def dothot(a, indices, in_size, out=None, dont_add=False):
"""
In:
a: a numpy array
indices: hot indices a K-hot encoded matrix
out:
out: a.T.dot(x), where x is a K-hot encoded matrix
"""
N, H = a.shape
_N, K = indices.shape
if _N != N:
raise ValueError( 'a.shape[0] != idx.shape[0]' )
bdim, gdim = Get_bdim_and_gdim1D(H)
if dont_add:
B = np.int32(1)
else:
B = np.int32(0)
if out is None:
out = cuda.zeros((H,in_size), dtype=np.float32)
if K > 1:
DotHot1_kernel.prepared_call(gdim, bdim,
a.gpudata, out.gpudata, indices.gpudata,
np.int32(K), np.int32(N), np.int32(H), np.int32(in_size), np.int32(B))
else:
DotHot2_kernel.prepared_call(gdim, bdim,
a.gpudata, out.gpudata, indices.gpudata,
np.int32(N), np.int32(H), np.int32(in_size), np.int32(B))
return out
def trans(self, fromd, tod, train=True):
self.bs = fromd.shape[1]
self.enc.reset_state()
self.mid.reset_state()
self.dec.reset_state()
self.zd = Variable(cuda.zeros((self.bs, self.nh), dtype=np.float32))
loss = 0
out = []
for wd in reversed(fromd):
#print wd
x = Variable(cuda.to_gpu(np.reshape(wd, (self.bs, 1))))
#print x.data
y = self.enc(self.embed(x))
self.mid.set_state(self.enc.c, self.enc.h)
y = self.mid(self.zd)
self.dec.set_state(self.mid.c, self.mid.h)
if train:
for wd in tod:
t = Variable(cuda.to_gpu(np.reshape(wd, (self.bs)))) # t.data)
x = self.out(self.dec(self.zd))
loss += F.softmax_cross_entropy(x, t)
else:
for i in xrange(100):
x = self.out(self.dec(self.zd))
ch = np.argmax(x.data[0]) # bs should be 1
out.append(ch)
if ch == 0:
break
return loss, out
def forward_gpu(self, x):
a, b = x
shape = self._output_shape(a, b)
ret = cuda.zeros(shape)
_batch_matmul_gpu(
a, b, transa=self.transa, transb=self.transb, out=ret)
return ret,
if len(args.init_from) > 0:
model = pickle.load(open(args.init_from, 'rb'))
else:
model = CharRNN(len(vocab), n_units)
if args.gpu >= 0:
cuda.init()
model.to_gpu()
optimizer = optimizers.RMSprop(lr=args.learning_rate, alpha=args.decay_rate, eps=1e-8)
optimizer.setup(model.collect_parameters())
whole_len = train_data.shape[0]
jump = whole_len / batchsize
cur_log_perp = cuda.zeros(())
epoch = 0
start_at = time.time()
cur_at = start_at
state = make_initial_state(n_units, batchsize=batchsize)
if args.gpu >= 0:
accum_loss = Variable(cuda.zeros(()))
for key, value in state.items():
value.data = cuda.to_gpu(value.data)
else:
accum_loss = Variable(np.zeros(()))
print 'going to train {} iterations'.format(jump * n_epochs)
for i in xrange(jump * n_epochs):
x_batch = np.array([train_data[(jump * j + i) % whole_len]
for j in xrange(batchsize)])
def zeros(shape, dtype=np.float32):
#return Variable(np.zeros(shape, dtype=dtype))
return Variable(cuda.zeros(shape, dtype=dtype))
if args.gpu >= 0:
cuda.init()
model.to_gpu()
optimizer = optimizers.RMSprop(lr=args.learning_rate, alpha=args.decay_rate, eps=1e-8)
optimizer.setup(model)
whole_len = train_data.shape[0]
jump = whole_len / batchsize
epoch = 0
start_at = time.time()
cur_at = start_at
state = make_initial_state(n_units, batchsize=batchsize)
if args.gpu >= 0:
accum_loss = Variable(cuda.zeros(()))
for key, value in state.items():
value.data = cuda.to_gpu(value.data)
else:
accum_loss = Variable(np.zeros(()).astype(np.float32))
print 'going to train {} iterations'.format(jump * n_epochs)
for i in xrange(jump * n_epochs):
x_batch = np.array([train_data[(jump * j + i) % whole_len]
for j in xrange(batchsize)])
y_batch = np.array([train_data[(jump * j + i + 1) % whole_len]
for j in xrange(batchsize)])
if args.gpu >=0:
x_batch = cuda.to_gpu(x_batch)
y_batch = cuda.to_gpu(y_batch)
def zeros(shape, dtype=np.float32):
#return Variable(np.zeros(shape, dtype=dtype))
return Variable(cuda.zeros(shape, dtype=dtype))
def _train(self, **kwargs):
gpu = -1 if "gpu" not in kwargs else kwargs["gpu"]
lr = 2e-3 if "lr" not in kwargs else kwargs["lr"]
lr_decay = 0.97 if "lr_decay" not in kwargs else kwargs["lr_decay"]
lr_decay_after=10 if "lr_decay_after" not in kwargs else kwargs["lr_decay_after"]
decay_rate = 0.95 if "decay_rate" not in kwargs else kwargs["decay_rate"]
dropout = 0.0 if "dropout" not in kwargs else kwargs["dropout"]
bprop_len = 50 if "bprop_len" not in kwargs else kwargs["bprop_len"]
batchsize = 50 if "batchsize" not in kwargs else kwargs["batchsize"]
grad_clip = 5 if "grad_clip" not in kwargs else kwargs["grad_clip"]
n_epochs = 5 if "epochs" not in kwargs else kwargs["epochs"]
if gpu >= 0:
cuda.get_device(gpu).use()
self.model.to_gpu()
optimizer = optimizers.RMSprop(lr=lr, alpha=decay_rate, eps=1e-8)
optimizer.setup(self.model)
train_data = self.dataset
whole_len = train_data.shape[0]
jump = whole_len // batchsize
epoch = 0
start_at = time.time()
cur_at = start_at
state = self.model.make_initial_state(batchsize=batchsize)
if gpu >= 0:
accum_loss = Variable(cuda.zeros(()))
for key, value in state.items():
value.data = cuda.to_gpu(value.data)#plist
else:
accum_loss = Variable(np.zeros((), dtype=np.float32))
print ('going to train {} iterations'.format(jump * n_epochs))
for i in range(jump * n_epochs):
x_batch = np.array([train_data[(jump * j + i) % whole_len]
for j in range(batchsize)])
y_batch = np.array([train_data[(jump * j + i + 1) % whole_len]
for j in range(batchsize)])
if gpu >=0:
x_batch = cuda.to_gpu(x_batch)
y_batch = cuda.to_gpu(y_batch)
state, loss_i = self.model.forward_one_step(x_batch, y_batch, state, dropout_ratio=dropout)
accum_loss += loss_i
if (i + 1) % bprop_len == 0: # Run truncated BPTT
now = time.time()
sys.stderr.write('\r{}/{}, train_loss = {}, time = {:.2f}'.format((i+1)//bprop_len,(jump*n_epochs)//bprop_len, accum_loss.data / bprop_len, now-cur_at))
sys.stderr.flush()
cur_at = now
optimizer.zero_grads()
accum_loss.backward()
accum_loss.unchain_backward() # truncate
if gpu >= 0:
accum_loss = Variable(cuda.zeros(()))
else:
accum_loss = Variable(np.zeros((), dtype=np.float32))
optimizer.clip_grads(grad_clip)
optimizer.update()
if (i + 1) % 10000 == 0:
pickle.dump(copy.deepcopy(self.model).to_cpu(), open(self.model_path, 'wb'))
if (i + 1) % jump == 0:
epoch += 1
if epoch >= lr_decay_after:
optimizer.lr *= lr_decay
print ('decayed learning rate by a factor {} to {}'.format(lr_decay, optimizer.lr))
sys.stdout.flush()
pickle.dump(copy.deepcopy(self.model).to_cpu(), open(self.model_path, 'wb'))
if args.optimizer == 'rmsprop':
optimizer = optimizers.RMSprop(lr=args.learning_rate,
alpha=args.decay_rate)
elif args.optimizer == 'adam':
optimizer = optimizers.Adam()
elif args.optimizers == 'adagrad':
optimizer = optimizers.AdaGrad(lr=args.learning_rate)
else:
error("unknown optimizer")
optimizer.setup(model.collect_parameters())
whole_len = train_data.shape[0]
jump = whole_len / batchsize
cur_log_perp = cuda.zeros(())
epoch = 0
start_at = time.time()
cur_at = start_at
if args.gpu >= 0:
accum_loss = Variable(cuda.zeros(()))
for key, value in state.items():
value.data = cuda.to_gpu(value.data)
else:
accum_loss = Variable(np.zeros(()))
print 'going to train {} iterations'.format(jump * n_epochs)
for i in xrange(jump * n_epochs):
x_batch = np.array(
[train_data[(jump * j + i) % whole_len] for j in xrange(batchsize)])
def reset_accum_loss(self):
if self.gpu >= 0:
self.accum_loss = Variable(cuda.zeros(()))
else:
self.accum_loss = Variable(np.zeros(()))
eps=1e-8)
optimizer.setup(model.collect_parameters())
whole_len = len(train_data)
whole_val_len = len(val_data)
epoch = 0
start_at = time.time()
cur_at = start_at
end_time = 0
state = make_initial_state(n_units)
train_loss_all = []
val_loss_all = []
iterations_count = 0
if args.gpu >= 0:
loss = Variable(cuda.zeros(()))
val_loss = Variable(cuda.zeros(()))
for key, value in state.items():
value.data = cuda.to_gpu(value.data)
else:
loss = Variable(np.zeros((), dtype=np.float32))
val_loss = Variable(np.zeros((), dtype=np.float32))
for i in xrange(whole_len * n_epochs):
for j in xrange(0, len(train_data[i % whole_len]) - 1):
x_t = np.array([train_data[i % whole_len][j]])
y_t = np.array([train_data[i % whole_len][j + 1]])
if args.gpu >= 0:
x_t = cuda.to_gpu(x_t)
y_t = cuda.to_gpu(y_t)
state, loss_i = model.forward_one_step(x_t,
optimizer = optimizers.RMSprop(lr=args.learning_rate, alpha=args.decay_rate, eps=1e-8)
optimizer.setup(model.collect_parameters())
whole_len = len(train_data)
whole_val_len = len(val_data)
epoch = 0
start_at = time.time()
cur_at = start_at
end_time = 0
state = make_initial_state(n_units)
train_loss_all = []
val_loss_all = []
iterations_count = 0
if args.gpu >= 0:
loss = Variable(cuda.zeros(()))
val_loss = Variable(cuda.zeros(()))
for key, value in state.items():
value.data = cuda.to_gpu(value.data)
else:
loss = Variable(np.zeros((), dtype=np.float32))
val_loss = Variable(np.zeros((), dtype=np.float32))
for i in xrange(whole_len * n_epochs):
for j in xrange(0, len(train_data[i % whole_len]) - 1):
x_t = np.array([train_data[i % whole_len][j]])
y_t = np.array([train_data[i % whole_len][j + 1]])
if args.gpu >= 0:
x_t = cuda.to_gpu(x_t)
y_t = cuda.to_gpu(y_t)
state, loss_i = model.forward_one_step(x_t, y_t, state, dropout_ratio=args.dropout)
def train(self, words, steps, batchsize=100, sequence_length=10):
""" Train the Predictor's model on words for steps number of steps. """
whole_len = len(words)
train_data = np.ndarray(whole_len, dtype=np.int32)
jumps = steps * sequence_length
# Initialize training data and maybe vocab.
if self.vocab is None:
vocab_initializing = True
self.vocab = {}
for i, word in enumerate(words):
if vocab_initializing:
if word not in self.vocab:
self.vocab[word] = len(self.vocab)
train_data[i] = self.vocab[word]
vocab_initializing = False
print 'corpus length:', len(words)
print 'self.vocab size:', len(self.vocab)
# Initialize base model (if we need to)
if self.model is None:
self.model = BaseRNN(len(self.vocab), self.units)
if self.gpu >= 0:
cuda.get_device(self.gpu).use()
self.model.to_self.gpu()
optimizer = optimizers.RMSprop(lr=self.settings.learning_rate,
alpha=self.settings.decay_rate,
eps=1e-8)
optimizer.setup(self.model)
jumpsPerEpoch = whole_len / batchsize
epoch = 0
start_at = time.time()
cur_at = start_at
state = make_initial_state(self.units, batchsize=batchsize)
if self.gpu >= 0:
accum_loss = Variable(cuda.zeros(()))
for _, value in state.items():
value.data = cuda.to_self.gpu(value.data)
else:
accum_loss = Variable(np.zeros((), dtype=np.float32))
print 'going to train {} iterations'.format(steps)
for i in xrange(jumps):
x_batch = np.array([
train_data[(jumpsPerEpoch * j + i) % whole_len]
for j in xrange(batchsize)
])
y_batch = np.array([
train_data[(jumpsPerEpoch * j + i + 1) % whole_len]
for j in xrange(batchsize)
])
if self.gpu >= 0:
x_batch = cuda.to_self.gpu(x_batch)
y_batch = cuda.to_self.gpu(y_batch)
state, loss_i = self.model.forward_one_step(
x_batch, y_batch, state, dropout_ratio=self.settings.dropout)
accum_loss += loss_i
if (i + 1) % sequence_length == 0:
now = time.time()
print '{}/{}, train_loss = {}, time = {:.2f}'.format(
(i + 1) / sequence_length, steps,
accum_loss.data / sequence_length, now - cur_at)
cur_at = now
optimizer.zero_grads()
accum_loss.backward()
accum_loss.unchain_backward() # truncate
if self.gpu >= 0:
accum_loss = Variable(cuda.zeros(()))
else:
accum_loss = Variable(np.zeros((), dtype=np.float32))
optimizer.clip_grads(self.settings.grad_clip)
optimizer.update()
if (i + 1) % jumpsPerEpoch == 0:
epoch += 1
if epoch >= self.settings.learning_rate_decay_after:
optimizer.lr *= self.settings.learning_rate_decay
print 'decayed self.settings.learning rate by a factor {} to {}'.format(
self.settings.learning_rate_decay, optimizer.lr)
def train(self, words, steps, batchsize=100, sequence_length=10):
""" Train the Predictor's model on words for steps number of steps. """
whole_len = len(words)
train_data = np.ndarray(whole_len, dtype=np.int32)
jumps = steps * sequence_length
# Initialize training data and maybe vocab.
if self.vocab is None:
vocab_initializing = True
self.vocab = {}
for i, word in enumerate(words):
if vocab_initializing:
if word not in self.vocab:
self.vocab[word] = len(self.vocab)
train_data[i] = self.vocab[word]
vocab_initializing = False
print 'corpus length:', len(words)
print 'self.vocab size:', len(self.vocab)
# Initialize base model (if we need to)
if self.model is None:
self.model = BaseRNN(len(self.vocab), self.units)
if self.gpu >= 0:
cuda.get_device(self.gpu).use()
self.model.to_self.gpu()
optimizer = optimizers.RMSprop(lr=self.settings.learning_rate,
alpha=self.settings.decay_rate,
eps=1e-8)
optimizer.setup(self.model)
jumpsPerEpoch = whole_len / batchsize
epoch = 0
start_at = time.time()
cur_at = start_at
state = make_initial_state(self.units, batchsize=batchsize)
if self.gpu >= 0:
accum_loss = Variable(cuda.zeros(()))
for _, value in state.items():
value.data = cuda.to_self.gpu(value.data)
else:
accum_loss = Variable(np.zeros((), dtype=np.float32))
print 'going to train {} iterations'.format(steps)
for i in xrange(jumps):
x_batch = np.array([train_data[(jumpsPerEpoch * j + i) % whole_len]
for j in xrange(batchsize)])
y_batch = np.array([train_data[(jumpsPerEpoch * j + i + 1) % whole_len]
for j in xrange(batchsize)])
if self.gpu >= 0:
x_batch = cuda.to_self.gpu(x_batch)
y_batch = cuda.to_self.gpu(y_batch)
state, loss_i = self.model.forward_one_step(x_batch,
y_batch,
state,
dropout_ratio=self.settings.dropout)
accum_loss += loss_i
if (i + 1) % sequence_length == 0:
now = time.time()
print '{}/{}, train_loss = {}, time = {:.2f}'.format((i+1)/sequence_length, steps, accum_loss.data / sequence_length, now-cur_at)
cur_at = now
optimizer.zero_grads()
accum_loss.backward()
accum_loss.unchain_backward() # truncate
if self.gpu >= 0:
accum_loss = Variable(cuda.zeros(()))
else:
accum_loss = Variable(np.zeros((), dtype=np.float32))
optimizer.clip_grads(self.settings.grad_clip)
optimizer.update()
if (i + 1) % jumpsPerEpoch == 0:
epoch += 1
if epoch >= self.settings.learning_rate_decay_after:
optimizer.lr *= self.settings.learning_rate_decay
print 'decayed self.settings.learning rate by a factor {} to {}'.format(self.settings.learning_rate_decay, optimizer.lr)
if args.gpu >= 0:
cuda.get_device(args.gpu).use()
model.to_gpu()
optimizer = optimizers.RMSprop(lr=args.learning_rate, alpha=args.decay_rate, eps=1e-8)
optimizer.setup(model)
whole_len = train_data.shape[0]
jump = int(whole_len / batchsize)
epoch = 0
start_at = time.time()
cur_at = start_at
state = make_initial_state(n_units, batchsize=batchsize)
if args.gpu >= 0:
accum_loss = Variable(cuda.zeros(()))
for key, value in list(state.items()):
value.data = cuda.to_gpu(value.data)
else:
accum_loss = Variable(np.zeros((), dtype=np.float32))
print('going to train {} iterations'.format(jump * n_epochs))
for i in range(jump * n_epochs):
x_batch = np.array([train_data[(jump * j + i) % whole_len]
for j in range(batchsize)])
y_batch = np.array([train_data[(jump * j + i + 1) % whole_len]
for j in range(batchsize)])
if args.gpu >=0:
x_batch = cuda.to_gpu(x_batch)
y_batch = cuda.to_gpu(y_batch)
def reset_accum_loss(self):
if self.gpu >= 0:
self.accum_loss = Variable(cuda.zeros(()))
else:
self.accum_loss = Variable(np.zeros(()))
def trainEvaluate(dataset, index):
"""現在のニューラルネットワーク評価処理"""
# モデルに影響を与えないようにコピーする
evaluator = s.dnn.model.copy() # to use different state
evaluator.reset_state() # initialize state
evaluator.train = False # dropout does nothing
evdnn = Dnn(evaluator, None)
# メモリ領域確保して学習＆教師データ取得
xa_cpu = evaluator.buildMiniBatchData(dataset, np.asarray([index]), rnnLen, rnnStep)
ta_cpu = np.zeros(shape=(rnnLen, 1, s.dnnOut), dtype=np.float32)
for i in range(rnnLen):
ta_cpu[i,0,:] = trainGetT(dataset, index + i * rnnStep)
if s.xp == np:
xa_gpu = xa_cpu
ta_gpu = ta_cpu
losses = np.zeros((rnnLen,), dtype=np.float32)
else:
xa_gpu = cuda.to_gpu(xa_cpu)
ta_gpu = cuda.to_gpu(ta_cpu)
losses = cuda.zeros((rnnLen,), dtype=np.float32)
# 必要ならグラフ表示初期化
if s.grEnable:
if s.xp == np:
ya_gpu = np.zeros((rnnLen,), dtype=np.float32)
else:
ya_gpu = cuda.zeros((rnnLen,), dtype=np.float32)
# グラフにデータを描画する
plt.title(s.trainDataFile + " : " + str(index)) # グラフタイトル
xvals = dataset[:, index : index + s.minEvalLen]
tvals = ta_cpu.reshape((rnnLen,))
glIn1.set_ydata(xvals[0])
glIn2.set_ydata(xvals[1])
glIn3.set_ydata(xvals[2])
glIn4.set_ydata(xvals[3])
glTeach.set_ydata(tvals)
subPlot1.set_ylim(f.npMaxMin(xvals))
# RNNを評価
for i in range(rnnLen):
y, loss = evdnn.evaluate(xa_gpu[i], ta_gpu[i])
losses[i : i + 1] = loss.data
if s.grEnable:
ya_gpu[i : i + 1] = y.data[0, 0 : 1]
# 必要ならグラフ表示
if s.grEnable:
if s.xp == np:
yvals = ya_gpu
else:
yvals = cuda.to_cpu(ya_gpu)
glOut.set_ydata(yvals)
subPlot2.set_ylim(f.npMaxMin([tvals, yvals]))
plt.draw()
plt.pause(0.001)
try:
return math.exp(float(losses.sum()) / rnnLen)
except Exception as e:
print("evaluate overflow")
return 0.0