vocab = Vocabulary()
vocab.load("data/vocabulary.json")
print("Loading model...", flush=True)
model = Seq2seqLSTM(vocab.size(), num_embed, num_hidden, num_layers)
model.load_parameters("model/seq2seq_lstm.params", ctx=context)
while True:
try:
source = input("> ")
except EOFError:
print("")
break
source = [vocab.char2idx(ch) for ch in source]
source = pad_sentence(
source, vocab,
[2**(i + 1) for i in range(int(math.log(sequence_length, 2)))])
print(source)
source = mx.nd.reverse(mx.nd.array(source, ctx=context), axis=0)
hidden = model.begin_state(func=mx.nd.zeros, batch_size=1, ctx=context)
hidden = model.encode(source.reshape((1, -1)).T, hidden)
sequences = [([vocab.char2idx("<GO>")], 0.0, hidden)]
while True:
candidates = []
for seq, score, hidden in sequences:
if seq[-1] == vocab.char2idx("<EOS>"):
candidates.append((seq, score, hidden))
else:
target = mx.nd.array([seq[-1]], ctx=context)
output, hidden = model.decode(
target.reshape((1, -1)).T, hidden)
def _handle_request(self):
m = self._path_pattern.match(self.path)
if not m or m.group(0) != self.path:
self.send_response(http.HTTPStatus.BAD_REQUEST)
self.end_headers()
return
if m.group(1) == "/coupletbot/say":
params = {}
if m.group(2):
for param in urllib.parse.unquote(m.group(2)[1:]).split("&"):
kv = self._param_pattern.match(param)
if kv:
params[kv.group(1)] = kv.group(2)
content = params["content"]
if not content:
self.send_response(http.HTTPStatus.BAD_REQUEST)
self.end_headers()
return
print(args.device_id, "say:", content)
source = [vocab.char2idx(ch) for ch in content]
source = pad_sentence(
source, vocab,
[2**(i + 1) for i in range(int(math.log(sequence_length, 2)))])
print(args.device_id, "tokenize:", source)
source = mx.nd.reverse(mx.nd.array(source, ctx=context), axis=0)
hidden = model.begin_state(func=mx.nd.zeros,
batch_size=1,
ctx=context)
enc_out, hidden = model.encode(source.reshape((1, -1)).T, hidden)
sequences = [([vocab.char2idx("<GO>")], 0.0, hidden)]
while True:
candidates = []
for seq, score, hidden in sequences:
if seq[-1] == vocab.char2idx("<EOS>"):
candidates.append((seq, score, hidden))
else:
target = mx.nd.array([seq[-1]], ctx=context)
output, hidden = model.decode(
target.reshape((1, -1)).T, hidden, enc_out)
probs = mx.nd.softmax(output, axis=1)
beam = probs.reshape((-1, )).topk(k=beam_size,
ret_typ="both")
for i in range(beam_size):
candidates.append(
(seq + [int(beam[1][i].asscalar())],
score + math.log(beam[0][i].asscalar()),
hidden))
if len(candidates) <= len(sequences):
break
sequences = sorted(candidates,
key=lambda tup: tup[1],
reverse=True)[:beam_size]
reply = ""
for token in sequences[0][0][1:-1]:
reply += vocab.idx2char(token)
print(args.device_id, "reply:", reply)
self.send_response(http.HTTPStatus.OK)
self.send_header("Content-Type", "text/plain; charset=utf-8")
self.send_header("Access-Control-Allow-Origin", "*")
self.send_header("Access-Control-Allow-Methods", "GET,POST")
self.send_header(
"Access-Control-Allow-Headers",
"Keep-Alive,User-Agent,Authorization,Content-Type")
self.end_headers()
self.wfile.write(reply.encode())
else:
self.send_response(http.HTTPStatus.NOT_FOUND)
self.end_headers()
return