Python Model.load_params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 load_params [as 别名]
def create_model(model_type, model_tree, freeze, dataset_dir, model_file, img_loader):
    cost = GeneralizedCost(costfunc=CrossEntropyMulti())

    if model_type == "alexnet":
        opt = create_alexnet_opt()
        layer_func = create_alexnet_layers
    elif model_type == "vgg":
        opt = create_vgg_opt()
        layer_func = create_vgg_layers
    else:
        raise NotImplementedError(model_type + " has not been implemented")

    if model_tree:
        ctree = ClassTaxonomy("Aves", "taxonomy_dict.p", dataset_dir)
        layers = created_branched(layer_func, ctree, img_loader)
        model = TaxonomicBranchModel(layers=layers)
    else:
        layers = layer_func(img_loader.nclass)
        model = Model(layers=layers)

    if freeze > 0:
        saved_model = Model(layers=layer_func(1000))
        saved_model.load_params(model_file)
        model.initialize(img_loader)
        model.initialized = False
        saved_lto = saved_model.layers.layers_to_optimize
        model_lto = model.layers.layers_to_optimize
        keep_length = len(saved_lto) - freeze * 2

        for i in range(len(saved_lto))[:keep_length]:
            model_lto[i].W[:] = saved_lto[i].W
            model_lto[i].optimize = False
        for i in range(len(model_lto))[keep_length:]:
            model_lto[i].optimize = True

        model.layers = FreezeSequential(layers)
        model.layers_to_optimize = model.layers.layers_to_optimize

    return model, cost, opt

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import load_params [as 别名]
        hidden_size,
        init_glorot,
        activation=Tanh(),
        gate_activation=Logistic(),
        reset_cells=True),
    RecurrentSum(),
    Dropout(keep=0.5),
    Affine(nclass, init_glorot, bias=init_glorot, activation=Softmax())
]

# load the weights
print("Initialized the models - ")
model_new = Model(layers=layers)
print("Loading the weights from {0}".format(args.model_weights))

model_new.load_params(args.model_weights)
model_new.initialize(dataset=(sentence_length, batch_size))

# setup buffers before accepting reviews
xdev = be.zeros((sentence_length, 1), dtype=np.int32)  # bsz is 1, feature size
xbuf = np.zeros((1, sentence_length), dtype=np.int32)
oov = 2
start = 1
index_from = 3
pad_char = 0
vocab, rev_vocab = pickle.load(open(args.vocab_file, 'rb'))

while True:
    line = input('Enter a Review from testData.tsv file \n')

    # clean the input

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import load_params [as 别名]
valmetric = TopKMisclassification(k=5)

# dummy optimizer for benchmarking
# training implementation coming soon
opt_gdm = GradientDescentMomentum(0.0, 0.0)
opt_biases = GradientDescentMomentum(0.0, 0.0)
opt = MultiOptimizer({'default': opt_gdm, 'Bias': opt_biases})

# setup cost function as CrossEntropy
cost = Multicost(costs=[GeneralizedCost(costfunc=CrossEntropyMulti()),
                        GeneralizedCost(costfunc=CrossEntropyMulti()),
                        GeneralizedCost(costfunc=CrossEntropyMulti())],
                 weights=[1, 0., 0.])  # We only want to consider the CE of the main path

assert os.path.exists(args.model_file), 'script requires the trained weights file'
model.load_params(args.model_file)
model.initialize(test, cost)


print 'running speed benchmark...'
model.benchmark(test, cost, opt)

print '\nCalculating performance on validation set...'
test.reset()
mets = model.eval(test, metric=valmetric)
print 'Validation set metrics:'
print 'LogLoss: %.2f, Accuracy: %.1f %% (Top-1), %.1f %% (Top-5)' % (mets[0],
                                                                     (1.0-mets[1])*100,
                                                                     (1.0-mets[2])*100)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import load_params [as 别名]
    MergeMultistream(layers=[image_path, sent_path], merge="recurrent"),
    Dropout(keep=0.5),
    LSTM(hidden_size, init, activation=Logistic(), gate_activation=Tanh(), reset_cells=True),
    Affine(train_set.vocab_size, init, bias=init2, activation=Softmax())
]

cost = GeneralizedCostMask(costfunc=CrossEntropyMulti(usebits=True))

# configure callbacks
checkpoint_model_path = "~/image_caption2.pickle"
if args.callback_args['save_path'] is None:
    args.callback_args['save_path'] = checkpoint_model_path

if args.callback_args['serialize'] is None:
    args.callback_args['serialize'] = 1

model = Model(layers=layers)

callbacks = Callbacks(model, train_set, **args.callback_args)

opt = RMSProp(decay_rate=0.997, learning_rate=0.0005, epsilon=1e-8, gradient_clip_value=1)

# train model
model.fit(train_set, optimizer=opt, num_epochs=num_epochs, cost=cost, callbacks=callbacks)

# load model (if exited) and evaluate bleu score on test set
model.load_params(checkpoint_model_path)
test_set = ImageCaptionTest(path=data_path)
sents, targets = test_set.predict(model)
test_set.bleu_score(sents, targets)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import load_params [as 别名]
    seq_len = 1

    if return_sequences is True:
        layers = [
            LSTM(hidden, init, activation=Logistic(), gate_activation=Tanh(), reset_cells=False),
            Affine(train_set.nfeatures, init, bias=init, activation=Identity())
        ]
    else:
        layers = [
            LSTM(hidden, init, activation=Logistic(), gate_activation=Tanh(), reset_cells=False),
            RecurrentLast(),
            Affine(train_set.nfeatures, init, bias=init, activation=Identity())
        ]

    model_new = Model(layers=layers)
    model_new.load_params(args.save_path)
    model_new.initialize(dataset=(train_set.nfeatures, seq_len))

    output = np.zeros((train_set.nfeatures, num_predict))
    seed = time_series.train[:seed_seq_len]

    x = model_new.be.empty((train_set.nfeatures, seq_len))
    for s_in in seed:
        x.set(s_in.reshape(train_set.nfeatures, seq_len))
        y = model_new.fprop(x, inference=False)

    for i in range(num_predict):
        # Take last prediction and feed into next fprop
        pred = y.get()[:, -1]
        output[:, i] = pred
        x[:] = pred.reshape(train_set.nfeatures, seq_len)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import load_params [as 别名]
    MergeMultistream(layers=[image_path, sent_path], merge="recurrent"),
    Dropout(keep=0.5),
    LSTM(hidden_size, init, activation=Logistic(), gate_activation=Tanh(), reset_cells=True),
    Affine(train_set.vocab_size, init, bias=init2, activation=Softmax())
]

cost = GeneralizedCostMask(costfunc=CrossEntropyMulti(usebits=True))

# configure callbacks
checkpoint_model_path = "~/image_caption2.pkl"
if args.callback_args['save_path'] is None:
    args.callback_args['save_path'] = checkpoint_model_path

if args.callback_args['serialize'] is None:
    args.callback_args['serialize'] = 1

model = Model(layers=layers)

callbacks = Callbacks(model, **args.callback_args)

opt = RMSProp(decay_rate=0.997, learning_rate=0.0005, epsilon=1e-8, gradient_clip_value=1)

# train model
model.fit(train_set, optimizer=opt, num_epochs=num_epochs, cost=cost, callbacks=callbacks)

# load model (if exited) and evaluate bleu score on test set
if os.path.exists(args.callback_args['save_path']):
    model.load_params(args.callback_args['save_path'])
sents, targets = test_set.predict(model)
test_set.bleu_score(sents, targets)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import load_params [as 别名]
param_file_name = home_dir + "/ubuntu/model/trained_bot_model_32x32.prm"
class_names = ["forward", "left", "right", "backward"]    # from ROBOT-C bot.c
nclasses = len(class_names)
size = H, W

be = gen_backend(backend='cpu', batch_size=1)    # NN backend
init_uni = Uniform(low=-0.1, high=0.1)           # Unnecessary NN weight initialization
bn = True                                        # enable NN batch normalization
layers = [Conv((5, 5, 16), init=init_uni, activation=Rectlin(), batch_norm=bn),
          Pooling((2, 2)),
          Conv((3, 3, 32), init=init_uni, activation=Rectlin(), batch_norm=bn),
          Pooling((2, 2)),
          Affine(nout=50, init=init_uni, activation=Rectlin(), batch_norm=bn),
          Affine(nout=nclasses, init=init_uni, activation=Softmax())]
model = Model(layers=layers)
model.load_params(param_file_name, load_states=False)

def usage():
    print "python connect_to_vex_cortex.py"
    print "  Raspberry Pi records video, commands from VEX Cortex 2.0"
    print "  -p " + file_name_prefix + ": file name prefix"
    print "  -d: display received commands for debug"
    print "  -w " + str(w) + ": video width"
    print "  -h " + str(h) + ": video height"
    print "  -f " + str(fps) + ": video FPS, 0 for camera default"
    print "  -q " + str(quality) + ": quality to record video, 1..40"
    print "  -b " + str(bitrate) + ": bitrate e.g. 15000000, 0 for unlimited"
    print "  -i " + str(iso) + ": ISO 0 | 100 ... 800, see picamera doc, 0 for camera default"
    print "  -m: horizontal mirror"
    print "  -v: vertical mirror"
    print "  -s: shut down system on exit (must run as super user)"

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

#.........这里部分代码省略.........
  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

    # calculate cost, just in case
    cost = self.cost.get_cost(preq, self.targets)
    assert cost.shape == (1,1)

    # clip errors
    if self.clip_error:
      self.be.clip(deltas, -self.clip_error, self.clip_error, out = deltas)

    # perform back-propagation of gradients
    self.model.bprop(deltas)

    # perform optimization
    self.optimizer.optimize(self.model.layers_to_optimize, epoch)

    # increase number of weight updates (needed for target clone interval)
    self.train_iterations += 1

    # calculate statistics
    if self.callback:
      self.callback.on_train(cost[0,0])

  def predict(self, states):
    # minibatch is full size, because Neon doesn't let change the minibatch size
    assert states.shape == ((self.batch_size, self.history_length,) + self.screen_dim)

    # calculate Q-values for the states
    self._setInput(states)
    qvalues = self.model.fprop(self.input, inference = True)
    assert qvalues.shape == (self.num_actions, self.batch_size)
    if logger.isEnabledFor(logging.DEBUG):
      logger.debug("Q-values: " + str(qvalues.asnumpyarray()[:,0]))

    # transpose the result, so that batch size is first dimension
    return qvalues.T.asnumpyarray()

  def load_weights(self, load_path):
    self.model.load_params(load_path)

  def save_weights(self, save_path):
    self.model.save_params(save_path)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import load_params [as 别名]
def test_model_serialize(backend_default, data):
    (X_train, y_train), (X_test, y_test), nclass = load_mnist(path=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(layers=layers)
    mlp.load_params(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']):
            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 np.allclose(_p, _p_e)
            elif isinstance(p, np.ndarray):
                assert np.allclose(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 load_params [as 别名]

#.........这里部分代码省略.........
        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)
        assert cost.shape == (1,1)
        # perform back-propagation of gradients
        self.model.bprop(errors)
        # perform optimization
        self.optimizer.optimize(self.model.layers_to_optimize, epoch)
        # increase number of weight updates (needed for target clone interval)
        self.update_iterations += 1
        if self.target_update_frequency and self.update_iterations % self.target_update_frequency == 0:
            self._copy_theta()
            _logger.info("Network update #%d: Cost = %s, Avg Max Q-value = %s" % (self.update_iterations, str(cost.asnumpyarray()[0][0]), str(maxpostq_avg)))
        # update statistics
        if self.callback:
            self.callback.from_learner(cost.asnumpyarray()[0,0], maxpostq_avg)

    def get_Q(self, state):
        """ Calculates the Q-values for one mini-batch.

        Args:
            state(numpy.ndarray): Single state, shape=(sequence_length,frame_width,frame_height).

        Returns:
            q_values (numpy.ndarray): Results for first element of mini-batch from one forward pass through the network, shape=(self.output_shape,)
        """
        _logger.debug("State shape = %s" % str(state.shape))
        # minibatch is full size, because Neon doesn't let change the minibatch size
        # so we need to run 32 forward steps to get the one we actually want
        self.dummy_batch[0] = state
        states = self.dummy_batch
        assert states.shape == ((self.batch_size, self.sequence_length,) + self.frame_dims)
        # calculate Q-values for the states
        self._prepare_network_input(states)
        qvalues = self.model.fprop(self.input, inference = True)
        assert qvalues.shape == (self.output_shape, self.batch_size)
        _logger.debug("Qvalues: %s" % (str(qvalues.asnumpyarray()[:,0])))
        return qvalues.asnumpyarray()[:,0]

    def _copy_theta(self):
        """ Copies the weights of the current network to the target network. """
        _logger.debug("Copying weights")
        pdict = self.model.get_description(get_weights=True, keep_states=True)
        self.target_model.deserialize(pdict, load_states=True)

    def save_weights(self, target_dir, epoch):
        """ Saves the current network parameters to disk.

        Args:
            target_dir (str): Directory where the network parameters are stored for each episode.
            epoch (int): Current epoch.
        """
        filename = "%s_%s_%s_%d.prm" % (str(self.args.game.lower()), str(self.args.net_type.lower()), str(self.args.optimizer.lower()), (epoch + 1))
        self.model.save_params(os.path.join(target_dir, filename))

    def load_weights(self, source_file):
        """ Loads the network parameters from a given file.

        Args:
            source_file (str): Complete path to a file with network parameters.
        """
        self.model.load_params(source_file)

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import load_params [as 别名]
from neon.layers import Conv, Affine, Pooling
from neon.initializers import Uniform
from neon.transforms.activation import Rectlin, Softmax
init_uni = Uniform(low=-0.1, high=0.1)
layers = [Conv(fshape=(5,5,16), init=init_uni, activation=Rectlin()),
          Pooling(fshape=2, strides=2),
          Conv(fshape=(5,5,32), init=init_uni, activation=Rectlin()),
          Pooling(fshape=2, strides=2),
          Affine(nout=500, init=init_uni, activation=Rectlin()),
          Affine(nout=10, init=init_uni, activation=Softmax())]

print("Before running this script, run my_cifar_train.py to train a CIFAR10 model")
print("Loading pre-trained CIFAR10 model")
from neon.models import Model
model = Model(layers)
model.load_params("cifar10_model.prm", load_states=False)

classes =["airplane", "automobile", "bird", "cat", "deer",
          "dog", "frog", "horse", "ship", "truck"]
nclass = len(classes)


# Sanity check 1
# an image of a frog from wikipedia
# image_source = "https://upload.wikimedia.org/wikipedia/commons/thumb/5/55/Atelopus_zeteki1.jpg/440px-Atelopus_zeteki1.jpg"
# import urllib
# urllib.urlretrieve(image_source, filename="image.jpg")

# crop and resize to 32x32
from PIL import Image
import numpy as np

# 需要导入模块: from neon.models import Model [as 别名]
# 或者: from neon.models.Model import load_params [as 别名]
# Now construct the network
from neon.layers import ColorNoise
#layers = [ColorNoise()]

# layers.append(Affine(nout=100, init=Kaiming(local=False), batch_norm=True, activation=Softmax()))

scales = [112, 128, 160, 240]

for scale in scales:
    print scale

    layers = []
    layers += [Conv(**conv_params(7, 32, 2))]
    for nfm, stride in zip(nfms, strides):
        layers.append(module_factory(nfm, stride))
    layers.append(Pooling(7, op='avg'))

    layers.append(Conv(fshape=(1,1,100), init=Kaiming(local=True), batch_norm=True))
    layers.append(Pooling(fshape='all', op='avg'))
    layers.append(Activation(Softmax()))

    model = Model(layers=layers)
    test = ImageLoader(set_name='validation', shuffle=False, do_transforms=False, inner_size=scale,
                       scale_range=scale, repo_dir=args.data_dir)

    model.load_params("/home/users/hunter/bigfeat_dropout.pkl")

    softmaxes = model.get_outputs(test)
    from neon.util.persist import save_obj
    save_obj(softmaxes, "bigfeat_dropout_SM_{}.pkl".format(scale))

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

#.........这里部分代码省略.........
        initializer = self.get_initializer(input_size = 7 * 7 * 64)
        layers.append(Affine(nout=512, init=initializer, bias=initializer, activation=Rectlin()))
        
        initializer = self.get_initializer(input_size = 512)
        layers.append(Affine(nout=max_action_no, init=initializer, bias=initializer))
        
        return layers        
        
    def clip_reward(self, reward):
        if reward > self.args.clip_reward_high:
            return self.args.clip_reward_high
        elif reward < self.args.clip_reward_low:
            return self.args.clip_reward_low
        else:
            return reward

    def set_input(self, data):
        if self.use_gpu_replay_mem:
            self.be.copy_transpose(data, self.input_uint8, axes=(1, 2, 3, 0))
            self.input[:] = self.input_uint8 / 255
        else:
            self.input.set(data.transpose(1, 2, 3, 0).copy())
            self.be.divide(self.input, 255, self.input)

    def predict(self, history_buffer):
        self.set_input(history_buffer)
        output  = self.train_net.fprop(self.input, inference=True)
        return output.T.asnumpyarray()[0]            

    def print_weights(self):
        pass

    def train(self, minibatch, replay_memory, learning_rate, debug):
        if self.args.prioritized_replay == True:
            prestates, actions, rewards, poststates, terminals, replay_indexes, heap_indexes, weights = minibatch
        else:
            prestates, actions, rewards, poststates, terminals = minibatch
        
        # Get Q*(s, a) with targetNet
        self.set_input(poststates)
        post_qvalue = self.target_net.fprop(self.input, inference=True).T.asnumpyarray()
        
        if self.args.double_dqn == True:
            # Get Q*(s, a) with trainNet
            post_qvalue2 = self.train_net.fprop(self.input, inference=True).T.asnumpyarray()
        
        # Get Q(s, a) with trainNet
        self.set_input(prestates)
        pre_qvalue = self.train_net.fprop(self.input, inference=False)
        
        label = pre_qvalue.asnumpyarray().copy()
        for i in range(0, self.train_batch_size):
            if self.args.clip_reward:
                reward = self.clip_reward(rewards[i])
            else:
                reward = rewards[i]
            if terminals[i]:
                label[actions[i], i] = reward
            else:
                if self.args.double_dqn == True:
                    max_index = np.argmax(post_qvalue2[i])
                    label[actions[i], i] = reward + self.discount_factor* post_qvalue[i][max_index]
                else:
                    label[actions[i], i] = reward + self.discount_factor* np.max(post_qvalue[i])

        # copy targets to GPU memory
        self.targets.set(label)
    
        delta = self.cost.get_errors(pre_qvalue, self.targets)
        
        if self.args.prioritized_replay == True:
            delta_value = delta.asnumpyarray()
            for i in range(self.train_batch_size):
                if debug:
                    print 'weight[%s]: %.5f, delta: %.5f, newDelta: %.5f' % (i, weights[i], delta_value[actions[i], i], weights[i] * delta_value[actions[i], i]) 
                replay_memory.update_td(heap_indexes[i], abs(delta_value[actions[i], i]))
                delta_value[actions[i], i] = weights[i] * delta_value[actions[i], i]
            delta.set(delta_value.copy())
          
        if self.args.clip_loss:
            self.be.clip(delta, -1.0, 1.0, out = delta)
                
        self.train_net.bprop(delta)
        self.optimizer.optimize(self.train_net.layers_to_optimize, epoch=0)

    def update_model(self):
        # have to serialize also states for batch normalization to work
        pdict = self.train_net.get_description(get_weights=True, keep_states=True)
        self.target_net.deserialize(pdict, load_states=True)
        #print ('Updated target model')

    def finish_train(self):
        self.running = False
    
    def load(self, file_name):
        self.train_net.load_params(file_name)
        self.update_model()
        
    def save(self, file_name):
        self.train_net.save_params(file_name)