def prepare_batch(encoder, batch):
global task2title_max_steps, embedding_size
(sents_tup, seq_lens), res_tup = batch
res = C(encode(encoder, res_tup))
sents = encode(encoder, sents_tup)
sents = sents.view(-1, task2title_max_steps, embedding_size)
#norm = sents.norm(p=2, dim=2, keepdim=True)
#return (sents.div(norm), seq_lens), res.div(res.norm(p=2, dim=1, keepdim=True))
return (sents, seq_lens), res
def forward(self, inp):
inp, lengths = inp
inp = C(inp)
out, imm = self.gru(inp)
# select only outputs at lengths[i]
padded_lengths = [
i * inp.size(1) + v - 1 for i, v in enumerate(lengths)
]
out_ = out.contiguous().view(-1, self.inner_size)[padded_lengths, :]
# then feed them to fully connected
out_ = self.linear(out_)
out_ = nn.functional.tanh(out_)
return out_.div(out_.norm(p=2, dim=1, keepdim=True))
# Combinatoric selections
from utils import C
limit = 1000000
candidate = []
for n in range(1, 101):
for r in range(1, n):
if C(n, r) > limit:
candidate.append((n,r))
print len(candidate)
"""
# dataset = 'debug_BA:train_size=1000,test_size=100,num_nodes_training=16,num_nodes_testing=64'
#dataset = 'debug_BA:train_size=1000,test_size=100,num_nodes_training=0,num_nodes_testing=0'
dataset = 'aids700nef'
parser.add_argument('--dataset', default=dataset)
dataset_version = None # 'v2'
parser.add_argument('--dataset_version', default=dataset_version)
filter_large_size = None
parser.add_argument('--filter_large_size', type=int, default=filter_large_size) # None or >= 1
select_node_pair = None
parser.add_argument('--select_node_pair', type=str, default=select_node_pair) # None or gid1_gid2
c = C()#counting
parser.add_argument('--node_fe_{}'.format(c.c()), default='one_hot')
# parser.add_argument('--node_fe_{}'.format(c.c()),
# default='local_degree_profile')
natts, eatts, tvt_options, align_metric_options, *_ = \
get_dataset_conf(dataset)
""" Must use exactly one alignment metric across the entire run. """
#align_metric = align_metric_options[0]
#if len(align_metric_options) == 2:
""" Choose which metric to use. """
#align_metric = 'ged'
align_metric = 'mcs'
parser.add_argument('--align_metric', default=align_metric)
def train(model, data_loader, encoder, training=None, testing=None):
global task2title_path, task2title_batch_size
num_batches = (data_loader.get_total_samples(source=training) +
task2title_batch_size - 1) // task2title_batch_size
if enable_cuda:
model = model.cuda()
model.train()
optimizer = optim.Adam(model.parameters())
best_loss = 1e8
with open(task2title_path + '.run.log',
'a') as logfile, open(task2title_path + '.vectors.log',
'a') as vector_file:
for epoch in range(num_epochs):
log('starting epoch ', epoch + 1, log_file=logfile)
total_loss = 0
last_saved = -save_backoff
for batchid, batch in enumerate(
data_loader.get_samples(batch_size=task2title_batch_size,
max_seq=task2title_max_steps,
source=training)):
steps, results = prepare_batch(encoder, batch)
predicted = model(steps)
optimizer.zero_grad()
length = len(batch[1][1])
y = torch.FloatTensor(length).fill_(1)
y = y.cuda() if enable_cuda else y
loss = nn.functional.cosine_embedding_loss(
predicted, results, C(y))
loss.backward()
torch.nn.utils.clip_grad_norm(model.parameters(),
gradient_clip)
optimizer.step()
this_step_loss = loss.sum().data[0]
total_loss += this_step_loss
if batchid % log_every == 0:
log("\tBatch {}/{}, average loss: {}, current loss: {}".
format(batchid, num_batches,
total_loss / (batchid + 1), this_step_loss),
log_file=logfile)
#log("\t\tPred norms: ", (predicted).norm(dim=1).data.tolist(), log_file=logfile)
#log("\t\tGT norms: ", (results).norm(dim=1).data.tolist(), log_file=logfile)
log("\t\tDiff norms: ",
(predicted - results).norm(dim=1).data.tolist(),
log_file=logfile)
if this_step_loss < best_loss and (last_saved +
save_backoff) <= batchid:
log("\t\tSaving best at epoch {}, batch {}, loss {}...".
format(epoch, batchid, this_step_loss),
log_file=logfile)
torch.save(model, task2title_path + ".best.pyt")
best_loss = this_step_loss
last_saved = batchid
if batchid % save_every == 0:
log("\t\tSaving regularly at epoch {}, batch {}...".format(
epoch, batchid),
log_file=logfile)
torch.save(model, task2title_path + ".regular.pyt")
# torch.save(model, task2title_path+".epoch-{}.pyt".format(epoch))
if testing:
model.eval()
for batchid, batch in enumerate(
data_loader.get_samples(batch_size=task2title_batch_size,
max_seq=task2title_max_steps,
source=testing)):
steps, results = prepare_batch(encoder, batch)
predicted = model(steps)
length = len(batch[1][1])
for v_id in range(length):
print('gt ',
results.data[v_id, :].tolist(),
file=vector_file)
print('pr ',
predicted.data[v_id, :].tolist(),
file=vector_file)
model.train()
return model
dataset:
(for MCS)
aids700nef linux imdbmulti redditmulti10k
"""
dataset = 'aids700nef'
parser.add_argument('--dataset', default=dataset)
filter_large_size = None
parser.add_argument('--filter_large_size', type=int,
default=filter_large_size) # None or >= 1
select_node_pair = None
parser.add_argument('--select_node_pair', type=str,
default=select_node_pair) # None or gid1_gid2
c = C()
parser.add_argument('--node_fe_{}'.format(c.c()), default='one_hot')
# parser.add_argument('--node_fe_{}'.format(c.c()),
# default='local_degree_profile')
natts, eatts, tvt_options, align_metric_options, *_ = \
get_dataset_conf(dataset)
""" Must use exactly one alignment metric across the entire run. """
align_metric = align_metric_options[0]
if len(align_metric_options) == 2:
""" Choose which metric to use. """
align_metric = 'ged'
# align_metric = 'mcs'
parser.add_argument('--align_metric', default=align_metric)
# generate names for each attribute
# used when output to excel
import utils
from utils import C
import xlwt
CTN = C(0, 0, 5, 0, 60)
B_BOTTOM = CTN.bmin
B_TOP = CTN.bmax
STEP = CTN.step
def generate_names():
names = ['R', 'G', 'B', 'RGB_STD']
for i in range(B_BOTTOM, B_TOP, STEP):
names.append('RGB_DIS_' + str(i))
names.append('DIS_STD')
for i in range(B_BOTTOM, B_TOP, STEP):
names.append('RGB_STD_' + str(i))
names.append('STD_STD')
names.append('MAX_DIS')
names.append('MIN_DIS')
for i in range(B_BOTTOM, B_TOP, STEP):
names.append('RGB_MEAN_1_' + str(i))
names.append('RGB_MEAN_2_' + str(i))
names.append('RGB_MEAN_3_' + str(i))
names.append('RGB_MEAN_4_' + str(i))
names.append('RGB_MEAN_5_' + str(i))
names.append('RGB_MEAN_6_' + str(i))
names.append('BI')
names.append('CI')
names.append('RI')