def visualize_attention_weights(self, epoch):
"""
Visualize attention of the model saved during an epoch
:param epoch: The epoch
"""
#TODO: Add a node name mapping for easily readable visualization
nodes = self._args.nodes
assert len(nodes) == 2
src, trg = nodes
utils.log(
f"Generating attention weight visualization between source = {src} and target = {trg} nodes"
)
self._load_best_model(best_epoch=epoch)
batch = {
"source_neighborhood": self._neighborhood[src].unsqueeze(0),
"target_neighborhood": self._neighborhood[trg].unsqueeze(0),
"source_mask": self._mask[src].unsqueeze(0),
"target_mask": self._mask[trg].unsqueeze(0),
}
self._model(**utils.to(batch, self._device))
src_mask = self._mask[src] == 0
trg_mask = self._mask[trg] == 0
source_atn = self._model.source_atn.squeeze().detach().cpu().numpy(
)[src_mask]
target_atn = self._model.target_atn.squeeze().detach().cpu().numpy(
)[trg_mask]
source_nbh = batch["source_neighborhood"].squeeze()[src_mask]
target_nbh = batch["source_neighborhood"].squeeze()[trg_mask]
params = self._params
data_dir = osp.join(
params["root"],
params["name"]) if "name" in params else params["root"]
src_path = osp.join(data_dir, "results", f"source_atn_vis_{src}.html")
trg_path = osp.join(data_dir, "results", f"target_atn_vis_{trg}.html")
utils.visualize_attention_weights(path=src_path,
neighborhood=source_nbh,
weights=source_atn,
node_id=src)
utils.visualize_attention_weights(path=trg_path,
neighborhood=target_nbh,
weights=target_atn,
node_id=trg)
def _fit(self, loader):
"""
Fits the GOAT model using data from a specified loader
:param loader: The loader
"""
sources, targets, source_atn, target_atn, source_features, target_features = [], [], [], [], [], []
for bc, batch in self._test_loader:
sources.append(batch["source"].cpu().numpy())
targets.append(batch["target"].cpu().numpy())
true_source_feature, true_target_feature, _, _ = self._model(
**utils.to(batch, self._device))
source_atn.append(
self._model.source_atn.squeeze().detach().cpu().numpy())
target_atn.append(
self._model.target_atn.squeeze().detach().cpu().numpy())
source_features.append(true_source_feature.detach().cpu().numpy())
target_features.append(true_target_feature.detach().cpu().numpy())
sources = np.concatenate(sources)
targets = np.concatenate(targets)
source_atn = np.concatenate([
utils.expand_if(atn_v, 0,
len(atn_v.shape) == 1) for atn_v in source_atn
])
target_atn = np.concatenate([
utils.expand_if(atn_v, 0,
len(atn_v.shape) == 1) for atn_v in target_atn
])
true_source_feature = np.concatenate([
utils.expand_if(f, 0,
len(f.shape) == 1) for f in source_features
])
true_target_feature = np.concatenate([
utils.expand_if(f, 0,
len(f.shape) == 1) for f in target_features
])
return sources, targets, source_atn, target_atn, true_source_feature, true_target_feature
def _fit_false_pairs(self, num_false_samples=None):
"""
Inference using false pairs. Useful for validating and testing the quality of link prediction.
:param num_false_samples: The number of false pairs.
"""
if num_false_samples:
false_pairs = self._negative_pairs[:num_false_samples]
else:
false_pairs = self._negative_pairs
num_false_samples = false_pairs.shape[0]
source_features, target_features = [], []
batch_size = self._args.batch
for i in range(0, num_false_samples, batch_size):
end = i + batch_size if num_false_samples - batch_size > i else num_false_samples
false_data = {
"source_neighborhood":
self._neighborhood[self._negative_pairs[i:end, 0]],
"target_neighborhood":
self._neighborhood[self._negative_pairs[i:end, 1]],
"source_mask":
self._mask[self._negative_pairs[i:end, 0]],
"target_mask":
self._mask[self._negative_pairs[i:end, 1]],
}
false_source_feature, false_target_feature, _, _ = self._model(
**utils.to(false_data, self._device))
source_features.append(false_source_feature.detach().cpu().numpy())
target_features.append(false_target_feature.detach().cpu().numpy())
false_source_feature = np.concatenate([
utils.expand_if(f, 0,
len(f.shape) == 1) for f in source_features
])
false_target_feature = np.concatenate([
utils.expand_if(f, 0,
len(f.shape) == 1) for f in target_features
])
return false_source_feature, false_target_feature
# Higher level interface for interacting with specific classes of events
import pyalm as alm
import cleanup
from utils import dictmapper, MappingRule as to
ContributorBase = dictmapper(
'ContributorBase', {
'contributor_role': ['contributor_role'],
'first_author': to(['first_author'], cleanup._text_to_bool),
'given_name': ['given_name'],
'sequence': ['sequence'],
'surname': ['surname']
})
class Contributor(ContributorBase):
def __repr__(self):
return '<%s: %s%s>' % (type(self).__name__, self.given_name,
self.surname)
CrossrefBase = dictmapper(
'CrossrefBase', {
'article_title': ['event', 'article_title'],
'doi': ['event', 'doi'],
'fl_count':
to(['event', 'fl_count'], cleanup._parse_numbers_to_int),
'issn': ['event', 'issn'],
'journal_abbreviation': ['event', 'journal_abbreviation'],
'journal_title': ['event', 'journal_title'],
'publication_type': ['event', 'publication_type'],
import requests
from utils import dictmapper, MappingRule as to
import cleanup
import config
BASE_HEADERS = {'Accept': 'application/json'}
MetricsBase = dictmapper(
'MetricsBase', {
'citations': to(['citations'], cleanup._parse_numbers_to_int),
'comments': to(['comments'], cleanup._parse_numbers_to_int),
'html': to(['html'], cleanup._parse_numbers_to_int),
'likes': to(['likes'], cleanup._parse_numbers_to_int),
'pdf': to(['pdf'], cleanup._parse_numbers_to_int),
'readers': to(['readers'], cleanup._parse_numbers_to_int),
'shares': to(['shares'], cleanup._parse_numbers_to_int),
'total': to(['total'], cleanup._parse_numbers_to_int)
})
class Metrics(MetricsBase):
def __repr__(self):
return """
<%s total:%s readers:%s pdf:%s html:%s comments:%s likes:%s>
""" % (type(self).__name__, self.total, self.readers, self.pdf, self.html,
self.comments, self.likes)
SourceBase = dictmapper(
'SourceBase', {
'name': ['name'],
# Higher level interface for interacting with specific classes of events
import pyalm as alm
import cleanup
from utils import dictmapper, MappingRule as to
ContributorBase = dictmapper('ContributorBase',
{
'contributor_role' : ['contributor_role'],
'first_author': to(['first_author'],
cleanup._text_to_bool),
'given_name': ['given_name'],
'sequence': ['sequence'],
'surname': ['surname']
}
)
class Contributor(ContributorBase):
def __repr__(self):
return '<%s: %s%s>' % (type(self).__name__, self.given_name, self.surname)
CrossrefBase = dictmapper('CrossrefBase',
{
'article_title' : ['event','article_title'],
'doi' : ['event','doi'],
'fl_count' : to(['event','fl_count'],
cleanup._parse_numbers_to_int),
'issn' : ['event','issn'],
'journal_abbreviation' : ['event','journal_abbreviation'],
'journal_title' : ['event','journal_title'],
'publication_type' : ['event','publication_type'],
import requests
from utils import dictmapper, MappingRule as to
import cleanup
import config
BASE_HEADERS = {'Accept': 'application/json'}
MetricsBase = dictmapper('MetricsBase',
{'citations': to(['citations'],
cleanup._parse_numbers_to_int),
'comments': to(['comments'],
cleanup._parse_numbers_to_int),
'html': to(['html'],
cleanup._parse_numbers_to_int),
'likes': to(['likes'],
cleanup._parse_numbers_to_int),
'pdf': to(['pdf'],
cleanup._parse_numbers_to_int),
'readers': to(['readers'],
cleanup._parse_numbers_to_int),
'shares': to(['shares'],
cleanup._parse_numbers_to_int),
'total': to(['total'],
cleanup._parse_numbers_to_int)
}
)
class Metrics(MetricsBase):
def __repr__(self):
return """
# Higher level interface for interacting with specific classes of events
import pyalm as alm
import cleanup
from utils import dictmapper, MappingRule as to
ContributorBase = dictmapper('ContributorBase',
{
'contributor_role' : ['contributor_role'],
'first_author': to(['first_author'],
cleanup._text_to_bool),
'given_name': ['given_name'],
'sequence': ['sequence'],
'surname': ['surname']
}
)
class Contributor(ContributorBase):
def __repr__(self):
return '<%s: %s%s>' % (type(self).__name__, self.given_name, self.surname)
CitationBase = dictmapper('CitationBase',
{
'article_title' : ['event','article_title'],
'doi' : ['event','doi'],
'fl_count' : to(['event','fl_count'],
cleanup._parse_numbers_to_int),
'issn' : ['event','issn'],
'journal_abbreviation' : ['event','journal_abbreviation'],
'journal_title' : ['event','journal_title'],
'publication_type' : ['event','publication_type'],
import requests
from utils import dictmapper, MappingRule as to
import cleanup
import config
BASE_HEADERS = {'Accept': 'application/json'}
MetricsBase = dictmapper('MetricsBase',
{'citations': to(['citations'],
cleanup._parse_numbers_to_int),
'comments': to(['comments'],
cleanup._parse_numbers_to_int),
'groups': to(['groups'],
cleanup._parse_numbers_to_int),
'likes': to(['likes'],
cleanup._parse_numbers_to_int),
'pdf': to(['pdf'],
cleanup._parse_numbers_to_int),
'shares': to(['shares'],
cleanup._parse_numbers_to_int),
'total': to(['total'],
cleanup._parse_numbers_to_int)
}
)
class Metrics(MetricsBase):
def __repr__(self):
return """
<%s total:%s shares:%s citations:%s comments:%s>
""" % (type(self).__name__, self.total, self.shares, self.citations, self.comments)
def fit(self, trainloader, validloader, num_epochs=1000):
# . . logs
logs = {}
# . . number of batches
num_train_batch = len(trainloader)
num_valid_batch = len(validloader)
# . . register num batches to logs
logs['num_train_batch'] = num_train_batch
logs['num_valid_batch'] = num_valid_batch
# . .
# . . set the callback handler
callback_handler = CallbackHandler(self.callbacks)
# . . keep track of the losses
train_losses = []
valid_losses = []
# . . call the callback function on_train_begin(): load the best model
callback_handler.on_train_begin(logs=logs, model=self.model)
for epoch in range(num_epochs):
#scheduler.step()
train_loss = 0.
valid_loss = 0.
# . . activate the training mode
self.model.train()
# . . get the next batch of training data
for batch, (center, left, right) in enumerate(trainloader):
# . . the training loss for the current batch
batch_loss = 0.
# . . send the batch to GPU
center, left, right = utils.to(center, self.device), utils.to(left, self.device), utils.to(right, self.device)
# . . zero the parameter gradients
self.optimizer.zero_grad()
# . . iterate over cameras
cameras = [center, left, right]
for camera in cameras:
image, steering = camera
# . . feed-forward network to predict the steering angle
angle = self.model(image)
#log_angle = torch.log10(1.0 + angle)
# . . calculate the loss function
loss = self.criterion(angle, steering.unsqueeze(1))
# . . backpropagate the loss
#loss.backward()
# . . backpropogate the scaled physics loss
with amp.scale_loss(loss, self.optimizer) as scaled_loss:
scaled_loss.backward()
# . . update weights
self.optimizer.step()
# . . training loss for the current batch: accumulate over cameras
batch_loss += loss.item()
# . . accumulate the training loss
train_loss += loss.item()
# . . register the batch training loss
logs['batch_loss'] = batch_loss
# . . call the callback functions on_epoch_end()
callback_handler.on_batch_end(batch, logs=logs, model=self.model)
# . . activate the evaluation (validation) mode
self.model.eval()
# . . turn off the gradient for performance
with torch.set_grad_enabled(False):
# . . get the next batch of validation data
for batch, (center, left, right) in enumerate(validloader):
# . . send the batch to GPU
center, left, right = utils.to(center, self.device), utils.to(left, self.device), utils.to(right, self.device)
# . . iterate over cameras
cameras = [center, left, right]
for camera in cameras:
image, steering = camera
# . . feed-forward network to predict the steering angle
angle = self.model(image)
# . . calculate the loss function
loss = self.criterion(angle, steering.unsqueeze(1))
# . . accumulate the training loss
valid_loss += loss.item()
# . . normalize the training and validation losses
train_loss /= (num_train_batch*3)
valid_loss /= (num_valid_batch*3)
# . . on epoch end
train_losses.append(train_loss)
valid_losses.append(valid_loss)
# . . update the epoch statistics (logs)
logs["train_loss"] = train_loss
logs["valid_loss"] = valid_loss
# . . call the callback functions on_epoch_end()
callback_handler.on_epoch_end(epoch, logs=logs, model=self.model)
# . . check if the training should continue
if self.model._stop_training:
break
# . . call the callback function on_train_end(): load the best model
callback_handler.on_train_end(logs=logs, model=self.model)
return train_losses, valid_losses
