def test_rnn():
def batch_check(x, modes, params):
for m, p in zip(modes, params):
#x.attach_grad()
#with mx.autograd.record():
y = npx.rnn(data=x, parameters=p, mode=m, \
state=np.random.normal(0, 1, (1, 4, 1)), \
state_size=1, num_layers=1)
assert y.shape == (INT_OVERFLOW, 4, 1)
assert type(y).__name__ == 'ndarray'
#y.backward()
#assert x.grad.shape == x.shape
#assert type(x.grad[0]).__name__ == 'ndarray'
data = np.random.normal(0, 1, (INT_OVERFLOW, 4, 4))
modes = ['rnn_relu', 'rnn_tanh', 'gru']
params = [np.random.normal(0, 1, (7,)), \
np.random.normal(0, 1, (7,)), \
np.random.normal(0, 1, (21,))]
batch_check(data, modes, params)
# check lstm seperately because it has an extra param
out = npx.rnn(data=data, parameters=np.random.normal(0, 1, (28,)), \
mode='lstm', \
state=np.random.normal(0, 1, (1, 4, 1)), \
state_cell=np.random.normal(0, 1, (1, 4, 1)), \
state_size=1, num_layers=1)
assert out.shape == (INT_OVERFLOW, 4, 1)
assert type(out[0]).__name__ == 'ndarray'
def batch_check(x, modes, params):
for m, p in zip(modes, params):
#x.attach_grad()
#with mx.autograd.record():
y = npx.rnn(data=x, parameters=p, mode=m, \
state=np.random.normal(0, 1, (1, 4, 1)), \
state_size=1, num_layers=1)
assert y.shape == (INT_OVERFLOW, 4, 1)
assert type(y).__name__ == 'ndarray'
def batch_check(x, modes, params):
state = np.random.normal(0, 1, (1, BAT, L_STA))
for m, p in zip(modes, params):
x.attach_grad()
with mx.autograd.record():
y = npx.rnn(data=x, parameters=p, mode=m, \
state=state, state_size=L_STA, num_layers=1)
assert y.shape == (L_SEQ, BAT, L_STA)
y.backward()
npx.waitall()
def test_rnn_gru():
L_SEQ, BAT, L_INP, L_STA = 2**20, 4, 2**10, 2
data = np.random.uniform(-1, 1, (L_SEQ, BAT, L_INP))
state = np.random.normal(0, 1, (1, BAT, L_STA))
params = np.random.normal(0, 1, (6168, ))
data.attach_grad()
with mx.autograd.record():
out = npx.rnn(data=data, parameters=params, mode='gru', \
state=state, state_size=L_STA, num_layers=1)
assert out.shape == (L_SEQ, BAT, L_STA)
out.backward()
npx.waitall()
