def init_forward(x_list, volatile=False):
init_h1 = np.zeros((batch_size, hidden_size), dtype=dt)
init_h2 = np.zeros((batch_size, hidden_size), dtype=dt)
if args.gpu >= 0:
init_h1 = cuda.to_gpu_async(init_h1)
init_h2 = cuda.to_gpu_async(init_h2)
x_list = cuda.to_gpu_async(x_list)
prev_h1 = V(init_h1, volatile=volatile)
prev_h2 = V(init_h2, volatile=volatile)
return prev_h1, prev_h2, x_list
def init_forward(x_list, volatile=False):
init_h1 = np.zeros((batch_size, hidden_size), dtype=dt)
init_h2 = np.zeros((batch_size, hidden_size), dtype=dt)
if args.gpu >= 0:
init_h1 = cuda.to_gpu_async(init_h1)
init_h2 = cuda.to_gpu_async(init_h2)
x_list = cuda.to_gpu_async(x_list)
prev_h1 = V(init_h1, volatile=volatile)
prev_h2 = V(init_h2, volatile=volatile)
return prev_h1, prev_h2, x_list
def forward(x_list, result, use_gpu=False):
init_h = np.zeros((batch_size, hidden_size), dtype=dt)
if use_gpu:
init_h = cuda.to_gpu(init_h)
prev_h = V(init_h)
for x in x_list:
if use_gpu:
x = cuda.to_gpu_async(x)
prev_h = forward_one_step(prev_h, x)
if use_gpu:
result = cuda.to_gpu_async(result)
t = V(result)
y = model.Wy(prev_h)
loss = F.mean_squared_error(y, t)
return loss, y
def forward(model, x_list, result, use_gpu=False):
init_h = np.zeros((batch_size,hidden_size), dtype=dt)
if use_gpu:
x_list = cuda.to_gpu_async(x_list)
result = cuda.to_gpu_async(result)
init_h = cuda.to_gpu(init_h)
prev_h = V(init_h)
for i in xrange(0, 784):
x = x_list[:,i]
prev_h = forward_one_step(model, prev_h, x)
t = V(result)
y = model.Wy(prev_h)
loss = F.softmax_cross_entropy(y, t)
accuracy = F.accuracy(y, t)
return loss, accuracy
def forward(x_list, result, use_gpu=False):
init_h = np.zeros((batch_size,hidden_size), dtype=dt)
if use_gpu:
init_h = cuda.to_gpu(init_h)
prev_h = V(init_h)
for x in x_list:
if use_gpu:
x = cuda.to_gpu_async(x)
prev_h = forward_one_step(prev_h, x)
if use_gpu:
result = cuda.to_gpu_async(result)
t = V(result)
y = model.Wy(prev_h)
loss = F.mean_squared_error(y, t)
return loss, y
def predict(x_list, pred_num=10, volatile=False):
prev_h1, prev_h2, x_list = init_forward(x_list)
preds = list()
pred = None
for cur_word in x_list:
prev_h1, prev_h2, new_loss, pred = forward_one_step(prev_h1, prev_h2, cur_word, cur_word, volatile=volatile)
for i in xrange(0, pred_num):
cur_word = np.array([pred])
if args.gpu >= 0:
cur_word = cuda.to_gpu_async(cur_word)
prev_h1, prev_h2, new_loss, pred = forward_one_step(prev_h1, prev_h2, cur_word, cur_word, volatile=volatile)
preds.append(pred)
return preds
def predict(x_list, pred_num=10, volatile=False):
prev_h1, prev_h2, x_list = init_forward(x_list)
preds = list()
pred = None
for cur_word in x_list:
prev_h1, prev_h2, new_loss, pred = forward_one_step(prev_h1,
prev_h2,
cur_word,
cur_word,
volatile=volatile)
for i in xrange(0, pred_num):
cur_word = np.array([pred])
if args.gpu >= 0:
cur_word = cuda.to_gpu_async(cur_word)
prev_h1, prev_h2, new_loss, pred = forward_one_step(prev_h1,
prev_h2,
cur_word,
cur_word,
volatile=volatile)
preds.append(pred)
return preds
def test_to_gpu_async_unavailable(self):
x = numpy.array([1])
if not cuda.available:
with self.assertRaises(RuntimeError):
cuda.to_gpu_async(x)
def test_to_gpu_async_unavailable(self):
x = numpy.array([1])
if not cuda.available:
with self.assertRaises(RuntimeError):
cuda.to_gpu_async(x)
