def eval_files(gold_file, test_file):
gold_trees = PhraseTree.load_treefile(gold_file)
test_trees = PhraseTree.load_treefile(test_file)
accuracy = FScore()
match_gold_trees = []
match_pred_trees = []
umatch_gold_trees = []
umatch_pred_trees = []
for gold, test in zip(gold_trees, test_trees):
if len(gold) == len(test):
match_gold_trees.append(gold)
match_pred_trees.append(test)
else:
umatch_gold_trees.append(gold)
umatch_pred_trees.append(test)
print("***eval matched pair***")
struct, label = eval_trees(match_gold_trees,
match_pred_trees,
verbose=True)
accuracy += label
print("***eval unmatched pair***")
_, label = eval_trees(umatch_gold_trees, umatch_pred_trees, verbose=True)
accuracy += label
print("sum struct score:", struct + _)
return accuracy
def vocab_init(fname, verbose=True):
"""
Learn vocabulary from file of strings.
"""
tag_freq = defaultdict(int)
trees = PhraseTree.load_treefile(fname)
for i, tree in enumerate(trees):
for (word, tag) in tree.sentence:
tag_freq[tag] += 1
if verbose:
print('\rTree {}'.format(i), end='')
sys.stdout.flush()
if verbose:
print('\r', end='')
tags = ['XX'] + sorted(tag_freq)
tdict = OrderedDict((t, i) for (i, t) in enumerate(tags))
if verbose:
print('Loading features from {}'.format(fname))
print('( {} tags)'.format(len(tdict), ))
return {
'tdict': tdict,
}
def vocab_init(fname, verbose=True):
"""
Learn vocabulary from file of strings.
"""
word_freq = defaultdict(int)
tag_freq = defaultdict(int)
label_freq = defaultdict(int)
trees = PhraseTree.load_treefile(fname)
for i, tree in enumerate(trees):
for (word, tag) in tree.sentence:
word_freq[word] += 1
tag_freq[tag] += 1
for action in Parser.gold_actions(tree):
if action.startswith('label-'):
label = action[6:]
label_freq[label] += 1
if verbose:
print('\rTree {}'.format(i), end='')
sys.stdout.flush()
if verbose:
print('\r', end='')
words = [
FeatureMapper.UNK,
FeatureMapper.START,
FeatureMapper.STOP,
] + sorted(word_freq)
wdict = OrderedDict((w,i) for (i,w) in enumerate(words))
tags = [
FeatureMapper.UNK,
FeatureMapper.START,
FeatureMapper.STOP,
] + sorted(tag_freq)
tdict = OrderedDict((t,i) for (i,t) in enumerate(tags))
labels = sorted(label_freq)
ldict = OrderedDict((l,i) for (i,l) in enumerate(labels))
if verbose:
print('Loading features from {}'.format(fname))
print('({} words, {} tags, {} nonterminal-chains)'.format(
len(wdict),
len(tdict),
len(ldict),
))
return {
'wdict': wdict,
'word_freq': word_freq,
'tdict': tdict,
'ldict': ldict,
}
def extract_origin_grammar(tree_file, out_file="grammar.out"):
grammar_dict = defaultdict(int)
trees = PhraseTree.load_treefile(tree_file)
for tree in trees:
tree.grammar(grammar_dict)
grammar_list = [grammar for grammar, val in grammar_dict.items()]
write_docs(fname=out_file, docs=grammar_list)
return grammar_dict
def gold_data_from_file(self, fname):
"""
Static oracle for file.
"""
trees = PhraseTree.load_treefile(fname)
result = []
for tree in trees:
sentence_data = self.gold_data(tree)
result.append(sentence_data)
return result
def train(
feature_mapper,
word_dims,
tag_dims,
lstm_units,
hidden_units,
epochs,
batch_size,
train_data_file,
dev_data_file,
model_save_file,
droprate,
unk_param,
alpha=1.0,
beta=0.0,
):
start_time = time.time()
fm = feature_mapper
word_count = fm.total_words()
tag_count = fm.total_tags()
network = Network(
word_count=word_count,
tag_count=tag_count,
word_dims=word_dims,
tag_dims=tag_dims,
lstm_units=lstm_units,
hidden_units=hidden_units,
struct_out=2,
label_out=fm.total_label_actions(),
droprate=droprate,
)
network.init_params()
print('Hidden units: {}, per-LSTM units: {}'.format(
hidden_units,
lstm_units,
))
print('Embeddings: word={} tag={}'.format(
(word_count, word_dims),
(tag_count, tag_dims),
))
print('Dropout rate: {}'.format(droprate))
print('Parameters initialized in [-0.01, 0.01]')
print('Random UNKing parameter z = {}'.format(unk_param))
print('Exploration: alpha={} beta={}'.format(alpha, beta))
training_data = fm.gold_data_from_file(train_data_file)
num_batches = -(-len(training_data) // batch_size)
print('Loaded {} training sentences ({} batches of size {})!'.format(
len(training_data),
num_batches,
batch_size,
))
parse_every = -(-num_batches // 4)
dev_trees = PhraseTree.load_treefile(dev_data_file)
print('Loaded {} validation trees!'.format(len(dev_trees)))
best_acc = FScore()
for epoch in xrange(1, epochs + 1):
print('........... epoch {} ...........'.format(epoch))
total_cost = 0.0
total_states = 0
training_acc = FScore()
np.random.shuffle(training_data)
for b in xrange(num_batches):
batch = training_data[(b * batch_size):((b + 1) * batch_size)]
explore = [
Parser.exploration(
example,
fm,
network,
alpha=alpha,
beta=beta,
) for example in batch
]
for (_, acc) in explore:
training_acc += acc
batch = [example for (example, _) in explore]
dynet.renew_cg()
network.prep_params()
errors = []
for example in batch:
## random UNKing ##
for (i, w) in enumerate(example['w']):
if w <= 2:
continue
freq = fm.word_freq_list[w]
drop_prob = unk_param / (unk_param + freq)
r = np.random.random()
if r < drop_prob:
example['w'][i] = 0
fwd, back = network.evaluate_recurrent(
example['w'],
example['t'],
)
for (left,
right), correct in example['struct_data'].items():
scores = network.evaluate_struct(
fwd, back, left, right)
probs = dynet.softmax(scores)
loss = -dynet.log(dynet.pick(probs, correct))
errors.append(loss)
total_states += len(example['struct_data'])
for (left,
right), correct in example['label_data'].items():
scores = network.evaluate_label(fwd, back, left, right)
probs = dynet.softmax(scores)
loss = -dynet.log(dynet.pick(probs, correct))
errors.append(loss)
total_states += len(example['label_data'])
batch_error = dynet.esum(errors)
total_cost += batch_error.scalar_value()
batch_error.backward()
network.trainer.update()
mean_cost = total_cost / total_states
print(
'\rBatch {} Mean Cost {:.4f} [Train: {}]'.format(
b,
mean_cost,
training_acc,
),
end='',
)
sys.stdout.flush()
if ((b + 1) % parse_every) == 0 or b == (num_batches - 1):
dev_acc = Parser.evaluate_corpus(
dev_trees,
fm,
network,
)
print(' [Val: {}]'.format(dev_acc))
if dev_acc > best_acc:
best_acc = dev_acc
network.save(model_save_file)
print(' [saved model: {}]'.format(model_save_file))
current_time = time.time()
runmins = (current_time - start_time) / 60.
print(' Elapsed time: {:.2f}m'.format(runmins))
else:
print('Must specify either --vocab-file or --train-data.')
print(' (Use -h or --help flag for full option list.)')
sys.exit()
if args.model is None:
print('Must specify --model or (or --write-vocab) parameter.')
print(' (Use -h or --help flag for full option list.)')
sys.exit()
if args.test is not None:
from phrase_tree import PhraseTree
from network import Network
from parser import Parser
test_trees = PhraseTree.load_treefile(args.test)
print('Loaded test trees from {}'.format(args.test))
network = Network.load(args.model)
print('Loaded model from: {}'.format(args.model))
accuracy = Parser.evaluate_corpus(test_trees, fm, network)
print('Accuracy: {}'.format(accuracy))
elif args.train is not None:
from network import Network
if args.np_seed is not None:
import numpy as np
np.random.seed(args.np_seed)
print('L2 regularization: {}'.format(args.dynet_l2))
Network.train(feature_mapper=fm,
import argparse
from phrase_tree import PhraseTree, FScore
parser = argparse.ArgumentParser()
parser.add_argument('--gold', type=str)
parser.add_argument('--pred', type=str)
args = parser.parse_args()
gold_trees = PhraseTree.load_treefile(args.gold)
pred_trees = PhraseTree.load_treefile(args.pred)
accuracy = FScore()
for gold, pred in zip(gold_trees, pred_trees):
local_accuracy = pred.compare(gold)
accuracy += local_accuracy
print(accuracy)
def gold_data_from_file(self, fname):
"""
Static oracle for file.
"""
trees = PhraseTree.load_treefile(fname)
return self.gold_data_from_trees(trees)
def train(
feature_mapper,
word_dims,
tag_dims,
lstm_units,
hidden_units,
epochs,
batch_size,
train_data_file,
dev_data_file,
model_save_file,
droprate,
unk_param,
alpha=1.0,
beta=0.0,
GPU=None,
):
start_time = time.time()
fm = feature_mapper
word_count = fm.total_words()
tag_count = fm.total_tags()
network = Network(
word_count=word_count,
tag_count=tag_count,
word_dims=word_dims,
tag_dims=tag_dims,
lstm_units=lstm_units,
hidden_units=hidden_units,
struct_out=2,
label_out=fm.total_label_actions(),
droprate=droprate,
GPU=GPU,
)
f_loss = nn.CrossEntropyLoss(size_average=False)
if GPU is not None:
f_loss = f_loss.cuda(GPU)
random.seed(1)
torch.manual_seed(1)
print('Hidden units: {}, per-LSTM units: {}'.format(
hidden_units,
lstm_units,
))
print('Embeddings: word={} tag={}'.format(
(word_count, word_dims),
(tag_count, tag_dims),
))
print('Dropout rate: {}'.format(droprate))
print('Parameters initialized in [-0.01, 0.01]')
print('Random UNKing parameter z = {}'.format(unk_param))
print('Exploration: alpha={} beta={}'.format(alpha, beta))
training_data = fm.gold_data_from_file(train_data_file)
num_batches = -(-len(training_data) // batch_size)
print('Loaded {} training sentences ({} batches of size {})!'.format(
len(training_data),
num_batches,
batch_size,
))
parse_every = -(-num_batches // 4)
dev_trees = PhraseTree.load_treefile(dev_data_file)
print('Loaded {} validation trees!'.format(len(dev_trees)))
best_acc = FScore()
network.init_hidden()
for epoch in xrange(1, epochs + 1):
print('........... epoch {} ...........'.format(epoch))
total_cost = 0.0
total_states = 0
training_acc = FScore()
np.random.shuffle(training_data)
for b in xrange(num_batches):
network.trainer.zero_grad()
batch = training_data[(b * batch_size):((b + 1) * batch_size)]
explore = [
Parser.exploration(
example,
fm,
network,
alpha=alpha,
beta=beta,
) for example in batch
]
for (_, acc) in explore:
training_acc += acc
batch = [example for (example, _) in explore]
errors = []
network.init_hidden()
for example in batch:
## random UNKing ##
for (i, w) in enumerate(example['w']):
if w <= 2:
continue
freq = fm.word_freq_list[w]
drop_prob = unk_param / (unk_param + freq)
r = np.random.random()
if r < drop_prob:
example['w'][i] = 0
fwd, back = network.evaluate_recurrent(
example['w'],
example['t'],
)
indices, targets = [], []
for (left,
right), correct in example['struct_data'].items():
indices.append((left, right))
targets.append(correct)
"""
print(example['w'])
print(indices)
print(targets)
raw_input()
"""
targets = autograd.Variable(torch.LongTensor(targets))
if network.GPU is not None:
targets = targets.cuda(network.GPU)
scores = network.evaluate_struct(fwd, back, indices)
for i in xrange(len(targets)):
score = scores[i]
target = targets[i]
loss = f_loss(score, target)
errors.append(loss)
total_states += len(example['struct_data'])
indices, targets = [], []
for (left,
right), correct in example['label_data'].items():
indices.append((left, right))
targets.append(correct)
targets = autograd.Variable(torch.LongTensor(targets))
if network.GPU is not None:
targets = targets.cuda(network.GPU)
scores = network.evaluate_label(fwd, back, indices)
for i in xrange(len(targets)):
score = scores[i]
target = targets[i]
loss = f_loss(score, target)
errors.append(loss)
total_states += len(example['label_data'])
batch_loss = torch.sum(torch.cat(errors))
#network.trainer.zero_grad()
batch_loss.backward()
network.trainer.step()
total_cost += batch_loss.data[0]
mean_cost = (total_cost / total_states)
print(
'\rBatch {} Mean Cost {:.4f} [Train: {}]'.format(
b,
mean_cost,
training_acc,
),
end='',
)
sys.stdout.flush()
if ((b + 1) % parse_every) == 0 or b == (num_batches - 1):
dev_acc = Parser.evaluate_corpus(
dev_trees,
fm,
network,
)
print(' [Val: {}]'.format(dev_acc))
if dev_acc > best_acc:
best_acc = dev_acc
network.save(model_save_file)
print(' [saved model: {}]'.format(model_save_file))
current_time = time.time()
runmins = (current_time - start_time) / 60.
print(' Elapsed time: {:.2f}m'.format(runmins))