Skip to content

Latest commit

 

History

History
156 lines (145 loc) · 5.83 KB

0620_NGCF.md

File metadata and controls

156 lines (145 loc) · 5.83 KB
Raw

NGCF

./run_ngcf.py

50 modelname = "NGCF" + \
51   "_bs_" + str(batch_size) + \
52   "_nemb_" + str(emb_dim) + \
53   "_layers_" + str(layers) + \
54   "_nodedr_" + str(node_dropout) + \
55   "_messdr_" + str(mess_dropout) + \
56   "_reg_" + str(reg) + \
57  "_lr_" + str(lr)

Create NGCF Model

Set values for early stopping

78 cur_best_loss, stopping_step, should_step = le3, 0, False
79 today = datetime.now()
  • cur_best_loss : ?
  • stopping_step : ?
  • should_stop : ?

Training

92 for epoch in range(args.n_epochs):
93    t1 = time()
94    loss = train(model, data_generator, optimizer) # ./utils/helper_functions.py
95    training_time = time() - t1
96    print("Epoch: {}/{}, Training Time: {:.2f}s, Loss: {:.4f}".
97        format(epoch, args.n_epochs, training_time, loss))

WAIT! - What is 'epoch'?

An epoch in machine learning means 'one complete pass of the training dataset through the algorithm'. This epochs number is an important hyperparameter for the algorithm. It specifies the number of epochs or complete passes of the entire training dataset passing through the training or learning process of the algorithm. With each epoch, the dataset’s internal model parameters are updated. Hence, a 1 batch epoch is called the batch gradient descent learning algorithm. Normally the batch size of an epoch is 1 or more and is always an integer value in what is epoch number. <Reference: Epoch in Machine Learning: A Simple Introduction(2021)>

./utils/load_data.py - def __init__

Manufacture Input Data(ratings.csv to ratings.txt)

22  df = pd.read_csv(path + '/ratings.csv', header=None)
23  f = open(path + 'ratings.txt', 'w')
24  temp = df[0][0]
25  f.write(str(int(temp)))
26  f.write(' ')
27  for i in range(len(df)):
28    if(df[0][i] != temp):
29      f.write('\n')
30      f.write(str(int(df[0][i])))
31      f.write(' ')
32      temp = df[0][i]
33    f.write(str(int(df[1][i])))
34    f.write(' ')
35  f.close()
36  train_file = path + '/ratings.txt'
37  test_file = '...' + './test.txt'
  • Origin Form - ratings.csv
userId,movieId,rating,timestamp
  • Transformed Form - ratings.txt
userId movieId1 movieId2 movieId3 ...

Search train_file for max userId/itemId(movieId)

47  with open(train_file) as f:
48    for l in f.readlines():
49      if len(l) > 0:
50        items = []
51        l = l.strip('\n').split(' ') # l = [userId, movieId1, movieId2, ...]
52        for i in l[1:]: # movieIds
53          try:
54            items.append(int(i)) # Append 'movieIds' to list 'items'
55          except Exception: 
56            continue
58        uid = int(l[0]) # 'userId' is equal to l[0]
59        self.exist_users.append(uid) # Append 'userId' to list 'exist_users'
60        self.n_items = max(self.n_items, max(items)) # Allocate a maximum number of items to 'n_items'
61        self.n_users = max(self.n_users, uid) # Allocate a maximum number of users to 'n_users'
62        self.n_train += len(items)

Create interactions/ratings matrix 'R' # dok = dictionary of keys

90  self.R_train = sp.dok_matrix((self.n_users, self.n_items), dtype = np.float32)

Create (self.n_users)x(self.n_items) matrix

Set value of a matrix 'R_train'

94  with open(train_file) as f_train:
95    with open(test_file) as f_test:
96      for l in f_train.readlines():
97        items = []
98        if len(l) == 0: break
99        l = l.strip('\n').split(' ') # l = [userId, movieId1, movieId2, movieId3, ...]
100       for i in l:
101         try:
102           items.append(int(i)) # Append movieIds to list 'items'
103         except:
104           continue
105       uid, train_itemss = items[0], items[1:] # train_itemss = [movieId1, movieId2, ...]
107       for i in train_itemss: # for-loop : movieIds 
108         self.R_train[uid, i] = 1. # Set 1 as a value of uid-th row and i-th col of R_train
109       self.train_items[uid] = train_itemss

./utils/helper_functions.py

37  def train(model, data_generator, optimizer):
38  """
39  Train the model PyTorch Style
41  Arguments:
43  model: PyTorch Model
44  data_generator: Data Object
45  optimizer: PyTorch Optimizer
46  """
47  model.train()
48  n_batch = data_generator.n_train // data_generator.batch_size + 1
49  running_loss = 0
50  for _ in range(n_batch):
51    u, i, j = data_generator.sample()
52    optimizer.zero_grad()
53    loss = model(u,i,j)
54    loss.backward()
55    optimizer.step()
56    running_loss += loss.item()
57  return running_loss

WAIT! - What is 'batch'?

The batch size is a hyperparameter that defines the number of samples to work through before updating the internal model parameters. <Reference: Difference Between a Batch and Epoch in a Neural Network>

WAIT! - What is 'loss'?

Typically, with neural networks, we seek to minimize the error. As such, the objective function is often referred to as a cost function or a loss function and the value calculated by the loss function is referred to as simply “loss.” <Reference: Loss and Loss Functions for Training Deep Learning Neural Networks>