def train_score(trial):
lr = trial.suggest_loguniform('lr', 1e-5, 1)
weight_decay = trial.suggest_loguniform('weight_decay', 1e-12, 1e-3)
config = {
'model': {
'layers': [
{'type': 'embed', 'node_dim': 8},
],
'readout': [
{'type': 'cga'}
]
},
'dataset': {
'key': 'rank18_reachability_0',
'competition': '2018',
'category': 'reachability',
'test_ratio': 0.2,
'min_tool_coverage': 0.8,
'ast_type': 'bag'
},
'train': {
'loss': "tasks::Rank_BCE",
'epoch': 20,
'batch': 32,
'shuffle': 42,
'augment': False,
'optimizer': {'type': 'torch::Adam', 'lr': lr,
'weight_decay': weight_decay},
'scheduler': {
'type': 'torch::StepLR', 'mode': 'epoch',
'step_size': 5, 'gamma': 0.1
},
'validate': {
'checkpoint_step': 0,
'score': 'spearmann',
'split': 0.03
}
}
}
train_test = build_model(config, trial, ret='spearmann')
with backend.openLocalSession() as sess:
mean = sess.run(train_test).join()
return mean
from taskflow import backend
from tasks import preparation as p
import argparse
from bson.objectid import ObjectId
parser = argparse.ArgumentParser()
parser.add_argument("tid")
args = parser.parse_args()
id = args.tid
id = ObjectId(id)
with backend.openLocalSession() as sess:
sess.run(p.ast_features_bag(id, 5000))
condition[key] = limit
filter = tu.filter_by_stat(competition, condition)
split = tu.train_test(dataset_key,
competition,
category=category,
test_ratio=0.2,
filter=filter)
cov = tu.tool_coverage(competition, filter=split[0], category=category)
tools = tu.covered_tools(dataset_key,
competition,
cov,
min_coverage=0.8)
train_index, test_index = split[0], split[1]
lr = svm_train_test(lr_key,
tools,
'norm_2018_2',
train_index,
test_index,
competition,
category,
C=[0.001, 0.01, 0.1, 1.0, 100, 1000])
with backend.openLocalSession(auto_join=True) as sess:
mean, _ = sess.run(lr)
res.append(mean)
print("Acc: %f (+- %f)" % (np.mean(res), np.std(res)))
def egin_execute(name,
dataset_key,
competition,
model_depth,
embed_size,
hidden_size,
batch_size,
loss_func,
epoch=50,
category=None):
model_layers = [{'type': 'embed', 'node_dim': embed_size}]
for i in range(model_depth):
model_layers.append({
'type': 'edge_gin',
'node_dim': embed_size,
'hidden': hidden_size,
'norm': True,
'dropout': 0.1
})
model_readout = [{'type': 'cga'} for i in range(len(model_layers))]
model = {'layers': model_layers, 'readout': model_readout}
config = {
'model': model,
'dataset': {
'key': dataset_key,
'competition': competition,
'category': category,
'test_ratio': 0.2,
'min_tool_coverage': 0.8,
'ast_type': 'bag'
},
'train': {
'loss': translate_loss(loss_func),
'epoch': epoch,
'batch': batch_size,
'shuffle': True,
'augment': False,
'optimizer': {
'type': 'torch::Adam',
'lr': 0.01,
'betas': [0.9, 0.98],
'eps': 1e-9
},
'scheduler': {
'type': 'torch::CosineAnnealingLR',
'mode': 'epoch',
'T_max': epoch,
'eta_min': 0.0001
},
'validate': {
'checkpoint_step': 0,
'score': 'spearmann',
'split': 0.1
}
},
'test': {
'type': 'category',
'scores': 'spearmann'
}
}
task = build_model(config)
with backend.openLocalSession() as sess:
return sess.run(task)
def cfg_egin_execute(config, env=None):
model_layers = [{'type': 'embed', 'node_dim': config['embed_size']}]
for i in range(config['model_depth']):
model_layers.append({
'type': 'edge_gin',
'node_dim': config['embed_size'],
'hidden': config['hidden_size'],
'norm': True,
'dropout': 0.1
})
model_readout = [{'type': 'cga'} for i in range(len(model_layers))]
model = {'layers': model_layers, 'readout': model_readout}
cfg = {
'model': model,
'dataset': {
'key': config['dataset_key'],
'competition': config['competition'],
'category': config['category'],
'test_ratio': 0.2,
'min_tool_coverage': 0.8,
'ast_type': 'bag'
},
'train': {
'loss': translate_loss(config['loss_func']),
'epoch': config['epoch'],
'batch': config['batch_size'],
'shuffle': True,
'augment': False,
'optimizer': {
'type': 'torch::Adam',
'lr': 0.01,
'betas': [0.9, 0.98],
'eps': 1e-9
},
'scheduler': {
'type': 'tasks::SuperConverge',
'mode': 'step',
'warmup': 40,
'd_model': 32
},
'validate': {
'checkpoint_step': 0,
'score': 'spearmann',
'split': 0.1
}
},
'test': {
'type': 'category',
'scores': 'spearmann'
}
}
if 'name' in config:
cfg['name'] = config['name']
task = build_model(cfg)
with backend.openLocalSession() as sess:
res = sess.run(task)
shutil.rmtree(env.get_cache_dir())
os.makedirs(env.get_cache_dir())
return res