class GRU4REC:
def __init__(self,
input_size,
hidden_size,
output_size,
num_layers=1,
optimizer_type='Adagrad',
lr=.01,
weight_decay=0,
momentum=0,
eps=1e-6,
loss_type='TOP1',
clip_grad=-1,
dropout_input=.0,
dropout_hidden=.5,
batch_size=50,
use_cuda=True,
time_sort=False,
pretrained=None):
""" The GRU4REC model
Args:
input_size (int): dimension of the gru input variables
hidden_size (int): dimension of the gru hidden units
output_size (int): dimension of the gru output variables
num_layers (int): the number of layers in the GRU
optimizer_type (str): optimizer type for GRU weights
lr (float): learning rate for the optimizer
weight_decay (float): weight decay for the optimizer
momentum (float): momentum for the optimizer
eps (float): eps for the optimizer
loss_type (str): type of the loss function to use
clip_grad (float): clip the gradient norm at clip_grad. No clipping if clip_grad = -1
dropout_input (float): dropout probability for the input layer
dropout_hidden (float): dropout probability for the hidden layer
batch_size (int): mini-batch size
use_cuda (bool): whether you want to use cuda or not
time_sort (bool): whether to ensure the the order of sessions is chronological (default: False)
pretrained (modules.layer.GRU): pretrained GRU layer, if it exists (default: None)
"""
# Initialize the GRU Layer
self.input_size = input_size
self.hidden_size = hidden_size
self.output_size = output_size
self.batch_size = batch_size
self.use_cuda = use_cuda
if pretrained is None:
self.gru = GRU(input_size,
hidden_size,
output_size,
num_layers,
dropout_input=dropout_input,
dropout_hidden=dropout_hidden,
use_cuda=use_cuda,
batch_size=batch_size)
else:
self.gru = pretrained
# Initialize the optimizer
self.optimizer_type = optimizer_type
self.weight_decay = weight_decay
self.momentum = momentum
self.lr = lr
self.eps = eps
self.optimizer = Optimizer(self.gru.parameters(),
optimizer_type=optimizer_type,
lr=lr,
weight_decay=weight_decay,
momentum=momentum,
eps=eps)
# Initialize the loss function
self.loss_type = loss_type
self.loss_fn = LossFunction(loss_type, use_cuda)
# gradient clipping(optional)
self.clip_grad = clip_grad
# etc
self.time_sort = time_sort
def train(self, n_epochs=10, save_dir='./models', model_name='GRU4REC'):
"""
Train the GRU4REC model on a pandas dataframe for several training epochs,
and store the intermediate models to the user-specified directory.
Args:
n_epochs (int): the number of training epochs to run
save_dir (str): the path to save the intermediate trained models
model_name (str): name of the model
"""
print(f'Model Name:{model_name}')
# Time the training process
start_time = time.time()
for epoch in range(n_epochs):
loss = self.run_epoch()
end_time = time.time()
wall_clock = (end_time - start_time) / 60
print(
f'Epoch:{epoch+1:2d}/Loss:{loss:0.3f}/TrainingTime:{wall_clock:0.3f}(min)'
)
start_time = time.time()
# Store the intermediate model
save_dir = Path(save_dir)
if not save_dir.exists(): save_dir.mkdir()
model_fname = f'{model_name}_{self.loss_type}_{self.optimizer_type}_{self.lr}_epoch{epoch+1:d}'
torch.save(self.gru.state_dict(), save_dir / model_fname)
def run_epoch(self):
""" Run a single training epoch """
self.gru.train()
# initialize
mb_losses = []
optimizer = self.optimizer
hidden = self.gru.init_hidden().data
# Start the training loop
loader = SessionDataLoader(df=self.df_train,
hidden=hidden,
session_key=self.session_key,
item_key=self.item_key,
time_key=self.time_key,
batch_size=self.batch_size,
training=self.gru.training,
time_sort=self.time_sort)
for input, target, hidden in loader.generate_batch():
if self.use_cuda:
input = input.cuda()
target = target.cuda()
# Go through the GRU layer
logit, hidden = self.gru(input, target, hidden)
######################## IMPORTANT #########################
# update the hidden state for the dataloader from the outside
#############################################################
loader.update_hidden(hidden.data)
# Calculate the mini-batch loss
mb_loss = self.loss_fn(logit)
mb_losses.append(mb_loss.data[0])
# flush the gradient b/f backprop
optimizer.zero_grad()
# Backprop
mb_loss.backward()
# Gradient Clipping(Optional)
if self.clip_grad != -1:
for p in self.gru.parameters():
p.grad.data.clamp_(max=self.clip_grad)
# Mini-batch GD
optimizer.step()
avg_epoch_loss = np.mean(mb_losses)
return avg_epoch_loss
def test(self, k=20, batch_size=50):
""" Model evaluation
Args:
k (int): the length of the recommendation list
batch_size (int): testing batch_size
Returns:
avg_recall: mean of the Recall@K over the session-parallel mini-batches
avg_mrr: mean of the MRR@K over the session-parallel mini-batches
"""
# set the gru layer into inference mode
self.gru.eval()
recalls = []
mrrs = []
hidden = self.gru.init_hidden().data
# Start the testing loop
loader = SessionDataLoader(df=self.df_test,
hidden=hidden,
session_key=self.session_key,
item_key=self.item_key,
time_key=self.time_key,
batch_size=batch_size,
training=self.gru.training,
time_sort=self.time_sort)
for input, target, hidden in loader.generate_batch():
if self.use_cuda:
input = input.cuda()
target = target.cuda()
# forward propagation
logit, hidden = self.gru(input, target, hidden)
# update the hidden state for the dataloader
loader.update_hidden(hidden.data)
# Evaluate the results
recall, mrr = E.evaluate(logit, target, k)
recalls.append(recall)
mrrs.append(mrr)
avg_recall = np.mean(recalls)
avg_mrr = np.mean(mrrs)
# reset the gru to a training mode
self.gru.train()
return avg_recall, avg_mrr
def init_data(self, df_train, df_test, session_key, time_key, item_key):
""" Initialize the training & test data.
The training/test set, session/time/item keys will be stored for later reuse.
Args:
df_train (pd.DataFrame): training set required to retrieve the training item indices.
df_test (pd.DataFrame): test set
session_key (str): session ID
time_key (str): time ID
item_key (str): item ID
"""
# Specify the identifiers
self.session_key = session_key
self.time_key = time_key
self.item_key = item_key
# Initialize the dataframes into adequate forms
self.df_train = self.init_df(df_train, session_key, time_key, item_key)
self.df_test = self.init_df(df_test,
session_key,
time_key,
item_key,
item_ids=df_train[item_key].unique())
@staticmethod
def init_df(df, session_key, time_key, item_key, item_ids=None):
""" Initialize the dataframe.
Involves the following steps:
1) Add new item indices to the dataframe
2) Sort the df
Args:
session_key: session identifier
time_key: timestamp
item_key: item identifier
item_ids: unique item ids. Should be `None` if the df is a training set, and should include the
ids for the items included in the training set if the df is a test set.
"""
# add item index column named "item_idx" to the df
if item_ids is None:
item_ids = df[item_key].unique() # unique item ids
item2idx = pd.Series(data=np.arange(len(item_ids)), index=item_ids)
df = pd.merge(df,
pd.DataFrame({
item_key: item_ids,
'item_idx': item2idx[item_ids].values
}),
on=item_key,
how='inner')
"""
Sort the df by time, and then by session ID. That is, df is sorted by session ID and
clicks within a session are next to each other, where the clicks within a session are time-ordered.
"""
df.sort_values([session_key, time_key], inplace=True)
return df
class GRU4REC:
def __init__(self, input_size, hidden_size, output_size, num_layers=1,
optimizer_type='Adagrad', lr=.01, weight_decay=0,
momentum=0, eps=1e-6, loss_type='TOP1',
clip_grad=-1, p_dropout_input=.0, p_dropout_hidden=.5,
batch_size=50, use_cuda=True, time_sort=False, pretrained=None):
""" The GRU4REC model
Args:
input_size (int): dimension of the gru input variables
hidden_size (int): dimension of the gru hidden units
output_size (int): dimension of the gru output variables
num_layers (int): the number of layers in the GRU
optimizer_type (str): optimizer type for GRU weights
lr (float): learning rate for the optimizer
weight_decay (float): weight decay for the optimizer
momentum (float): momentum for the optimizer
eps (float): eps for the optimizer
loss_type (str): type of the loss function to use
clip_grad (float): clip the gradient norm at clip_grad. No clipping if clip_grad = -1
p_dropout_input (float): dropout probability for the input layer
p_dropout_hidden (float): dropout probability for the hidden layer
batch_size (int): mini-batch size
use_cuda (bool): whether you want to use cuda or not
time_sort (bool): whether to ensure the the order of sessions is chronological (default: False)
pretrained (modules.layer.GRU): pretrained GRU layer, if it exists (default: None)
"""
# Initialize the GRU Layer
self.input_size = input_size
self.hidden_size = hidden_size
self.output_size = output_size
self.batch_size = batch_size
self.use_cuda = use_cuda
self.device = torch.device('cuda' if use_cuda else 'cpu')
if pretrained is None:
self.gru = GRU(input_size, hidden_size, output_size, num_layers,
p_dropout_input=p_dropout_input,
p_dropout_hidden=p_dropout_hidden,
batch_size=batch_size,
use_cuda=use_cuda)
else:
self.gru = pretrained
# Initialize the optimizer
self.optimizer_type = optimizer_type
self.weight_decay = weight_decay
self.momentum = momentum
self.lr = lr
self.eps = eps
self.optimizer = Optimizer(self.gru.parameters(),
optimizer_type=optimizer_type,
lr=lr,
weight_decay=weight_decay,
momentum=momentum,
eps=eps)
# Initialize the loss function
self.loss_type = loss_type
self.loss_fn = LossFunction(loss_type)
# gradient clipping(optional)
self.clip_grad = clip_grad
# etc
self.time_sort = time_sort
def run_epoch(self, dataset, k=20, training=True):
""" Run a single training epoch """
start_time = time.time()
# initialize
losses = []
recalls = []
mrrs = []
optimizer = self.optimizer
hidden = self.gru.init_hidden()
if not training:
self.gru.eval()
device = self.device
def reset_hidden(hidden, mask):
"""Helper function that resets hidden state when some sessions terminate"""
if len(mask) != 0:
hidden[:, mask, :] = 0
return hidden
# Start the training loop
loader = SessionDataLoader(dataset, batch_size=self.batch_size)
for input, target, mask in loader:
input = input.to(device)
target = target.to(device)
# device_desktop.to(device)
# reset the hidden states if some sessions have just terminated
hidden = reset_hidden(hidden, mask).detach()
# Go through the GRU layer
logit, hidden = self.gru(input, target, hidden)
# Output sampling
logit_sampled = logit[:, target.view(-1)]
# Calculate the mini-batch loss
loss = self.loss_fn(logit_sampled)
with torch.no_grad():
recall, mrr = E.evaluate(logit, target, k)
losses.append(loss.item())
recalls.append(recall)
mrrs.append(mrr)
# Gradient Clipping(Optional)
if self.clip_grad != -1:
for p in self.gru.parameters():
p.grad.data.clamp_(max=self.clip_grad)
# Mini-batch GD
if training:
# Backprop
loss.backward()
optimizer.step()
optimizer.zero_grad() # flush the gradient after the optimization
results = dict()
results['loss'] = np.mean(losses)
results['recall'] = np.mean(recalls)
results['mrr'] = np.mean(mrrs)
end_time = time.time()
results['time'] = (end_time - start_time) / 60
if not training:
self.gru.train()
return results
def train(self, dataset, k=20, n_epochs=10, save_dir='./models', save=True, model_name='GRU4REC'):
"""
Train the GRU4REC model on a pandas dataframe for several training epochs,
and store the intermediate models to the user-specified directory.
Args:
n_epochs (int): the number of training epochs to run
save_dir (str): the path to save the intermediate trained models
model_name (str): name of the model
"""
print(f'Training {model_name}...')
results = self.run_epoch(dataset, k=k, training=True)
results = [f'{k}:{v:.3f}' for k, v in results.items()]
tmp0=float(results[1].split(":")[1])
res0=results
print(f'epoch:{1:2d}/{"/".join(results)}')
for epoch in range(n_epochs-1):
results = self.run_epoch(dataset, k=k, training=True)
results = [f'{k}:{v:.3f}' for k, v in results.items()]
tmp=float(results[1].split(":")[1])
if tmp>tmp0:
tmp0=tmp
print(f'epoch:{epoch+1:2d}/{"/".join(results)}')
else:
print("The end of training~")
break
# Store the intermediate model
if save:
save_dir = Path(save_dir)
if not save_dir.exists(): save_dir.mkdir()
model_fname = f'{model_name}_{self.loss_type}_{self.optimizer_type}_{self.lr}_epoch{epoch+1:d}'
torch.save(self.gru.state_dict(), save_dir/model_fname)
def test(self, dataset, k=20):
""" Model evaluation
Args:
k (int): the length of the recommendation list
Returns:
avg_loss: mean of the losses over the session-parallel minibatches
avg_recall: mean of the Recall@K over the session-parallel mini-batches
avg_mrr: mean of the MRR@K over the session-parallel mini-batches
wall_clock: time took for testing
"""
results = self.run_epoch(dataset, k=k, training=False)
results = [f'{k}:{v:.3f}' for k, v in results.items()]
print(results)
class GRU4REC:
def __init__(self, input_size, hidden_size, output_size, num_layers=1,
optimizer_type='Adagrad', lr=.05, weight_decay=0,
momentum=0, eps=1e-6, loss_type='TOP1',
clip_grad=-1, dropout_input=.0, dropout_hidden=.5,
batch_size=50, use_cuda=True, time_sort=False, pretrained=None):
""" The GRU4REC model
Args:
input_size (int): dimension of the gru input variables
hidden_size (int): dimension of the gru hidden units
output_size (int): dimension of the gru output variables
num_layers (int): the number of layers in the GRU
optimizer_type (str): optimizer type for GRU weights
lr (float): learning rate for the optimizer
weight_decay (float): weight decay for the optimizer
momentum (float): momentum for the optimizer
eps (float): eps for the optimizer
loss_type (str): type of the loss function to use
clip_grad (float): clip the gradient norm at clip_grad. No clipping if clip_grad = -1
dropout_input (float): dropout probability for the input layer
dropout_hidden (float): dropout probability for the hidden layer
batch_size (int): mini-batch size
use_cuda (bool): whether you want to use cuda or not
time_sort (bool): whether to ensure the the order of sessions is chronological (default: False)
pretrained (modules.layer.GRU): pretrained GRU layer, if it exists (default: None)
"""
# Initialize the GRU Layer
self.input_size = input_size
self.hidden_size = hidden_size
self.output_size = output_size
self.batch_size = batch_size
self.use_cuda = use_cuda
if pretrained is None:
self.gru = GRU(input_size, hidden_size, output_size, num_layers,
dropout_input=dropout_input,
dropout_hidden=dropout_hidden,
use_cuda=use_cuda,
batch_size=batch_size)
else:
self.gru = pretrained
# Initialize the optimizer
self.optimizer_type = optimizer_type
self.weight_decay = weight_decay
self.momentum = momentum
self.lr = lr
self.eps = eps
self.optimizer = Optimizer(self.gru.parameters(),
optimizer_type=optimizer_type,
lr=lr,
weight_decay=weight_decay,
momentum=momentum,
eps=eps)
# Initialize the loss function
self.loss_type = loss_type
self.loss_fn = LossFunction(loss_type, use_cuda)
# gradient clipping(optional)
self.clip_grad = clip_grad
# etc
self.time_sort = time_sort
def train(self, df, session_key, time_key, item_key, n_epochs=10, save_dir='./models', model_name='GRU4REC'):
"""
Train the GRU4REC model on a pandas dataframe for several training epochs,
and store the intermediate models to the user-specified directory.
Args:
df (pd.DataFrame): training dataset
session_key (str): session ID
time_key (str): time ID
item_key (str): item ID
n_epochs (int): the number of training epochs to run
save_dir (str): the path to save the intermediate trained models
model_name (str): name of the model
"""
df, click_offsets, session_idx_arr = GRU4REC.init_data(df, session_key, time_key, item_key,
time_sort=self.time_sort)
# Time the training process
start_time = time.time()
for epoch in range(n_epochs):
loss = self.run_epoch(df, click_offsets, session_idx_arr)
end_time = time.time()
wall_clock = (end_time - start_time) / 60
print(f'Epoch:{epoch+1:2d}/Loss:{loss:0.3f}/TrainingTime:{wall_clock:0.3f}(min)')
start_time = time.time()
# Store the intermediate model
save_dir = Path(save_dir)
if not save_dir.exists(): save_dir.mkdir()
model_fname = f'{model_name}_{self.loss_type}_{self.optimizer_type}_{self.lr}_epoch{epoch+1:d}'
torch.save(self.gru.state_dict(), save_dir/model_fname)
def run_epoch(self, df, click_offsets, session_idx_arr):
""" Runs a single training epoch """
mb_losses = []
# initializations
iters = np.arange(self.batch_size)
maxiter = iters.max()
start = click_offsets[session_idx_arr[iters]]
end = click_offsets[session_idx_arr[iters]+1]
# initialize the hidden state
hidden = self.gru.init_hidden().data
optimizer = self.optimizer
# Start the training loop
finished = False
while not finished:
minlen = (end-start).min()
# Item indices(for embedding) for clicks where the first sessions start
idx_target = df.item_idx.values[start]
for i in range(minlen - 1):
# Build inputs, targets, and hidden states
idx_input = idx_target
idx_target = df.item_idx.values[start + i + 1]
input = torch.LongTensor(idx_input) #(B) At first, input is a Tensor
target = Variable(torch.LongTensor(idx_target)) #(B)
if self.use_cuda:
input = input.cuda()
target = target.cuda()
# Now, convert into an embedded Variable
embedded = self.gru.emb(input)
hidden = Variable(hidden)
# Go through the GRU layer
logit, hidden = self.gru(embedded, target, hidden)
# Calculate the mini-batch loss
mb_loss = self.loss_fn(logit)
mb_losses.append(mb_loss.data[0])
# flush the gradient b/f backprop
optimizer.zero_grad()
# Backprop
mb_loss.backward()
# Gradient Clipping(Optional)
if self.clip_grad != -1:
for p in self.gru.parameters():
p.grad.data.clamp_(max=self.clip_grad)
# Mini-batch GD
optimizer.step()
# Detach the hidden state for later reuse
hidden = hidden.data
# click indices where a particular session meets second-to-last element
start = start + (minlen - 1)
# figure out how many sessions should terminate
mask = np.arange(len(iters))[(end-start)<=1]
for idx in mask:
maxiter += 1
if maxiter >= len(click_offsets)-1:
finished = True
break
# update the next starting/ending point
iters[idx] = maxiter
start[idx] = click_offsets[session_idx_arr[maxiter]]
end[idx] = click_offsets[session_idx_arr[maxiter]+1]
# reset the rnn hidden state to zero after transition
if len(mask) != 0:
hidden[:, mask, :] = 0
avg_epoch_loss = np.mean(mb_losses)
return avg_epoch_loss
def predict(self, input, target, hidden):
""" Forward propagation for testing
Args:
input (B,C): torch.LongTensor. The one-hot embedding for the item indices
target (B,): a Variable that stores the indices for the next items
hidden: previous hidden state
Returns:
logits (B,C): logits for the next items
hidden: next hidden state
"""
# convert the item indices into embeddings
embedded = self.gru.emb(input, volatile=True)
hidden = Variable(hidden, volatile=True)
# forward propagation
logits, hidden = self.gru(embedded, target, hidden)
return logits, hidden
def test(self, df_train, df_test, session_key, time_key, item_key,
k=20, batch_size=50):
""" Model evaluation
Args:
df_train (pd.DataFrame): training set required to retrieve the training item indices.
df_test (pd.DataFrame): test set
session_key (str): session ID
time_key (str): time ID
item_key (str): item ID
k (int): the length of the recommendation list
batch_size (int): testing batch_size
Returns:
avg_recall: mean of the Recall@K over the session-parallel mini-batches
avg_mrr: mean of the MRR@K over the session-parallel mini-batches
"""
# set the gru layer into inference mode
if self.gru.training:
self.gru.switch_mode()
recalls = []
mrrs = []
# initializations
# Build item2idx from train data.
iids = df_train[item_key].unique() # unique item ids
item2idx = pd.Series(data=np.arange(len(iids)), index=iids)
df_test = pd.merge(df_test,
pd.DataFrame({item_key: iids,
'item_idx': item2idx[iids].values}),
on=item_key,
how='inner')
# Sort the df by time, and then by session ID.
df_test.sort_values([session_key, time_key], inplace=True)
# Return the offsets of the beginning clicks of each session IDs
click_offsets = GRU4REC.get_click_offsets(df_test, session_key)
session_idx_arr = GRU4REC.order_session_idx(df_test, session_key, time_key, time_sort=self.time_sort)
iters = np.arange(batch_size)
maxiter = iters.max()
start = click_offsets[session_idx_arr[iters]]
end = click_offsets[session_idx_arr[iters]+1]
hidden = self.gru.init_hidden().data
# Start the training loop
finished = False
while not finished:
minlen = (end-start).min()
# Item indices(for embedding) for clicks where the first sessions start
idx_target = df_test.item_idx.values[start]
for i in range(minlen - 1):
# Build inputs, targets, and hidden states
idx_input = idx_target
idx_target = df_test.item_idx.values[start + i + 1]
input = torch.LongTensor(idx_input) #(B) At first, input is a Tensor
target = Variable(torch.LongTensor(idx_target), volatile=True) # (B)
if self.use_cuda:
input = input.cuda()
target = target.cuda()
logit, hidden = self.predict(input, target, hidden)
recall, mrr = evaluate(logit, target, k)
recalls.append(recall)
mrrs.append(mrr)
# Detach the hidden state for later reuse
hidden = hidden.data
# click indices where a particular session meets second-to-last element
start = start + (minlen - 1)
# see if how many sessions should terminate
mask = np.arange(len(iters))[(end-start)<=1]
for idx in mask:
maxiter += 1
if maxiter >= len(click_offsets)-1:
finished = True
break
# update the next starting/ending point
iters[idx] = maxiter
start[idx] = click_offsets[session_idx_arr[maxiter]]
end[idx] = click_offsets[session_idx_arr[maxiter]+1]
# reset the rnn hidden state to zero after transition
if len(mask)!= 0:
hidden[:,mask,:] = 0
avg_recall = np.mean(recalls)
avg_mrr = np.mean(mrrs)
# reset the gru to a training mode
self.gru.switch_mode()
return avg_recall, avg_mrr
@staticmethod
def init_data(df, session_key, time_key, item_key, time_sort):
"""
Initialize the data.
"""
# add item indices to the dataframe
df = GRU4REC.add_item_indices(df, item_key)
"""
Sort the df by time, and then by session ID. That is, df is sorted by session ID and
clicks within a session are next to each other, where the clicks within a session are time-ordered.
"""
df.sort_values([session_key, time_key], inplace=True)
click_offsets = GRU4REC.get_click_offsets(df, session_key)
session_idx_arr = GRU4REC.order_session_idx(df, session_key, time_key, time_sort=time_sort)
return df, click_offsets, session_idx_arr
@staticmethod
def add_item_indices(df, item_key):
"""
Adds an item index column named "item_idx" to the df.
Args:
df: pd.DataFrame to add the item indices to
Returns:
df: copy of the original df with item indices
"""
iids = df[item_key].unique() # unique item ids
item2idx = pd.Series(data=np.arange(len(iids)), index=iids)
df = pd.merge(df,
pd.DataFrame({item_key: iids,
'item_idx': item2idx[iids].values}),
on=item_key,
how='inner')
return df
@staticmethod
def get_click_offsets(df, session_key):
"""
Return the offsets of the beginning clicks of each session IDs,
where the offset is calculated against the first click of the first session ID.
"""
offsets = np.zeros(df[session_key].nunique()+1, dtype=np.int32)
# group & sort the df by session_key and get the offset values
offsets[1:] = df.groupby(session_key).size().cumsum()
return offsets
@staticmethod
def order_session_idx(df, session_key, time_key, time_sort=False):
""" Order the session indices """
if time_sort:
# starting time for each sessions, sorted by session IDs
sessions_start_time = df.groupby(session_key)[time_key].min().values
# order the session indices by session starting times
session_idx_arr = np.argsort(sessions_start_time)
else:
session_idx_arr = np.arange(df[session_key].nunique())
return session_idx_arr
class GRU4REC:
def __init__(self,
input_size,
hidden_size,
output_size,
num_layers=1,
optimizer_type='Adagrad',
lr=.01,
weight_decay=0,
momentum=0,
eps=1e-6,
loss_type='TOP1',
clip_grad=-1,
dropout_input=.0,
dropout_hidden=.5,
batch_size=50,
use_cuda=True,
time_sort=False,
pretrained=None,
n_sample=2048,
sample_alpha=0.75,
sample_store=10000000,
bpreg=1.0):
""" The GRU4REC model
Args:
input_size (int): dimension of the gru input variables
hidden_size (int): dimension of the gru hidden units
output_size (int): dimension of the gru output variables
num_layers (int): the number of layers in the GRU
optimizer_type (str): optimizer type for GRU weights
lr (float): learning rate for the optimizer
weight_decay (float): weight decay for the optimizer
momentum (float): momentum for the optimizer
eps (float): eps for the optimizer
loss_type (str): type of the loss function to use
clip_grad (float): clip the gradient norm at clip_grad. No clipping if clip_grad = -1
dropout_input (float): dropout probability for the input layer
dropout_hidden (float): dropout probability for the hidden layer
batch_size (int): mini-batch size
use_cuda (bool): whether you want to use cuda or not
time_sort (bool): whether to ensure the the order of sessions is chronological (default: False)
pretrained (modules.layer.GRU): pretrained GRU layer, if it exists (default: None)
"""
# Initialize the GRU Layer
self.input_size = input_size
self.hidden_size = hidden_size
self.output_size = output_size
self.batch_size = batch_size
self.use_cuda = use_cuda
self.device = torch.device('cuda' if use_cuda else 'cpu')
###修改###
self.n_sample = n_sample
self.sample_alpha = sample_alpha
self.sample_store = sample_store
self.bpreg = bpreg
###修改###
if pretrained is None:
self.gru = GRU(input_size,
hidden_size,
output_size,
num_layers,
dropout_input=dropout_input,
dropout_hidden=dropout_hidden,
batch_size=batch_size,
use_cuda=use_cuda)
else:
self.gru = pretrained
# Initialize the optimizer
self.optimizer_type = optimizer_type
self.weight_decay = weight_decay
self.momentum = momentum
self.lr = lr
self.eps = eps
self.optimizer = Optimizer(self.gru.parameters(),
optimizer_type=optimizer_type,
lr=lr,
weight_decay=weight_decay,
momentum=momentum,
eps=eps)
# Initialize the loss function
self.loss_type = loss_type
self.loss_fn = LossFunction(loss_type, use_cuda)
# gradient clipping(optional)
self.clip_grad = clip_grad
# etc
self.time_sort = time_sort
def generate_neg_samples(self, n_items, pop, length):
if self.sample_alpha:
sample = np.searchsorted(pop,
np.random.rand(self.n_sample * length))
else:
# 从 np.arange(n_items) 中产生一个size为n_sample * length的随机采样
sample = np.random.choice(n_items, size=self.n_sample * length)
if length > 1:
sample = sample.reshape((length, self.n_sample))
return sample
def run_epoch(self, dataset, k=20, training=True):
""" Run a single training epoch """
start_time = time.time()
# initialize
losses = []
recalls = []
mrrs = []
##增加###
zippers = []
##增加###
optimizer = self.optimizer
hidden = self.gru.init_hidden()
if not training:
self.gru.eval()
device = self.device
def reset_hidden(hidden, mask):
"""Helper function that resets hidden state when some sessions terminate"""
if len(mask) != 0:
hidden[:, mask, :] = 0
return hidden
# Start the training loop
loader = SessionDataLoader(dataset, batch_size=self.batch_size)
# 一个bach一个bach的迭代,每次迭代是一个input:tensor([ 31, 26, 27, 29, 24]);一个output:tensor([ 31, 26, 28, 17, 24])
#
if training == True:
n_items = len(dataset.items)
# sampling 增加额外负样本采样
if self.n_sample > 0:
pop = dataset.df.groupby(
'ItemId').size() # item的流行度supp,数据如下格式
# ItemId
# 214507331 1
# 214507365 1
# 将sample_alpha设置为1会导致基于流行度的采样，将其设置为0会导致均匀采样
pop = pop[
dataset.itemmap[dataset.item_key].
values].values**self.sample_alpha # item选择作为样本的概率为supp ^ sample_alpha
pop = pop.cumsum() / pop.sum()
pop[-1] = 1
if self.sample_store:
generate_length = self.sample_store // self.n_sample
if generate_length <= 1:
sample_store = 0
print('No example store was used')
else:
neg_samples = self.generate_neg_samples(
n_items, pop, generate_length)
sample_pointer = 0
else:
print('No example store was used')
for input, target, mask in loader:
input = input.to(device)
target = target.to(device)
# print(input)
# print(target)
#额外的 SAMPLING THE OUTPUT
if self.n_sample > 0 and training:
if self.sample_store:
if sample_pointer == generate_length:
neg_samples = self.generate_neg_samples(
n_items, pop, generate_length)
sample_pointer = 0
sample = neg_samples[sample_pointer]
sample_pointer += 1
else:
sample = self.generate_neg_samples(pop, 1)
y = torch.LongTensor(np.hstack([target, sample]))
else:
y = target #不增加额外采样
# reset the hidden states if some sessions have just terminated
hidden = reset_hidden(hidden, mask).detach()
# Go through the GRU layer
logit, hidden = self.gru(input, target, hidden)
# Output sampling #理解,很重要！！！！！！！
y = y.to(device)
logit_sampled = logit[:, y]
# Calculate the mini-batch loss
loss = self.loss_fn(logit_sampled)
with torch.no_grad():
recall, mrr = E.evaluate(logit, target, k)
losses.append(loss.item())
recalls.append(recall)
mrrs.append(mrr)
# Gradient Clipping(Optional)
if self.clip_grad != -1:
for p in self.gru.parameters():
p.grad.data.clamp_(max=self.clip_grad)
# Mini-batch GD
if training:
# Backprop
loss.backward()
optimizer.step()
optimizer.zero_grad(
) # flush the gradient after the optimization
results = dict()
results['loss'] = np.mean(losses)
results['recall'] = np.mean(recalls)
results['mrr'] = np.mean(mrrs)
end_time = time.time()
results['time'] = (end_time - start_time) / 60
if not training:
self.gru.train()
return results
def train(self,
dataset,
k=20,
n_epochs=10,
save_dir='./models',
save=True,
model_name='GRU4REC'):
"""
Train the GRU4REC model on a pandas dataframe for several training epochs,
and store the intermediate models to the user-specified directory.
Args:
n_epochs (int): the number of training epochs to run
save_dir (str): the path to save the intermediate trained models
model_name (str): name of the model
"""
print(f'Training {model_name}...')
for epoch in range(n_epochs):
results = self.run_epoch(dataset, k=k, training=True)
results = [f'{k}:{v:.3f}' for k, v in results.items()]
print(f'epoch:{epoch+1:2d}/{"/".join(results)}')
# Store the intermediate model
if save:
save_dir = Path(save_dir)
if not save_dir.exists(): save_dir.mkdir()
model_fname = f'{model_name}_{self.loss_type}_{self.optimizer_type}_{self.lr}_epoch{epoch+1:d}'
torch.save(self.gru.state_dict(), save_dir / model_fname)
def test(self, dataset, k=20):
""" Model evaluation
Args:
k (int): the length of the recommendation list
Returns:
avg_loss: mean of the losses over the session-parallel minibatches
avg_recall: mean of the Recall@K over the session-parallel mini-batches
avg_mrr: mean of the MRR@K over the session-parallel mini-batches
wall_clock: time took for testing
"""
results = self.run_epoch(dataset, k=k, training=False)
results = [f'{k}:{v:.3f}' for k, v in results.items()]
print(f'Test result: {"/".join(results)}')
class GRU4REC:
def __init__(self,
input_size,
if_embedding,
embedding_size,
hidden_size,
output_size,
num_layers=1,
optimizer_type='Adagrad',
lr=.01,
weight_decay=0,
momentum=0,
eps=1e-6,
loss_type='TOP1',
clip_grad=-1,
dropout_input=.0,
dropout_hidden=.5,
batch_size=50,
use_cuda=True,
cuda_id=1,
compress=False,
time_sort=False,
pretrained=None):
""" The GRU4REC model
Args:
input_size (int): dimension of the gru input variables
hidden_size (int): dimension of the gru hidden units
output_size (int): dimension of the gru output variables
num_layers (int): the number of layers in the GRU
optimizer_type (str): optimizer type for GRU weights
lr (float): learning rate for the optimizer
weight_decay (float): weight decay for the optimizer
momentum (float): momentum for the optimizer
eps (float): eps for the optimizer
loss_type (str): type of the loss function to use
clip_grad (float): clip the gradient norm at clip_grad. No clipping if clip_grad = -1
dropout_input (float): dropout probability for the input layer
dropout_hidden (float): dropout probability for the hidden layer
batch_size (int): mini-batch size
use_cuda (bool): whether you want to use cuda or not
time_sort (bool): whether to ensure the the order of sessions is chronological (default: False)
pretrained (modules.layer.GRU): pretrained GRU layer, if it exists (default: None)
"""
# Initialize the GRU Layer
self.input_size = input_size
self.if_embedding = if_embedding
self.embedding_size = embedding_size
self.hidden_size = hidden_size
self.output_size = output_size
self.batch_size = batch_size
self.use_cuda = use_cuda
self.cuda_id = cuda_id
self.device = torch.device(
'cuda:%d' % cuda_id if use_cuda else 'cpu'
) # must specify cuda_id or it will be torch.cuda.current_device()
print(self.device)
if pretrained is None:
self.gru = GRU(input_size,
if_embedding,
embedding_size,
hidden_size,
output_size,
num_layers,
dropout_input=dropout_input,
dropout_hidden=dropout_hidden,
batch_size=batch_size,
use_cuda=use_cuda,
cuda_id=cuda_id)
else:
self.gru = pretrained
# Initialize the optimizer
self.optimizer_type = optimizer_type
self.weight_decay = weight_decay
self.momentum = momentum
self.lr = lr
self.eps = eps
self.compress = compress
self.compression_scheduler = None
if self.compress:
# Create a CompressionScheduler and configure it from a YAML schedule file
source = self.compress
self.compression_scheduler = distiller.config.file_config(
self.gru, None, self.compress)
self.optimizer = Optimizer(self.gru.parameters(),
optimizer_type=optimizer_type,
lr=lr,
weight_decay=weight_decay,
momentum=momentum,
eps=eps)
# Initialize the loss function
self.loss_type = loss_type
self.loss_fn = LossFunction(loss_type, use_cuda, cuda_id)
# gradient clipping(optional)
self.clip_grad = clip_grad
# etc
self.time_sort = time_sort
def run_epoch(self,
dataset,
k=20,
training=True,
compression_scheduler=None,
epoch=0):
""" Run a single training epoch """
start_time = time.time()
# initialize
losses = []
recalls = []
mrrs = []
optimizer = self.optimizer
hidden = self.gru.init_hidden()
if not training:
self.gru.eval()
device = self.device
def reset_hidden(hidden, mask):
"""Helper function that resets hidden state when some sessions terminate"""
if len(mask) != 0:
hidden[:, mask, :] = 0
return hidden
# Start the training loop
loader = SessionDataLoader(dataset, batch_size=self.batch_size)
batch_id = 0
steps_per_epoch = math.ceil(len(dataset.items) / self.batch_size)
for input, target, mask in loader:
batch_id = batch_id + 1
input = input.to(device)
target = target.to(device)
# reset the hidden states if some sessions have just terminated
hidden = reset_hidden(hidden, mask).detach()
if compression_scheduler:
compression_scheduler.on_minibatch_begin(
epoch,
minibatch_id=batch_id,
minibatches_per_epoch=steps_per_epoch)
# Go through the GRU layer
logit, hidden = self.gru(input, target, hidden)
# Output sampling
logit_sampled = logit[:, target.view(-1)]
# Calculate the mini-batch loss
loss = self.loss_fn(logit_sampled)
with torch.no_grad():
recall, mrr = E.evaluate(logit, target, k)
losses.append(loss.item())
recalls.append(recall)
mrrs.append(mrr)
# Gradient Clipping(Optional)
if self.clip_grad != -1:
for p in self.gru.parameters():
p.grad.data.clamp_(max=self.clip_grad)
# Mini-batch GD
if training:
if compression_scheduler:
# Before running the backward phase, we allow the scheduler to modify the loss
# (e.g. add regularization loss)
loss = compression_scheduler.before_backward_pass(
epoch,
minibatch_id=batch_id,
minibatches_per_epoch=steps_per_epoch,
loss=loss,
return_loss_components=False)
# Backprop
loss.backward()
optimizer.step()
optimizer.zero_grad(
) # flush the gradient after the optimization
if compression_scheduler:
compression_scheduler.on_minibatch_end(
epoch,
minibatch_id=batch_id,
minibatches_per_epoch=steps_per_epoch)
results = dict()
#ipdb.set_trace()
results['loss'] = np.mean(losses)
results['recall'] = np.mean(recalls)
results['mrr'] = np.mean(mrrs)
end_time = time.time()
results['time'] = (end_time - start_time) / 60
if not training:
self.gru.train()
return results
def train(self,
dataset,
k=20,
n_epochs=20,
save_dir='./models',
save=True,
model_name='GRU4REC'):
"""
Train the GRU4REC model on a pandas dataframe for several training epochs,
and store the intermediate models to the user-specified directory.
Args:
n_epochs (int): the number of training epochs to run
save_dir (str): the path to save the intermediate trained models
model_name (str): name of the model
"""
print(f'Training {model_name}...')
for epoch in range(n_epochs):
if self.compression_scheduler:
self.compression_scheduler.on_epoch_begin(epoch)
results = self.run_epoch(
dataset,
k=k,
training=True,
compression_scheduler=self.compression_scheduler,
epoch=epoch)
results = [f'{k}:{v:.3f}' for k, v in results.items()]
print(f'epoch:{epoch+1:2d}/{"/".join(results)}')
t, total = distiller.weights_sparsity_tbl_summary(
self.gru, return_total_sparsity=True)
print("\nParameters:\n" + str(t))
print('Total sparsity: {:0.2f}\n'.format(total))
if self.compression_scheduler:
self.compression_scheduler.on_epoch_end(epoch)
# Store the intermediate model
if save:
save_dir = Path(save_dir)
if not save_dir.exists(): save_dir.mkdir()
model_fname = f'{model_name}_{self.loss_type}_{self.optimizer_type}_{self.lr}_epoch{epoch+1:d}'
torch.save(self.gru.state_dict(), save_dir / model_fname)
def test(self, dataset, k=20):
""" Model evaluation
Args:
k (int): the length of the recommendation list
Returns:
avg_loss: mean of the losses over the session-parallel minibatches
avg_recall: mean of the Recall@K over the session-parallel mini-batches
avg_mrr: mean of the MRR@K over the session-parallel mini-batches
wall_clock: time took for testing
"""
results = self.run_epoch(dataset, k=k, training=False)
results = [f'{k}:{v:.3f}' for k, v in results.items()]
print(f'Test result: {"/".join(results)}')