def run_lstm():
del_shared()
n_in = X.shape[-1]
n_hid = 20
n_out = y.shape[-1]
random_state = np.random.RandomState(42)
h_init = np.zeros((minibatch_size, 2 * n_hid)).astype("float32")
h0 = tensor.fmatrix()
l1 = lstm_fork([X_sym], [n_in], n_hid, name="l1",
random_state=random_state)
def step(in_t, h_tm1):
h_t = lstm(in_t, h_tm1, n_hid, name="rec", random_state=random_state)
return h_t
h, _ = theano.scan(step, sequences=[l1], outputs_info=[h0])
h_o = slice_state(h, n_hid)
pred = linear([h_o], [n_hid], n_out, name="l2", random_state=random_state)
cost = ((y_sym - pred) ** 2).sum()
params = list(get_params().values())
grads = tensor.grad(cost, params)
learning_rate = 0.000000000001
opt = sgd(params, learning_rate)
updates = opt.updates(params, grads)
f = theano.function([X_sym, y_sym, h0], [cost, h], updates=updates,
mode="FAST_COMPILE")
f(X, y, h_init)
def test_feedforward_theano_mix():
del_shared()
minibatch_size = 100
random_state = np.random.RandomState(1999)
X_sym = tensor.fmatrix()
y_sym = tensor.fmatrix()
l1_o = linear([X_sym], [X.shape[1]],
proj_dim=20,
name='l1',
random_state=random_state)
l1_o = .999 * l1_o
y_pred = softmax([l1_o], [20],
proj_dim=n_classes,
name='out',
random_state=random_state)
cost = categorical_crossentropy(y_pred, y_sym).mean()
params = list(get_params().values())
grads = theano.grad(cost, params)
learning_rate = 0.001
opt = sgd(params, learning_rate)
updates = opt.updates(params, grads)
fit_function = theano.function([X_sym, y_sym], [cost],
updates=updates,
mode="FAST_COMPILE")
cost_function = theano.function([X_sym, y_sym], [cost],
mode="FAST_COMPILE")
train_itr = minibatch_iterator([X, y], minibatch_size, axis=0)
valid_itr = minibatch_iterator([X, y], minibatch_size, axis=0)
X_train, y_train = next(train_itr)
X_valid, y_valid = next(valid_itr)
fit_function(X_train, y_train)
cost_function(X_valid, y_valid)
def test_fixed_projection():
random_state = np.random.RandomState(1999)
rand_projection = random_state.randn(64, 12)
rand_dim = rand_projection.shape[1]
out = fixed_projection([X_sym], [X.shape[1]], rand_projection, 'proj1')
out2 = fixed_projection([X_sym], [X.shape[1]],
rand_projection,
'proj2',
pre=rand_projection[:, 0])
out3 = fixed_projection([X_sym], [X.shape[1]],
rand_projection,
'proj3',
post=rand_projection[0])
final = linear([out2], [rand_dim], 5, 'linear', random_state=random_state)
# Test that it compiles with and without bias
f = theano.function([X_sym], [out, out2, out3, final], mode="FAST_COMPILE")
# Test updates
params = list(get_params().values())
grads = tensor.grad(final.mean(), params)
opt = sgd(params, .1)
updates = opt.updates(params, grads)
f2 = theano.function([X_sym], [out2, final], updates=updates)
ret = f(np.ones_like(X))[0]
assert ret.shape[1] != X.shape[1]
ret2 = f(np.ones_like(X))[1]
assert ret.shape[1] != X.shape[1]
out1, final1 = f2(X)
out2, final2 = f2(X)
# Make sure fixed basis is unchanged
assert_almost_equal(out1, out2)
# Make sure linear layer is updated
assert_raises(AssertionError, assert_almost_equal, final1, final2)
def step(in_t, h_tm1):
h_t = lstm(in_t, h_tm1, [n_hid], n_hid, name="lstm_l1", random_state=random_state)
return h_t
h, _ = theano.scan(step, sequences=[in_fork], outputs_info=[h0])
h_o = slice_state(h, n_hid)
y_pred = softmax([h_o], [n_hid], n_classes, name="h2", random_state=random_state)
loss = categorical_crossentropy(y_pred, y_sym)
cost = loss.mean(axis=1).sum(axis=0)
params = list(get_params().values())
params = params
grads = tensor.grad(cost, params)
learning_rate = 0.0001
opt = adam(params, learning_rate)
updates = opt.updates(params, grads)
fit_function = theano.function([X_sym, y_sym, h0], [cost, h], updates=updates)
cost_function = theano.function([X_sym, y_sym, h0], [cost, h])
predict_function = theano.function([X_sym, h0], [y_pred, h])
def train_loop(itr):
mb = next(itr)
X_mb, y_mb = mb[:-1], mb[1:]
random_state=random_state)
return h_t
h, _ = theano.scan(step, sequences=[in_fork], outputs_info=[h0])
h_o = slice_state(h, n_hid)
y_pred = softmax([h_o], [n_hid],
n_classes,
name="h2",
random_state=random_state)
loss = categorical_crossentropy(y_pred, y_sym)
cost = loss.mean(axis=1).sum(axis=0)
params = list(get_params().values())
params = params
grads = tensor.grad(cost, params)
learning_rate = 0.0001
opt = adam(params, learning_rate)
updates = opt.updates(params, grads)
fit_function = theano.function([X_sym, y_sym, h0], [cost, h], updates=updates)
cost_function = theano.function([X_sym, y_sym, h0], [cost, h])
predict_function = theano.function([X_sym, h0], [y_pred, h])
def train_loop(itr):
mb = next(itr)
X_mb, y_mb = mb[:-1], mb[1:]
