diff --git a/deepy/dataset/__init__.py b/deepy/dataset/__init__.py
index 9302fd6..61e7fd7 100644
--- a/deepy/dataset/__init__.py
+++ b/deepy/dataset/__init__.py
@@ -10,4 +10,6 @@
 from seq_mini_batch import SequentialMiniBatches
 from binarized_mnist import BinarizedMnistDataset
 from bunch_seq import BunchSequences
-from ondisk_dataset import OnDiskDataset
\ No newline at end of file
+from ondisk_dataset import OnDiskDataset
+from data_processor import DataProcessor
+from padding import pad_dataset
\ No newline at end of file
diff --git a/deepy/dataset/data_processor.py b/deepy/dataset/data_processor.py
new file mode 100644
index 0000000..708b625
--- /dev/null
+++ b/deepy/dataset/data_processor.py
@@ -0,0 +1,10 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+
+class DataProcessor(object):
+    """
+    An abstract class for data processor.
+    """
+    def process(self, split, epoch, dataset):
+        return dataset
\ No newline at end of file
diff --git a/deepy/dataset/ondisk_dataset.py b/deepy/dataset/ondisk_dataset.py
index c92844e..56fc49b 100644
--- a/deepy/dataset/ondisk_dataset.py
+++ b/deepy/dataset/ondisk_dataset.py
@@ -6,6 +6,7 @@
 
 import types
 from . import Dataset
+from data_processor import DataProcessor
 from deepy.utils import FakeGenerator, StreamPickler, global_rand
 
 import logging as loggers
@@ -19,7 +20,10 @@ class OnDiskDataset(Dataset):
     """
 
     def __init__(self, train_path, valid_path=None, test_path=None, train_size=None,
-                 cached=False, post_processing=None, shuffle_memory=False, curriculum=None):
+                 cached=False, post_processing=None, shuffle_memory=False, data_processor=None):
+        """
+        :type data_processor: DataProcessor
+        """
         self._train_path = train_path
         self._valid_path = valid_path
         self._test_path = test_path
@@ -28,48 +32,41 @@ def __init__(self, train_path, valid_path=None, test_path=None, train_size=None,
         self._cached_train_data = None
         self._post_processing = post_processing if post_processing else lambda x: x
         self._shuffle_memory = shuffle_memory
-        self._curriculum = curriculum
-        self._curriculum_count = 0
-        if curriculum and not callable(curriculum):
-            raise Exception("curriculum function must be callable")
-        if curriculum and not cached:
-            raise Exception("curriculum learning needs training data to be cached")
+        self._epoch = 0
+        self._data_processor = data_processor
+        if data_processor and not isinstance(data_processor, DataProcessor):
+            raise Exception("data_processor must be an instance of DataProcessor.")
         if self._cache_on_memory:
             logging.info("Cache on memory")
             self._cached_train_data = list(map(self._post_processing, StreamPickler.load(open(self._train_path))))
             self._train_size = len(self._cached_train_data)
-            # if self._shuffle_memory:
-            #     logging.info("Shuffle on-memory data")
-            #     global_rand.shuffle(self._cached_train_data)
+            if self._shuffle_memory:
+                logging.info("Shuffle on-memory data")
+                global_rand.shuffle(self._cached_train_data)
 
-    def curriculum_train_data(self):
-        self._curriculum_count += 1
-        logging.info("curriculum learning: round {}".format(self._curriculum_count))
-        return self._curriculum(self._cached_train_data, self._curriculum_count)
+    def _process_data(self, split, epoch, dataset):
+        if self._data_processor:
+            return self._data_processor.process(split, epoch, dataset)
+        else:
+            return dataset
 
     def generate_train_data(self):
-        for data in StreamPickler.load(open(self._train_path)):
+        self._epoch += 1
+        data_source = self._cached_train_data if self._cache_on_memory else StreamPickler.load(open(self._train_path))
+        for data in self._process_data('train', self._epoch, data_source):
             yield self._post_processing(data)
 
     def generate_valid_data(self):
-        for data in StreamPickler.load(open(self._valid_path)):
+        data_source = StreamPickler.load(open(self._valid_path))
+        for data in self._process_data('valid', self._epoch, data_source):
             yield self._post_processing(data)
 
     def generate_test_data(self):
-        for data in StreamPickler.load(open(self._test_path)):
+        data_source = StreamPickler.load(open(self._test_path))
+        for data in self._process_data('test', self._epoch, data_source):
             yield self._post_processing(data)
 
     def train_set(self):
-        if self._cache_on_memory:
-            if self._shuffle_memory:
-                logging.info("shuffle on-memory data")
-                global_rand.shuffle(self._cached_train_data)
-            if self._curriculum:
-                if not isinstance(self._curriculum(self._cached_train_data, 1), types.GeneratorType):
-                    raise Exception("Curriculum function must be a generator.")
-                return FakeGenerator(self, "curriculum_train_data")
-            else:
-                return self._cached_train_data
         if not self._train_path:
             return None
         return FakeGenerator(self, "generate_train_data")
diff --git a/deepy/dataset/padding.py b/deepy/dataset/padding.py
index 4cb3316..d5ba304 100644
--- a/deepy/dataset/padding.py
+++ b/deepy/dataset/padding.py
@@ -4,7 +4,7 @@
 import numpy as np
 from deepy.utils import FLOATX
 
-def pad_dataset(subset, side, length):
+def pad_dataset(subset, side="right", length=-1):
     """
     Pad data set to specified length.
     Parameters:
diff --git a/deepy/layers/block.py b/deepy/layers/block.py
index 76fbe74..4d9e0e3 100644
--- a/deepy/layers/block.py
+++ b/deepy/layers/block.py
@@ -51,6 +51,14 @@ def compute_tensor(self, x):
     def compute_test_tesnor(self, x):
         return x
 
+    def load_params(self, path, exclude_free_params=False):
+        """
+        Load parameters to the block.
+        """
+        from deepy.networks.comp_graph import ComputationalGraph
+        model = ComputationalGraph(blocks=[self])
+        model.load_params(path, exclude_free_params=exclude_free_params)
+
     @property
     def all_parameters(self):
         return self.parameters
\ No newline at end of file
diff --git a/deepy/layers/functions.py b/deepy/layers/functions.py
index 9dcd357..b96def7 100644
--- a/deepy/layers/functions.py
+++ b/deepy/layers/functions.py
@@ -9,7 +9,10 @@ def concatenate(vars, axis=-1):
     """
     A utility function of concatenate.
     """
-    return Concatenate(axis=axis).compute(*vars)
+    concat_var = Concatenate(axis=axis).compute(*vars)
+    if axis == -1 or axis == vars[0].tensor.ndim - 1:
+        concat_var.output_dim = sum([x.output_dim for x in vars], 0)
+    return concat_var
 
 @neural_computation
 def ifelse(condition, then_branch, else_branch):
diff --git a/deepy/layers/layer.py b/deepy/layers/layer.py
index 1e176d0..033a5e2 100644
--- a/deepy/layers/layer.py
+++ b/deepy/layers/layer.py
@@ -93,10 +93,10 @@ def compute(self, *inputs, **kwargs):
         output = self.compute_tensor(*[t.tensor for t in inputs], **train_kwargs)
         test_output = self.compute_test_tesnor(*[t.test_tensor for t in inputs], **test_kwargs)
 
-        if type(output) != list:
+        if type(output) != list and type(output) != tuple:
             return NeuralVariable(output, test_output, self.output_dim)
         else:
-            return [NeuralVariable(*item) for item in zip(self.output_dims, output, test_output)]
+            return [NeuralVariable(*item) for item in zip(output, test_output, self.output_dims)]
 
     def prepare(self):
         """
diff --git a/deepy/layers/recurrent.py b/deepy/layers/recurrent.py
index a895995..afb39a2 100644
--- a/deepy/layers/recurrent.py
+++ b/deepy/layers/recurrent.py
@@ -75,6 +75,22 @@ def merge_inputs(self, input_var, additional_inputs=None):
     def prepare(self):
         pass
 
+    @neural_computation
+    def compute_step(self, state, lstm_cell=None, input=None, additional_inputs=None):
+        """
+        Compute one step in the RNN.
+        :return: one variable for RNN and GRU, multiple variables for LSTM
+        """
+        input_map = self.merge_inputs(input, additional_inputs=additional_inputs)
+        input_map.update({"state": state, "lstm_cell": lstm_cell})
+        output_map = self.compute_new_state(input_map)
+        outputs = [output_map.pop("state")]
+        outputs += output_map.values()
+        if len(outputs) == 1:
+            return outputs[0]
+        else:
+            return outputs
+
     @neural_computation
     def get_initial_states(self, input_var):
         """
diff --git a/deepy/networks/comp_graph.py b/deepy/networks/comp_graph.py
index 7dc7389..b38aea3 100644
--- a/deepy/networks/comp_graph.py
+++ b/deepy/networks/comp_graph.py
@@ -11,7 +11,7 @@ class ComputationalGraph(NeuralNetwork):
     """
 
     def __init__(self, input_dim=0, model=None, input_tensor=None, monitors=None,
-                 cost=None, output=None, outputs=None, blocks=None, input_vars=None, target_vars=None):
+                 cost=None, output=None, outputs=None, blocks=None, input_vars=None, target_vars=None, output_map=None):
         """
         Create a basic network.
 
@@ -41,6 +41,9 @@ def __init__(self, input_dim=0, model=None, input_tensor=None, monitors=None,
             if not output and not cost:
                 self._test_output = None
             self._test_outputs = [o.test_tensor for o in outputs]
+
+        self.output_map = output_map if output_map else {}
+
         if monitors:
             if type(monitors) == dict:
                 monitors = monitors.items()
diff --git a/deepy/preprocessing/__init__.py b/deepy/preprocessing/__init__.py
new file mode 100644
index 0000000..7cb5f9a
--- /dev/null
+++ b/deepy/preprocessing/__init__.py
@@ -0,0 +1,4 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+from padding import pad_sequence
\ No newline at end of file
diff --git a/deepy/preprocessing/padding.py b/deepy/preprocessing/padding.py
new file mode 100644
index 0000000..4dce63d
--- /dev/null
+++ b/deepy/preprocessing/padding.py
@@ -0,0 +1,26 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+import numpy as np
+from itertools import izip, izip_longest
+
+def pad_sequence(batch, pad_value=0, output_mask=True, length=None):
+    if length:
+        max_len = length
+    else:
+        max_len = max(map(len, batch))
+    mask = None
+    if output_mask:
+        mask = []
+        for i in range(len(batch)):
+            mask.append([1] * len(batch[i]) + [0] * (max_len - len(batch[i])))
+        mask = np.array(mask, dtype="float32")
+    if length:
+        new_batch = []
+        for i in range(len(batch)):
+            new_row = list(batch[i]) + [pad_value] * (max_len - len(batch[i]))
+            new_batch.append(new_row)
+        new_batch = np.array(new_batch)
+    else:
+        new_batch = np.array(list(izip(*izip_longest(*batch, fillvalue=pad_value))))
+    return new_batch, mask
\ No newline at end of file
diff --git a/deepy/trainers/base.py b/deepy/trainers/base.py
index 42c0e96..a6da4a3 100644
--- a/deepy/trainers/base.py
+++ b/deepy/trainers/base.py
@@ -230,7 +230,7 @@ def _run_train(self, epoch, train_set, train_size=None):
         self.last_run_costs = costs
         return costs
 
-    def _run_valid(self, epoch, valid_set, dry_run=False):
+    def _run_valid(self, epoch, valid_set, dry_run=False, save_path=None):
         """
         Run one valid iteration, return true if to continue training.
         """
@@ -246,9 +246,10 @@ def _run_valid(self, epoch, valid_set, dry_run=False):
                 self.best_cost = J
                 self.best_epoch = epoch
 
-            if self.config.auto_save and self._skip_batches == 0:
+            save_path = save_path if save_path else self.config.auto_save
+            if save_path and self._skip_batches == 0:
                 self.network.train_logger.record_progress(self._progress)
-                self.network.save_params(self.config.auto_save, new_thread=True)
+                self.network.save_params(save_path, new_thread=True)
 
         info = ' '.join('%s=%.2f' % el for el in costs)
         epoch_str = "epoch=%d" % (epoch + 1)
diff --git a/deepy/utils/activations.py b/deepy/utils/activations.py
index 91b4fc3..d3b0d31 100644
--- a/deepy/utils/activations.py
+++ b/deepy/utils/activations.py
@@ -49,7 +49,8 @@ def compose(a, b):
             'theano_softmax': T.nnet.softmax,
 
             # shorthands
-            'relu': lambda z: z * (z > 0),
+            'relu': lambda z: T.nnet.relu(z),
+            'leaky_relu': lambda z: T.nnet.relu(z, 0.01),
             'trel': lambda z: z * (z > 0) * (z < 1),
             'trec': lambda z: z * (z > 1),
             'tlin': lambda z: z * (abs(z) > 1),
diff --git a/deepy/utils/decorations.py b/deepy/utils/decorations.py
index 622ddfa..88017e4 100644
--- a/deepy/utils/decorations.py
+++ b/deepy/utils/decorations.py
@@ -1,13 +1,14 @@
 #!/usr/bin/env python
 # -*- coding: utf-8 -*-
 
+from theano.tensor.var import TensorVariable
+
 def convert_to_theano_var(obj):
     """
     Convert neural vars to theano vars.
     :param obj: NeuralVariable or list or dict or tuple
     :return: theano var, test var, tensor found, neural var found
     """
-    from theano.tensor.var import TensorVariable
     from deepy.layers.neural_var import NeuralVariable
     if type(obj) == tuple:
         return tuple(convert_to_theano_var(list(obj)))
@@ -39,6 +40,18 @@ def convert_to_theano_var(obj):
         return obj.tensor, obj.test_tensor, False, True
     elif type(obj) == TensorVariable:
         return obj, obj, True, False
+    elif type(obj) == slice:
+        normal_args = []
+        test_args = []
+        theano_var_found = False
+        neural_var_found = False
+        for arg in [obj.start, obj.stop, obj.step]:
+            normal_var, test_var, tensor_found, neural_found = convert_to_theano_var(arg)
+            normal_args.append(normal_var)
+            test_args.append(test_var)
+            if tensor_found: theano_var_found = True
+            if neural_found: neural_var_found = True
+        return slice(*normal_args), slice(*test_args), theano_var_found, neural_var_found
     else:
         return obj, obj, False, False
 
@@ -74,7 +87,6 @@ def neural_computation(original_func, prefer_tensor=False):
     """
 
     def wrapper(*args, **kwargs):
-
         normal_args, test_args, tensor_found_in_args, neural_found_in_args = convert_to_theano_var(args)
         normal_kwargs, test_kwargs, tensor_found_in_kwargs, neural_found_in_kwargs = convert_to_theano_var(kwargs)
 
@@ -90,10 +102,15 @@ def wrapper(*args, **kwargs):
             # No neural variables are inputted, so output tensors
             return normal_result
         else:
-            # Output neural variables
+            # Output neural variables, auto set output_dim
             test_result = original_func(*test_args, **test_kwargs)
-            return convert_to_neural_var(normal_result, test_result)
-
+            result_var = convert_to_neural_var(normal_result, test_result)
+            if (isinstance(normal_result, TensorVariable) and
+                    hasattr(normal_result.tag, "test_value") and
+                    hasattr(normal_result.tag.test_value, "shape") and
+                    normal_result.tag.test_value.shape):
+                result_var.output_dim = normal_result.tag.test_value.shape[-1]
+            return result_var
     return wrapper
 
 def neural_computation_prefer_tensor(original_func):
diff --git a/deepy/utils/neural_tensor.py b/deepy/utils/neural_tensor.py
index 60fb1d5..96626c9 100644
--- a/deepy/utils/neural_tensor.py
+++ b/deepy/utils/neural_tensor.py
@@ -1,8 +1,10 @@
 #!/usr/bin/env python
 # -*- coding: utf-8 -*-
 
+import theano.tensor as T
 from theano import tensor as theano_tensor
 from decorations import neural_computation
+from deepy.layers.neural_var import NeuralVariable
 
 class NeuralTensorNet(object):
 
@@ -17,8 +19,12 @@ def wrapper(*args, **kwargs):
 class NeuralTensor(object):
     """
     A class for exporting Theano tensor operations to neural variables.
+
     """
 
+    def constant(self, value, dtype="float32", dim=None):
+        return NeuralVariable(T.constant(value, dtype=dtype), dim=dim)
+
     def __getattr__(self, func_name):
         global deepy_nnet
         @neural_computation