Python Model.fprop方法代码示例

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
def test_conv_rnn(backend_default):
    train_shape = (1, 17, 142)

    be = NervanaObject.be
    inp = be.array(be.rng.randn(np.prod(train_shape), be.bsz))
    delta = be.array(be.rng.randn(10, be.bsz))

    init_norm = Gaussian(loc=0.0, scale=0.01)
    bilstm = DeepBiLSTM(128, init_norm, activation=Rectlin(), gate_activation=Rectlin(),
                        depth=1, reset_cells=True)
    birnn_1 = DeepBiRNN(128, init_norm, activation=Rectlin(),
                        depth=1, reset_cells=True, batch_norm=False)
    birnn_2 = DeepBiRNN(128, init_norm, activation=Rectlin(),
                        depth=2, reset_cells=True, batch_norm=False)
    bibnrnn = DeepBiRNN(128, init_norm, activation=Rectlin(),
                        depth=1, reset_cells=True, batch_norm=True)
    birnnsum = DeepBiRNN(128, init_norm, activation=Rectlin(),
                         depth=1, reset_cells=True, batch_norm=False, bi_sum=True)
    rnn = Recurrent(128, init=init_norm, activation=Rectlin(), reset_cells=True)
    lstm = LSTM(128, init_norm, activation=Rectlin(), gate_activation=Rectlin(), reset_cells=True)
    gru = GRU(128, init_norm, activation=Rectlin(), gate_activation=Rectlin(), reset_cells=True)

    rlayers = [bilstm, birnn_1, birnn_2, bibnrnn, birnnsum, rnn, lstm, gru]

    for rl in rlayers:
        layers = [
                    Conv((2, 2, 4), init=init_norm, activation=Rectlin(),
                         strides=dict(str_h=2, str_w=4)),
                    Pooling(2, strides=2),
                    Conv((3, 3, 4), init=init_norm, batch_norm=True, activation=Rectlin(),
                         strides=dict(str_h=1, str_w=2)),
                    rl,
                    RecurrentMean(),
                    Affine(nout=10, init=init_norm, activation=Rectlin()),
                ]
        model = Model(layers=layers)
        cost = GeneralizedCost(costfunc=CrossEntropyBinary())
        model.initialize(train_shape, cost)
        model.fprop(inp)
        model.bprop(delta)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
def test_model_get_outputs(backend):
    (X_train, y_train), (X_test, y_test), nclass = load_mnist()
    train_set = DataIterator(X_train[:backend.bsz * 3])

    init_norm = Gaussian(loc=0.0, scale=0.1)

    layers = [Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin()),
              Affine(nout=10, init=init_norm, activation=Logistic(shortcut=True))]
    mlp = Model(layers=layers)
    out_list = []
    for x, t in train_set:
        x = mlp.fprop(x)
        out_list.append(x.get().T.copy())
    ref_output = np.vstack(out_list)

    train_set.reset()
    output = mlp.get_outputs(train_set)
    assert np.allclose(output, ref_output)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
def test_model_get_outputs(backend_default, data):
    dataset = MNIST(path=data)
    train_set = dataset.train_iter

    init_norm = Gaussian(loc=0.0, scale=0.1)

    layers = [Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin()),
              Affine(nout=10, init=init_norm, activation=Logistic(shortcut=True))]
    mlp = Model(layers=layers)
    out_list = []
    mlp.initialize(train_set)
    for x, t in train_set:
        x = mlp.fprop(x)
        out_list.append(x.get().T.copy())
    ref_output = np.vstack(out_list)

    train_set.reset()
    output = mlp.get_outputs(train_set)
    assert allclose_with_out(output, ref_output[:output.shape[0], :])

    # test model benchmark inference
    mlp.benchmark(train_set, inference=True, niterations=5)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
def test_model_get_outputs(backend_default, data):
    (X_train, y_train), (X_test, y_test), nclass = load_mnist(path=data)
    train_set = ArrayIterator(X_train[:backend_default.bsz * 3])

    init_norm = Gaussian(loc=0.0, scale=0.1)

    layers = [Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin()),
              Affine(nout=10, init=init_norm, activation=Logistic(shortcut=True))]
    mlp = Model(layers=layers)
    out_list = []
    mlp.initialize(train_set)
    for x, t in train_set:
        x = mlp.fprop(x)
        out_list.append(x.get().T.copy())
    ref_output = np.vstack(out_list)

    train_set.reset()
    output = mlp.get_outputs(train_set)
    assert np.allclose(output, ref_output)

    # test model benchmark inference
    mlp.benchmark(train_set, inference=True, niterations=5)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
layers = [
    LSTM(hidden_size, init, activation=Logistic(), gate_activation=Tanh()),
    Affine(len(train_set.vocab), init, bias=init, activation=Softmax())
]
model_new = Model(layers=layers)
model_new.load_params(args.save_path)
model_new.initialize(dataset=(train_set.shape[0], time_steps))

# Generate text
text = []
seed_tokens = list('ROMEO:')

x = model_new.be.zeros((len(train_set.vocab), time_steps))

for s in seed_tokens:
    x.fill(0)
    x[train_set.token_to_index[s], 0] = 1
    y = model_new.fprop(x)

for i in range(num_predict):
    # Take last prediction and feed into next fprop
    pred = sample(y.get()[:, -1])
    text.append(train_set.index_to_token[int(pred)])

    x.fill(0)
    x[int(pred), 0] = 1
    y = model_new.fprop(x)

neon_logger.display(''.join(seed_tokens + text))

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]

#.........这里部分代码省略.........
    layers.append(Conv((3, 3, 64), strides=1, init=init_norm, activation=Rectlin(), batch_norm=self.batch_norm))
    # The final hidden layer is fully-connected and consists of 512 rectifier units.
    layers.append(Affine(nout=512, init=init_norm, activation=Rectlin(), batch_norm=self.batch_norm))
    # The output layer is a fully-connected linear layer with a single output for each valid action.
    layers.append(Affine(nout=num_actions, init = init_norm))
    return layers

  def _setInput(self, states):
    # change order of axes to match what Neon expects
    states = np.transpose(states, axes = (1, 2, 3, 0))
    # copy() shouldn't be necessary here, but Neon doesn't work otherwise
    self.input.set(states.copy())
    # normalize network input between 0 and 1
    self.be.divide(self.input, 255, self.input)

  def train(self, minibatch, epoch):
    # expand components of minibatch
    prestates, actions, rewards, poststates, terminals = minibatch
    assert len(prestates.shape) == 4
    assert len(poststates.shape) == 4
    assert len(actions.shape) == 1
    assert len(rewards.shape) == 1
    assert len(terminals.shape) == 1
    assert prestates.shape == poststates.shape
    assert prestates.shape[0] == actions.shape[0] == rewards.shape[0] == poststates.shape[0] == terminals.shape[0]

    if self.target_steps and self.train_iterations % self.target_steps == 0:
      # have to serialize also states for batch normalization to work
      pdict = self.model.get_description(get_weights=True, keep_states=True)
      self.target_model.deserialize(pdict, load_states=True)

    # feed-forward pass for poststates to get Q-values
    self._setInput(poststates)
    postq = self.target_model.fprop(self.input, inference = True)
    assert postq.shape == (self.num_actions, self.batch_size)

    # calculate max Q-value for each poststate
    maxpostq = self.be.max(postq, axis=0).asnumpyarray()
    assert maxpostq.shape == (1, self.batch_size)

    # feed-forward pass for prestates
    self._setInput(prestates)
    preq = self.model.fprop(self.input, inference = False)
    assert preq.shape == (self.num_actions, self.batch_size)

    # make copy of prestate Q-values as targets
    # It seems neccessary for cpu backend.
    targets = preq.asnumpyarray().copy()

    # clip rewards between -1 and 1
    rewards = np.clip(rewards, self.min_reward, self.max_reward)

    # update Q-value targets for actions taken
    for i, action in enumerate(actions):
      if terminals[i]:
        targets[action, i] = float(rewards[i])
      else:
        targets[action, i] = float(rewards[i]) + self.discount_rate * maxpostq[0,i]

    # copy targets to GPU memory
    self.targets.set(targets)

    # calculate errors
    deltas = self.cost.get_errors(preq, self.targets)
    assert deltas.shape == (self.num_actions, self.batch_size)
    #assert np.count_nonzero(deltas.asnumpyarray()) == 32

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
train.reset()
# get 1 image
for im, l in train:
    break
train.exit_batch_provider()
with open('im1.pkl', 'w') as fid:
    pickle.dump((im.get(), l.get()), fid)
im_save = im.get().copy()
if args.resume:
    with open('im1.pkl', 'r') as fid:
        (im2, l2) = pickle.load(fid)
    im.set(im2)
    l.set(l2)

# run fprop and bprop on this minibatch save the results
out_fprop = model.fprop(im)
out_fprop_save = [x.get() for x in out_fprop]
im.set(im_save)
out_fprop = model.fprop(im)
out_fprop_save2 = [x.get() for x in out_fprop]
for x, y in zip(out_fprop_save, out_fprop_save2):
    assert np.max(np.abs(x-y)) == 0.0, '2 fprop iterations do not match'

# run fit fot 1 minibatch
# have to do this by hand
delta = model.cost.get_errors(im, l)
model.bprop(delta)
if args.resume:
    model.optimizer = opt
model.optimizer.optimize(model.layers_to_optimize, epoch=model.epoch_index)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
class DeepQNetwork:
  def __init__(self, num_actions, args):
    # create Neon backend
    self.be = gen_backend(backend = args.backend,
                 batch_size = args.batch_size,
                 rng_seed = args.random_seed,
                 device_id = args.device_id,
                 default_dtype = np.dtype(args.datatype).type,
                 stochastic_round = args.stochastic_round)

    # create model
    layers = self.createLayers(num_actions)
    self.model = Model(layers = layers)
    self.cost = GeneralizedCost(costfunc = SumSquared())
    self.optimizer = RMSProp(learning_rate = args.learning_rate, 
        decay_rate = args.rmsprop_decay_rate, 
        stochastic_round = args.stochastic_round)

    # create target model
    self.target_steps = args.target_steps
    self.train_iterations = 0
    if self.target_steps:
      self.target_model = Model(layers = self.createLayers(num_actions))
      self.save_weights_path = args.save_weights_path
    else:
      self.target_model = self.model

    # remember parameters
    self.num_actions = num_actions
    self.batch_size = args.batch_size
    self.discount_rate = args.discount_rate
    self.history_length = args.history_length
    self.screen_dim = (args.screen_height, args.screen_width)
    self.clip_error = args.clip_error

    # prepare tensors once and reuse them
    self.input_shape = (self.history_length,) + self.screen_dim + (self.batch_size,)
    self.tensor = self.be.empty(self.input_shape)
    self.tensor.lshape = self.input_shape # needed for convolutional networks
    self.targets = self.be.empty((self.num_actions, self.batch_size))

    self.callback = None

  def createLayers(self, num_actions):
    # create network
    init_norm = Gaussian(loc=0.0, scale=0.01)
    layers = []
    # The first hidden layer convolves 32 filters of 8x8 with stride 4 with the input image and applies a rectifier nonlinearity.
    layers.append(Conv((8, 8, 32), strides=4, init=init_norm, activation=Rectlin()))
    # The second hidden layer convolves 64 filters of 4x4 with stride 2, again followed by a rectifier nonlinearity.
    layers.append(Conv((4, 4, 64), strides=2, init=init_norm, activation=Rectlin()))
    # This is followed by a third convolutional layer that convolves 64 filters of 3x3 with stride 1 followed by a rectifier.
    layers.append(Conv((3, 3, 64), strides=1, init=init_norm, activation=Rectlin()))
    # The final hidden layer is fully-connected and consists of 512 rectifier units.
    layers.append(Affine(nout=512, init=init_norm, activation=Rectlin()))
    # The output layer is a fully-connected linear layer with a single output for each valid action.
    layers.append(Affine(nout = num_actions, init = init_norm))
    return layers

  def setTensor(self, states):
    # change order of axes to match what Neon expects
    states = np.transpose(states, axes = (1, 2, 3, 0))
    # copy() shouldn't be necessary here, but Neon doesn't work otherwise
    self.tensor.set(states.copy())
    # normalize network input between 0 and 1
    self.be.divide(self.tensor, 255, self.tensor)

  def train(self, minibatch, epoch):
    # expand components of minibatch
    prestates, actions, rewards, poststates, terminals = minibatch
    assert len(prestates.shape) == 4
    assert len(poststates.shape) == 4
    assert len(actions.shape) == 1
    assert len(rewards.shape) == 1
    assert len(terminals.shape) == 1
    assert prestates.shape == poststates.shape
    assert prestates.shape[0] == actions.shape[0] == rewards.shape[0] == poststates.shape[0] == terminals.shape[0]

    if self.target_steps and self.train_iterations % self.target_steps == 0:
      # HACK: push something through network, so that weights exist
      self.model.fprop(self.tensor)
      # HACK: serialize network to disk and read it back to clone
      filename = os.path.join(self.save_weights_path, "target_network.pkl")
      save_obj(self.model.serialize(keep_states = False), filename)
      self.target_model.load_weights(filename)

    # feed-forward pass for poststates to get Q-values
    self.setTensor(poststates)
    postq = self.target_model.fprop(self.tensor, inference = True)
    assert postq.shape == (self.num_actions, self.batch_size)

    # calculate max Q-value for each poststate
    maxpostq = self.be.max(postq, axis=0).asnumpyarray()
    assert maxpostq.shape == (1, self.batch_size)

    # feed-forward pass for prestates
    self.setTensor(prestates)
    preq = self.model.fprop(self.tensor, inference = False)
    assert preq.shape == (self.num_actions, self.batch_size)

#.........这里部分代码省略.........

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
# parse the command line arguments
parser = NeonArgparser(__doc__)
parser.add_argument("hdf5")
parser.add_argument("model_pkl")
args = parser.parse_args()

model = Model(args.model_pkl)

h5s = [h5py.File(args.hdf5)]
num_moves = sum(h['X'].shape[0] for h in h5s)
print("Found {} HDF5 files with {} moves".format(len(h5s), num_moves))
inputs = HDF5Iterator([h['X'] for h in h5s],
                      [h['y'] for h in h5s],
                      ndata=(1024 * 1024))

out_predict = h5s[0].require_dataset("predictions", (num_moves, 362), dtype=np.float32)
out_score = h5s[0].require_dataset("scores", (num_moves,), dtype=np.float32)
out_max = h5s[0].require_dataset("best", (num_moves,), dtype=np.float32)


model.initialize(inputs)
for indata, actual, sl in inputs.predict():
    prediction = model.fprop(indata, inference=False).get().T
    actual = actual.astype(int)
    actual_idx = actual[:, 0] * 19 + actual[:, 1]
    actual_idx[actual_idx < 0] = 361
    out_predict[sl, :] = prediction
    out_score[sl] = prediction[range(prediction.shape[0]), actual_idx]
    out_max[sl] = prediction.max(axis=1)
    print (sl)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
def test_model_serialize(backend_default, data):
    dataset = MNIST(path=data)
    (X_train, y_train), (X_test, y_test), nclass = dataset.load_data()
    train_set = ArrayIterator(
        [X_train, X_train], y_train, nclass=nclass, lshape=(1, 28, 28))

    init_norm = Gaussian(loc=0.0, scale=0.01)

    # initialize model
    path1 = Sequential([Conv((5, 5, 16), init=init_norm, bias=Constant(0), activation=Rectlin()),
                        Pooling(2),
                        Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin())])
    path2 = Sequential([Affine(nout=100, init=init_norm, bias=Constant(0), activation=Rectlin()),
                        Dropout(keep=0.5),
                        Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin())])
    layers = [MergeMultistream(layers=[path1, path2], merge="stack"),
              Affine(nout=20, init=init_norm, batch_norm=True, activation=Rectlin()),
              Affine(nout=10, init=init_norm, activation=Logistic(shortcut=True))]

    tmp_save = 'test_model_serialize_tmp_save.pickle'
    mlp = Model(layers=layers)
    mlp.optimizer = GradientDescentMomentum(learning_rate=0.1, momentum_coef=0.9)
    mlp.cost = GeneralizedCost(costfunc=CrossEntropyBinary())
    mlp.initialize(train_set, cost=mlp.cost)
    n_test = 3
    num_epochs = 3
    # Train model for num_epochs and n_test batches
    for epoch in range(num_epochs):
        for i, (x, t) in enumerate(train_set):
            x = mlp.fprop(x)
            delta = mlp.cost.get_errors(x, t)
            mlp.bprop(delta)
            mlp.optimizer.optimize(mlp.layers_to_optimize, epoch=epoch)
            if i > n_test:
                break

    # Get expected outputs of n_test batches and states of all layers
    outputs_exp = []
    pdicts_exp = [l.get_params_serialize() for l in mlp.layers_to_optimize]
    for i, (x, t) in enumerate(train_set):
        outputs_exp.append(mlp.fprop(x, inference=True))
        if i > n_test:
            break

    # Serialize model
    mlp.save_params(tmp_save, keep_states=True)

    # Load model
    mlp = Model(tmp_save)

    mlp.initialize(train_set)
    outputs = []
    pdicts = [l.get_params_serialize() for l in mlp.layers_to_optimize]
    for i, (x, t) in enumerate(train_set):
        outputs.append(mlp.fprop(x, inference=True))
        if i > n_test:
            break

    # Check outputs, states, and params are the same
    for output, output_exp in zip(outputs, outputs_exp):
        assert allclose_with_out(output.get(), output_exp.get())

    for pd, pd_exp in zip(pdicts, pdicts_exp):
        for s, s_e in zip(pd['states'], pd_exp['states']):
            if isinstance(s, list):  # this is the batch norm case
                for _s, _s_e in zip(s, s_e):
                    assert allclose_with_out(_s, _s_e)
            else:
                assert allclose_with_out(s, s_e)
        for p, p_e in zip(pd['params'], pd_exp['params']):
            assert type(p) == type(p_e)
            if isinstance(p, list):  # this is the batch norm case
                for _p, _p_e in zip(p, p_e):
                    assert allclose_with_out(_p, _p_e)
            elif isinstance(p, np.ndarray):
                assert allclose_with_out(p, p_e)
            else:
                assert p == p_e

    os.remove(tmp_save)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
model_desc = ModelDescription(load_obj(args.save_model_file))
for layer in segnet_model.layers_to_optimize:
    name = layer.name
    trained_layer = model_desc.getlayer(name)
    layer.load_weights(trained_layer)

fig = plt.figure()
if args.display:
    plt.ion()

im1 = None
im2 = None

cnt = 1
for x, t in test_set:
    z = segnet_model.fprop(x).get()

    z = np.argmax(z.reshape((c, h, w)), axis=0)
    t = np.argmax(t.get().reshape((c, h, w)), axis=0)

    # calculate the misclass rate
    acc = (np.where(z == t)[0].size / float(z.size))*100.0

    plt.subplot(2,1,1);
    if im1 is None:
        im1 = plt.imshow(t);plt.title('Truth')
    else:
        im1.set_data(t)

    plt.subplot(2,1,2);
    if im2 is None:

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]

#.........这里部分代码省略.........

    def _prepare_network_input(self, states):
        """ Transforms and normalizes the states from one minibatch.

        Args:
            states (): a set of states with the size of minibatch
        """
        _logger.debug("Normalizing and transforming input")
        # change order of axes to match what Neon expects
        states = np.transpose(states, axes = (1, 2, 3, 0))
        # copy() shouldn't be necessary here, but Neon doesn't work otherwise
        self.input.set(states.copy())
        # normalize network input between 0 and 1
        self.be.divide(self.input, self.grayscales, self.input)

    def train(self, minibatch, epoch):
        """ Prepare, perform and document a complete train step for one minibatch.

        Args:
            minibatch (numpy.ndarray): Mini-batch of states, shape=(batch_size,sequence_length,frame_width,frame_height)
            epoch (int): Current train epoch
        """
        _logger.debug("Complete trainig step for one minibatch")
        prestates, actions, rewards, poststates, terminals = minibatch
        assert len(prestates.shape) == 4
        assert len(poststates.shape) == 4
        assert len(actions.shape) == 1
        assert len(rewards.shape) == 1
        assert len(terminals.shape) == 1
        assert prestates.shape == poststates.shape
        assert prestates.shape[0] == actions.shape[0] == rewards.shape[0] == poststates.shape[0] == terminals.shape[0]
        # feed-forward pass for poststates to get Q-values
        self._prepare_network_input(poststates)
        postq = self.target_model.fprop(self.input, inference = True)
        assert postq.shape == (self.output_shape, self.batch_size)
        # calculate max Q-value for each poststate
        maxpostq = self.be.max(postq, axis=0).asnumpyarray()
        assert maxpostq.shape == (1, self.batch_size)
        # average maxpostq for stats
        maxpostq_avg = maxpostq.mean()
        # feed-forward pass for prestates
        self._prepare_network_input(prestates)
        preq = self.model.fprop(self.input, inference = False)
        assert preq.shape == (self.output_shape, self.batch_size)
        # make copy of prestate Q-values as targets
        targets = preq.asnumpyarray()
        # clip rewards between -1 and 1
        rewards = np.clip(rewards, self.min_reward, self.max_reward)
        # update Q-value targets for each state only at actions taken
        for i, action in enumerate(actions):
            if terminals[i]:
                targets[action, i] = float(rewards[i])
            else:
                targets[action, i] = float(rewards[i]) + self.discount_rate * maxpostq[0,i]
        # copy targets to GPU memory
        self.targets.set(targets)
        # calculate errors
        errors = self.cost_func.get_errors(preq, self.targets)
        assert errors.shape == (self.output_shape, self.batch_size)
        # average error where there is a error (should be 1 in every row)
        #TODO: errors_avg = np.sum(errors)/np.size(errors[errors>0.])
        # clip errors
        if self.clip_error:
            self.be.clip(errors, -self.clip_error, self.clip_error, out = errors)
        # calculate cost, just in case
        cost = self.cost_func.get_cost(preq, self.targets)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
num_predict = 1000

layers = [
    LSTM(hidden_size, init, Logistic(), Tanh()),
    Affine(len(train_set.vocab), init, bias=init, activation=Softmax())
]
model = Model(layers=layers)
model.load_weights(args.save_path)

# Generate text
text = []
seed_tokens = list('ROMEO:')

x = be.zeros((len(train_set.vocab), time_steps))

for s in seed_tokens:
    x.fill(0)
    x[train_set.token_to_index[s], 0] = 1
    y = model.fprop(x)

for i in range(num_predict):
    # Take last prediction and feed into next fprop
    pred = sample(y.get()[:, -1])
    text.append(train_set.index_to_token[int(pred)])

    x.fill(0)
    x[pred, 0] = 1
    y = model.fprop(x)

print ''.join(seed_tokens + text)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
def test_model_serialize(backend):
    (X_train, y_train), (X_test, y_test), nclass = load_mnist()
    train_set = DataIterator([X_train, X_train], y_train, nclass=nclass, lshape=(1, 28, 28))

    init_norm = Gaussian(loc=0.0, scale=0.01)

    # initialize model
    path1 = [Conv((5, 5, 16), init=init_norm, bias=Constant(0), activation=Rectlin()),
             Pooling(2),
             Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin())]
    path2 = [Dropout(keep=0.5),
             Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin())]
    layers = [MergeConcat([path1, path2]),
              Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin()),
              BatchNorm(),
              Affine(nout=10, init=init_norm, activation=Logistic(shortcut=True))]

    tmp_save = 'test_model_serialize_tmp_save.pickle'
    mlp = Model(layers=layers)
    mlp.optimizer = GradientDescentMomentum(learning_rate=0.1, momentum_coef=0.9)
    mlp.cost = GeneralizedCost(costfunc=CrossEntropyBinary())

    n_test = 3
    num_epochs = 3
    # Train model for num_epochs and n_test batches
    for epoch in range(num_epochs):
        for i, (x, t) in enumerate(train_set):
            x = mlp.fprop(x)
            delta = mlp.cost.get_errors(x, t)
            mlp.bprop(delta)
            mlp.optimizer.optimize(mlp.layers_to_optimize, epoch=epoch)
            if i > n_test:
                break

    # Get expected outputs of n_test batches and states of all layers
    outputs_exp = []
    pdicts_exp = [l.get_params_serialize() for l in mlp.layers_to_optimize]
    for i, (x, t) in enumerate(train_set):
        outputs_exp.append(mlp.fprop(x, inference=True))
        if i > n_test:
            break

    # Serialize model
    save_obj(mlp.serialize(keep_states=True), tmp_save)

    # Load model
    mlp = Model(layers=layers)
    mlp.load_weights(tmp_save)

    outputs = []
    pdicts = [l.get_params_serialize() for l in mlp.layers_to_optimize]
    for i, (x, t) in enumerate(train_set):
        outputs.append(mlp.fprop(x, inference=True))
        if i > n_test:
            break

    # Check outputs, states, and params are the same
    for output, output_exp in zip(outputs, outputs_exp):
        assert np.allclose(output.get(), output_exp.get())

    for pd, pd_exp in zip(pdicts, pdicts_exp):
        for s, s_e in zip(pd['states'], pd_exp['states']):
            if isinstance(s, list):  # this is the batch norm case
                for _s, _s_e in zip(s, s_e):
                    assert np.allclose(_s, _s_e)
            else:
                assert np.allclose(s, s_e)
        for p, p_e in zip(pd['params'], pd_exp['params']):
            if isinstance(p, list):  # this is the batch norm case
                for _p, _p_e in zip(p, p_e):
                    assert np.allclose(_p, _p_e)
            else:
                assert np.allclose(p, p_e)

    os.remove(tmp_save)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import fprop [as 别名]
metric = Accuracy()

##########################################################################

model = Model(layers=layers)
optimizer = Adagrad(learning_rate=0.01, clip_gradients=clip_gradients)
callbacks = Callbacks(model, train_set, args, eval_set=valid_set)

model.load_weights(os.path.join(args.data_dir, '128128_49_model_e2.pkl'))

print "Test  Accuracy - ", 100 * model.eval(valid_set, metric=metric)
print "Train Accuracy - ", 100 * model.eval(train_set, metric=metric)

# output result directly
for x, y in valid_set:
    x = model.fprop(x, inference=True)
    print(x.get())
    print(y.get())
    break

#########################################################################
# continue training
# optimizer = Adagrad(learning_rate=0.01, clip_gradients=clip_gradients)
# callbacks = Callbacks(model, train_set, args, eval_set=valid_set)

# import ipdb; ipdb.set_trace()

# re-allocate output memories for each layer
# model.initialized = False
# model.initialize(train_set, cost=cost)