch.tensor(
self.emb_matrix, device=device, dtype=dtype, requires_grad=True
)
optimizer = torch.optim.Adam([Gu, Gi, E], lr=self.learning_rate)
for epoch in range(1, self.n_epochs + 1):
sum_loss = 0.0
count = 0
progress_bar = tqdm(
total=train_set.num_batches(self.batch_size),
desc="Epoch {}/{}".format(epoch, self.n_epochs),
disable=not self.verbose,
)
for batch_u, batch_i, batch_j in train_set.uij_iter(
self.batch_size, shuffle=True
):
gamma_u = Gu[batch_u]
gamma_i = Gi[batch_i]
gamma_j = Gi[batch_j]
feat_i = F[batch_i]
feat_j = F[batch_j]
gamma_diff = gamma_i - gamma_j
feat_diff = feat_i - feat_j
Xuij = _inner(gamma_u, gamma_diff) + _inner(gamma_u, feat_diff.mm(E))
log_likelihood = torch.nn.functional.logsigmoid(Xuij).sum()
# adversarial part
feat_i.retain_grad()
feat_j.retain_grad()
log_likelihood.backward(retain_graph=True)
feat_i_delta = feat_i.grad
feat_j_delta = feat_j.grad
adv_feat_diff = feat_diff + (feat_i_delta - feat_j_delta)
adv_Xuij = _inner(gamma_u, gamma_diff) + _inner(
gamma_u, adv_feat_diff.mm(E)
)
adv_log_likelihood = torch.nn.functional.logsigmoid(adv_Xuij).sum()
reg = (
_l2_loss(gamma_u, gamma_i, gamma_j) * self.lambda_w
+ _l2_loss(E) * self.lambda_e
)
loss = -log_likelihood - self.lambda_adv * adv_log_likelihood + reg
optimizer.zero_grad()
loss.backward()
optimizer.step()
sum_loss += loss.data.item()
count += len(batch_u)
if count % (self.batch_size * 10) == 0:
progress_bar.set_postfix(loss=(sum_loss / count))
progress_bar.update(1)
progress_bar.close()
print("Optimization finished!")
self.gamma_user = Gu.data.cpu().numpy()
self.gamma_item = Gi.data.cpu().numpy()
self.emb_matrix = E.data.cpu().numpy()
# pre-computed for faster evaluation
self.theta_item = F.mm(E).data.cpu().numpy()
[docs]
def score(self, user_idx, item_idx=None):
"""Predict the scores/ratings of a user for an item.
Parameters
----------
user_idx: int, required
The index of the user for whom to perform score prediction.
item_idx: int, optional, default: None
The index of the item for which to perform score prediction.
If None, scores for all known items will be returned.
Returns
-------
res : A scalar or a Numpy array
Relative scores that the user gives to the item or to all known items
"""
if item_idx is None:
known_item_scores = np.zeros(self.gamma_item.shape[0], dtype=np.float32)
fast_dot(self.gamma_user[user_idx], self.gamma_item, known_item_scores)
fast_dot(self.gamma_user[user_idx], self.theta_item, known_item_scores)
return known_item_scores
else:
item_score = np.dot(self.gamma_item[item_idx], self.gamma_user[user_idx])
item_score += np.dot(self.theta_item[item_idx], self.gamma_user[user_idx])
return item_score
[docs]
def get_vector_measure(self):
"""Getting a valid choice of vector measurement in ANNMixin._measures.
Returns
-------
measure: MEASURE_DOT
Dot product aka. inner product
"""
return MEASURE_DOT
[docs]
def get_user_vectors(self):
"""Getting a matrix of user vectors serving as query for ANN search.
Returns
-------
out: numpy.array
Matrix of user vectors for all users available in the model.
"""
user_vectors = np.concatenate((self.gamma_user, self.gamma_user), axis=1)
return user_vectors
[docs]
def get_item_vectors(self):
"""Getting a matrix of item vectors used for building the index for ANN search.
Returns
-------
out: numpy.array
Matrix of item vectors for all items available in the model.
"""
item_vectors = np.concatenate((self.gamma_item, self.theta_item), axis=1)
return item_vectors
