def __init__(self, vocabularySize, config_param):
    self.vocabularySize = vocabularySize
    self.config = config_param

    self._inputX = tf.placeholder(tf.int32, [self.config.batch_size, self.config.sequence_size], "InputsX")
    self._inputTargetsY = tf.placeholder(tf.int32, [self.config.batch_size, self.config.sequence_size], "InputTargetsY")


    #Converting Input in an Embedded form
    with tf.device("/cpu:0"): #Tells Tensorflow what GPU to use specifically
      embedding = tf.get_variable("embedding", [self.vocabularySize, self.config.embeddingSize])
      embeddingLookedUp = tf.nn.embedding_lookup(embedding, self._inputX)
      inputs = tf.split(1, self.config.sequence_size, embeddingLookedUp)
      inputTensorsAsList = [tf.squeeze(input_, [1]) for input_ in inputs]


    #Define Tensor RNN
    singleRNNCell = rnn_cell.BasicRNNCell(self.config.hidden_size)
    self.multilayerRNN =  rnn_cell.MultiRNNCell([singleRNNCell] * self.config.num_layers)
    self._initial_state = self.multilayerRNN.zero_state(self.config.batch_size, tf.float32)

    #Defining Logits
    hidden_layer_output, states = rnn.rnn(self.multilayerRNN, inputTensorsAsList, initial_state=self._initial_state)
    hidden_layer_output = tf.reshape(tf.concat(1, hidden_layer_output), [-1, self.config.hidden_size])
    self._logits = tf.nn.xw_plus_b(hidden_layer_output, tf.get_variable("softmax_w", [self.config.hidden_size, self.vocabularySize]), tf.get_variable("softmax_b", [self.vocabularySize]))
    self._predictionSoftmax = tf.nn.softmax(self._logits)

    #Define the loss
    loss = seq2seq.sequence_loss_by_example([self._logits], [tf.reshape(self._inputTargetsY, [-1])], [tf.ones([self.config.batch_size * self.config.sequence_size])], self.vocabularySize)
    self._cost = tf.div(tf.reduce_sum(loss), self.config.batch_size)

    self._final_state = states[-1]

    def __init__(self, args, infer=False):
        self.args = args
        if infer:
            args.batch_size = 1
            args.seq_length = 1

        additional_cell_args = {}
        if args.model == 'rnn':
            cell_fn = rnn_cell.BasicRNNCell
        elif args.model == 'gru':
            cell_fn = rnn_cell.GRUCell
        elif args.model == 'lstm':
            cell_fn = rnn_cell.BasicLSTMCell
        elif args.model == 'gridlstm':
            cell_fn = grid_rnn.Grid2LSTMCell
            additional_cell_args.update({'use_peepholes': True, 'forget_bias': 1.0})
        elif args.model == 'gridgru':
            cell_fn = grid_rnn.Grid2GRUCell
        else:
            raise Exception("model type not supported: {}".format(args.model))

        cell = cell_fn(args.rnn_size, **additional_cell_args)

        self.cell = cell = rnn_cell.MultiRNNCell([cell] * args.num_layers)

        self.input_data = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
        self.targets = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
        self.initial_state = cell.zero_state(args.batch_size, tf.float32)

        with tf.variable_scope('rnnlm'):
            softmax_w = tf.get_variable("softmax_w", [args.rnn_size, args.vocab_size])
            softmax_b = tf.get_variable("softmax_b", [args.vocab_size])
            with tf.device("/cpu:0"):
                embedding = tf.get_variable("embedding", [args.vocab_size, args.rnn_size])
                inputs = tf.split(1, args.seq_length, tf.nn.embedding_lookup(embedding, self.input_data))
                inputs = [tf.squeeze(input_, [1]) for input_ in inputs]

        def loop(prev, _):
            prev = tf.nn.xw_plus_b(prev, softmax_w, softmax_b)
            prev_symbol = tf.stop_gradient(tf.argmax(prev, 1))
            return tf.nn.embedding_lookup(embedding, prev_symbol)

        outputs, last_state = seq2seq.rnn_decoder(inputs, self.initial_state, cell,
                                                  loop_function=loop if infer else None, scope='rnnlm')
        output = tf.reshape(tf.concat(1, outputs), [-1, args.rnn_size])
        self.logits = tf.nn.xw_plus_b(output, softmax_w, softmax_b)
        self.probs = tf.nn.softmax(self.logits)
        loss = seq2seq.sequence_loss_by_example([self.logits],
                                                [tf.reshape(self.targets, [-1])],
                                                [tf.ones([args.batch_size * args.seq_length])],
                                                args.vocab_size)
        self.cost = tf.reduce_sum(loss) / args.batch_size / args.seq_length
        self.final_state = last_state
        self.lr = tf.Variable(0.0, trainable=False)
        tvars = tf.trainable_variables()
        grads, _ = tf.clip_by_global_norm(tf.gradients(self.cost, tvars),
                                          args.grad_clip)
        optimizer = tf.train.AdamOptimizer(self.lr)
        self.train_op = optimizer.apply_gradients(zip(grads, tvars))

    def __init__(self, is_training, config):
        self.batch_size = batch_size = config.batch_size  # size for mini batch training
        self.num_steps = num_steps = config.num_steps  # maximum number of training iteration?
        size = config.hidden_size  # state size
        feature_size = config.feature_size

        self._input_data = tf.placeholder(tf.int32, [batch_size, num_steps, feature_size])
        self._targets = tf.placeholder(tf.int32, [batch_size, num_steps, feature_size])

        basic_cell = rnn_cell.BasicLSTMCell(size)
        if is_training and config.keep_prob < 1:  # use dropout
            basic_cell = rnn_cell.DropoutWrapper(basic_cell, output_keep_prob=config.keep_prob)
        cell = rnn_cell.MultiRNNCell([basic_cell] * config.num_layers)  # multiple layers
        self._initial_state = cell.zero_state(batch_size, tf.float32)

        inputs = self._input_data
        print inputs
        print "haha"
        if is_training and config.keep_prob < 1:
            inputs = tf.nn.dropout(inputs, config.keep_prob)

        #        inputs = [tf.squeeze(input_, [1])
        #                   for input_ in tf.split(1, num_steps, inputs)]
        #        outputs, states = rnn.rnn(
        #            cell, inputs, initial_state=self._initial_state)
        #
        outputs = []
        states = []
        state = self._initial_state
        with tf.variable_scope("RNN"):
            for time_step in range(num_steps):
                if time_step > 0:
                    tf.get_variable_scope().reuse_variables()
                (cell_output, state) = cell(inputs[:, time_step, :], state)
                outputs.append(cell_output)
                states.append(state)

        print outputs
        output = tf.reshape(tf.concat(1, outputs), [-1, size])
        print output
        logits = tf.nn.xw_plus_b(
            output, tf.get_variable("softmax_w", [size, feature_size]), tf.get_variable("softmax_b", [feature_size])
        )

        loss = seq2seq.sequence_loss_by_example(
            [logits], [tf.reshape(self._targets, [-1])], [tf.ones([batch_size * num_steps])], feature_size
        )
        self._cost = cost = tf.reduce_sum(loss) / batch_size
        self._final_state = states[-1]

        if not is_training:
            return

        self._lr = tf.Variable(0.0, trainable=False)
        tvars = tf.trainable_variables()
        grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars), config.max_grad_norm)
        optimizer = tf.train.GradientDescentOptimizer(self.lr)
        self._train_op = optimizer.apply_gradients(zip(grads, tvars))

  def testSequenceLossByExample(self):
    with self.test_session() as sess:
      output_classes = 5
      logits = [tf.constant(i + 0.5, shape=[2, output_classes])
                for i in xrange(3)]
      targets = [tf.constant(i, tf.int32, shape=[2]) for i in xrange(3)]
      weights = [tf.constant(1.0, shape=[2]) for i in xrange(3)]

      average_loss_per_example = seq2seq.sequence_loss_by_example(
          logits, targets, weights, output_classes,
          average_across_timesteps=True)
      res = sess.run(average_loss_per_example)
      self.assertAllClose(res, np.asarray([1.609438, 1.609438]))

      loss_per_sequence = seq2seq.sequence_loss_by_example(
          logits, targets, weights, output_classes,
          average_across_timesteps=False)
      res = sess.run(loss_per_sequence)
      self.assertAllClose(res, np.asarray([4.828314, 4.828314]))

  def __init__(self, is_training, config):
    self.batch_size = batch_size = config.batch_size
    self.num_steps = num_steps = config.num_steps
    size = config.hidden_size
    output_size = config.output_size

    self._input_data = tf.placeholder(tf.float32, [batch_size, num_steps, size])
    self._targets = tf.placeholder(tf.int32, [batch_size, num_steps])


    lstm_cell = rnn_cell.BasicLSTMCell(size, forget_bias=0.0)
    if is_training and config.keep_prob < 1:
      lstm_cell = rnn_cell.DropoutWrapper(
          lstm_cell, output_keep_prob=config.keep_prob)
    cell = rnn_cell.MultiRNNCell([lstm_cell] * config.num_layers)

    self._initial_state = cell.zero_state(batch_size, tf.float32)

    inputs = self._input_data


    outputs = []
    states = []
    state = self._initial_state
    with tf.variable_scope("RNN"):
      for time_step in range(num_steps):
        if time_step > 0: tf.get_variable_scope().reuse_variables()
        (cell_output, state) = cell(inputs[:, time_step, :], state)
        outputs.append(cell_output)
        states.append(state)

    output = tf.reshape(tf.concat(1, outputs), [-1, size])
    logits = tf.nn.xw_plus_b(output,
                             tf.get_variable("softmax_w", [size, output_size]),
                             tf.get_variable("softmax_b", [output_size]))
    loss = seq2seq.sequence_loss_by_example([logits],
                                            [tf.reshape(self._targets, [-1])],
                                            [tf.ones([batch_size * num_steps])],
                                            output_size)
    self._cost = cost = tf.reduce_sum(loss) / batch_size
    self._final_state = states[-1]
    self._output = output
    self._logits = logits

    if not is_training:
      return

    self._lr = tf.Variable(0.0, trainable=False)
    tvars = tf.trainable_variables()
    grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars),
                                      config.max_grad_norm)
    optimizer = tf.train.GradientDescentOptimizer(self.lr)
    self._train_op = optimizer.apply_gradients(zip(grads, tvars))

  def __init__(self, args, sampling=False):
    self.args = args
    if sampling:
      args.batch_size = 1
      args.seq_length = 1

    basic_cell = rnn_cell.BasicLSTMCell(args.rnn_size)
    self.cell = rnn_cell.MultiRNNCell([basic_cell] * args.num_layers)

    self.input_data = tf.placeholder(tf.int32,
                                     [args.batch_size, args.seq_length])
    self.targets = tf.placeholder(tf.int32,
                                  [args.batch_size, args.seq_length])
    self.initial_state = self.cell.zero_state(args.batch_size, tf.float32)

    with tf.variable_scope('rnnlm'):
      softmax_w = tf.get_variable("softmax_w",
                                  [args.rnn_size, args.vocab_size])
      softmax_b = tf.get_variable("softmax_b",
                                  [args.vocab_size])
      with tf.device("/cpu:0"):
        embedding = tf.get_variable("embedding",
                                    [args.vocab_size, args.rnn_size])
        inputs = tf.split(1, args.seq_length,
                          tf.nn.embedding_lookup(embedding, self.input_data))
        inputs = [tf.squeeze(input_, [1]) for input_ in inputs]

    def loop(prev, _):
      prev = tf.nn.xw_plus_b(prev, softmax_w, softmax_b)
      prev_symbol = tf.stop_gradient(tf.argmax(prev, 1))
      return tf.nn.embedding_lookup(embedding, prev_symbol)

    outputs, last_state = seq2seq.rnn_decoder(
      inputs, self.initial_state, self.cell,
      loop_function=loop if sampling else None, scope='rnnlm')

    output = tf.reshape(tf.concat(1, outputs), [-1, args.rnn_size])
    self.logits = tf.nn.xw_plus_b(output, softmax_w, softmax_b)
    self.probs = tf.nn.softmax(self.logits)
    loss = seq2seq.sequence_loss_by_example(
      [self.logits],
      [tf.reshape(self.targets, [-1])],
      [tf.ones([args.batch_size * args.seq_length])],
      args.vocab_size)
    self.cost = tf.reduce_sum(loss) / args.batch_size / args.seq_length
    self.final_state = last_state
    self.lr = tf.Variable(0.0, trainable=False)
    tvars = tf.trainable_variables()
    grads, _ = tf.clip_by_global_norm(tf.gradients(self.cost, tvars),
                                      args.grad_clip)
    optimizer = tf.train.AdamOptimizer(self.lr)
    self.train_op = optimizer.apply_gradients(zip(grads, tvars))

    def __init__(self, conf):
        self.conf = conf

        cell_fw = BasicLSTMCell(self.conf.rnn_size)
        cell_bw = BasicLSTMCell(self.conf.rnn_size)
        
        if conf.keep_prob < 1.0 and not conf.infer:
            cell_fw = DropoutWrapper(cell_fw, output_keep_prob=conf.keep_prob)
            cell_bw = DropoutWrapper(cell_bw, output_keep_prob=conf.keep_prob)
        self.cell_fw = cell_fw = MultiRNNCell([cell_fw] * self.conf.num_layers)
        self.cell_bw = cell_bw = MultiRNNCell([cell_bw] * self.conf.num_layers)
        
        self.input_data = tf.placeholder(tf.int32, [self.conf.batch_size, self.conf.seq_length])
        self.targets = tf.placeholder(tf.int32, [self.conf.batch_size, self.conf.seq_length])
    
        self.initial_state_fw = cell_fw.zero_state(self.conf.batch_size, tf.float32)
        
        self.initial_state_bw = cell_bw.zero_state(self.conf.batch_size, tf.float32)
        with tf.variable_scope('rnn'):
            softmax_w = tf.get_variable("softmax_w", [self.conf.rnn_size*2, self.conf.output_size])
            softmax_b = tf.get_variable("softmax_b", [self.conf.output_size])
        
        embedding = tf.get_variable("embedding", [self.conf.nerloader.vocab_size, self.conf.rnn_size])
        _inputs = tf.nn.embedding_lookup(embedding, self.input_data)
        if conf.keep_prob < 1.0 and not conf.infer:
            _inputs = tf.nn.dropout(_inputs,conf.keep_prob)
        inputs = tf.split(1, conf.seq_length, _inputs)
        inputs = [tf.squeeze(input_, [1]) for input_ in inputs]
            
        outputs_bi = rnn.bidirectional_rnn(cell_fw, cell_bw, inputs, initial_state_fw=self.initial_state_fw, initial_state_bw=self.initial_state_bw, scope='rnn')
        output = tf.reshape(tf.concat(1, outputs_bi), [-1, self.conf.rnn_size*2])
        self.logits = tf.nn.xw_plus_b(output, softmax_w, softmax_b)
        self.probs = tf.nn.softmax(self.logits)

        self.loss_weights = [tf.ones([self.conf.batch_size * self.conf.seq_length])]

        loss = seq2seq.sequence_loss_by_example([self.logits],
                [tf.reshape(self.targets, [-1])],
                self.loss_weights)
        self.cost = (tf.reduce_sum(loss) / self.conf.batch_size / self.conf.seq_length)
        tf.scalar_summary("loss",self.cost)
        self.out = output
        self.lr = tf.Variable(0.0, trainable=False)
        tvars = tf.trainable_variables()
        grads, _ = tf.clip_by_global_norm(tf.gradients(self.cost, tvars),
                self.conf.grad_clip)
        optimizer = tf.train.AdamOptimizer(self.lr)
        self.train_op = optimizer.apply_gradients(zip(grads, tvars))
        self.merged_summary_op = tf.merge_all_summaries()

#.........这里部分代码省略.........
            out_proj = hidden_size
            if hidden_proj > 0:
                out_proj = hidden_proj

            with tf.device("/cpu:0"):
                w = tf.get_variable("proj_w", [out_proj, vocab_size])
                w_t = tf.transpose(w)
                b = tf.get_variable("proj_b", [vocab_size])
            self.output_projection = (w, b)

            sampled_softmax = False

            # Sampled softmax only makes sense if we sample less than vocabulary size.
            if 0 < num_samples < vocab_size:
                sampled_softmax = True

                def sampled_loss(logits, labels):
                    with tf.device("/cpu:0"):
                        labels = tf.reshape(labels, [-1, 1])
                        losses = tf.nn.sampled_softmax_loss(w_t, b, logits, labels, num_samples, vocab_size)
                        return losses

                loss_function = sampled_loss

            with tf.device("/cpu:0"):
                # input come as one big tensor so we have to split it into a list of tensors to run the rnn cell
                embedding = tf.Variable(
                    tf.random_uniform(
                        [vocab_size, proj_size],
                        minval=-init_scale, maxval=init_scale
                    ),
                    name="embedding"
                )
                # embedding = tf.get_variable("embedding", [vocab_size, proj_size])

                inputs_train = [tf.nn.embedding_lookup(embedding, i) for i in self.input_data_train]
                inputs_valid = [tf.nn.embedding_lookup(embedding, i) for i in self.input_data_valid]

            with tf.variable_scope("RNN", initializer=initializer):

                if attentive:
                    outputs_train, state_train, _ = lm_ops.apply_attentive_lm(
                        self.cell, inputs_train, sequence_length=array_ops.squeeze(math_ops.add_n(self.mask_train)),
                        projection_attention_f=projection_attention_f, output_form=output_form,
                        dropout=self.dropout_feed, initializer=initializer, dtype=tf.float32
                    )

                    outputs_valid, state_valid, _ = lm_ops.apply_attentive_lm(
                        self.cell, inputs_valid, sequence_length=array_ops.squeeze(math_ops.add_n(self.mask_valid)),
                        projection_attention_f=projection_attention_f, output_form=output_form,
                        dropout=self.dropout_feed, initializer=initializer, dtype=tf.float32
                    )

                else:
                   outputs_train, state_train = lm_ops.apply_lm(
                       self.cell, inputs_train, sequence_length=math_ops.add_n(self.mask_train),
                       dropout=self.dropout_feed, dtype=tf.float32
                   )

                   outputs_valid, state_valid = lm_ops.apply_lm(
                       self.cell, inputs_valid, sequence_length=math_ops.add_n(self.mask_valid),
                       dropout=self.dropout_feed, dtype=tf.float32
                   )

                if sampled_softmax is False:
                    logits_train = [tf.nn.xw_plus_b(o, self.output_projection[0], self.output_projection[1])
                                    for o in outputs_train]
                    logits_valid = [tf.nn.xw_plus_b(o, self.output_projection[0], self.output_projection[1])
                                    for o in outputs_valid]
                else:
                    logits_train = outputs_train
                    logits_valid = outputs_valid

            loss_train = seq2seq.sequence_loss_by_example(
                logits_train, self.targets_train, self.mask_train, average_across_timesteps=True
            )

            loss_valid = seq2seq.sequence_loss_by_example(
                logits_valid, self.targets_valid, self.mask_valid, average_across_timesteps=True
            )

            self._cost_train = cost = tf.reduce_sum(loss_train) / float(batch_size)
            self._final_state_train = state_train

            self._cost_valid = tf.reduce_sum(loss_valid) / float(batch_size)
            self._final_state_valid = state_valid

            if not is_training:
                return

            tvars = tf.trainable_variables()
            grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars),
                                              max_grad_norm)

            opt = optimization_ops.get_optimizer(optimizer, learning_rate)
            self._train_op = opt.apply_gradients(zip(grads, tvars), global_step=self.global_step)
            self._valid_op = tf.no_op()

            self.saver = tf.train.Saver(tf.all_variables())
            self.saver_best = tf.train.Saver(tf.all_variables())

    def __init__(self, is_training, config):
        """constructs a graph"""
        self.batch_size = batch_size = config.batch_size
        self.num_steps = num_steps = config.num_steps
        size = config.hidden_size
        vocab_size = config.vocab_size
        self._input_data = tf.placeholder(tf.int32, [batch_size, num_steps],
                                          name="input_data")
        self._targets = tf.placeholder(tf.int32, [batch_size, num_steps],
                                       name="targets")

        # here it is
        lstm_cell = rnn_cell.BasicLSTMCell(size, forget_bias=1.0)
        if is_training and config.keep_prob < 1:
            lstm_cell = rnn_cell.DropoutWrapper(
                lstm_cell, output_keep_prob=config.keep_prob)
        cell = rnn_cell.MultiRNNCell([lstm_cell] * config.num_layers)

        self._initial_state = cell.zero_state(batch_size, tf.float32)

        # do an embedding (always on cpu)
        with tf.device("/cpu:0"):
            embedding = tf.get_variable("embedding", [vocab_size, size])
            inputs = tf.split(
                1, num_steps, tf.nn.embedding_lookup(embedding, self._input_data))
            inputs = [tf.squeeze(input_, [1]) for input_ in inputs]

        if is_training and config.keep_prob < 1:
            inputs = [tf.nn.dropout(input_, config.keep_prob) for input_ in inputs]

        from tensorflow.models.rnn import rnn
        outputs, states = rnn.rnn(cell, inputs, initial_state=self._initial_state)

        # reshape
        outputs = tf.reshape(tf.concat(1, outputs), [-1, size])

        logits = tf.nn.xw_plus_b(outputs,
                                 tf.get_variable("softmax_W", [size,vocab_size]),
                                 tf.get_variable("softmax_b", [vocab_size]))
        self._softmax_out = tf.nn.softmax(logits) # this is just used for sampling
        loss = seq2seq.sequence_loss_by_example([logits],
                                                [tf.reshape(self._targets,[-1])],
                                                [tf.ones([batch_size * num_steps])],
                                                vocab_size)
        self._cost = cost = tf.div(tf.reduce_sum(loss),
                                   tf.constant(batch_size, dtype=tf.float32))
        self._final_state = states[-1]

        if not is_training:
            return # don't need to optimisation ops

        self._lr = tf.Variable(0.0, trainable=False)
        tvars = tf.trainable_variables()
        grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars),
                                          config.max_grad_norm)
        # actually the simple guy does good
        # with the grad clipping and the lr schedule and whatnot
#ftrl?
        #optimizer = tf.train.GradientDescentOptimizer(self.lr)
        optimizer = tf.train.FtrlOptimizer(self.lr)
        self._train_op = optimizer.apply_gradients(zip(grads, tvars))

		inputs_dis    = [tf.matmul(tf.squeeze(i, [1]), embedding) for i in inputs_dis]

state   = initial_state_dis
outputs = []

for i, inp in enumerate(inputs_dis):
	if i > 0:
		tf.get_variable_scope().reuse_variables()
	output, state = cell_dis(inp, state)
	outputs.append(output)
last_state = state

output_tf   = tf.reshape(tf.concat(1, outputs), [-1, args.rnn_size])
logits = tf.nn.xw_plus_b(output_tf, softmax_w, softmax_b)
probs  = tf.nn.softmax(logits)

loss = seq2seq.sequence_loss_by_example(
	[logits],
	[tf.reshape(targets, [-1])], 
	[tf.ones([args.batch_size * args.seq_length])],
	2)

cost = tf.reduce_sum(loss) / args.batch_size / args.seq_length

final_state = last_state
lr          = tf.Variable(0.0, trainable = False)
tvars 	         = tf.trainable_variables()
grads, _         = tf.clip_by_global_norm(tf.gradients(cost, tvars, aggregation_method = 2), args.grad_clip)
optimizer        = tf.train.AdamOptimizer(lr)
train_op    = optimizer.apply_gradients(zip(grads, tvars))

#.........这里部分代码省略.........
        # outputs_forward = tf.reshape(tf.concat(1, outputs), [-1, self.lstm_size],
        #                    name="reshape-outputs_forward")

        # outputs_backward = tf.reshape(tf.concat(1, outputs_rev), [-1, self.lstm_size],
        #                    name="reshape-outputs_backward")

        # forward_w = tf.get_variable("forward_w", [self.lstm_size, self.lstm_size])
        # backward_w = tf.get_variable("backward_w", [self.lstm_size, self.lstm_size])
        # non_linearity_bias = tf.get_variable("non_linearity_b", [self.lstm_size])

        outputs_bidi = [tf.concat(1, [o1, o2]) for o1, o2 in zip(outputs, reversed(outputs_rev))]

        # output = tf.tanh(tf.matmul(outputs_forward, forward_w) + tf.matmul(outputs_backward, backward_w) + non_linearity_bias)
        output = tf.reshape(tf.concat(1, outputs_bidi), [-1, 2 * self.lstm_size], name="reshape-outputs_bidi")
        output_dropped = tf.nn.dropout(output, self.dropout_prob[1])

        # We are computing only the logits, not the actual softmax -- while
        # computing the loss, it is done by the sequence_loss_by_example and
        # during the runtime classification, the argmax over logits is enough.

        softmax_w = tf.get_variable("softmax_w", [2 * self.lstm_size, len(tagset)])
        logits_flatten = tf.nn.xw_plus_b(output_dropped, softmax_w, tf.get_variable("softmax_b", [len(tagset)]))
        # tf.get_variable_scope().reuse_variables()
        regularize.append(softmax_w)

        self.logits = tf.reshape(logits_flatten, [-1, num_steps, len(tagset)], name="reshape-logits")
        estimated_tags_flat = tf.to_int32(tf.argmax(logits_flatten, dimension=1))
        self.last_state = last_state

        # output maks: compute loss only if it insn't a padded word (i.e. zero index)
        output_mask = tf.reshape(tf.to_float(tf.not_equal(self.tags, 0)), [-1])

        gt_tags_flat = tf.reshape(self.tags, [-1])
        tagging_loss = seq2seq.sequence_loss_by_example(
            logits=[logits_flatten], targets=[gt_tags_flat], weights=[output_mask]
        )

        tagging_accuracy = tf.reduce_sum(
            tf.to_float(tf.equal(estimated_tags_flat, gt_tags_flat)) * output_mask
        ) / tf.reduce_sum(output_mask)
        tf.scalar_summary("train_accuracy", tagging_accuracy, collections=["train"])
        tf.scalar_summary("dev_accuracy", tagging_accuracy, collections=["dev"])

        self.cost = tf.reduce_mean(tagging_loss)

        tf.scalar_summary("train_tagging_loss", tf.reduce_mean(tagging_loss), collections=["train"])
        tf.scalar_summary("dev_tagging_loss", tf.reduce_mean(tagging_loss), collections=["dev"])

        if generate_lemmas:
            with tf.variable_scope("decoder"):
                self.lemma_chars = tf.placeholder(tf.int32, [None, num_steps, num_chars + 2], name="lemma_chars")

                lemma_state_size = self.lstm_size

                lemma_w = tf.Variable(tf.random_uniform([lemma_state_size, len(alphabet)], 0.5), name="state_to_char_w")
                lemma_b = tf.Variable(tf.fill([len(alphabet)], -math.log(len(alphabet))), name="state_to_char_b")
                lemma_char_embeddings = tf.Variable(
                    tf.random_uniform(
                        [len(alphabet), lemma_state_size / (2 if supply_form_characters_to_lemma else 1)], -0.5, 0.5
                    ),
                    name="char_embeddings",
                )

                lemma_char_inputs = [
                    tf.squeeze(input_, [1])
                    for input_ in tf.split(

    def __init__(self, CellType, is_training, config):
        self.batch_size = batch_size = config.batch_size
        self.num_steps = num_steps = config.num_steps
        size = config.hidden_size
        vocab_size = config.vocab_size

        self.input_data = tf.placeholder(tf.int32, [batch_size, num_steps], name="input_data")
        self.targets = tf.placeholder(tf.int32, [batch_size, num_steps], name="targets")

        lstm_cell = CellType(size)
        if is_training and config.keep_prob < 1:
            lstm_cell = rnn_cell.DropoutWrapper(lstm_cell, output_keep_prob=config.keep_prob)
        cell = rnn_cell.MultiRNNCell([lstm_cell] * config.num_layers)
        self.initial_state = cell.zero_state(batch_size, tf.float32)

        # initializer used for reusable variable initializer (see `get_variable`)
        initializer = tf.random_uniform_initializer(-config.init_scale, config.init_scale)

        with tf.device("/cpu:0"):
            embedding = tf.get_variable("embedding", [vocab_size, size], initializer=initializer)
            inputs = tf.nn.embedding_lookup(embedding, self.input_data)

        if is_training and config.keep_prob < 1:
            inputs = tf.nn.dropout(inputs, config.keep_prob)

        outputs = []
        states = []
        state = self.initial_state

        with tf.variable_scope("RNN", initializer=initializer):
            for time_step in range(num_steps):
                if time_step > 0:
                    tf.get_variable_scope().reuse_variables()

                inputs_slice = inputs[:,time_step,:]
                (cell_output, state) = cell(inputs_slice, state)

                outputs.append(cell_output)
                states.append(state)

        self.final_state = states[-1]

        output = tf.reshape(tf.concat(1, outputs), [-1, size])
        w = tf.get_variable("softmax_w",
                                    [size, vocab_size],
                                    initializer=initializer)
        b = tf.get_variable("softmax_b", [vocab_size], initializer=initializer)

        logits = tf.nn.xw_plus_b(output, w, b) # compute logits for loss
        targets = tf.reshape(self.targets, [-1]) # reshape our target outputs
        weights = tf.ones([batch_size * num_steps]) # used to scale the loss average

        # computes loss and performs softmax on our fully-connected output layer
        loss = sequence_loss_by_example([logits], [targets], [weights], vocab_size)
        self.cost = cost = tf.div(tf.reduce_sum(loss), batch_size, name="cost")

        if is_training:
            # setup learning rate variable to decay
            self.lr = tf.Variable(1.0, trainable=False)

            # define training operation and clip the gradients
            tvars = tf.trainable_variables()
            grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars), config.max_grad_norm)
            optimizer = tf.train.GradientDescentOptimizer(self.lr)
            self.train_op = optimizer.apply_gradients(zip(grads, tvars), name="train")
        else:
            # if this model isn't for training (i.e. testing/validation) then we don't do anything here
            self.train_op = tf.no_op()

    def __init__(self, args, infer=False):
        self.args = args
        if infer:
            args.batch_size = 1
            args.seq_length = 1

        if args.model == 'rnn': cell_fn = jzRNNCell
        elif args.model == 'gru': cell_fn = jzGRUCell
        elif args.model == 'lstm': cell_fn = jzLSTMCell
        else: raise Exception("model type not supported: {}".format(args.model))

        if args.activation == 'tanh': cell_af = tf.tanh
        elif args.activation == 'sigmoid': cell_af = tf.sigmoid
        elif args.activation == 'relu': cell_af = tf.nn.relu
        else: raise Exception("activation function not supported: {}".format(args.activation))

        self.input_data = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
        self.targets = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])

        with tf.variable_scope('rnnlm'):
            if not args.bidirectional:
                softmax_w = tf.get_variable("softmax_w", [args.rnn_size, args.vocab_size])
            else:
                softmax_w = tf.get_variable("softmax_w", [args.rnn_size*2, args.vocab_size])
            softmax_b = tf.get_variable("softmax_b", [args.vocab_size])
            with tf.device("/cpu:0"):
                embedding = tf.get_variable("embedding", [args.vocab_size, args.rnn_size])
                inputs = tf.split(1, args.seq_length, tf.nn.embedding_lookup(embedding, self.input_data))
                inputs = [tf.nn.dropout(tf.squeeze(input_, [1]),args.dropout) for input_ in inputs]

        # one-directional RNN (nothing changed here..)
        if not args.bidirectional:
            cell = cell_fn(args.rnn_size,activation=cell_af)
            self.cell = cell = rnn_cell.MultiRNNCell([cell] * args.num_layers)
            self.initial_state = cell.zero_state(args.batch_size, tf.float32)
            def loop(prev, _):
                prev = tf.matmul(prev, softmax_w) + softmax_b
                prev_symbol = tf.stop_gradient(tf.argmax(prev, 1))
                return tf.nn.embedding_lookup(embedding, prev_symbol)
            outputs, last_state = seq2seq.rnn_decoder(inputs, self.initial_state, cell, loop_function=loop if infer else None, scope='rnnlm')
            output = tf.reshape(tf.concat(1, outputs), [-1, args.rnn_size])

        # bi-directional RNN
        else:
            lstm_fw = cell_fn(args.rnn_size,activation=cell_af)
            lstm_bw = cell_fn(args.rnn_size,activation=cell_af)
            self.lstm_fw = lstm_fw = rnn_cell.MultiRNNCell([lstm_fw]*args.num_layers)
            self.lstm_bw = lstm_bw = rnn_cell.MultiRNNCell([lstm_bw]*args.num_layers)
            self.initial_state_fw = lstm_fw.zero_state(args.batch_size,tf.float32)
            self.initial_state_bw = lstm_bw.zero_state(args.batch_size,tf.float32)
            outputs,_,_ = rnn.bidirectional_rnn(lstm_fw, lstm_bw, inputs,
                                            initial_state_fw=self.initial_state_fw,
                                            initial_state_bw=self.initial_state_bw,
                                                sequence_length=args.batch_size) 
            output = tf.reshape(tf.concat(1, outputs), [-1, args.rnn_size*2])

        self.logits = tf.matmul(tf.nn.dropout(output,args.dropout), softmax_w) + softmax_b
        self.probs = tf.nn.softmax(self.logits)
        loss = seq2seq.sequence_loss_by_example([self.logits],
                [tf.reshape(self.targets, [-1])],
                [tf.ones([args.batch_size * args.seq_length])],
                args.vocab_size)
        self.cost = tf.reduce_sum(loss) / args.batch_size / args.seq_length
        self.final_state = last_state
        self.lr = tf.Variable(0.0, trainable=False)
        tvars = tf.trainable_variables()
        grads, _ = tf.clip_by_global_norm(tf.gradients(self.cost, tvars),
                args.grad_clip)
        optimizer = tf.train.AdamOptimizer(self.lr)
        self.train_op = optimizer.apply_gradients(zip(grads, tvars))

    def __init__(self, args, infer=False):
        self.args = args
        # if infer:
        #     args.batch_size = 1
        #     args.seq_length = 1

        if args.model == 'rnn':
            cell_fn = rnn_cell.BasicRNNCell
        elif args.model == 'gru':
            cell_fn = rnn_cell.GRUCell
        elif args.model == 'lstm':
            cell_fn = rnn_cell.BasicLSTMCell
        else:
            raise Exception("model type not supported: {}".format(args.model))

        cell = cell_fn(args.rnn_size)

        self.cell = cell = rnn_cell.MultiRNNCell([cell] * args.num_layers)

        #self.seq_length = tf.placeholder(tf.int32)
        #args.seq_length = self.seq_length
        self.input_data = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
        self.targets = tf.placeholder(tf.int32, [args.batch_size])
        self.initial_state = cell.zero_state(args.batch_size, tf.float32)

        with tf.variable_scope('rnnlm'):
            softmax_w = tf.get_variable("softmax_w", [args.rnn_size, args.vocab_size])
            softmax_b = tf.get_variable("softmax_b", [args.vocab_size])
            with tf.device("/cpu:0"):
                embedding = tf.get_variable("embedding", [args.vocab_size, args.rnn_size])
                inputs = tf.split(1, args.seq_length, tf.nn.embedding_lookup(embedding, self.input_data))
                # len(inputs)==args.seq_length, shape(inputs[0])==(args.batch_size, args.rnn_size)
                inputs = [tf.squeeze(input_, [1]) for input_ in inputs]

        def loop(prev, _):
            return None  # TODO
            prev = tf.nn.xw_plus_b(prev, softmax_w, softmax_b)
            prev_symbol = tf.stop_gradient(tf.argmax(prev, 1))
            return tf.nn.embedding_lookup(embedding, prev_symbol)

        # len(outputs)==args.seq_length, shape(outputs[0])==(args.batch_size, args.rnn_size)
        outputs, states = seq2seq.rnn_decoder(inputs, self.initial_state, cell, loop_function=loop if infer else None, scope='rnnlm')
        # # shape(output) = (batch_size*seq_length, rnn_size)
        # output = tf.reshape(tf.concat(1, outputs), [-1, args.rnn_size])
        def handle_outputs(use_lastone=True):
            """ Shape of return is [batch_size, rnn_size].
            """
            if use_lastone:
                return outputs[-1]
            output = tf.add_n(outputs)
            output = tf.div(output, len(outputs))
            return output
        output = handle_outputs(use_lastone=False)
        # shape(logits) = (batch_size, vocab_size)
        self.logits = tf.nn.xw_plus_b(output, softmax_w, softmax_b)
        self.probs = tf.nn.softmax(self.logits)
        loss = seq2seq.sequence_loss_by_example([self.logits],
                [tf.reshape(self.targets, [-1])],
                [tf.ones([args.batch_size])],
                args.vocab_size)
        self.cost = tf.reduce_sum(loss) / args.batch_size
        _ = tf.scalar_summary('cost', self.cost)

        # Evaluate accuracy
        correct_pred = tf.equal(tf.cast(tf.argmax(self.logits, 1), tf.int32), tf.reshape(self.targets, [-1]))
        self.accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
        _ = tf.scalar_summary('accuracy', self.accuracy)

        self.final_state = states
        self.lr = tf.Variable(0.0, trainable=False)
        tvars = tf.trainable_variables()
        grads, _ = tf.clip_by_global_norm(tf.gradients(self.cost, tvars),
                args.grad_clip)
        optimizer = tf.train.AdamOptimizer(self.lr)
        self.train_op = optimizer.apply_gradients(zip(grads, tvars))

        inputs = tf.split(1, seq_length, tf.nn.embedding_lookup(embedding, input_data))
        inputs = [tf.squeeze(input_, [1]) for input_ in inputs]
# Loop function for seq2seq
def loop(prev, _):
    prev = tf.nn.xw_plus_b(prev, softmax_w, softmax_b)
    prev_symbol = tf.stop_gradient(tf.argmax(prev, 1))
    return tf.nn.embedding_lookup(embedding, prev_symbol)
# Output of RNN
outputs, last_state = seq2seq.rnn_decoder(inputs, initial_state, cell, loop_function=None, scope='rnnlm')
output = tf.reshape(tf.concat(1, outputs), [-1, rnn_size])
logits = tf.nn.xw_plus_b(output, softmax_w, softmax_b)
# Next word probability
probs = tf.nn.softmax(logits)
# Define LOSS
loss = seq2seq.sequence_loss_by_example([logits], # Input
    [tf.reshape(targets, [-1])], # Target
    [tf.ones([batch_size * seq_length])], # Weight
    vocab_size)
# Define Optimizer
cost = tf.reduce_sum(loss) / batch_size / seq_length
final_state = last_state
lr = tf.Variable(0.0, trainable=False)
tvars = tf.trainable_variables()
grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars), grad_clip)
_optm = tf.train.AdamOptimizer(lr)
optm = _optm.apply_gradients(zip(grads, tvars))

print ("Network Ready")


# In[ ]:

    def __init__(self, args, infer=False):
        self.args = args
        if infer:
            args.batch_size = 1
            args.seq_length = 1

        if args.model == 'rnn':
            cell_fn = rnn_cell.BasicRNNCell
        elif args.model == 'gru':
            cell_fn = rnn_cell.GRUCell
        elif args.model == 'lstm':
            cell_fn = rnn_cell.BasicLSTMCell
        else:
            raise Exception("model type not supported: {}".format(args.model))

        cell = cell_fn(args.rnn_size)

        self.cell = cell = rnn_cell.MultiRNNCell([cell] * args.num_layers)

        # create tensorflow placeholder
        self.input_data = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
        self.targets = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
        # Initial state of the cell memory.
        self.initial_state = cell.zero_state(args.batch_size, tf.float32)

        # create namespace for shareable variables (variable name = "rnnlm/softmax_w")
        with tf.variable_scope('rnnlm'):
            # create (or get) a variable with shape [rnn_size, vocab_size]
            softmax_w = tf.get_variable("softmax_w", [args.rnn_size, args.vocab_size])
            softmax_b = tf.get_variable("softmax_b", [args.v