def next(self):
seqs = []
try:
while len(seqs) < self.batch_size:
line = next(self.txt_file).strip()
seq, _ = parse_input(self.state, self.indx, line, raise_unk=self.raise_unk,
unk_sym=self.unk_sym, null_sym=self.null_sym)
seqs.append(seq)
return self._pack(seqs)
except StopIteration:
if not seqs:
raise StopIteration()
return self._pack(seqs)
def main():
args = parse_args()
state = prototype_state()
with open(args.state) as src:
state.update(cPickle.load(src))
state.update(eval("dict({})".format(args.changes)))
state['sort_k_batches'] = 1
state['shuffle'] = False
state['use_infinite_loop'] = False
state['force_enc_repr_cpu'] = False
logging.basicConfig(level=getattr(logging, state['level']), format="%(asctime)s: %(name)s: %(levelname)s: %(message)s")
rng = numpy.random.RandomState(state['seed'])
enc_dec = RNNEncoderDecoder(state, rng, skip_init=True, compute_alignment=True)
enc_dec.build()
lm_model = enc_dec.create_lm_model()
lm_model.load(args.model_path)
indx_word_src = cPickle.load(open(state['word_indx'],'rb'))
indx_word_trgt = cPickle.load(open(state['word_indx_trgt'], 'rb'))
if args.mode == "batch":
data_given = args.src or args.trg
txt = data_given and not (args.src.endswith(".h5") and args.trg.endswith(".h5"))
if data_given and not txt:
state['source'] = [args.src]
state['target'] = [args.trg]
if not data_given and not txt:
logger.info("Using the training data")
if txt:
data_iter = BatchBiTxtIterator(state,
args.src, indx_word_src, args.trg, indx_word_trgt,
state['bs'], raise_unk=not args.allow_unk)
data_iter.start()
else:
data_iter = get_batch_iterator(state)
data_iter.start(0)
score_file = open(args.scores, "w") if args.scores else sys.stdout
scorer = enc_dec.create_scorer(batch=True)
count = 0
n_samples = 0
logger.info('Scoring phrases')
for i, batch in enumerate(data_iter):
if batch == None:
continue
if args.n_batches >= 0 and i == args.n_batches:
break
if args.y_noise:
y = batch['y']
random_words = numpy.random.randint(0, 100, y.shape).astype("int64")
change_mask = numpy.random.binomial(1, args.y_noise, y.shape).astype("int64")
y = change_mask * random_words + (1 - change_mask) * y
batch['y'] = y
st = time.time()
[scores] = scorer(batch['x'], batch['y'],
batch['x_mask'], batch['y_mask'])
if args.print_probs:
scores = numpy.exp(scores)
up_time = time.time() - st
for s in scores:
print >>score_file, "{:.5e}".format(float(s))
n_samples += batch['x'].shape[1]
count += 1
if count % 100 == 0:
score_file.flush()
logger.debug("Scores flushed")
logger.debug("{} batches, {} samples, {} per sample; example scores: {}".format(
count, n_samples, up_time/scores.shape[0], scores[:5]))
logger.info("Done")
score_file.flush()
elif args.mode == "interact":
scorer = enc_dec.create_scorer()
while True:
try:
compute_probs = enc_dec.create_probs_computer()
src_line = raw_input('Source sequence: ')
trgt_line = raw_input('Target sequence: ')
src_seq = parse_input(state, indx_word_src, src_line, raise_unk=not args.allow_unk,
unk_sym=state['unk_sym_source'], null_sym=state['null_sym_source'])
trgt_seq = parse_input(state, indx_word_trgt, trgt_line, raise_unk=not args.allow_unk,
unk_sym=state['unk_sym_target'], null_sym=state['null_sym_target'])
print "Binarized source: ", src_seq
print "Binarized target: ", trgt_seq
probs = compute_probs(src_seq, trgt_seq)
print "Probs: {}, cost: {}".format(probs, -numpy.sum(numpy.log(probs)))
except Exception:
traceback.print_exc()
elif args.mode == "txt":
assert args.src and args.trg
scorer = enc_dec.create_scorer()
src_file = open(args.src, "r")
trg_file = open(args.trg, "r")
compute_probs = enc_dec.create_probs_computer(return_alignment=True)
try:
numpy.set_printoptions(precision=3, linewidth=150, suppress=True)
i = 0
while True:
src_line = next(src_file).strip()
trgt_line = next(trg_file).strip()
src_seq, src_words = parse_input(state,
indx_word_src, src_line, raise_unk=not args.allow_unk,
unk_sym=state['unk_sym_source'], null_sym=state['null_sym_source'])
trgt_seq, trgt_words = parse_input(state,
indx_word_trgt, trgt_line, raise_unk=not args.allow_unk,
unk_sym=state['unk_sym_target'], null_sym=state['null_sym_target'])
probs, alignment = compute_probs(src_seq, trgt_seq)
if args.verbose:
print "Probs: ", probs.flatten()
if alignment.ndim == 3:
print "Alignment:".ljust(20), src_line, "<eos>"
for i, word in enumerate(trgt_words):
print "{}{}".format(word.ljust(20), alignment[i, :, 0])
print "Generated by:"
for i, word in enumerate(trgt_words):
j = numpy.argmax(alignment[i, :, 0])
print "{} <--- {}".format(word,
src_words[j] if j < len(src_words) else "<eos>")
i += 1
if i % 100 == 0:
sys.stdout.flush()
logger.debug(i)
print -numpy.sum(numpy.log(probs))
except StopIteration:
pass
else:
raise Exception("Unknown mode {}".format(args.mode))
def main():
args = parse_args()
state = getattr(experiments.nmt, args.state_fn)()
if hasattr(args, 'state') and args.state:
with open(args.state) as src:
state.update(cPickle.load(src))
state.update(eval("dict({})".format(args.changes)))
assert state['enc_rec_layer'] == "RecursiveConvolutionalLayer", "Only works with gated recursive convolutional encoder"
logging.basicConfig(level=getattr(logging, state['level']), format="%(asctime)s: %(name)s: %(levelname)s: %(message)s")
rng = numpy.random.RandomState(state['seed'])
enc_dec = RNNEncoderDecoder(state, rng, skip_init=True)
enc_dec.build()
lm_model = enc_dec.create_lm_model()
lm_model.load(args.model_path)
indx_word = cPickle.load(open(state['word_indx'],'rb'))
idict_src = cPickle.load(open(state['indx_word'],'r'))
x = TT.lvector()
h = TT.tensor3()
proj_x = theano.function([x], enc_dec.encoder.input_embedders[0](
enc_dec.encoder.approx_embedder(x)).out, name='proj_x')
new_h, gater = enc_dec.encoder.transitions[0].step_fprop(
None, h, return_gates = True)
step_up = theano.function([h], [new_h, gater], name='gater_step')
while True:
try:
seqin = raw_input('Input Sequence: ')
seq,parsed_in = parse_input(state, indx_word, seqin, idx2word=idict_src)
print "Parsed Input:", parsed_in
except Exception:
print "Exception while parsing your input:"
traceback.print_exc()
continue
# get the initial embedding
new_h = proj_x(seq)
new_h = new_h.reshape(new_h.shape[0], 1, new_h.shape[1])
nodes = numpy.arange(len(seq)).tolist()
node_idx = len(seq)-1
rules = []
nodes_level = copy.deepcopy(nodes)
G = nx.DiGraph()
input_nodes = []
merge_nodes = []
aggregate_nodes = []
nidx = 0
vpos = 0
nodes_pos = {}
nodes_labels = {}
# input nodes
for nn in nodes[:-1]:
nidx += 1
G.add_node(nn, pos=(nidx, 0), ndcolor="blue", label="%d"%nn)
nodes_pos[nn] = (nidx, vpos)
nodes_labels[nn] = idict_src[seq[nidx-1]]
input_nodes.append(nn)
node_idx = len(seq) - 1
vpos += 6
for dd in xrange(len(seq)-1):
new_h, gater = step_up(new_h)
decisions = numpy.argmax(gater, -1)
new_nodes_level = numpy.zeros(len(seq) - (dd+1))
hpos = float(len(seq)+1) - 0.5 * (dd+1)
last_node = True
for nn in xrange(len(seq)-(dd+1)):
hpos -= 1
if not last_node:
# merge nodes
node_idx += 1
G.add_node(node_idx, ndcolor="red", label="m")
nodes_labels[node_idx] = ""
nodes_pos[node_idx] = (hpos, vpos)
G.add_edge(nodes_level[-(nn+1)], node_idx, weight=gater[-(nn+1),0,0])
G.add_edge(nodes_level[-(nn+2)], node_idx, weight=gater[-(nn+1),0,0])
merge_nodes.append(node_idx)
merge_node = node_idx
# linear aggregation nodes
node_idx += 1
G.add_node(node_idx, ndcolor="red", label="")
nodes_labels[node_idx] = "$+$"
nodes_pos[node_idx] = (hpos, vpos+6)
G.add_edge(merge_node, node_idx, weight=gater[-(nn+1),0,0])
G.add_edge(nodes_level[-(nn+2)], node_idx, weight=gater[-(nn+1),0,1])
G.add_edge(nodes_level[-(nn+1)], node_idx, weight=gater[-(nn+1),0,2])
aggregate_nodes.append(node_idx)
new_nodes_level[-(nn+1)] = node_idx
last_node = False
nodes_level = copy.deepcopy(new_nodes_level)
vpos += 12
# TODO: Show only strong edges.
threshold = float(raw_input('Threshold: '))
edges = [(u,v,d) for (u,v,d) in G.edges(data=True) if d['weight'] > threshold]
#edges = G.edges(data=True)
use_weighting = raw_input('Color according to weight [Y/N]: ')
if use_weighting == 'Y':
cm = plt.get_cmap('binary')
cNorm = colors.Normalize(vmin=0., vmax=1.)
scalarMap = cmx.ScalarMappable(norm=cNorm, cmap=cm)
colorList = [scalarMap.to_rgba(d['weight']) for (u,v,d) in edges]
else:
colorList = 'k'
nx.draw_networkx_nodes(G, pos=nodes_pos, nodelist=input_nodes, node_color='white', alpha=1., edge_color='white')
nx.draw_networkx_nodes(G, pos=nodes_pos, nodelist=merge_nodes, node_color='blue', alpha=0.8, node_size=20)
nx.draw_networkx_nodes(G, pos=nodes_pos, nodelist=aggregate_nodes, node_color='red', alpha=0.8, node_size=80)
nx.draw_networkx_edges(G, pos=nodes_pos, edge_color=colorList, edgelist=edges)
nx.draw_networkx_labels(G,pos=nodes_pos,labels=nodes_labels,font_family='sans-serif')
plt.axis('off')
figname = raw_input('Save to: ')
if figname[-3:] == "pdf":
plt.savefig(figname, type='pdf')
else:
plt.savefig(figname)
plt.close()
G.clear()
def main():
args = parse_args()
state = prototype_state()
with open(args.state) as src:
state.update(cPickle.load(src))
state.update(eval("dict({})".format(args.changes)))
logging.basicConfig(level=getattr(logging, state['level']), format="%(asctime)s: %(name)s: %(levelname)s: %(message)s")
rng = numpy.random.RandomState(state['seed'])
enc_dec = RNNEncoderDecoder(state, rng, skip_init=True)
enc_dec.build()
lm_model = enc_dec.create_lm_model()
lm_model.load(args.model_path)
indx_word = cPickle.load(open(state['word_indx'],'rb'))
sampler = None
beam_search = None
if args.beam_search:
beam_search = BeamSearch(enc_dec)
beam_search.compile()
else:
sampler = enc_dec.create_sampler(many_samples=True)
idict_src = cPickle.load(open(state['indx_word'],'r'))
if args.source and args.trans:
# Actually only beam search is currently supported here
assert beam_search
assert args.beam_size
fsrc = open(args.source, 'r')
ftrans = open(args.trans, 'w')
start_time = time.time()
n_samples = args.beam_size
total_cost = 0.0
logging.debug("Beam size: {}".format(n_samples))
for i, line in enumerate(fsrc):
seqin = line.strip()
seq, parsed_in = parse_input(state, indx_word, seqin, idx2word=idict_src)
if args.verbose:
print "Parsed Input:", parsed_in
trans, costs, _ = sample(lm_model, seq, n_samples, sampler=sampler,
beam_search=beam_search, ignore_unk=args.ignore_unk, normalize=args.normalize)
best = numpy.argmin(costs)
print >>ftrans, trans[best]
if args.verbose:
print "Translation:", trans[best]
total_cost += costs[best]
if (i + 1) % 100 == 0:
ftrans.flush()
logger.debug("Current speed is {} per sentence".
format((time.time() - start_time) / (i + 1)))
print "Total cost of the translations: {}".format(total_cost)
fsrc.close()
ftrans.close()
else:
while True:
try:
seqin = raw_input('Input Sequence: ')
n_samples = int(raw_input('How many samples? '))
alpha = None
if not args.beam_search:
alpha = float(raw_input('Inverse Temperature? '))
seq,parsed_in = parse_input(state, indx_word, seqin, idx2word=idict_src)
print "Parsed Input:", parsed_in
except Exception:
print "Exception while parsing your input:"
traceback.print_exc()
continue
sample(lm_model, seq, n_samples, sampler=sampler,
beam_search=beam_search,
ignore_unk=args.ignore_unk, normalize=args.normalize,
alpha=alpha, verbose=True)
def process_sentence(source_sentence, model, max_phrase_length, n_samples,
copy_UNK_words, add_period, normalize, reverse_score):
#Setting up comp_score function
logger.debug("setting up comp_score function")
[lm_model, enc_dec, indx_word_src, indx_word_trgt, state, \
lm_model_fr_2_en, enc_dec_fr_2_en, state_fr2en] = model
eol_src = state['null_sym_source']
src_seq = parse_input(state, indx_word_src, source_sentence)
if src_seq[-1] == eol_src:
src_seq = src_seq[:-1]
n_s = len(src_seq)
#Create sorted phrase lists
tiled_source_phrase_list = []
index_order_list = []
for i in xrange(n_s):
for j in numpy.arange(i, min(i+max_phrase_length, n_s)):
index_order_list.append([i, j])
logger.debug("sorting list")
index_order_list.sort(key=lambda (i, j): (j - i))
logger.debug("creating phrase lists")
if add_period:
period_src = indx_word_src['.']
for i, j in index_order_list:
tiled_source_phrase_list.append(numpy.hstack((src_seq[i:j+1], period_src, eol_src)))
else:
for i, j in index_order_list:
tiled_source_phrase_list.append(numpy.hstack((src_seq[i:j+1], eol_src)))
#Compute nested score dictionary
logger.debug("computing nested score dictionary")
score_dict = {}
trans = {}
for phrase_idx in xrange(0, len(index_order_list)):
logger.debug("{0} out of {1}".format(phrase_idx, len(index_order_list)))
i, j = index_order_list[phrase_idx]
logger.debug("Translating phrase : {}".format(" ".join(source_sentence.strip().split()[i:j+1])))
if copy_UNK_words == True:
phrase_to_translate = tiled_source_phrase_list[phrase_idx]
n_UNK_words = numpy.sum([word == 1 for word in phrase_to_translate])
if n_UNK_words >= 1 and n_UNK_words == len(phrase_to_translate) - 1:
suggested_translation = " ".join(source_sentence.strip().split()[i:j+1])
trans[i, j] = suggested_translation
score = .0001
score_dict[i, j] = score
if n_UNK_words >= 1 and n_UNK_words != len(phrase_to_translate) -1:
suggested_translation = "WILL NOT BE USED"
trans[i, j] = suggested_translation
score = 1e9
score_dict[i, j] = score
if n_UNK_words == 0:
suggested_translation, score = sample_targets(
input_phrase= \
tiled_source_phrase_list[phrase_idx],
model=model,
n_samples=n_samples,
reverse_score=reverse_score,
normalize=normalize
)
trans[i, j] = suggested_translation
score_dict[i, j] = score
else:
phrase_to_translate = tiled_source_phrase_list[phrase_idx]
suggested_translation, score = sample_targets(
input_phrase=phrase_to_translate,
model=model, n_samples=n_samples,
reverse_score=reverse_score,
normalize=normalize
)
trans[i, j] = suggested_translation
score_dict[i, j] = score
#Remove the period at the end if not last word
#Lower case first word if not first word
if add_period:
for phrase_idx in xrange(0, len(index_order_list)):
i, j = index_order_list[phrase_idx]
if i != 0:
trans[i, j] = " ".join([trans[i,j][0].lower()] + [trans[i,j][1:]])
if j != len(src_seq) - 1:
last_word = trans[i,j].strip().split()[-1]
if last_word == '.':
trans[i,j] = " ".join(trans[i,j].strip().split()[:-1])
#Translation of full sentence without segmentation
logger.debug("Translating full sentence")
phrase_to_translate = numpy.hstack((src_seq, eol_src))
full_translation, __ = sample_targets(input_phrase=phrase_to_translate,
model=model,
n_samples=n_samples,
reverse_score=reverse_score,
normalize=normalize
)
logger.debug("Translation output:".format(full_translation))
return trans, score_dict, full_translation
