def dmn_mid(args):
#assert args.word_vector_size in [50, 100, 200, 300]
network_name = args.prefix + '%s.mh%d.n%d.bs%d%s%s%s.babi%s' % (
args.network,
args.memory_hops,
args.dim,
args.batch_size,
".na" if args.normalize_attention else "",
".bn" if args.batch_norm else "",
(".d" + str(args.dropout)) if args.dropout>0 else "",
args.babi_id)
babi_train_raw, babi_test_raw = utils.get_babi_raw(args.babi_id, args.babi_test_id)
#word2vec = utils.load_glove(args.word_vector_size)
word2vec = {}
args_dict = dict(args._get_kwargs())
args_dict['babi_train_raw'] = babi_train_raw
args_dict['babi_test_raw'] = babi_test_raw
args_dict['word2vec'] = word2vec
if args.network == 'dmn_tied':
import dmn_tied
if (args.batch_size != 1):
print "==> not using minibatch training in this mode"
args.batch_size = 1
dmn = dmn_tied.DMN_tied(**args_dict)
elif args.network == 'dmn_untied':
import dmn_untied
if (args.batch_size != 1):
print "==> not using minibatch training in this mode"
args.batch_size = 1
dmn = dmn_untied.DMN_untied(**args_dict)
else:
raise Exception("No such network known: " + args.network)
if args.load_state != "":
dmn.load_state(args.load_state)
return args, network_name, dmn
def dmn_mid(args):
#assert args.word_vector_size in [50, 100, 200, 300]
network_name = args.prefix + '%s.mh%d.n%d.bs%d%s%s%s.babi%s' % (
args.network, args.memory_hops, args.dim, args.batch_size, ".na"
if args.normalize_attention else "", ".bn" if args.batch_norm else "",
(".d" + str(args.dropout)) if args.dropout > 0 else "", args.babi_id)
babi_train_raw, babi_test_raw = utils.get_babi_raw(args.babi_id,
args.babi_test_id)
#word2vec = utils.load_glove(args.word_vector_size)
word2vec = {}
args_dict = dict(args._get_kwargs())
args_dict['babi_train_raw'] = babi_train_raw
args_dict['babi_test_raw'] = babi_test_raw
args_dict['word2vec'] = word2vec
if args.network == 'dmn_tied':
import dmn_tied
if (args.batch_size != 1):
print "==> not using minibatch training in this mode"
args.batch_size = 1
dmn = dmn_tied.DMN_tied(**args_dict)
elif args.network == 'dmn_untied':
import dmn_untied
if (args.batch_size != 1):
print "==> not using minibatch training in this mode"
args.batch_size = 1
dmn = dmn_untied.DMN_untied(**args_dict)
else:
raise Exception("No such network known: " + args.network)
if args.load_state != "":
dmn.load_state(args.load_state)
return args, network_name, dmn
parser.add_argument('--normalize_attention', type=bool, default=False, help='flag for enabling softmax on attention vector')
parser.add_argument('--log_every', type=int, default=1, help='print information every x iteration')
parser.add_argument('--save_every', type=int, default=1, help='save state every x epoch')
parser.add_argument('--prefix', type=str, default="", help='optional prefix of network name')
parser.add_argument('--no-shuffle', dest='shuffle', action='store_false')
parser.add_argument('--babi_test_id', type=int, default=-1, help='babi_id of test set')
parser.set_defaults(shuffle=True)
args = parser.parse_args()
assert args.word_vector_size in [50, 100, 200, 300]
network_name = args.prefix + '%s.mh%d.n%d.bs%d%s.babi%s' % (args.network, args.memory_hops, args.dim,
args.batch_size, ".na" if args.normalize_attention else "", args.babi_id)
babi_train_raw, babi_test_raw = utils.get_babi_raw(args.babi_id, args.babi_test_id)
word2vec = utils.load_glove(args.word_vector_size)
args_dict = dict(args._get_kwargs())
args_dict['babi_train_raw'] = babi_train_raw
args_dict['babi_test_raw'] = babi_test_raw
args_dict['word2vec'] = word2vec
# init class
if args.network == 'dmn_batch':
import dmn_batch
dmn = dmn_batch.DMN_batch(**args_dict)
elif args.network == 'dmn_basic':
for i in range(0,l):
loss,pred = self.f(test_inp[i],test_q[i],test_mask[i],test_ans[i])
a_loss = a_loss + loss
y_true.append(test_ans[i])
y_pred.append(pred.argmax(axis=0))
accuracy = sum([1 if t == p else 0 for t, p in zip(y_true, y_pred)])
print "accuracy: %.2f percent" % (accuracy * 100.0 / l)
def shuffle(train_input,train_q,train_answer,train_input_mask,train_sf):
print "==> Shuffling the train set"
combined = zip(train_input,train_q,train_answer,train_input_mask,train_sf)
random.shuffle(combined)
train_input, train_q, train_answer, train_input_mask, train_sf = zip(*combined)
return train_input,train_q,train_answer,train_input_mask,train_sf
babi_train_raw, babi_test_raw = utils.get_babi_raw(task,task)
train_input, train_q, train_answer, train_input_mask, train_sf = _process_input(babi_train_raw)
test_input, test_q, test_answer, test_input_mask, test_sf = _process_input(babi_test_raw)
vocab_size = len(vocab)
a1 = train_input
a2 = train_q
a3 = train_input_mask
a4 = train_answer
dmn = DMN(word_vector_size,vocab_size)
dmn.load_state('states_orig/states7/DMN_orig.epoch20')
#dmn.train(a1,a2,a4,a3,train_sf)
dmn.test(a1,a2,a4,a3)
dmn.test(test_input,test_q,test_answer,test_input_mask)
word_vector_size=word_vector_size,
to_return="word2vec") for w in q
]
utils.process_word(
word=x["A"].lower(), # TODO: add .lower() here
word2vec=word2vec,
vocab=vocab,
ivocab=ivocab,
word_vector_size=word_vector_size,
to_return="index")
gate_len.append(len(x["S"]))
return sent_len, gate_len
word_vector_size = 50
word2vec = utils.load_glove(word_vector_size)
for babi_id in range(1, 21):
babi_id = str(babi_id)
print "processing babi." + babi_id
vocab = {}
ivocab = {}
babi_train_raw, babi_test_raw = utils.get_babi_raw(babi_id, babi_id)
train_len, train_gate = _process_input(babi_train_raw, word2vec, vocab,
ivocab, word_vector_size)
test_len, test_gate = _process_input(babi_test_raw, word2vec, vocab,
ivocab, word_vector_size)
print len(vocab)
print max(train_gate), min(train_gate), max(test_gate), min(test_gate)
print max(train_len), min(train_len), max(test_len), min(test_len)
sent_len.append(len(inp_vector))
q_vector = [utils.process_word(word = w.lower(),
word2vec = word2vec,
vocab = vocab,
ivocab = ivocab,
word_vector_size = word_vector_size,
to_return = "word2vec") for w in q]
utils.process_word(word = x["A"].lower(), # TODO: add .lower() here
word2vec = word2vec,
vocab = vocab,
ivocab = ivocab,
word_vector_size = word_vector_size,
to_return = "index")
gate_len.append(len(x["S"]))
return sent_len, gate_len
word_vector_size=50
word2vec = utils.load_glove(word_vector_size)
for babi_id in range(1,21):
babi_id = str(babi_id)
print "processing babi."+babi_id
vocab = {}
ivocab = {}
babi_train_raw, babi_test_raw = utils.get_babi_raw(babi_id, babi_id)
train_len, train_gate = _process_input(babi_train_raw,word2vec,vocab,ivocab,word_vector_size)
test_len, test_gate = _process_input(babi_test_raw,word2vec,vocab,ivocab,word_vector_size)
print len(vocab)
print max(train_gate), min(train_gate), max(test_gate), min(test_gate)
print max(train_len),min(train_len),max(test_len),min(test_len)
for i, w in enumerate(split_context):
if w == k:
split_context[i] = v
for i, w in enumerate(split_question):
if w == k:
split_question[i] = v
if (correct_answer == k):
data["A"] = v
if (prediction == k):
data["P"] = v
data["C"] = " ".join(split_context)
data["Q"] = " ".join(split_question).replace("@placeholder", "________")
return data
if __name__ == "__main__":
babi_train_raw, babi_test_raw = utils.get_babi_raw("cnn", "")
word2vec = utils.load_glove(word_vector_size)
dmn = dmn_smooth.DMN_smooth(
None, babi_test_raw, word2vec, word_vector_size, dim, mode,
answer_module, input_mask_mode, memory_hops, l2, normalize_attention,
batch_norm, dropout, learning_rate)
dmn.load_state(state)
print "==> running server"
app.run(debug=True, use_reloader=False)
