def test_ocr_rnn():
length = 5
carry_size = 3
class_count = 4
inputs = np.zeros((1, length, 4))
def rnn():
return Rnn(*GRUCell(carry_size, zeros))
net = Sequential(
rnn(),
rnn(),
rnn(),
lambda x: np.reshape(x, (-1, carry_size)
), # -> same weights for all time steps
Dense(class_count, zeros, zeros),
softmax,
lambda x: np.reshape(x, (-1, length, class_count)))
params = net.init_parameters(PRNGKey(0), inputs)
assert len(params) == 4
cell = params.rnn0.gru_cell
assert len(cell) == 3
assert np.array_equal(np.zeros((7, 3)), cell.update_kernel)
assert np.array_equal(np.zeros((7, 3)), cell.reset_kernel)
assert np.array_equal(np.zeros((7, 3)), cell.compute_kernel)
out = net.apply(params, inputs)
assert np.array_equal(.25 * np.ones((1, 5, 4)), out)
def conv_block(inputs):
main = Sequential(Conv(filters1, (1, 1), strides), BatchNorm(), relu,
Conv(filters2, (ks, ks),
padding='SAME'), BatchNorm(), relu,
Conv(filters3, (1, 1)), BatchNorm())
shortcut = Sequential(Conv(filters3, (1, 1), strides), BatchNorm())
return relu(sum((main(inputs), shortcut(inputs))))
def res_layer(inputs):
"""
From original doc string:
The layer contains a gated filter that connects to dense output
and to a skip connection:
|-> [gate] -| |-> 1x1 conv -> skip output
| |-> (*) -|
input -|-> [filter] -| |-> 1x1 conv -|
| |-> (+) -> dense output
|------------------------------------|
Where `[gate]` and `[filter]` are causal convolutions with a
non-linear activation at the output
"""
gated = Sequential(
Conv1D(dilation_channels, (filter_width, ), dilation=(dilation, )),
sigmoid)(inputs)
filtered = Sequential(
Conv1D(dilation_channels, (filter_width, ), dilation=(dilation, )),
np.tanh)(inputs)
p = gated * filtered
out = Conv1D(residual_channels, (1, ), padding='SAME')(p)
# Add the transformed output of the resblock to the sliced input:
sliced_inputs = lax.dynamic_slice(
inputs, [0, inputs.shape[1] - out.shape[1], 0],
[inputs.shape[0], out.shape[1], inputs.shape[2]])
new_out = sum(out, sliced_inputs)
skip = Conv1D(residual_channels, (1, ),
padding='SAME')(skip_slice(p, output_width))
return new_out, skip
def test_regularized_submodule():
net = Sequential(Conv(2, (1, 1)), relu, Conv(2, (1, 1)), relu, flatten,
L2Regularized(Sequential(Dense(2), relu, Dense(2), np.sum), .1))
input = np.ones((1, 3, 3, 1))
params = net.init_parameters(input, key=PRNGKey(0))
assert (2, 2) == params.regularized.model.dense1.kernel.shape
out = net.apply(params, input)
assert () == out.shape
def test_submodule_reuse():
inputs = np.zeros((1, 2))
layer = Dense(5)
net1 = Sequential(layer, Dense(2))
net2 = Sequential(layer, Dense(3))
layer_params = layer.init_parameters(inputs, key=PRNGKey(0))
net1_params = net1.init_parameters(inputs,
key=PRNGKey(1),
reuse={layer: layer_params})
net2_params = net2.init_parameters(inputs,
key=PRNGKey(2),
reuse={layer: layer_params})
out1 = net1.apply(net1_params, inputs)
assert out1.shape == (1, 2)
out2 = net2.apply(net2_params, inputs)
assert out2.shape == (1, 3)
assert_dense_parameters_equal(layer_params, net1_params[0])
assert_dense_parameters_equal(layer_params, net2_params[0])
new_layer_params = layer.init_parameters(inputs, key=PRNGKey(3))
combined_params = net1.parameters_from(
{
net1: net1_params,
layer: new_layer_params
}, inputs)
assert_dense_parameters_equal(new_layer_params, combined_params.dense0)
assert_dense_parameters_equal(net1_params.dense1, combined_params.dense1)
def test_reparametrized_submodule():
net = Sequential(
Conv(2, (3, 3)), relu, Conv(2, (3, 3)), relu, flatten,
Reparametrized(Sequential(Dense(2), relu, Dense(2)), Scaled))
input = np.ones((1, 3, 3, 1))
params = net.init_parameters(PRNGKey(0), input)
assert (2, 2) == params.reparametrized.model.dense1.kernel.shape
out = net.apply(params, input)
assert (1, 2) == out.shape
def test_pool_shape(Pool):
conv = Conv(2,
filter_shape=(3, 3),
padding='SAME',
kernel_init=zeros,
bias_init=zeros)
inputs = np.zeros((1, 5, 5, 2))
pooled = Sequential(conv, Pool(window_shape=(1, 1), strides=(2, 2)))
params = pooled.init_parameters(PRNGKey(0), inputs)
out = pooled.apply(params, inputs)
assert np.array_equal(np.zeros((1, 3, 3, 2)), out)
def test_input_dependent_nested_modules():
@parametrized
def layer(inputs):
return Dense(inputs.shape[0])(inputs)
net = Sequential(Dense(3), layer)
inputs = np.zeros((5, 3))
params = net.init_parameters(inputs, key=PRNGKey(0))
out = net.apply(params, inputs)
assert (5, 5) == out.shape
def test_reuse_api():
inputs = np.zeros((1, 2))
net = Dense(5)
net_params = net.init_parameters(inputs, key=PRNGKey(0))
# train net params...
transfer_net = Sequential(net, relu, Dense(2))
transfer_net_params = transfer_net.init_parameters(inputs, key=PRNGKey(1),
reuse={net: net_params})
assert net_params == transfer_net_params.dense0
def test_external_param_sharing():
layer = Dense(2, zeros, zeros)
shared_net = Sequential(layer, layer)
inputs = np.zeros((1, 2))
params = shared_net.init_parameters(inputs, key=PRNGKey(0))
assert_parameters_equal(((np.zeros((2, 2)), np.zeros(2)), ), params)
out = shared_net.apply(params, inputs)
assert np.array_equal(np.zeros((1, 2)), out)
out = shared_net.apply(params, inputs, jit=True)
assert np.array_equal(np.zeros((1, 2)), out)
def test_Sequential_graceful_update_message():
message = 'Call like Sequential(Dense(10), relu), without "[" and "]". ' \
'(Or pass iterables with Sequential(*layers).)'
try:
Sequential([Dense(2), relu])
assert False
except ValueError as e:
assert message == str(e)
try:
Sequential(Dense(2) for _ in range(3))
assert False
except ValueError as e:
assert message == str(e)
def test_flatten_shape():
conv = Conv(2,
filter_shape=(3, 3),
padding='SAME',
kernel_init=zeros,
bias_init=zeros)
inputs = np.zeros((1, 5, 5, 2))
params = conv.init_parameters(PRNGKey(0), inputs)
out = conv.apply(params, inputs)
assert np.array_equal(np.zeros((1, 5, 5, 2)), out)
flattened = Sequential(conv, flatten)
out = flattened.apply_from({conv: params}, inputs)
assert np.array_equal(np.zeros((1, 50)), out)
def test_diamond_shared_submodules():
p = Parameter(lambda rng: np.ones(()))
a = Sequential(p)
b = Sequential(p)
@parametrized
def net(inputs):
return a(inputs), b(inputs)
params = net.init_parameters(PRNGKey(0), np.zeros(()))
assert 1 == len(params)
assert np.array_equal(np.ones(()), params)
a, b = net.apply(params, np.zeros(()))
assert np.array_equal(np.ones(()), a)
assert np.array_equal(np.ones(()), b)
def test_collection_input(type):
@parametrized
def net(inputs):
assert isinstance(inputs, type)
return inputs[0] * inputs[1] * parameter((), zeros)
inputs = type((np.zeros(2), np.zeros(2)))
params = net.init_parameters(inputs, key=PRNGKey(0))
out = net.apply(params, inputs)
assert np.array_equal(np.zeros(2), out)
net = Sequential(net)
params = net.init_parameters(inputs, key=PRNGKey(0))
out = net.apply(params, inputs)
assert np.array_equal(np.zeros(2), out)
def identity_block(inputs):
main = Sequential(Conv(filters1, (1, 1)), BatchNorm(), relu,
Conv(filters2, (ks, ks), padding='SAME'),
BatchNorm(), relu, Conv(inputs.shape[3], (1, 1)),
BatchNorm())
return relu(sum((main(inputs), inputs)))
def main(batch_size=256, env_name="CartPole-v1"):
env = gym.make(env_name)
policy = Sequential(Dense(64), relu, Dense(env.action_space.n))
@parametrized
def loss(observations, actions, rewards_to_go):
logprobs = log_softmax(policy(observations))
action_logprobs = logprobs[np.arange(logprobs.shape[0]), actions]
return -np.mean(action_logprobs * rewards_to_go, axis=0)
opt = Adam()
shaped_loss = loss.shaped(np.zeros((1, ) + env.observation_space.shape),
np.array([0]), np.array([0]))
@jit
def sample_action(state, key, observation):
loss_params = opt.get_parameters(state)
logits = policy.apply_from({shaped_loss: loss_params}, observation)
return sample_categorical(key, logits)
rng_init, rng = random.split(PRNGKey(0))
state = opt.init(shaped_loss.init_parameters(key=rng_init))
returns, observations, actions, rewards_to_go = [], [], [], []
for i in range(250):
while len(observations) < batch_size:
observation = env.reset()
episode_done = False
rewards = []
while not episode_done:
rng_step, rng = random.split(rng)
action = sample_action(state, rng_step, observation)
observations.append(observation)
actions.append(action)
observation, reward, episode_done, info = env.step(int(action))
rewards.append(reward)
returns.append(onp.sum(rewards))
rewards_to_go += list(onp.flip(onp.cumsum(onp.flip(rewards))))
print(f'Batch {i}, recent mean return: {onp.mean(returns[-100:]):.1f}')
state = opt.update(loss.apply,
state,
np.array(observations[:batch_size]),
np.array(actions[:batch_size]),
np.array(rewards_to_go[:batch_size]),
jit=True)
observations = observations[batch_size:]
actions = actions[batch_size:]
rewards_to_go = rewards_to_go[batch_size:]
def test_params_from_shared_submodules2():
sublayer = Dense(2)
a = Sequential(sublayer, relu)
b = Sequential(sublayer, np.sum)
@parametrized
def net(inputs):
return a(inputs), b(inputs)
inputs = np.zeros((1, 3))
a_params = a.init_parameters(PRNGKey(0), inputs)
out = a.apply(a_params, inputs)
params = net.parameters_from({a: a_params}, inputs)
assert_dense_params_equal(a_params.dense, params.sequential0.dense)
assert_dense_params_equal(a_params.dense, params.sequential1.dense)
# TODO parameters are duplicated, optimization with weight sharing is wrong:
# TODO instead: assert 1 == len(params)
out_, _ = net.apply(params, inputs)
assert np.array_equal(out, out_)
def wavenet(inputs):
hidden = Conv1D(residual_channels, (initial_filter_width, ))(inputs)
out = np.zeros((hidden.shape[0], out_width, residual_channels),
'float32')
for dilation in dilations:
res = ResLayer(dilation_channels, residual_channels, filter_width,
dilation, out_width)(hidden)
hidden, out_partial = res
out += out_partial
return Sequential(relu, Conv1D(skip_channels, (1, )), relu,
Conv1D(3 * nr_mix, (1, )))(out)
def test_ocr_rnn():
length = 5
carry_size = 3
class_count = 4
inputs = jnp.zeros((1, length, 4))
def rnn():
return Rnn(*GRUCell(carry_size, zeros))
net = Sequential(
rnn(),
rnn(),
rnn(),
lambda x: jnp.reshape(x, (-1, carry_size)
), # -> same weights for all time steps
Dense(class_count, zeros, zeros),
softmax,
lambda x: jnp.reshape(x, (-1, length, class_count)))
params = net.init_parameters(inputs, key=PRNGKey(0))
assert len(params) == 4
cell = params.rnn0.gru_cell
assert len(cell) == 3
assert jnp.array_equal(jnp.zeros((7, 3)), cell.update_kernel)
assert jnp.array_equal(jnp.zeros((7, 3)), cell.reset_kernel)
assert jnp.array_equal(jnp.zeros((7, 3)), cell.compute_kernel)
out = net.apply(params, inputs)
@parametrized
def cross_entropy(images, targets):
prediction = net(images)
return jnp.mean(-jnp.sum(targets * jnp.log(prediction), (1, 2)))
opt = optimizers.RmsProp(0.003)
state = opt.init(cross_entropy.init_parameters(inputs, out,
key=PRNGKey(0)))
state = opt.update(cross_entropy.apply, state, inputs, out)
opt.update(cross_entropy.apply, state, inputs, out, jit=True)
def test_L2Regularized_sequential():
loss = Sequential(Dense(1, ones, ones), relu, Dense(1, ones, ones), sum)
reg_loss = L2Regularized(loss, scale=2)
inputs = np.ones(1)
params = reg_loss.init_parameters(PRNGKey(0), inputs)
assert np.array_equal(np.ones((1, 1)), params.model.dense0.kernel)
assert np.array_equal(np.ones((1, 1)), params.model.dense1.kernel)
reg_loss_out = reg_loss.apply(params, inputs)
assert 7 == reg_loss_out
def test_parameters_from_shared_submodules():
sublayer = Dense(2)
a = Sequential(sublayer, relu)
b = Sequential(sublayer, np.sum)
@parametrized
def net(inputs):
return a(inputs) * b(inputs)
inputs = np.zeros((1, 3))
a_params = a.init_parameters(inputs, key=PRNGKey(0))
out = a.apply(a_params, inputs)
params = net.parameters_from({a: a_params}, inputs)
assert_parameters_equal(a_params.dense.kernel,
params.sequential0.dense.kernel)
assert_parameters_equal((), params.sequential1)
out = net.apply(params, inputs)
out_ = net.apply_from({a: a_params}, inputs)
assert np.array_equal(out, out_)
out_ = net.apply_from({a: a_params}, inputs, jit=True)
assert np.array_equal(out, out_)
out_ = net.apply_from({a.shaped(inputs): a_params}, inputs)
assert np.array_equal(out, out_)
out_ = net.apply_from({a.shaped(inputs): a_params}, inputs, jit=True)
assert np.array_equal(out, out_)
out_ = net.shaped(inputs).apply_from({a: a_params})
assert np.array_equal(out, out_)
out_ = net.shaped(inputs).apply_from({a: a_params}, jit=True)
assert np.array_equal(out, out_)
out_ = net.shaped(inputs).apply_from({a.shaped(inputs): a_params})
assert np.array_equal(out, out_)
out_ = net.shaped(inputs).apply_from({a.shaped(inputs): a_params},
jit=True)
assert np.array_equal(out, out_)
def ResNet50(num_classes):
return Sequential(
GeneralConv(('HWCN', 'OIHW', 'NHWC'), 64, (7, 7), (2, 2), 'SAME'),
BatchNorm(), relu, MaxPool((3, 3), strides=(2, 2)),
ConvBlock(3, [64, 64, 256], strides=(1, 1)),
IdentityBlock(3, [64, 64]), IdentityBlock(3, [64, 64]),
ConvBlock(3, [128, 128, 512]), IdentityBlock(3, [128, 128]),
IdentityBlock(3, [128, 128]), IdentityBlock(3, [128, 128]),
ConvBlock(3, [256, 256, 1024]), IdentityBlock(3, [256, 256]),
IdentityBlock(3, [256, 256]), IdentityBlock(3, [256, 256]),
IdentityBlock(3, [256, 256]), IdentityBlock(3, [256, 256]),
ConvBlock(3, [512, 512, 2048]), IdentityBlock(3, [512, 512]),
IdentityBlock(3, [512, 512]), AvgPool((7, 7)), flatten,
Dense(num_classes), logsoftmax)
def test_parameters_from_subsubmodule():
subsublayer = Dense(2)
sublayer = Sequential(subsublayer, relu)
net = Sequential(sublayer, np.sum)
inputs = np.zeros((1, 3))
params = net.init_parameters(inputs, key=PRNGKey(0))
out = net.apply(params, inputs)
subsublayer_params = subsublayer.init_parameters(inputs, key=PRNGKey(0))
params_ = net.parameters_from({subsublayer: subsublayer_params}, inputs)
assert_dense_parameters_equal(subsublayer_params, params_[0][0])
out_ = net.apply(params_, inputs)
assert out.shape == out_.shape
out_ = net.apply_from({subsublayer: subsublayer_params}, inputs)
assert out.shape == out_.shape
out_ = net.apply_from({subsublayer: subsublayer_params}, inputs, jit=True)
assert out.shape == out_.shape
def main():
# TODO https://github.com/JuliusKunze/jaxnet/issues/4
print("Sorry, this example does not work yet work with the new jax version.")
return
train, test = read_dataset()
_, length, x_size = train.data.shape
class_count = train.target.shape[2]
carry_size = 200
batch_size = 10
def rnn():
return Rnn(*GRUCell(carry_size=carry_size,
param_init=lambda rng, shape: random.normal(rng, shape) * 0.01))
net = Sequential(
rnn(),
rnn(),
rnn(),
lambda x: np.reshape(x, (-1, carry_size)), # -> same weights for all time steps
Dense(out_dim=class_count),
softmax,
lambda x: np.reshape(x, (-1, length, class_count)))
@parametrized
def cross_entropy(images, targets):
prediction = net(images)
return np.mean(-np.sum(targets * np.log(prediction), (1, 2)))
@parametrized
def error(inputs, targets):
prediction = net(inputs)
return np.mean(np.not_equal(np.argmax(targets, 2), np.argmax(prediction, 2)))
opt = optimizers.RmsProp(0.003)
batch = train.sample(batch_size)
params = cross_entropy.init_parameters(random.PRNGKey(0), batch.data, batch.target)
state = opt.init(params)
for epoch in range(10):
params = get_params(state)
e = error.apply_from({cross_entropy: params}, test.data, test.target, jit=True)
print(f'Epoch {epoch} error {e * 100:.1f}')
break # TODO https://github.com/JuliusKunze/jaxnet/issues/2
for _ in range(100):
batch = train.sample(batch_size)
state = opt.update(cross_entropy.apply, state, batch.data, batch.target, jit=True)
def test_submodule_reuse():
inputs = np.zeros((1, 2))
layer = Dense(5)
net1 = Sequential(layer, Dense(2))
net2 = Sequential(layer, Dense(3))
layer_params = layer.init_parameters(PRNGKey(0), inputs)
net1_params = net1.init_parameters(PRNGKey(1), inputs, reuse={layer: layer_params})
net2_params = net2.init_parameters(PRNGKey(2), inputs, reuse={layer: layer_params})
out1 = net1.apply(net1_params, inputs)
assert out1.shape == (1, 2)
out2 = net2.apply(net2_params, inputs)
assert out2.shape == (1, 3)
assert_dense_params_equal(layer_params, net1_params[0])
assert_dense_params_equal(layer_params, net2_params[0])
def test_parameters_from_sharing_between_multiple_parents():
a = Dense(2)
b = Sequential(a, np.sum)
@parametrized
def net(inputs):
return a(inputs), b(inputs)
inputs = np.zeros((1, 3))
a_params = a.init_parameters(inputs, key=PRNGKey(0))
out = a.apply(a_params, inputs)
params = net.parameters_from({a: a_params}, inputs)
assert_dense_parameters_equal(a_params, params.dense)
assert_parameters_equal((), params.sequential)
assert 2 == len(params)
out_, _ = net.apply(params, inputs)
assert np.array_equal(out, out_)
def test_mnist_vae():
@parametrized
def encode(input):
input = Sequential(Dense(5), relu, Dense(5), relu)(input)
mean = Dense(10)(input)
variance = Sequential(Dense(10), softplus)(input)
return mean, variance
decode = Sequential(Dense(5), relu, Dense(5), relu, Dense(5 * 5))
@parametrized
def elbo(key, images):
mu_z, sigmasq_z = encode(images)
logits_x = decode(gaussian_sample(key, mu_z, sigmasq_z))
return bernoulli_logpdf(logits_x, images) - gaussian_kl(mu_z, sigmasq_z)
params = elbo.init_parameters(PRNGKey(0), np.zeros((32, 5 * 5)), key=PRNGKey(0))
assert (5, 10) == params.encode.sequential1.dense.kernel.shape
def test_no_reuse():
inputs = np.zeros((1, 2))
layer = Dense(5)
net1 = Sequential(layer, Dense(2))
p1 = net1.init_parameters(inputs, key=PRNGKey(0))
net2 = Sequential(layer, Dense(3))
p2 = net2.init_parameters(inputs, key=PRNGKey(1))
assert p1[0].kernel.shape == p2[0].kernel.shape
assert p1[0].bias.shape == p2[0].bias.shape
assert not np.array_equal(p1[0][0], p2[0][0])
assert not np.array_equal(p1[0][1], p2[0][1])
def main():
train, test = read_dataset()
_, length, x_size = train.data.shape
class_count = train.target.shape[2]
carry_size = 200
batch_size = 10
def rnn():
return Rnn(*GRUCell(carry_size=carry_size,
param_init=lambda key, shape: random.normal(key, shape) * 0.01))
net = Sequential(
rnn(),
rnn(),
rnn(),
lambda x: np.reshape(x, (-1, carry_size)), # -> same weights for all time steps
Dense(out_dim=class_count),
softmax,
lambda x: np.reshape(x, (-1, length, class_count)))
@parametrized
def cross_entropy(images, targets):
prediction = net(images)
return np.mean(-np.sum(targets * np.log(prediction), (1, 2)))
@parametrized
def error(inputs, targets):
prediction = net(inputs)
return np.mean(np.not_equal(np.argmax(targets, 2), np.argmax(prediction, 2)))
opt = optimizers.RmsProp(0.003)
batch = train.sample(batch_size)
state = opt.init(cross_entropy.init_parameters(batch.data, batch.target, key=PRNGKey(0)))
for epoch in range(10):
params = opt.get_parameters(state)
e = error.apply_from({cross_entropy: params}, test.data, test.target, jit=True)
print(f'Epoch {epoch} error {e * 100:.1f}')
for _ in range(100):
batch = train.sample(batch_size)
state = opt.update(cross_entropy.apply, state, batch.data, batch.target, jit=True)
def test_readme():
net = Sequential(Dense(1024), relu, Dense(1024), relu, Dense(4), log_softmax)
@parametrized
def loss(inputs, targets):
return -np.mean(net(inputs) * targets)
def next_batch(): return np.zeros((3, 784)), np.zeros((3, 4))
params = loss.init_parameters(*next_batch(), key=PRNGKey(0))
print(params.sequential.dense2.bias) # [-0.01101029, -0.00749435, -0.00952365, 0.00493979]
assert np.allclose([-0.01101029, -0.00749435, -0.00952365, 0.00493979],
params.sequential.dense2.bias)
out = loss.apply(params, *next_batch())
assert () == out.shape
out_ = loss.apply(params, *next_batch(), jit=True)
assert out.shape == out_.shape
