"""
Module containining class definitions of trainers for skipgram models,
which are partly used by Deep Graph Kernels
Author: Paul Scherer
"""
import torch
import torch.optim as optim
from torch.utils.data import DataLoader
from tqdm import tqdm
# Internal
from .skipgram_data_reader import SkipgramCorpus, InMemorySkipgramCorpus
from .skipgram import Skipgram
from .utils import save_subgraph_embeddings
[docs]class Trainer(object):
"""Handles corpus construction (hard drive version), skipgram initialization and training.
Paramaters
----------
corpus_dir : str
path to directory containing graph files
extension : str
extension used in graph documents produced after decomposition stage
max_files : int
the maximum number of graph files to consider, default of 0 uses all files
output_fh : str
the path to the file where embeddings should be saved
emb_dimension : int (default=128)
the desired dimension of the embeddings
batch_size : int (default=32)
the desired batch size
epochs : int (default=100)
the desired number of epochs for which the network should be trained
initial_lr : float (default=1e-3)
the initial learning rate
min_count : int (default=1)
the minimum number of times a pattern should occur across the dataset to
be considered part of the substructure pattern vocabulary
Returns
-------
self : Trainer
A Trainer instance
"""
def __init__(self, corpus_dir, extension, max_files, window_size, output_fh, emb_dimension=128, batch_size=32, epochs=100, initial_lr=1e-3, min_count=1):
self.corpus = SkipgramCorpus(corpus_dir, extension, max_files, min_count, window_size)
self.dataloader = DataLoader(self.corpus, batch_size, shuffle=False, num_workers=4, collate_fn = self.corpus.collate)
self.corpus_dir = corpus_dir
self.extension = extension
self.max_files = max_files
self.output_fh = output_fh
self.emb_dimension = emb_dimension
self.batch_size = batch_size
self.epochs = epochs
self.initial_lr = initial_lr
self.min_count = min_count
self.window_size = window_size
self.num_targets = self.corpus.num_subgraphs # the special feature here is that we are learning subgraph reps
self.vocab_size = self.corpus.num_subgraphs
self.skipgram = Skipgram(self.num_targets, self.vocab_size, self.emb_dimension)
if torch.cuda.is_available():
self.device = torch.device("cuda")
self.skipgram.cuda()
else:
self.device = torch.device("cpu")
[docs] def train(self):
"""Train the network with the settings used to initialise the Trainer"""
for epoch in range(self.epochs):
print("### Epoch: " + str(epoch))
optimizer = optim.Adagrad(self.skipgram.parameters(), lr=self.initial_lr)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, len(self.dataloader))
running_loss = 0.0
for sample_batched in tqdm(self.dataloader):
if len(sample_batched[0]) > 1:
pos_target = sample_batched[0].to(self.device)
pos_context = sample_batched[1].to(self.device)
neg_context = sample_batched[2].to(self.device)
optimizer.zero_grad()
loss = self.skipgram.forward(pos_target, pos_context, neg_context) # the loss is integrated into the forward function
loss.backward()
optimizer.step()
scheduler.step()
running_loss = running_loss * 0.9 + loss.item() * 0.1
print(" Loss: " + str(running_loss))
final_embeddings = self.skipgram.target_embeddings.weight.cpu().data.numpy()
save_subgraph_embeddings(self.corpus, final_embeddings, self.output_fh)
return(final_embeddings)
[docs]class InMemoryTrainer(object):
"""Handles corpus construction (in-memory version), PVDBOW initialization and training.
Paramaters
----------
corpus_dir : str
path to directory containing graph files
extension : str
extension used in graph documents produced after decomposition stage
max_files : int
the maximum number of graph files to consider, default of 0 uses all files
output_fh : str
the path to the file where embeddings should be saved
emb_dimension : int (default=128)
the desired dimension of the embeddings
batch_size : int (default=32)
the desired batch size
epochs : int (default=100)
the desired number of epochs for which the network should be trained
initial_lr : float (default=1e-3)
the initial learning rate
min_count : int (default=1)
the minimum number of times a pattern should occur across the dataset to
be considered part of the substructure pattern vocabulary
Returns
-------
self : InMemoryTrainer
A trainer instance which has the dataset stored in memory for fast access
"""
def __init__(self, corpus_dir, extension, max_files, window_size, output_fh, emb_dimension=128, batch_size=32, epochs=100, initial_lr=1e-3, min_count=1):
self.corpus = InMemorySkipgramCorpus(corpus_dir, extension, max_files, min_count, window_size)
self.dataloader = DataLoader(self.corpus, batch_size, shuffle=False, num_workers=4, pin_memory=True, collate_fn = self.corpus.collate)
self.corpus_dir = corpus_dir
self.extension = extension
self.max_files = max_files
self.output_fh = output_fh
self.emb_dimension = emb_dimension
self.batch_size = batch_size
self.epochs = epochs
self.initial_lr = initial_lr
self.min_count = min_count
self.window_size = window_size
self.num_targets = self.corpus.num_subgraphs # the special feature here is that we are learning subgraph reps
self.vocab_size = self.corpus.num_subgraphs
self.skipgram = Skipgram(self.num_targets, self.vocab_size, self.emb_dimension)
if torch.cuda.is_available():
self.device = torch.device("cuda")
self.skipgram.cuda()
else:
self.device = torch.device("cpu")
[docs] def train(self):
"""Train the network with the settings used to initialise the Trainer"""
for epoch in range(self.epochs):
print("### Epoch: " + str(epoch))
optimizer = optim.Adagrad(self.skipgram.parameters(), lr=self.initial_lr)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, len(self.dataloader))
running_loss = 0.0
for sample_batched in tqdm(self.dataloader):
if len(sample_batched[0]) > 1:
pos_target = sample_batched[0].to(self.device)
pos_context = sample_batched[1].to(self.device)
neg_context = sample_batched[2].to(self.device)
optimizer.zero_grad()
loss = self.skipgram.forward(pos_target, pos_context, neg_context) # the loss is integrated into the forward function
loss.backward()
optimizer.step()
scheduler.step()
running_loss = running_loss * 0.9 + loss.item() * 0.1
print(" Loss: " + str(running_loss))
final_embeddings = self.skipgram.target_embeddings.weight.cpu().data.numpy()
save_subgraph_embeddings(self.corpus, final_embeddings, self.output_fh)
return(final_embeddings)
if __name__ == '__main__':
pass