def fit(self, trainset):
AlgoBase.fit(self, trainset)
numUsers = trainset.n_users
numItems = trainset.n_items
trainingMatrix = np.zeros([numUsers, numItems, 10], dtype = np.float32)
for (uid, iid, rating) in trainset.all_ratings():
adjustedRating = int(float(rating)*2.0) - 1
trainingMatrix[int(uid), int(iid), adjustedRating] = 1
trainingMatrix = np.reshape(trainingMatrix,[trainingMatrix.shape[0], - 1])
rbm = RBM(trainingMatrix.shape[1], hiddenDimensions = self.hiddenDim, learningRate = self.learningRate, batchSize = self.batchSize)
rbm.Train(trainingMatrix)
self.predictedRatings = np.zeros([numUsers, numItems], dtype = np.float32)
for uiid in range(trainset.n_users):
if(uiid % 50 == 0):
print("Procissing user ", uiid)
recs = rbm.GetRecommendations([trainingMatrix[uiid]])
recs = np.reshape(recs, [numItems, 10])
for itemID, rec in enumerate(recs):
normalized = self.softmax(rec)
rating = np.average(np.arange(10), weights = normalized)
self.predictedRatings[uiid,itemID] = (rating + 1)* 0.5
return self
def fit(self, trainset):
AlgoBase.fit(self, trainset)
numUsers = trainset.n_users
numItems = trainset.n_items
trainingMatrix = np.zeros([numUsers, numItems, 10], dtype=np.float32)
for (uid, iid, rating) in trainset.all_ratings():
adjustedRating = int(float(rating)*2.0) - 1
trainingMatrix[int(uid), int(iid), adjustedRating] = 1
# Flatten to a 2D array, with nodes for each possible rating type on each possible item, for every user.
trainingMatrix = np.reshape(trainingMatrix, [trainingMatrix.shape[0], -1])
# Create an RBM with (num items * rating values) visible nodes
rbm = RBM(trainingMatrix.shape[1], hiddenDimensions=self.hiddenDim, learningRate=self.learningRate, batchSize=self.batchSize, epochs=self.epochs)
rbm.Train(trainingMatrix)
self.predictedRatings = np.zeros([numUsers, numItems], dtype=np.float32)
for uiid in range(trainset.n_users):
if (uiid % 50 == 0):
print("Processing user ", uiid)
recs = rbm.GetRecommendations([trainingMatrix[uiid]])
recs = np.reshape(recs, [numItems, 10])
for itemID, rec in enumerate(recs):
# The obvious thing would be to just take the rating with the highest score:
#rating = rec.argmax()
# ... but this just leads to a huge multi-way tie for 5-star predictions.
# The paper suggests performing normalization over K values to get probabilities
# and take the expectation as your prediction, so we'll do that instead:
normalized = self.softmax(rec)
rating = np.average(np.arange(10), weights=normalized)
self.predictedRatings[uiid, itemID] = (rating + 1) * 0.5
return self
def fit(self, trainset):
AlgoBase.fit(self, trainset)
numUsers = trainset.n_users
numItems = trainset.n_items
trainingMatrix = np.zeros([numUsers, numItems, 10], dtype=np.float32)
for (uid, iid, rating) in trainset.all_ratings():
adjustedRating = int(float(rating) * 2.0) - 1
trainingMatrix[int(uid), int(iid), adjustedRating] = 1
#ユーザーごとにアイテムの可能な評価スコアに対して、ノードに2次元配列で平坦化する
trainingMatrix = np.reshape(trainingMatrix,
[trainingMatrix.shape[0], -1])
#可視化できるノードにアイテムの整数値と評価値を掛け算して、RBMを新たに構築する
rbm = RBM(trainingMatrix.shape[1],
hiddenDimensions=self.hiddenDim,
learningRate=self.learningRate,
batchSize=self.batchSize,
epochs=self.epochs)
rbm.Train(trainingMatrix)
self.predictedRatings = np.zeros([numUsers, numItems],
dtype=np.float32)
for uiid in range(trainset.n_users):
if (uiid % 50 == 0):
print("Processing user ", uiid)
recs = rbm.GetRecommendations([trainingMatrix[uiid]])
recs = np.reshape(recs, [numItems, 10])
for itemID, rec in enumerate(recs):
#評価値に最も高いスコアを付与して明示する
#rating = rec.argmax()
#5つ星に複数の評価値を導入する
#考えうるk値の正規化を行う
#予測する際に、期待値を取る
normalized = self.softmax(rec)
rating = np.average(np.arange(10), weights=normalized)
self.predictedRatings[uiid, itemID] = (rating + 1) * 0.5
return self
def main(args):
train, test = read_data(args.in_dir)
train, val = split_data(train)
print('\n\nTrain: ', train.shape)
print('Val: ', val.shape)
print('Test: ', test.shape)
# RBM object
rbm = RBM(args.num_hidden, val.shape[1], args.lr, args.n, args.batch_size,
args.epochs)
# Train RBM
train_loss, val_loss = rbm.Train(train, val)
# Create output dir if it doesn't exist
if not os.path.exists(args.out_dir):
os.makedirs(args.out_dir)
# Plot error
plot_error(train_loss, val_loss, args.out_dir)
# Performance on Test set
error_test = rbm.reconstruction_error(test.T)
print("\n\n\nReconstruction error...\n")
print("Train : ", train_loss[-1])
print("Val : ", val_loss[-1])
print("Test : ", error_test)
# For viewing reconstruction
reconstruct_images(rbm, test, (args.image_height, args.image_width),
args.out_dir)
# Saving the model learned weights
pickle.dump([rbm.W, rbm.b_h, rbm.b_v],
open(args.out_dir + '\\weights.pkl', 'wb'))
print(f"\n\nRBM weights saved in {args.out_dir}\\weights.pkl")
# Load up common data set for the recommender algorithms
ml = LoadData()
#to train the model
#trainset= evaluationData.build_full_trainset()
numUsers = ml.getNumberUsers()
numItems = ml.getNumberItems()
trainingMatrix = np.zeros([numUsers + 1, numItems + 1, 10], dtype=np.float32)
ratings = ml.all_ratings()
for i in ratings:
uid = i[0]
iid = i[1]
rate = i[2]
adjustedRating = int(float(rate) * 2.0) - 1
trainingMatrix[int(uid), int(iid), adjustedRating] = 1
# Flatten to a 2D array, with nodes for each possible rating type on each possible item, for every user.
trainingMatrix = np.reshape(trainingMatrix, [trainingMatrix.shape[0], -1])
epochs = 22
hiddenDim = 100
learningRate = 0.001
batchSize = 100
# Create an RBM with (num items * rating values) visible nodes
rbm_recommend = RBM(trainingMatrix.shape[1],
hiddenDimensions=hiddenDim,
learningRate=learningRate,
batchSize=batchSize,
epochs=epochs)
rbm_recommend.Train(trainingMatrix)
