def learnParamsAdaGrad(self,
decoder,
corpus,
param_file,
loss_func,
num_passes=10,
oracle_len='nolen'):
"""
learn parameters using Structured Perceptron, Ramp Loss, (Hinge??)
"""
logger.debug('start learning parameters...')
shuffle(corpus) # shuffle corpus
avg_weights = FeatureVector()
curr_instances = 0
node_perf = PerfScore(0.0, 0.0, 0.0) # node performance
edge_perf = PerfScore(0.0, 0.0, 0.0) # edge performance
eta = 1.0 # stepsize
l2reg = 0.0 #
node_cost_scaling = 1.0 # cost scaling factor
edge_cost_scaling = 1.0 # cost scaling factor
sumSq = FeatureVector()
for curr_num_passes in xrange(1, num_passes + 1):
logger.debug('#curr_num_passes#: %d' % curr_num_passes)
for instance in corpus:
curr_instances += 1
logger.debug('processing instance %d...' % curr_instances)
# perceptron loss
if loss_func.startswith('perceptron'):
gradient, selected_nodes, selected_edges, score_pred = decoder.decode(
instance, oracle_len)
plus_feats, oracle_nodes, oracle_edges, score_true = decoder.oracle(
instance)
curr_loss = score_pred - score_true # @UnusedVariable
gradient -= plus_feats
decoder.weights -= eta * gradient
# ramp loss + cost-augmented decoding
if loss_func.startswith('ramp'):
node_cost, edge_cost = node_cost_scaling, edge_cost_scaling
gradient, _, _, score_plus_cost = decoder.decode(
instance, oracle_len, node_cost, edge_cost)
plus_feats, selected_nodes, selected_edges, score_minus_cost = decoder.decode(
instance, oracle_len, -1.0 * node_cost,
-1.0 * edge_cost)
_, oracle_nodes, oracle_edges, score_true = decoder.oracle(
instance)
curr_loss = score_plus_cost - score_minus_cost # @UnusedVariable
gradient -= plus_feats
for k, v in gradient.iteritems():
if v == 0.0: continue
sumSq[k] = sumSq.get(k, 0.0) + v * v
decoder.weights[k] = decoder.weights.get(
k, 0.0) - eta * v / sqrt(sumSq[k])
if l2reg != 0.0:
for k, v in decoder.weights.iteritems():
if v == 0.0: continue
value = l2reg * v
sumSq[k] = sumSq.get(k, 0.0) + value * value
decoder.weights[k] = v - eta * value / sqrt(
sumSq[k])
# hinge-loss + cost-augmented decoding
if loss_func.startswith('hinge'):
node_cost, edge_cost = node_cost_scaling, edge_cost_scaling
gradient, selected_nodes, selected_edges, score_plus_cost = decoder.decode(
instance, oracle_len, node_cost, edge_cost)
plus_feats, oracle_nodes, oracle_edges, score_true = decoder.oracle(
instance)
curr_loss = score_plus_cost - score_true # @UnusedVariable
gradient -= plus_feats
for k, v in gradient.iteritems():
if v == 0.0: continue
sumSq[k] = sumSq.get(k, 0.0) + v * v
decoder.weights[k] = decoder.weights.get(
k, 0.0) - eta * v / sqrt(sumSq[k])
if l2reg != 0.0:
for k, v in decoder.weights.iteritems():
if v == 0.0: continue
value = l2reg * v
sumSq[k] = sumSq.get(k, 0.0) + value * value
decoder.weights[k] = v - eta * value / sqrt(
sumSq[k])
# use gold nodes and edges to calculate P/R/F
num_gold_nodes, num_gold_edges = instance.gold
# P/R/F scores of nodes and edges, for current instance
# Edge recall can not reach %100 since decoding produces only tree structure
intersect_nodes = set(selected_nodes) & set(oracle_nodes)
curr_node_perf = getPRFScores(len(intersect_nodes),
len(selected_nodes),
num_gold_nodes)
logPRFScores('train_node', curr_node_perf)
intersect_edges = set(selected_edges) & set(oracle_edges)
curr_edge_perf = getPRFScores(len(intersect_edges),
len(selected_edges),
num_gold_edges)
logPRFScores('train_edge', curr_edge_perf)
# P/R/F scores of nodes and edges, averaged across all curr_instances
node_perf = PerfScore(
*[sum(x) for x in zip(node_perf, curr_node_perf)])
edge_perf = PerfScore(
*[sum(x) for x in zip(edge_perf, curr_edge_perf)])
logPRFScores(
'train_node_avg',
PerfScore(node_perf.prec / curr_instances,
node_perf.rec / curr_instances,
node_perf.fscore / curr_instances))
logPRFScores(
'train_edge_avg',
PerfScore(edge_perf.prec / curr_instances,
edge_perf.rec / curr_instances,
edge_perf.fscore / curr_instances))
# averaging weight vectors
avg_weights += decoder.weights
# output averaged weight vectors to file
curr_weights = FeatureVector()
curr_weights += avg_weights * (1 / curr_num_passes)
if param_file:
with codecs.open(param_file, 'w', 'utf-8') as outfile:
outfile.write('#curr_num_passes#: %d\n' % curr_num_passes)
outfile.write('%s\n' % curr_weights.toString())
final_weights = FeatureVector()
final_weights += avg_weights * (1 / num_passes)
return final_weights
curr_filename = inst.filename
my_nodes, s_nodes = inst.nodes
my_edges, s_edges = inst.edges
# logger.debug('extracting features for file: %s' % curr_filename)
# for k_edge, v_edge in my_edges.iteritems():
# for tag in [0,1]:
# feat_vec += feat_extr.getEdgeFeats(k_edge, v_edge, tag, curr_filename, my_nodes, my_edges)
logger.debug('extracting features for file: %s' % curr_filename)
for k_node, v_node in my_nodes.iteritems():
for tag in [0,1]:
feat_vec += feat_extr.getNodeFeats(k_node, v_node, tag, curr_filename, my_nodes, my_edges)
with codecs.open('output_file', 'w', 'utf-8') as outfile:
outfile.write('%s\n' % feat_vec.toString())
if __name__ == '__main__':
input_dir = '/Users/user/Data/SemanticSumm/Proxy/gold/split/dev/'
body_file = 'aligned-amr-release-1.0-dev-proxy-body.txt'
summ_file = 'aligned-amr-release-1.0-dev-proxy-summary.txt'
corpus = buildCorpus(os.path.join(input_dir, body_file),
os.path.join(input_dir, summ_file))
feat_extr = FeatureExtractor()
feat_vec = FeatureVector()
for inst in corpus:
curr_filename = inst.filename
my_nodes, s_nodes = inst.nodes
my_edges, s_edges = inst.edges
# logger.debug('extracting features for file: %s' % curr_filename)
# for k_edge, v_edge in my_edges.iteritems():
# for tag in [0,1]:
# feat_vec += feat_extr.getEdgeFeats(k_edge, v_edge, tag, curr_filename, my_nodes, my_edges)
logger.debug('extracting features for file: %s' % curr_filename)
for k_node, v_node in my_nodes.iteritems():
for tag in [0, 1]:
feat_vec += feat_extr.getNodeFeats(k_node, v_node, tag,
curr_filename, my_nodes,
my_edges)
with codecs.open('output_file', 'w', 'utf-8') as outfile:
outfile.write('%s\n' % feat_vec.toString())
def learnParamsAdaGrad(self, decoder, corpus, param_file, loss_func, num_passes=10, oracle_len='nolen'):
"""
learn parameters using Structured Perceptron, Ramp Loss, (Hinge??)
"""
logger.debug('start learning parameters...')
shuffle(corpus) # shuffle corpus
avg_weights = FeatureVector()
curr_instances = 0
node_perf = PerfScore(0.0, 0.0, 0.0) # node performance
edge_perf = PerfScore(0.0, 0.0, 0.0) # edge performance
eta = 1.0 # stepsize
l2reg = 0.0 #
node_cost_scaling = 1.0 # cost scaling factor
edge_cost_scaling = 1.0 # cost scaling factor
sumSq = FeatureVector()
for curr_num_passes in xrange(1, num_passes+1):
logger.debug('#curr_num_passes#: %d' % curr_num_passes)
for instance in corpus:
curr_instances += 1
logger.debug('processing instance %d...' % curr_instances)
# perceptron loss
if loss_func.startswith('perceptron'):
gradient, selected_nodes, selected_edges, score_pred = decoder.decode(instance, oracle_len)
plus_feats, oracle_nodes, oracle_edges, score_true = decoder.oracle(instance)
curr_loss = score_pred - score_true # @UnusedVariable
gradient -= plus_feats
decoder.weights -= eta * gradient
# ramp loss + cost-augmented decoding
if loss_func.startswith('ramp'):
node_cost, edge_cost = node_cost_scaling, edge_cost_scaling
gradient, _, _, score_plus_cost = decoder.decode(instance, oracle_len, node_cost, edge_cost)
plus_feats, selected_nodes, selected_edges, score_minus_cost = decoder.decode(instance, oracle_len, -1.0 * node_cost, -1.0 * edge_cost)
_, oracle_nodes, oracle_edges, score_true = decoder.oracle(instance)
curr_loss = score_plus_cost - score_minus_cost # @UnusedVariable
gradient -= plus_feats
for k, v in gradient.iteritems():
if v == 0.0: continue
sumSq[k] = sumSq.get(k, 0.0) + v * v
decoder.weights[k] = decoder.weights.get(k, 0.0) - eta * v / sqrt(sumSq[k])
if l2reg != 0.0:
for k, v in decoder.weights.iteritems():
if v == 0.0: continue
value = l2reg * v
sumSq[k] = sumSq.get(k, 0.0) + value * value
decoder.weights[k] = v - eta * value / sqrt(sumSq[k])
# hinge-loss + cost-augmented decoding
if loss_func.startswith('hinge'):
node_cost, edge_cost = node_cost_scaling, edge_cost_scaling
gradient, selected_nodes, selected_edges, score_plus_cost = decoder.decode(instance, oracle_len, node_cost, edge_cost)
plus_feats, oracle_nodes, oracle_edges, score_true = decoder.oracle(instance)
curr_loss = score_plus_cost - score_true # @UnusedVariable
gradient -= plus_feats
for k, v in gradient.iteritems():
if v == 0.0: continue
sumSq[k] = sumSq.get(k, 0.0) + v * v
decoder.weights[k] = decoder.weights.get(k, 0.0) - eta * v / sqrt(sumSq[k])
if l2reg != 0.0:
for k, v in decoder.weights.iteritems():
if v == 0.0: continue
value = l2reg * v
sumSq[k] = sumSq.get(k, 0.0) + value * value
decoder.weights[k] = v - eta * value / sqrt(sumSq[k])
# use gold nodes and edges to calculate P/R/F
num_gold_nodes, num_gold_edges = instance.gold
# P/R/F scores of nodes and edges, for current instance
# Edge recall can not reach %100 since decoding produces only tree structure
intersect_nodes = set(selected_nodes) & set(oracle_nodes)
curr_node_perf = getPRFScores(len(intersect_nodes), len(selected_nodes), num_gold_nodes)
logPRFScores('train_node', curr_node_perf)
intersect_edges = set(selected_edges) & set(oracle_edges)
curr_edge_perf = getPRFScores(len(intersect_edges), len(selected_edges), num_gold_edges)
logPRFScores('train_edge', curr_edge_perf)
# P/R/F scores of nodes and edges, averaged across all curr_instances
node_perf = PerfScore(*[sum(x) for x in zip(node_perf, curr_node_perf)])
edge_perf = PerfScore(*[sum(x) for x in zip(edge_perf, curr_edge_perf)])
logPRFScores('train_node_avg',
PerfScore(node_perf.prec/curr_instances, node_perf.rec/curr_instances,
node_perf.fscore/curr_instances))
logPRFScores('train_edge_avg',
PerfScore(edge_perf.prec/curr_instances, edge_perf.rec/curr_instances,
edge_perf.fscore/curr_instances))
# averaging weight vectors
avg_weights += decoder.weights
# output averaged weight vectors to file
curr_weights = FeatureVector()
curr_weights += avg_weights * (1/curr_num_passes)
if param_file:
with codecs.open(param_file, 'w', 'utf-8') as outfile:
outfile.write('#curr_num_passes#: %d\n' % curr_num_passes)
outfile.write('%s\n' % curr_weights.toString())
final_weights = FeatureVector()
final_weights += avg_weights * (1/num_passes)
return final_weights
