def main():
args = parse_args()
state = prototype_phrase_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")
server_address = ('', args.port)
httpd = ThreadedHTTPServer(server_address, MTReqHandler)
#httpd = BaseHTTPServer.HTTPServer(server_address, MTReqHandler)
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'))
tokenizer_cmd = [os.getcwd() + '/tokenizer.perl', '-l', 'en', '-q', '-']
detokenizer_cmd = [
os.getcwd() + '/detokenizer.perl', '-l', 'fr', '-q', '-'
]
sampler = Sampler(state,
lm_model,
indx_word,
idict_src,
beam_search=beam_search,
tokenizer_cmd=tokenizer_cmd,
detokenizer_cmd=detokenizer_cmd)
httpd.sampler = sampler
print 'Server starting..'
httpd.serve_forever()
'''
def main():
args = parse_args()
state = prototype_phrase_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"
)
server_address = ("", args.port)
httpd = BaseHTTPServer.HTTPServer(server_address, MTReqHandler)
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"))
tokenizer_cmd = [os.getcwd() + "/tokenizer.perl", "-l", "en", "-q", "-"]
detokenizer_cmd = [os.getcwd() + "/detokenizer.perl", "-l", "fr", "-q", "-"]
sampler = Sampler(
state,
lm_model,
indx_word,
idict_src,
beam_search=beam_search,
tokenizer_cmd=tokenizer_cmd,
detokenizer_cmd=detokenizer_cmd,
)
httpd.sampler = sampler
print "Server starting.."
httpd.serve_forever()
"""
def main():
args = parse_args()
state = prototype_phrase_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")
if 'rolling_vocab' not in state:
state['rolling_vocab'] = 0
if 'save_algo' not in state:
state['save_algo'] = 0
if 'save_gs' not in state:
state['save_gs'] = 0
if 'save_iter' not in state:
state['save_iter'] = -1
if 'var_src_len' not in state:
state['var_src_len'] = False
with open(args.topn_file, 'rb') as f:
topn = cPickle.load(
f
) # Load dictionary (source word index : list of target word indices)
if args.less_transfer:
for elt in topn:
topn[elt] = topn[
elt][:args.num_ttables] # Take the first args.num_ttables only
else:
for elt in topn:
topn[elt] = set(
topn[elt][:args.num_ttables]
) # Take the first args.num_ttables only and convert list to set
num_models = len(args.models)
rng = numpy.random.RandomState(state['seed'])
enc_decs = []
lm_models = []
original_W_0_dec_approx_embdr = []
original_W2_dec_deep_softmax = []
original_b_dec_deep_softmax = []
for i in xrange(num_models):
enc_decs.append(RNNEncoderDecoder(state, rng, skip_init=True))
enc_decs[i].build()
lm_models.append(enc_decs[i].create_lm_model())
lm_models[i].load(args.models[i])
original_W_0_dec_approx_embdr.append(lm_models[i].params[
lm_models[i].name2pos['W_0_dec_approx_embdr']].get_value())
original_W2_dec_deep_softmax.append(lm_models[i].params[
lm_models[i].name2pos['W2_dec_deep_softmax']].get_value())
original_b_dec_deep_softmax.append(lm_models[i].params[
lm_models[i].name2pos['b_dec_deep_softmax']].get_value())
# On GPU, this will free memory for the next models
# Additional gains could be made by rolling the source vocab
lm_models[i].params[
lm_models[i].name2pos['W_0_dec_approx_embdr']].set_value(
numpy.zeros((1, 1), dtype=numpy.float32))
lm_models[i].params[
lm_models[i].name2pos['W2_dec_deep_softmax']].set_value(
numpy.zeros((1, 1), dtype=numpy.float32))
lm_models[i].params[
lm_models[i].name2pos['b_dec_deep_softmax']].set_value(
numpy.zeros((1), dtype=numpy.float32))
indx_word = cPickle.load(open(state['word_indx'], 'rb')) #Source w2i
sampler = None
beam_search = None
if args.beam_search:
beam_search = BeamSearch(enc_decs)
beam_search.compile()
else:
raise NotImplementedError
#sampler = enc_dec.create_sampler(many_samples=True)
idict_src = cPickle.load(open(state['indx_word'], 'r')) #Source i2w
original_target_i2w = lm_models[0].word_indxs.copy()
# I don't think that we need target_word2index
max_words = len(original_b_dec_deep_softmax[0])
if args.less_transfer:
# Use OrderedDict instead of set for reproducibility
d = OrderedDict() # Up to now
D = OrderedDict() # Full
C = OrderedDict() # Allowed to reject
prev_line = 0
logger.info("%d" % prev_line)
D_dict = OrderedDict()
output = False
for i in xrange(args.num_common):
D[i] = 0
C[i] = 0
null_unk_indices = [state['null_sym_target'], state['unk_sym_target']]
update_dicts(null_unk_indices, d, D, C, args.num_common)
with open(args.source, 'r') as f:
for i, line in enumerate(f):
seqin = line.strip()
seq, parsed_in = parse_input(
state, indx_word, seqin,
idx2word=idict_src) # seq is the ndarray of indices
indices = []
for elt in seq[:-1]: # Exclude the EOL token
if elt != 1: # Exclude OOV (1 will not be a key of topn)
indices.extend(
topn[elt]
) # Add topn best unigram translations for each source word
output = update_dicts(indices, d, D, C, args.num_common)
if (i % args.change_every
) == 0 and args.change_every > 0 and i > 0:
output = True
if output:
D_dict[prev_line] = D.copy(
) # Save dictionary for the lines preceding this one
prev_line = i
logger.info("%d" % i)
output = False
d = OrderedDict()
if args.no_reset:
C = D.copy()
else:
D = OrderedDict() # Full
C = OrderedDict() # Allowed to reject
for i in xrange(args.num_common):
D[i] = 0
C[i] = 0
null_unk_indices = [
state['null_sym_target'], state['unk_sym_target']
]
update_dicts(null_unk_indices, d, D, C, args.num_common)
update_dicts(
indices, d, D, C, args.num_common
) # Assumes you cannot fill d with only 1 line
D_dict[prev_line] = D.copy()
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) # seq is the ndarray of indices
# For now, keep all input words in the model.
# In the future, we may want to filter them to save on memory, but this isn't really much of an issue now
if args.verbose:
print "Parsed Input:", parsed_in
if args.less_transfer:
if i in D_dict:
indices = D_dict[i].keys()
eos_id = indices.index(state['null_sym_target']
) # Find new eos and unk positions
unk_id = indices.index(state['unk_sym_target'])
for j in xrange(num_models):
lm_models[j].params[lm_models[j].name2pos[
'W_0_dec_approx_embdr']].set_value(
original_W_0_dec_approx_embdr[j][indices])
lm_models[j].params[lm_models[j].name2pos[
'W2_dec_deep_softmax']].set_value(
original_W2_dec_deep_softmax[j][:, indices])
lm_models[j].params[lm_models[j].name2pos[
'b_dec_deep_softmax']].set_value(
original_b_dec_deep_softmax[j][indices])
lm_models[0].word_indxs = dict([
(k, original_target_i2w[index])
for k, index in enumerate(indices)
]) # target index2word
trans, costs, _ = sample(lm_models[0],
seq,
n_samples,
sampler=sampler,
beam_search=beam_search,
ignore_unk=args.ignore_unk,
normalize=args.normalize,
normalize_p=args.normalize_p,
eos_id=eos_id,
unk_id=unk_id,
final=True,
wp=args.wp)
else:
# Extract the indices you need
indices = set()
for elt in seq[:-1]: # Exclude the EOL token
if elt != 1: # Exclude OOV (1 will not be a key of topn)
indices = indices.union(
topn[elt]
) # Add topn best unigram translations for each source word
num_common_words = args.num_common
while True:
if num_common_words >= max_words:
final = True
num_common_words = max_words
else:
final = False
if args.final: # No matter the number of words
final = True
indices = indices.union(set(
xrange(num_common_words))) # Add common words
indices = list(
indices) # Convert back to list for advanced indexing
eos_id = indices.index(state['null_sym_target']
) # Find new eos and unk positions
unk_id = indices.index(state['unk_sym_target'])
# Set the target word matrices and biases
for j in xrange(num_models):
lm_models[j].params[lm_models[j].name2pos[
'W_0_dec_approx_embdr']].set_value(
original_W_0_dec_approx_embdr[j][indices])
lm_models[j].params[lm_models[j].name2pos[
'W2_dec_deep_softmax']].set_value(
original_W2_dec_deep_softmax[j][:, indices])
lm_models[j].params[lm_models[j].name2pos[
'b_dec_deep_softmax']].set_value(
original_b_dec_deep_softmax[j][indices])
lm_models[0].word_indxs = dict([
(k, original_target_i2w[index])
for k, index in enumerate(indices)
]) # target index2word
try:
trans, costs, _ = sample(lm_models[0],
seq,
n_samples,
sampler=sampler,
beam_search=beam_search,
ignore_unk=args.ignore_unk,
normalize=args.normalize,
normalize_p=args.normalize_p,
eos_id=eos_id,
unk_id=unk_id,
final=final)
break # Breaks only if it succeeded (If final=True, will always succeed)
except RuntimeError:
indices = set(indices)
num_common_words *= 2
if not args.n_best:
best = numpy.argmin(costs)
print >> ftrans, trans[best]
else:
order = numpy.argsort(costs)
best = order[0]
for elt in order:
print >> ftrans, str(
i + args.start) + ' ||| ' + trans[elt] + ' ||| ' + str(
costs[elt])
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:
raise NotImplementedError
def main():
args = parse_args()
state = prototype_phrase_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")
if 'rolling_vocab' not in state:
state['rolling_vocab'] = 0
if 'save_algo' not in state:
state['save_algo'] = 0
if 'save_gs' not in state:
state['save_gs'] = 0
if 'save_iter' not in state:
state['save_iter'] = -1
if 'var_src_len' not in state:
state['var_src_len'] = False
if args.num_common and args.num_ttables and args.topn_file:
with open(args.topn_file, 'rb') as f:
topn = cPickle.load(f) # Load dictionary (source word index : list of target word indices)
for elt in topn:
topn[elt] = topn[elt][:args.num_ttables] # Take the first args.num_ttables only
num_models = len(args.models)
rng = numpy.random.RandomState(state['seed'])
enc_decs = []
lm_models = []
alignment_fns = []
if args.num_common and args.num_ttables and args.topn_file:
original_W_0_dec_approx_embdr = []
original_W2_dec_deep_softmax = []
original_b_dec_deep_softmax = []
for i in xrange(num_models):
enc_decs.append(RNNEncoderDecoder(state, rng, skip_init=True, compute_alignment=True))
enc_decs[i].build()
lm_models.append(enc_decs[i].create_lm_model())
lm_models[i].load(args.models[i])
alignment_fns.append(theano.function(inputs=enc_decs[i].inputs, outputs=[enc_decs[i].alignment], name="alignment_fn"))
if args.num_common and args.num_ttables and args.topn_file:
original_W_0_dec_approx_embdr.append(lm_models[i].params[lm_models[i].name2pos['W_0_dec_approx_embdr']].get_value())
original_W2_dec_deep_softmax.append(lm_models[i].params[lm_models[i].name2pos['W2_dec_deep_softmax']].get_value())
original_b_dec_deep_softmax.append(lm_models[i].params[lm_models[i].name2pos['b_dec_deep_softmax']].get_value())
lm_models[i].params[lm_models[i].name2pos['W_0_dec_approx_embdr']].set_value(numpy.zeros((1,1), dtype=numpy.float32))
lm_models[i].params[lm_models[i].name2pos['W2_dec_deep_softmax']].set_value(numpy.zeros((1,1), dtype=numpy.float32))
lm_models[i].params[lm_models[i].name2pos['b_dec_deep_softmax']].set_value(numpy.zeros((1), dtype=numpy.float32))
if args.mapping:
with open(args.mapping, 'rb') as f:
mapping = cPickle.load(f)
heuristic = args.heuristic
else:
heuristic = 0
mapping = None
word2idx_src = cPickle.load(open(state['word_indx'], 'rb'))
idict_src = cPickle.load(open(state['indx_word'], 'r'))
word2idx_trg = cPickle.load(open(state['word_indx_trgt'], 'rb'))
idict_trg = cPickle.load(open(state['indx_word_target'], 'r'))
word2idx_trg['<eos>'] = state['null_sym_target']
word2idx_trg[state['oov']] = state['unk_sym_target'] # 'UNK' may be in the vocabulary. Now points to the right index.
idict_trg[state['null_sym_target']] = '<eos>'
idict_trg[state['unk_sym_target']] = state['oov']
if args.num_common and args.num_ttables and args.topn_file:
# Use OrderedDict instead of set for reproducibility
d = OrderedDict() # Up to now
D = OrderedDict() # Full
C = OrderedDict() # Allowed to reject
prev_line = 0
logger.info("%d" % prev_line)
D_dict = OrderedDict()
output = False
for i in xrange(args.num_common):
D[i] = 0
C[i] = 0
null_unk_indices = [state['null_sym_target'],state['unk_sym_target']]
update_dicts(null_unk_indices, d, D, C, args.num_common)
with open(args.source, 'r') as f:
for i, line in enumerate(f):
seqin = line.strip()
seq, _ = parse_input(state, word2idx_src, seqin) # seq is the ndarray of indices
indices = []
for elt in seq[:-1]: # Exclude the EOL token
if elt != 1 and elt in topn: # Exclude OOV (1 will not be a key of topn)
indices.extend(topn[elt]) # Add topn best unigram translations for each source word
update_dicts(indices, d, D, C, args.num_common)
if (i % args.change_every) == 0 and args.change_every > 0 and i > 0:
D_dict[prev_line] = D.copy() # Save dictionary for the lines preceding this one
prev_line = i
logger.info("%d" % i)
output = False
d = OrderedDict()
if args.no_reset:
C = D.copy()
else:
D = OrderedDict() # Full
C = OrderedDict() # Allowed to reject
for i in xrange(args.num_common):
D[i] = 0
C[i] = 0
null_unk_indices = [state['null_sym_target'], state['unk_sym_target']]
update_dicts(null_unk_indices, d, D, C, args.num_common)
update_dicts(indices, d, D, C, args.num_common) # Assumes you cannot fill d with only 1 line
D_dict[prev_line] = D.copy()
start_time = time.time()
if args.source and args.trans and args.new_trans:
with open(args.source, 'r') as src_file:
with open(args.trans, 'r') as trans_file:
with open(args.new_trans, 'w') as new_trans_file:
if not (args.num_common and args.num_ttables and args.topn_file):
eos_id = state['null_sym_target']
unk_id = state['unk_sym_target']
new_word2idx_trg = word2idx_trg
prev_i = -1
if args.n_best:
full_trans_line = trans_file.readline()
if full_trans_line == '':
raise IOError("File is empty")
full_trans_line = full_trans_line.split('|||')
n_best_start = int(full_trans_line[0].strip())
trans_file.seek(0)
while True:
if args.n_best:
full_trans_line = trans_file.readline()
if full_trans_line == '':
break
full_trans_line = full_trans_line.split('|||')
i = int(full_trans_line[0].strip()) - n_best_start
trans_line = full_trans_line[1].strip()
else:
trans_line = trans_file.readline()
if trans_line == '':
break
i = prev_i + 1
if i == (prev_i + 1):
prev_i = i
if (i % args.change_every) == 0 and i > 0:
hard_alignments = compute_alignment(src_seqs, trg_seqs, alignment_fns, args.batchsize)
replace_unknown_words(
src_word_seqs, trg_seqs, trg_word_seqs,
hard_alignments, heuristic, mapping, unk_id,
new_trans_file, args.n_best, full_trans_lines)
if (i % 100 == 0) and i > 0:
new_trans_file.flush()
logger.debug("Current speed is {} per sentence".
format((time.time() - start_time) / i))
src_line = src_file.readline()
src_seq, src_words = parse_input(state, word2idx_src, src_line.strip())
src_words.append('<eos>')
if (i % args.change_every) == 0:
src_seqs = []
src_word_seqs = []
trg_seqs = []
trg_word_seqs = []
full_trans_lines = [] # Only used with n-best lists
if args.num_common and args.num_ttables and args.topn_file:
indices = D_dict[i].keys()
eos_id = indices.index(state['null_sym_target']) # Find new eos and unk positions
unk_id = indices.index(state['unk_sym_target'])
for j in xrange(num_models):
lm_models[j].params[lm_models[j].name2pos['W_0_dec_approx_embdr']].set_value(original_W_0_dec_approx_embdr[j][indices])
lm_models[j].params[lm_models[j].name2pos['W2_dec_deep_softmax']].set_value(original_W2_dec_deep_softmax[j][:, indices])
lm_models[j].params[lm_models[j].name2pos['b_dec_deep_softmax']].set_value(original_b_dec_deep_softmax[j][indices])
new_word2idx_trg = dict([(idict_trg[index], k) for k, index in enumerate(indices)])
elif i != prev_i:
raise ValueError("prev_i: %d, i: %d" % (prev_i, i))
trans_seq, trans_words = parse_output(new_word2idx_trg, trans_line.strip(), eos_id=eos_id, unk_id=unk_id)
trans_words.append('<eos>')
src_seqs.append(src_seq)
src_word_seqs.append(src_words)
trg_seqs.append(trans_seq)
trg_word_seqs.append(trans_words)
if args.n_best:
full_trans_lines.append(full_trans_line)
# Out of loop
hard_alignments = compute_alignment(src_seqs, trg_seqs, alignment_fns, args.batchsize)
replace_unknown_words(src_word_seqs, trg_seqs, trg_word_seqs,
hard_alignments, heuristic, mapping, unk_id,
new_trans_file, args.n_best, full_trans_lines)
else:
raise NotImplementedError
def main():
args = parse_args()
state = prototype_phrase_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")
if 'rolling_vocab' not in state:
state['rolling_vocab'] = 0
if 'save_algo' not in state:
state['save_algo'] = 0
if 'save_gs' not in state:
state['save_gs'] = 0
if 'save_iter' not in state:
state['save_iter'] = -1
if 'var_src_len' not in state:
state['var_src_len'] = False
with open(args.topn_file, 'rb') as f:
topn = cPickle.load(f) # Load dictionary (source word index : list of target word indices)
if args.less_transfer:
for elt in topn:
topn[elt] = topn[elt][:args.num_ttables] # Take the first args.num_ttables only
else:
for elt in topn:
topn[elt] = set(topn[elt][:args.num_ttables]) # Take the first args.num_ttables only and convert list to set
num_models = len(args.models)
rng = numpy.random.RandomState(state['seed'])
enc_decs = []
lm_models = []
original_W_0_dec_approx_embdr = []
original_W2_dec_deep_softmax = []
original_b_dec_deep_softmax = []
for i in xrange(num_models):
enc_decs.append(RNNEncoderDecoder(state, rng, skip_init=True))
enc_decs[i].build()
lm_models.append(enc_decs[i].create_lm_model())
lm_models[i].load(args.models[i])
original_W_0_dec_approx_embdr.append(lm_models[i].params[lm_models[i].name2pos['W_0_dec_approx_embdr']].get_value())
original_W2_dec_deep_softmax.append(lm_models[i].params[lm_models[i].name2pos['W2_dec_deep_softmax']].get_value())
original_b_dec_deep_softmax.append(lm_models[i].params[lm_models[i].name2pos['b_dec_deep_softmax']].get_value())
# On GPU, this will free memory for the next models
# Additional gains could be made by rolling the source vocab
lm_models[i].params[lm_models[i].name2pos['W_0_dec_approx_embdr']].set_value(numpy.zeros((1,1), dtype=numpy.float32))
lm_models[i].params[lm_models[i].name2pos['W2_dec_deep_softmax']].set_value(numpy.zeros((1,1), dtype=numpy.float32))
lm_models[i].params[lm_models[i].name2pos['b_dec_deep_softmax']].set_value(numpy.zeros((1), dtype=numpy.float32))
indx_word = cPickle.load(open(state['word_indx'],'rb')) #Source w2i
sampler = None
beam_search = None
if args.beam_search:
beam_search = BeamSearch(enc_decs)
beam_search.compile()
else:
raise NotImplementedError
#sampler = enc_dec.create_sampler(many_samples=True)
idict_src = cPickle.load(open(state['indx_word'],'r')) #Source i2w
original_target_i2w = lm_models[0].word_indxs.copy()
# I don't think that we need target_word2index
max_words = len(original_b_dec_deep_softmax[0])
if args.less_transfer:
# Use OrderedDict instead of set for reproducibility
d = OrderedDict() # Up to now
D = OrderedDict() # Full
C = OrderedDict() # Allowed to reject
prev_line = 0
logger.info("%d" % prev_line)
D_dict = OrderedDict()
output = False
for i in xrange(args.num_common):
D[i] = 0
C[i] = 0
null_unk_indices = [state['null_sym_target'],state['unk_sym_target']]
update_dicts(null_unk_indices, d, D, C, args.num_common)
with open(args.source, 'r') as f:
for i, line in enumerate(f):
seqin = line.strip()
seq, parsed_in = parse_input(state, indx_word, seqin, idx2word=idict_src) # seq is the ndarray of indices
indices = []
for elt in seq[:-1]: # Exclude the EOL token
if elt != 1: # Exclude OOV (1 will not be a key of topn)
indices.extend(topn[elt]) # Add topn best unigram translations for each source word
output = update_dicts(indices, d, D, C, args.num_common)
if (i % args.change_every) == 0 and args.change_every > 0 and i > 0:
output = True
if output:
D_dict[prev_line] = D.copy() # Save dictionary for the lines preceding this one
prev_line = i
logger.info("%d" % i)
output = False
d = OrderedDict()
if args.no_reset:
C = D.copy()
else:
D = OrderedDict() # Full
C = OrderedDict() # Allowed to reject
for i in xrange(args.num_common):
D[i] = 0
C[i] = 0
null_unk_indices = [state['null_sym_target'], state['unk_sym_target']]
update_dicts(null_unk_indices, d, D, C, args.num_common)
update_dicts(indices, d, D, C, args.num_common) # Assumes you cannot fill d with only 1 line
D_dict[prev_line] = D.copy()
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) # seq is the ndarray of indices
# For now, keep all input words in the model.
# In the future, we may want to filter them to save on memory, but this isn't really much of an issue now
if args.verbose:
print "Parsed Input:", parsed_in
if args.less_transfer:
if i in D_dict:
indices = D_dict[i].keys()
eos_id = indices.index(state['null_sym_target']) # Find new eos and unk positions
unk_id = indices.index(state['unk_sym_target'])
for j in xrange(num_models):
lm_models[j].params[lm_models[j].name2pos['W_0_dec_approx_embdr']].set_value(original_W_0_dec_approx_embdr[j][indices])
lm_models[j].params[lm_models[j].name2pos['W2_dec_deep_softmax']].set_value(original_W2_dec_deep_softmax[j][:, indices])
lm_models[j].params[lm_models[j].name2pos['b_dec_deep_softmax']].set_value(original_b_dec_deep_softmax[j][indices])
lm_models[0].word_indxs = dict([(k, original_target_i2w[index]) for k, index in enumerate(indices)]) # target index2word
trans, costs, _ = sample(lm_models[0], seq, n_samples, sampler=sampler,
beam_search=beam_search, ignore_unk=args.ignore_unk, normalize=args.normalize,
normalize_p=args.normalize_p, eos_id=eos_id, unk_id=unk_id, final=True, wp=args.wp)
else:
# Extract the indices you need
indices = set()
for elt in seq[:-1]: # Exclude the EOL token
if elt != 1: # Exclude OOV (1 will not be a key of topn)
indices = indices.union(topn[elt]) # Add topn best unigram translations for each source word
num_common_words = args.num_common
while True:
if num_common_words >= max_words:
final = True
num_common_words = max_words
else:
final = False
if args.final: # No matter the number of words
final = True
indices = indices.union(set(xrange(num_common_words))) # Add common words
indices = list(indices) # Convert back to list for advanced indexing
eos_id = indices.index(state['null_sym_target']) # Find new eos and unk positions
unk_id = indices.index(state['unk_sym_target'])
# Set the target word matrices and biases
for j in xrange(num_models):
lm_models[j].params[lm_models[j].name2pos['W_0_dec_approx_embdr']].set_value(original_W_0_dec_approx_embdr[j][indices])
lm_models[j].params[lm_models[j].name2pos['W2_dec_deep_softmax']].set_value(original_W2_dec_deep_softmax[j][:, indices])
lm_models[j].params[lm_models[j].name2pos['b_dec_deep_softmax']].set_value(original_b_dec_deep_softmax[j][indices])
lm_models[0].word_indxs = dict([(k, original_target_i2w[index]) for k, index in enumerate(indices)]) # target index2word
try:
trans, costs, _ = sample(lm_models[0], seq, n_samples, sampler=sampler,
beam_search=beam_search, ignore_unk=args.ignore_unk, normalize=args.normalize,
normalize_p=args.normalize_p, eos_id=eos_id, unk_id=unk_id, final=final)
break # Breaks only if it succeeded (If final=True, will always succeed)
except RuntimeError:
indices = set(indices)
num_common_words *= 2
if not args.n_best:
best = numpy.argmin(costs)
print >>ftrans, trans[best]
else:
order = numpy.argsort(costs)
best = order[0]
for elt in order:
print >>ftrans, str(i+args.start) + ' ||| ' + trans[elt] + ' ||| ' + str(costs[elt])
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:
raise NotImplementedError
def main():
args = parse_args()
state = prototype_phrase_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 main():
args = parse_args()
state = prototype_phrase_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")
if 'rolling_vocab' not in state:
state['rolling_vocab'] = 0
if 'save_algo' not in state:
state['save_algo'] = 0
if 'save_gs' not in state:
state['save_gs'] = 0
if 'save_iter' not in state:
state['save_iter'] = -1
if 'var_src_len' not in state:
state['var_src_len'] = False
if args.num_common and args.num_ttables and args.topn_file:
with open(args.topn_file, 'rb') as f:
topn = cPickle.load(f) # Load dictionary (source word index : list of target word indices)
for elt in topn:
topn[elt] = topn[elt][:args.num_ttables] # Take the first args.num_ttables only
num_models = len(args.models)
rng = numpy.random.RandomState(state['seed'])
enc_decs = []
lm_models = []
alignment_fns = []
if args.num_common and args.num_ttables and args.topn_file:
original_W_0_dec_approx_embdr = []
original_W2_dec_deep_softmax = []
original_b_dec_deep_softmax = []
for i in xrange(num_models):
enc_decs.append(RNNEncoderDecoder(state, rng, skip_init=True, compute_alignment=True))
enc_decs[i].build()
lm_models.append(enc_decs[i].create_lm_model())
lm_models[i].load(args.models[i])
alignment_fns.append(theano.function(inputs=enc_decs[i].inputs, outputs=[enc_decs[i].alignment], name="alignment_fn"))
if args.num_common and args.num_ttables and args.topn_file:
original_W_0_dec_approx_embdr.append(lm_models[i].params[lm_models[i].name2pos['W_0_dec_approx_embdr']].get_value())
original_W2_dec_deep_softmax.append(lm_models[i].params[lm_models[i].name2pos['W2_dec_deep_softmax']].get_value())
original_b_dec_deep_softmax.append(lm_models[i].params[lm_models[i].name2pos['b_dec_deep_softmax']].get_value())
lm_models[i].params[lm_models[i].name2pos['W_0_dec_approx_embdr']].set_value(numpy.zeros((1,1), dtype=numpy.float32))
lm_models[i].params[lm_models[i].name2pos['W2_dec_deep_softmax']].set_value(numpy.zeros((1,1), dtype=numpy.float32))
lm_models[i].params[lm_models[i].name2pos['b_dec_deep_softmax']].set_value(numpy.zeros((1), dtype=numpy.float32))
if args.mapping:
with open(args.mapping, 'rb') as f:
mapping = cPickle.load(f)
heuristic = args.heuristic
else:
heuristic = 0
mapping = None
word2idx_src = cPickle.load(open(state['word_indx'], 'rb'))
idict_src = cPickle.load(open(state['indx_word'], 'r'))
word2idx_trg = cPickle.load(open(state['word_indx_trgt'], 'rb'))
idict_trg = cPickle.load(open(state['indx_word_target'], 'r'))
word2idx_trg['<eos>'] = state['null_sym_target']
word2idx_trg[state['oov']] = state['unk_sym_target'] # 'UNK' may be in the vocabulary. Now points to the right index.
idict_trg[state['null_sym_target']] = '<eos>'
idict_trg[state['unk_sym_target']] = state['oov']
if args.num_common and args.num_ttables and args.topn_file:
# Use OrderedDict instead of set for reproducibility
d = OrderedDict() # Up to now
D = OrderedDict() # Full
C = OrderedDict() # Allowed to reject
prev_line = 0
logger.info("%d" % prev_line)
D_dict = OrderedDict()
output = False
for i in xrange(args.num_common):
D[i] = 0
C[i] = 0
null_unk_indices = [state['null_sym_target'],state['unk_sym_target']]
update_dicts(null_unk_indices, d, D, C, args.num_common)
with open(args.source, 'r') as f:
for i, line in enumerate(f):
seqin = line.strip()
seq, _ = parse_input(state, word2idx_src, seqin) # seq is the ndarray of indices
indices = []
for elt in seq[:-1]: # Exclude the EOL token
if elt != 1: # Exclude OOV (1 will not be a key of topn)
indices.extend(topn[elt]) # Add topn best unigram translations for each source word
update_dicts(indices, d, D, C, args.num_common)
if (i % args.change_every) == 0 and args.change_every > 0 and i > 0:
D_dict[prev_line] = D.copy() # Save dictionary for the lines preceding this one
prev_line = i
logger.info("%d" % i)
output = False
d = OrderedDict()
if args.no_reset:
C = D.copy()
else:
D = OrderedDict() # Full
C = OrderedDict() # Allowed to reject
for i in xrange(args.num_common):
D[i] = 0
C[i] = 0
null_unk_indices = [state['null_sym_target'], state['unk_sym_target']]
update_dicts(null_unk_indices, d, D, C, args.num_common)
update_dicts(indices, d, D, C, args.num_common) # Assumes you cannot fill d with only 1 line
D_dict[prev_line] = D.copy()
start_time = time.time()
if args.source and args.trans and args.new_trans:
with open(args.source, 'r') as src_file:
with open(args.trans, 'r') as trans_file:
with open(args.new_trans, 'w') as new_trans_file:
if not (args.num_common and args.num_ttables and args.topn_file):
eos_id = state['null_sym_target']
unk_id = state['unk_sym_target']
new_word2idx_trg = word2idx_trg
prev_i = -1
if args.n_best:
full_trans_line = trans_file.readline()
if full_trans_line == '':
raise IOError("File is empty")
full_trans_line = full_trans_line.split('|||')
n_best_start = int(full_trans_line[0].strip())
trans_file.seek(0)
while True:
if args.n_best:
full_trans_line = trans_file.readline()
if full_trans_line == '':
break
full_trans_line = full_trans_line.split('|||')
i = int(full_trans_line[0].strip()) - n_best_start
trans_line = full_trans_line[1].strip()
else:
trans_line = trans_file.readline()
if trans_line == '':
break
i = prev_i + 1
if i == (prev_i + 1):
prev_i = i
if (i % args.change_every) == 0 and i > 0:
hard_alignments = compute_alignment(src_seqs, trg_seqs, alignment_fns, args.batchsize)
replace_unknown_words(
src_word_seqs, trg_seqs, trg_word_seqs,
hard_alignments, heuristic, mapping, unk_id,
new_trans_file, args.n_best, full_trans_lines)
if (i % 100 == 0) and i > 0:
new_trans_file.flush()
logger.debug("Current speed is {} per sentence".
format((time.time() - start_time) / i))
src_line = src_file.readline()
src_seq, src_words = parse_input(state, word2idx_src, src_line.strip())
src_words.append('<eos>')
if (i % args.change_every) == 0:
src_seqs = []
src_word_seqs = []
trg_seqs = []
trg_word_seqs = []
full_trans_lines = [] # Only used with n-best lists
if args.num_common and args.num_ttables and args.topn_file:
indices = D_dict[i].keys()
eos_id = indices.index(state['null_sym_target']) # Find new eos and unk positions
unk_id = indices.index(state['unk_sym_target'])
for j in xrange(num_models):
lm_models[j].params[lm_models[j].name2pos['W_0_dec_approx_embdr']].set_value(original_W_0_dec_approx_embdr[j][indices])
lm_models[j].params[lm_models[j].name2pos['W2_dec_deep_softmax']].set_value(original_W2_dec_deep_softmax[j][:, indices])
lm_models[j].params[lm_models[j].name2pos['b_dec_deep_softmax']].set_value(original_b_dec_deep_softmax[j][indices])
new_word2idx_trg = dict([(idict_trg[index], k) for k, index in enumerate(indices)])
elif i != prev_i:
raise ValueError("prev_i: %d, i: %d" % (prev_i, i))
trans_seq, trans_words = parse_output(new_word2idx_trg, trans_line.strip(), eos_id=eos_id, unk_id=unk_id)
trans_words.append('<eos>')
src_seqs.append(src_seq)
src_word_seqs.append(src_words)
trg_seqs.append(trans_seq)
trg_word_seqs.append(trans_words)
if args.n_best:
full_trans_lines.append(full_trans_line)
# Out of loop
hard_alignments = compute_alignment(src_seqs, trg_seqs, alignment_fns, args.batchsize)
replace_unknown_words(src_word_seqs, trg_seqs, trg_word_seqs,
hard_alignments, heuristic, mapping, unk_id,
new_trans_file, args.n_best, full_trans_lines)
else:
raise NotImplementedError
def main():
args = parse_args()
state = prototype_phrase_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, rng)
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)
trgt_seq = parse_input(state, indx_word_trgt, trgt_line, raise_unk=not args.allow_unk)
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)
trgt_seq, trgt_words = parse_input(state, indx_word_trgt, trgt_line, raise_unk=not args.allow_unk)
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 = prototype_phrase_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")
if 'rolling_vocab' not in state:
state['rolling_vocab'] = 0
if 'save_algo' not in state:
state['save_algo'] = 0
if 'save_gs' not in state:
state['save_gs'] = 0
if 'save_iter' not in state:
state['save_iter'] = -1
if 'var_src_len' not in state:
state['var_src_len'] = False
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, rng)
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)
trgt_seq = parse_input(state, indx_word_trgt, trgt_line, raise_unk=not args.allow_unk)
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)
trgt_seq, trgt_words = parse_input(state,
indx_word_trgt, trgt_line, raise_unk=not args.allow_unk)
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 = prototype_phrase_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")
if 'rolling_vocab' not in state:
state['rolling_vocab'] = 0
if 'save_algo' not in state:
state['save_algo'] = 0
if 'save_gs' not in state:
state['save_gs'] = 0
if 'save_iter' not in state:
state['save_iter'] = -1
if 'var_src_len' not in state:
state['var_src_len'] = False
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,
normalize_p=args.normalize_p)
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, normalize_p=args.normalize_p,
alpha=alpha, verbose=True)
