def compare_sampler_grads(lds, num_samples, seed):
init_params, pair_params, node_params = lds
messages, _ = natural_filter_forward_general(init_params, pair_params,
node_params)
def fun1(messages):
npr.seed(seed)
samples = natural_sample_backward_general(messages, pair_params,
num_samples)
return np.sum(np.sin(samples))
grads1 = grad(fun1)(messages)
messages, _ = _natural_filter_forward_general(init_params, pair_params,
node_params)
def fun2(messages):
npr.seed(seed)
samples = _natural_sample_backward(messages, pair_params,
num_samples)
return np.sum(np.sin(samples))
grads2 = grad(fun2)(messages)
unpack_dense_grads = lambda x: interleave(*map(lambda y: zip(*y), x))
assert allclose(grads1, unpack_dense_grads(grads2))
def natural_filter_forward_general(init_params, pair_params, node_params):
init_params = _canonical_init_params(init_params)
node_params = zip(*_canonical_node_params(node_params))
pair_params = _repeat_param(pair_params, len(node_params) - 1)
def unit(J, h, logZ):
return [(J, h)], logZ
def bind(result, step):
messages, lognorm = result
new_message, term = step(messages[-1])
return messages + [new_message], lognorm + term
condition = lambda node_param: lambda (J, h): natural_condition_on_general(J, h, *node_param)
predict = lambda pair_param: lambda (J, h): natural_predict(J, h, *pair_param)
steps = interleave(map(condition, node_params), map(predict, pair_params))
messages, lognorm = monad_runner(bind)(unit(*init_params), steps)
lognorm += natural_lognorm(*messages[-1])
return messages, lognorm
def natural_filter_forward(natparam, data):
T, p = data.shape
init_params, pair_params, node_params = _unpack_repeated(natparam, T)
def unit(J, h):
return [(J, h)], 0.
def bind(result, step):
messages, lognorm = result
new_message, term = step(messages[-1])
return messages + [new_message], lognorm + term
condition = lambda node_param, y: lambda (J, h): natural_condition_on(J, h, y, *node_param)
predict = lambda pair_param: lambda (J, h): natural_predict(J, h, *pair_param)
steps = interleave(map(condition, node_params, data), map(predict, pair_params))
J_init, h_init, logZ_init = init_params
messages, lognorm = monad_runner(bind)(unit(J_init, h_init), steps)
lognorm += natural_lognorm(*messages[-1]) + logZ_init
return messages, lognorm - T*p/2*np.log(2*np.pi)
def compare_sampler_grads(lds, num_samples, seed):
init_params, pair_params, node_params = lds
messages, _ = natural_filter_forward_general(
init_params, pair_params, node_params)
def fun1(messages):
npr.seed(seed)
samples = natural_sample_backward_general(messages, pair_params, num_samples)
return np.sum(np.sin(samples))
grads1 = grad(fun1)(messages)
messages, _ = _natural_filter_forward_general(
init_params, pair_params, node_params)
def fun2(messages):
npr.seed(seed)
samples = _natural_sample_backward(messages, pair_params, num_samples)
return np.sum(np.sin(samples))
grads2 = grad(fun2)(messages)
unpack_dense_grads = lambda x: interleave(*map(lambda y: zip(*y), x))
assert allclose(grads1, unpack_dense_grads(grads2))
def natural_filter_forward_general(init_params, pair_params, node_params):
init_params = _canonical_init_params(init_params)
node_params = zip(*_canonical_node_params(node_params))
pair_params = _repeat_param(pair_params, len(node_params) - 1)
def unit(J, h, logZ):
return [(J, h)], logZ
def bind(result, step):
messages, lognorm = result
new_message, term = step(messages[-1])
return messages + [new_message], lognorm + term
condition = lambda node_param: lambda (J, h): natural_condition_on_general(
J, h, *node_param)
predict = lambda pair_param: lambda (J, h): natural_predict(
J, h, *pair_param)
steps = interleave(map(condition, node_params), map(predict, pair_params))
messages, lognorm = monad_runner(bind)(unit(*init_params), steps)
lognorm += natural_lognorm(*messages[-1])
return messages, lognorm
def unpack_dense_messages(messages):
(J_predict, h_predict), (J_filtered, h_filtered) = messages
prediction_messages = zip(J_predict, h_predict)
filtered_messages = zip(J_filtered, h_filtered)
return interleave(prediction_messages, filtered_messages)
def unpack_dense_messages(messages):
(J_predict, h_predict), (J_filtered, h_filtered) = messages
prediction_messages = zip(J_predict, h_predict)
filtered_messages = zip(J_filtered, h_filtered)
return interleave(prediction_messages, filtered_messages)