diff --git a/.gitignore b/.gitignore
index a0be001..169b41d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -13,3 +13,4 @@ build/
 *.os
 *.a
 *.so
+pydoc/_build
diff --git a/README.md b/README.md
index 87c2f6d..e8adff3 100644
--- a/README.md
+++ b/README.md
@@ -68,10 +68,7 @@ After building the executables, you can run two simple test runs as follows:
 
 To build the Python extension, run
 
-    python setup.py build
-    sudo python setup.py install
-
-(this is currently broken)
+    pip install .
 
 # Documentation / Examples
 
@@ -165,9 +162,10 @@ storage format.
 
 # Python API
 
-The `clstm.i` file implements a simple Python interface to clstm, plus
-a wrapper that makes an INetwork mostly a replacement for the lstm.py
-implementation from ocropy.
+The source code includes a Python interface to clstm (via Cython). Currently
+it only exposes the `CLSTMOCR` class for OCR training and prediction.
+To install it, just make sure you have the above dependencies and
+Cython (>=0.23) installed and run `pip install .`.
 
 # Comand Line Drivers
 
diff --git a/_clstm.pxd b/_clstm.pxd
new file mode 100644
index 0000000..9dae05c
--- /dev/null
+++ b/_clstm.pxd
@@ -0,0 +1,81 @@
+from cpython.ref cimport PyObject
+from libc.stddef cimport wchar_t
+from libcpp.vector cimport vector
+from libcpp.string cimport string
+from libcpp.memory cimport shared_ptr
+
+
+cdef extern from "<string>" namespace "std":
+    cppclass wstring:
+        cppclass iterator:
+            iterator()
+            wchar_t* operator*()
+            iterator(iterator &)
+            iterator operator++()
+            iterator operator--()
+            iterator operator==(iterator)
+            iterator operator!=(iterator)
+        iterator begin()
+        iterator end()
+
+
+cdef extern from "pyextra_defs.h":
+    cdef Py_ssize_t Unicode_AsWideChar(PyObject* ustr, Py_ssize_t length,
+                                       wchar_t* wchars)
+
+
+cdef extern from "pstring.h":
+    wstring utf8_to_utf32(string s)
+
+
+cdef extern from "clstm.h":
+    cdef double levenshtein[A, B](A a, B b)
+
+
+cdef extern from "clstm.h" namespace "ocropus":
+    cdef cppclass Assoc:
+        Assoc()
+        Assoc(string &s)
+        bint contains(string &key, bint parent = true)
+        string get(string &key)
+        string get(string &key, string default)
+        void set(string &key, string value)
+
+    cdef cppclass INetwork:
+        Assoc attr
+
+    ctypedef shared_ptr[INetwork] Network
+
+
+cdef extern from "tensor.h" namespace "ocropus":
+    cppclass TensorMap2:
+        pass
+
+    cdef cppclass Tensor2:
+        int dims[2]
+        float *ptr
+        void resize(int i, int j)
+        void put(float val, int i, int j)
+        float get(int i, int j)
+        TensorMap2 map()
+
+cdef extern from "clstmhl.h" namespace "ocropus":
+    struct CharPrediction:
+        int i
+        int x
+        wchar_t c
+        float p
+
+    # NOTE: The content of `codec` should be the utf-32 characters that the
+    #       network is supposed to learn, encoded as integers
+    cppclass CLSTMOCR:
+        int target_height
+        Network net
+        bint maybe_load(string &fname)
+        bint maybe_save(string &fname)
+        void createBidi(vector[int] codec, int nhidden)
+        void setLearningRate(float learning_rate, float momentum)
+        string train_utf8(TensorMap2 imgdata, string &target)
+        string predict_utf8(TensorMap2 imgdata)
+        void predict(vector[CharPrediction] &preds, TensorMap2 imgdata)
+        string aligned_utf8()
diff --git a/clstm.i b/clstm.i
deleted file mode 100644
index 9fb45ce..0000000
--- a/clstm.i
+++ /dev/null
@@ -1,304 +0,0 @@
-// -*- C++ -*-
-
-%{
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#pragma GCC diagnostic ignored "-Wuninitialized"
-#pragma GCC diagnostic ignored "-Wunused-but-set-variable"
-#define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION
-%}
-
-%module(docstring="C-version of the ocropy LSTM implementation") clstm;
-%feature("autodoc",1);
-%include "typemaps.i"
-%include "std_string.i"
-%include "std_wstring.i"
-%include "std_shared_ptr.i"
-%include "std_vector.i"
-%shared_ptr(ITrainable)
-%shared_ptr(INetwork)
-#ifdef SWIGPYTHON
-%include "cstring.i"
-#endif
-
-%{
-#include <memory>
-#include <iostream>
-#include "clstm.h"
-#include "clstm_compute.h"
-using namespace ocropus;
-using namespace std;
-%}
-
-typedef float Float;
-using std::string;
-
-#ifdef SWIGPYTHON
-%exception {
-    try {
-        $action
-    }
-    catch(const char *s) {
-        PyErr_SetString(PyExc_IndexError,s);
-        return NULL;
-    }
-    catch(...) {
-        PyErr_SetString(PyExc_IndexError,"unknown exception in iulib");
-        return NULL;
-    }
-}
-#endif
-
-%{
-#include "numpy/arrayobject.h"
-%}
-
-%init %{
-import_array();
-%}
-
-/* create simple interface definitions for the built-in Sequence types */
-
-struct Classes {
-    Classes();
-    ~Classes();
-    %rename(__getitem__) operator[];
-    int operator[](int i);
-    int size();
-    void resize(int);
-};
-%extend Classes {
-    void __setitem__(int i,int value) {
-        (*$self)[i] = value;
-    }
-}
-
-struct Batch {
-  void resize(int,int);
-  void setZero(int,int);
-  int rows();
-  int cols();
-  float &v(int,int);
-  float &d(int,int);
-};
-
-struct Params {
-  void resize(int,int);
-  void setZero(int,int);
-  int rows();
-  int cols();
-  float &v(int,int);
-  float &d(int,int);
-};
-
-
-struct Sequence {
-    Sequence();
-    ~Sequence();
-    int size();
-    int rows();
-    int cols();
-    %rename(__getitem__) operator[];
-    Batch &operator[](int i);
-};
-
-struct Assoc {
-  string get(string key);
-  string get(string key, string dflt);
-  void set(string key, string dflt);
-};
-
-struct Codec {
-  std::vector<int> codec;
-  int size() { return codec.size(); }
-  void set(const vector<int> &data);
-  wchar_t decode(int cls);
-  std::wstring decode(Classes &cs);
-  void encode(Classes &cs, const std::wstring &s);
-private:
-  void operator=(const Codec &);
-};
-
-struct INetwork;
-typedef std::shared_ptr<INetwork> Network;
-%template(vectornet) std::vector<std::shared_ptr<INetwork> >;
-
-struct INetwork {
-    string kind;
-    Assoc attr;
-    virtual void setLearningRate(Float lr, Float momentum) = 0;
-    virtual void forward() = 0;
-    virtual void backward() = 0;
-    virtual void initialize();
-    virtual ~INetwork();
-    Sequence inputs;
-    Sequence outputs;
-    std::vector<std::shared_ptr<INetwork> > sub;
-    Codec codec;
-    Codec icodec;
-    virtual int ninput();
-    virtual int noutput();
-    virtual void add(std::shared_ptr<INetwork> net);
-};
-
-void sgd_update(Network net);
-void set_inputs(Network net, Sequence &inputs);
-void set_targets(Network net, Sequence &targets);
-void set_classes(Network net, Classes &classes);
-void mktargets(Sequence &seq, Classes &targets, int ndim);
-
-std::shared_ptr<INetwork> make_layer(string);
-std::shared_ptr<INetwork> make_net_init(string,string);
-
-#if 0
-%rename(seq_forward) forward_algorithm;
-void forward_algorithm(Mat &lr,Mat &lmatch,double skip=-5.0);
-%rename(seq_forwardbackward) forwardbackward;
-void forwardbackward(Mat &both,Mat &lmatch);
-#endif
-
-%rename(seq_ctc_align) ctc_align_targets;
-void ctc_align_targets(Sequence &posteriors,Sequence &outputs,Sequence &targets);
-void mktargets(Sequence &seq, Classes &targets, int ndim);
-
-void save_net(const string &file, Network net);
-Network load_net(const string &file);
-
-%rename(network_info) network_info_as_strings;
-
-%inline %{
-int string_edit_distance(string a, string b) {
-    return levenshtein(a, b);
-}
-
-string network_info_as_string(Network net) {
-    string result = "";
-    walk_networks(net, [&result] (string s, INetwork *net) {
-        double lr = net->attr.get("learning_rate","-1");
-        double momentum = net->attr.get("momentum","-1");
-        result += s + ": " + to_string(lr);
-        result += string(" ") + to_string(momentum);
-        result += string(" ") + to_string(net->ninput());
-        result += string(" ") + to_string(net->noutput());
-        result += "\n";
-    });
-    return result;
-}
-
-string sequence_info(Sequence &seq) {
-    string result = "";
-    result += to_string(seq.size());
-    result += string(":") + (seq.size()>0?to_string(seq[0].rows()):"*");
-    result += string(":") + (seq.size()>0?to_string(seq[0].cols()):"*");
-#if 0
-    // FIXME
-    double lo = 1e99, hi = -1e99;
-    for (int t=0;t<seq.size(); t++) {
-        lo = fmin(lo, minimum(seq[t].V()));
-        hi = fmax(hi, maximum(seq[t].V()));
-    }
-    result += "[" + to_string(lo) + "," + to_string(hi) + "]";
-#endif
-    return result;
-}
-
-%}
-
-#ifdef SWIGPYTHON
-%{
-#include "numpyarray.h"
-%}
-%inline %{
-void sequence_of_array(Sequence &a,PyObject *object_) {
-    npa_float np(object_);
-    if(np.rank()!=3) throw "rank must be 3";
-    int N = np.dim(0);
-    int d = np.dim(1);
-    int bs = np.dim(2);
-    a.resize(N,d,bs);
-    for(int t=0;t<N;t++) {
-        for(int i=0; i<d; i++)
-            for(int b=0; b<bs; b++)
-                a[t].v(i,b) = np(t,i,b);
-    }
-}
-
-void d_sequence_of_array(Sequence &a,PyObject *object_) {
-    npa_float np(object_);
-    if(np.rank()!=3) throw "rank must be 3";
-    int N = np.dim(0);
-    int d = np.dim(1);
-    int bs = np.dim(2);
-    if (a.size() != N) throw "size mismatch";
-    for(int t=0;t<N;t++) {
-        for(int i=0; i<d; i++)
-            for(int b=0; b<bs; b++)
-                a[t].d(i,b) = np(t,i,b);
-    }
-}
-
-void array_of_sequence(PyObject *object_,Sequence &a) {
-    npa_float np(object_);
-    int N = a.size();
-    if (N==0) throw "empty sequence";
-    int d = a[0].rows();
-    if (d==0) throw "empty feature vector";
-    int bs = a[0].cols();
-    if (bs==0) throw "empty batch";
-    np.resize(N,d,bs);
-    for(int t=0; t<N; t++) {
-        for(int i=0; i<d; i++)
-            for(int b=0; b<bs; b++)
-                np(t,i,b) = a[t].v(i,b);
-    }
-}
-
-void array_of_d_sequence(PyObject *object_,Sequence &a) {
-    npa_float np(object_);
-    int N = a.size();
-    if (N==0) throw "empty sequence";
-    int d = a[0].rows();
-    if (d==0) throw "empty feature vector";
-    int bs = a[0].cols();
-    if (bs==0) throw "empty batch";
-    np.resize(N,d,bs);
-    for(int t=0; t<N; t++) {
-        for(int i=0; i<d; i++)
-            for(int b=0; b<bs; b++)
-                np(t,i,b) = a[t].d(i,b);
-    }
-}
-%}
-
-%pythoncode %{
-import numpy
-
-def Sequence_array(self):
-    a = numpy.zeros(1,'f')
-    array_of_sequence(a, self)
-    return a
-Sequence.array = Sequence_array
-
-def Sequence_aset(self, a):
-    sequence_of_array(self, a)
-Sequence.aset = Sequence_aset
-
-def Sequence_darray(self):
-    a = numpy.zeros(1,'f')
-    array_of_d_sequence(a, self)
-    return a
-Sequence.darray = Sequence_darray
-
-def Sequence_dset(self, a):
-    d_sequence_of_array(self, a)
-Sequence.dset = Sequence_dset
-
-def ctcalign(outputs_,targets_):
-    outputs = Sequence()
-    targets = Sequence()
-    outputs.aset(outputs_)
-    targets.aset(targets_)
-    posteriors = Sequence()
-    seq_ctc_align(posteriors,outputs,targets)
-    return posteriors.array()
-%}
-#endif
diff --git a/clstmocrtrain.cc b/clstmocrtrain.cc
index eb31249..f59b99a 100644
--- a/clstmocrtrain.cc
+++ b/clstmocrtrain.cc
@@ -164,8 +164,8 @@ int main1(int argc, char **argv) {
 
     if (test_trigger(trial)) {
       auto tse = test_set_error(clstm, testset);
-      double errors = tse.first;
-      double count = tse.second;
+      double count = tse.first;
+      double errors = tse.second;
       test_error = errors / count;
       print("ERROR", trial, test_error, "   ", errors, count);
       if (test_error < best_error) {
diff --git a/pyclstm.pyx b/pyclstm.pyx
new file mode 100644
index 0000000..3a8633a
--- /dev/null
+++ b/pyclstm.pyx
@@ -0,0 +1,238 @@
+from cpython.version cimport PY_MAJOR_VERSION
+from cpython.ref cimport PyObject
+from libc.stddef cimport wchar_t
+from libc.stdlib cimport malloc, free
+from libcpp.vector cimport vector
+from libcpp.set cimport set
+from libcpp.string cimport string
+
+cimport _clstm
+
+from collections import namedtuple
+
+
+CharPrediction = namedtuple("CharPrediction",
+                            ("x_position", "char", "confidence"))
+
+
+cdef extern from "Python.h":
+    ctypedef PyObject PyUnicodeObject
+    Py_ssize_t PyUnicode_AsWideChar(PyUnicodeObject *o, wchar_t *w,
+                                    Py_ssize_t size)
+    PyObject* PyUnicode_FromWideChar(wchar_t *w, Py_ssize_t size)
+
+
+cpdef double levenshtein(unicode a, unicode b):
+    """ Determine the Levenshtein-distance between two unicode strings.
+
+    :type a:    unicode
+    :type b:    unicode
+    :rtype:     int
+    """
+    return _clstm.levenshtein[string, string](
+        a.encode('utf8'), b.encode('utf8'))
+
+
+cdef safe_str(in_str):
+    cdef bint needs_encode = (
+        (PY_MAJOR_VERSION < 3 and not isinstance(in_str, str)) or
+        not isinstance(in_str, bytes))
+    if needs_encode:
+        return in_str.encode('utf8')
+    else:
+        return in_str
+
+
+cdef load_img(img, _clstm.Tensor2 *data):
+    """ Copy image data from a PIL image to an Eigen tensor.
+
+    :param img:     Image
+    :type img:      :py:class:`PIL.Image.Image`
+    """
+    if hasattr(img, 'width'):
+        width, height = img.width, img.height
+    elif hasattr(img, 'size'):
+        width, height = img.size
+    data.resize(width, height)
+    imgdata = img.load()
+    for i in range(width):
+        for j in range(height):
+            px = imgdata[i, j]
+            # Pillow returns pixels as [0, 255], but we need [0, 1]
+            if isinstance(px, tuple):
+                # For color images, we use the mean across all channels
+                px = (sum(px)/len(px))/255.
+            else:
+                px = px/255.
+            px = -px + 1.
+            data[0].put(px, i, j)
+
+
+cdef load_nparray(npdata, _clstm.Tensor2 *data):
+    """ Copy image data from a numpy array to an Eigen tensor.
+
+    This assumes that all pixels are floating point numbers in the range [0, 1]
+    where 0 is white and 1 is black.
+
+    Currently this copies all data, but it should be possible to make Eigen
+    use the numpy-allocated data.
+    """
+    data.resize(npdata.shape[0], npdata.shape[1])
+    for i in range(npdata.shape[0]):
+        for j in range(npdata.shape[1]):
+            data[0].put(npdata[i,j], i, j)
+
+
+cdef class ClstmOcr:
+    """ An OCR engine based on CLSTM, operating on line images.
+
+    Use this class to either train your own OCR model or to load a
+    pre-trained model from disk.
+
+    For training, set your parameters with :py:meth:`prepare_training`, and
+    then iteratively supply a line image (:py:class:`PIL.Image.Image` or
+    :py:class:`numpy.ndarray`) and the ground truth for the line to
+    :py:meth:`train`. Once finished with training, call :py:meth:`save`
+    to persist the trained model to disk.
+
+    For prediction, two methods are available. The simplest,
+    :py:meth:`recognize` takes a line image (see above) and returns the
+    recognized text as a string. If more information about the recognized
+    text is needed, use :py:meth:`recognize_chars`, which returns a generator
+    that yields :py:class:`CharPrediction` objects that contain information
+    about each character (x-offset, confidence and recognized character).
+    """
+    cdef _clstm.CLSTMOCR *_ocr
+
+    def __cinit__(self, fname=None):
+        """ Initialize the OCR engine, optionally loading a model from disk.
+
+        :param fname:   Path to pre-trained model on disk
+        :type fname:    str/unicode/bytes
+        """
+        self._ocr = new _clstm.CLSTMOCR()
+        if fname:
+            fname = safe_str(fname)
+            self.load(fname)
+
+    cpdef load(self, fname):
+        """ Load a pre-trained model from disk.
+
+        :param fname:   Path to pre-trained model on disk
+        :type fname:    str/unicode/bytes
+        """
+        fname = safe_str(fname)
+        cdef bint rv = self._ocr.maybe_load(fname)
+        if not rv:
+            raise IOError("Could not load model from {}".format(fname))
+
+    cpdef save(self, fname):
+        """ Save the model to disk.
+
+        :param fname:   Path to store model in
+        :type fname:    str/unicode/bytes
+        """
+        fname = safe_str(fname)
+        cdef bint rv = self._ocr.maybe_save(fname)
+        if not rv:
+            raise IOError("Could not save model to {}".format(fname))
+
+    cpdef prepare_training(self, graphemes, int num_hidden=100,
+                           float learning_rate=0.0001, float momentum=0.9):
+        """ Prepare training by setting the graphemes and hyperparameters.
+
+        :param graphemes:   Iterable of graphemes that are to be recognized
+                            by the OCR model, must not have duplicates
+        :type graphemes:    iterable of str/unicode
+        :param num_hidden:  Number of hidden units in the LSTM layers, larger
+                            values require more storage/memory and take longer
+                            for training and recognition, so try to find
+                            a good performance/cost tradeoff.
+        :type num_hidden:   int
+        :param learning_rate:   Learning rate for the model training
+        :type learning_rate:    float
+        :param momentum:        Momentum for the model training
+        :type momentum:         float
+        """
+        graphemes_str = u"".join(sorted(graphemes))
+        cdef vector[int] codec
+        cdef Py_ssize_t length = len(graphemes_str)
+        cdef wchar_t *wchars = <wchar_t *>malloc(length * sizeof(wchar_t))
+        cdef Py_ssize_t number_written = _clstm.Unicode_AsWideChar(
+            <PyObject*>graphemes_str, length, wchars)
+        codec.push_back(0)
+        for i in range(length-1):
+            codec.push_back(<int>(wchars[i]))
+        self._ocr.createBidi(codec, num_hidden)
+        self._ocr.setLearningRate(learning_rate, momentum)
+
+    def aligned(self):
+        """ Get the aligned output of the last trained sample.
+
+        :rtype:     unicode
+        """
+        return self._ocr.aligned_utf8().decode('utf8')
+
+    def train(self, img, unicode text):
+        """ Train the model with a line image and its ground truth.
+
+        :param img:     The line image for the ground truth
+        :type img:      :py:class:`PIL.Image.Image`/:py:class:`numpy.ndarray`
+        :param text:    The ground truth text for the line image
+        :type text:     unicode/str
+        :returns:       The recognized text for the line image, can be used
+                        to estimate error against the ground truth
+                        (via :py:func:`levenshtein`)
+        :rtype:         unicode
+        """
+        cdef _clstm.Tensor2 data
+        if hasattr(img, 'shape'):
+            load_nparray(img, &data)
+        else:
+            load_img(img, &data)
+        return self._ocr.train_utf8(
+            data.map(), text.encode('utf8')).decode('utf8')
+
+    def recognize(self, img):
+        """ Recognize the text on the line image.
+
+        :param img:     The line image for the ground truth
+        :type img:      :py:class:`PIL.Image.Image`/:py:class:`numpy.ndarray`
+        :returns:       The recognized text for the line
+        :rtype:         unicode
+        """
+        cdef _clstm.Tensor2 data
+        if hasattr(img, 'shape'):
+            load_nparray(img, &data)
+        else:
+            load_img(img, &data)
+        return self._ocr.predict_utf8(data.map()).decode('utf8')
+
+    def recognize_chars(self, img):
+        """ Recognize the characters on the line, along with their position
+            and confidence.
+
+        :param img:     The line image for the ground truth
+        :type img:      :py:class:`PIL.Image.Image`/:py:class:`numpy.ndarray`
+        :returns:       The recognized text for the line, represented as
+                        information about its composing characters.
+        :rtype:         generator that yield :py:class:`CharPrediction`
+        """
+        cdef _clstm.Tensor2 data
+        cdef vector[_clstm.CharPrediction] preds
+        cdef vector[_clstm.CharPrediction].iterator pred_it
+        cdef wchar_t[2] cur_char
+        if hasattr(img, 'shape'):
+            load_nparray(img, &data)
+        else:
+            load_img(img, &data)
+        self._ocr.predict(preds, data.map())
+        for i in range(preds.size()):
+            cur_char[0] = preds[i].c
+            yield CharPrediction(
+                preds[i].x,
+                <unicode>PyUnicode_FromWideChar(cur_char, 1),
+                preds[i].p)
+
+    def __dealloc__(self):
+        del self._ocr
diff --git a/pyclstm_cli/__init__.py b/pyclstm_cli/__init__.py
new file mode 100644
index 0000000..fdff37f
--- /dev/null
+++ b/pyclstm_cli/__init__.py
@@ -0,0 +1 @@
+from .train import OcrTrainer
diff --git a/pyclstm_cli/train.py b/pyclstm_cli/train.py
new file mode 100644
index 0000000..3a699dc
--- /dev/null
+++ b/pyclstm_cli/train.py
@@ -0,0 +1,219 @@
+from __future__ import division
+
+from itertools import chain
+import argparse
+import glob
+import logging
+import os
+import random
+import re
+import sys
+
+import pyclstm
+from PIL import Image
+
+logging.basicConfig(level=logging.INFO, format="%(asctime)s\t%(message)s")
+
+class OcrTrainer:
+
+    def __init__(self,
+                 imgdir,
+                 gtdir=None,
+                 model_name=None,
+                 train_ratio=.8,
+                 max_iterations=2000000,
+                 num_hidden=50,
+                 learning_rate=0.0001,
+                 momentum=0.9):
+        """Initialize a trainer.
+
+        :param imgdir:          Basedirectory containing '*/*.bin.png' images
+        :type imgdir:           str/unicode
+        :param max_iterations:  Maximum iterations
+        :type max_iterations:   int
+        :param gtdir:           Base directory containing '*/*.gt.txt' transcriptions
+        :type imgdir:           str/unicode
+        :param train_ratio:     Ratio of ground truth touse for training
+        :type train_ratio:      float
+        :param model_name:      Model basename. 'foo' => 'foo.clstm' or 'foo-1000.clstm'
+        :type model_name:       str/unicode
+        :param num_hidden:      See pyclstm.pyx
+        :param learning_rate:   See pyclstm.pyx
+        :param momentum:        See pyclstm.pyx
+        """
+        if not gtdir:
+            gtdir = imgdir
+        self.model_name = model_name if model_name else re.sub(r"[^A-Za-z0-9_-]", "", imgdir)
+        self.imgdir = imgdir
+        all_imgs = [Image.open(p)
+                    for p in sorted(glob.glob("{0}/*/*.bin.png".format(imgdir)))
+                   ]
+        all_texts = [open(p).read().strip()
+                     for p in sorted(glob.glob("{0}/*/*.gt.txt".format(gtdir)))]
+        if sys.version_info <= (3,):
+            all_texts = [t.decode('utf8') for t in all_texts]
+        all_data = list(zip(all_imgs, all_texts))
+        nr_train = int(len(all_data) * train_ratio)
+        self.train_data = all_data[:nr_train]
+        self.test_data = all_data[nr_train:]
+        self.graphemes = set(chain.from_iterable(all_texts))
+        self.ocr = pyclstm.ClstmOcr()
+        self.max_iterations = max_iterations
+        self.num_hidden = num_hidden
+        self.learning_rate = learning_rate
+        self.momentum = momentum
+
+    def save_model(self, suffix=''):
+        """
+        Save the model
+
+        :param suffix: Optional suffix to save intermediary results.
+        :param suffix: str/unicode
+        """
+        fname = self.model_name
+        if suffix != '': fname += '-' + suffix
+        fname += '.clstm'
+        logging.info("Saving model to {}".format(fname))
+        self.ocr.save(fname)
+
+    def calculate_error(self):
+        """
+        Calculate error rate by recognizing test_data and comparing by
+        levenshtein distance
+        """
+        logging.info("Calculating error")
+        errors = 0
+        chars = 0
+        for img, txt in self.test_data:
+            out = self.ocr.recognize(img)
+            errors += pyclstm.levenshtein(txt, out)
+            chars += len(txt)
+        return errors / chars
+
+    def train(self,
+              log_out=True,
+              log_every=10,
+              save_every=-1,
+              calculate_every=1000):
+        """Train from OCR ground truth data.
+
+        :param log_out:         Whether to log 'OUT'
+        :param log_every:       Log 'TRN'/'ALN' every N iterations
+        :param save_every:      Save every N iterations with suffix -N.clst
+        :param calculate_every: Calculate error rate every N iterations and save
+        """
+        logprog = lambda i, err, msg: logging.info("[{:6d} / {:2.2f}%] {:s}".format(i, (err) * 100, msg))
+        self.ocr.prepare_training(
+            self.graphemes,
+            num_hidden=self.num_hidden,
+            learning_rate=self.learning_rate,
+            momentum=self.momentum)
+
+        logging.info("Learning from {} [n-train={},n-test={}]".format(
+            self.imgdir, len(self.train_data), len(self.test_data)))
+        if save_every > -1: logging.info("Saving every {}th iteration".format(save_every))
+        logging.info("Evaluating every {}th iteration".format(calculate_every))
+        logging.info("Logging everyy {}th iteration".format(log_every))
+        err = 1.
+        for i in range(self.max_iterations):
+            img, txt = random.choice(self.train_data)
+            out = self.ocr.train(img, txt)
+            if i == 0: continue
+            if not i % log_every:
+                aligned = self.ocr.aligned()
+                logprog(i, err, "TRN: {}".format(txt))
+                logprog(i, err, "ALN: {}".format(aligned))
+                if log_out: logprog(i, err, "OUT: {}".format(out))
+            if not i % calculate_every:
+                cur_err = self.calculate_error()
+                if cur_err != err:
+                    diff = 100*(err-cur_err)
+                    logging.info("=== {}ed by {:.2f}%".format("Improv" if diff>0 else "Degrad", diff))
+                    self.save_model()
+                    err = cur_err
+                else:
+                    logging.info("No change in error rate")
+            elif save_every > 0 and not i % save_every:
+                self.save_model(suffix="{:05d}".format(i))
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('imgdir', help='Directory matching "*/*.bin.png"')
+    parser.add_argument(
+        '--model-name', help='Name of the model file without ".clstm". Default: Generated from input name')
+    parser.add_argument(
+        '--gtdir', help='Directory matching "*/*.gt.txt". Default: imgdir')
+    parser.add_argument(
+        '--learning-rate',
+        metavar='RATE',
+        type=float,
+        default=0.0001,
+        help="Learning rate for the model training. Default: %(default)s")
+    parser.add_argument(
+        '--momentum',
+        type=float,
+        default=0.9,
+        help="Momentum for the model training. Default: %(default)s")
+    parser.add_argument(
+        '--num-hidden',
+        metavar='NUM',
+        type=int,
+        default=100,
+        help="Number of hidden units in the LSTM layers, larger "
+        "values require more storage/memory and take longer "
+        "for training and recognition, so try to find "
+        "a good performance/cost tradeoff. Default: %(default)s")
+    parser.add_argument(
+        '--max-iterations',
+        metavar='N',
+        type=int,
+        default=2000000,
+        help="Maximum iterations. Default: %(default)s")
+    parser.add_argument(
+        '--train-ratio',
+        metavar='RATIO',
+        type=float,
+        default=0.8,
+        help="Ratio of ground truth to be used for training. Default: %(default)s")
+    parser.add_argument(
+        '--log-out',
+        action='store_true',
+        default=False,
+        help="Log 'OUT' messages. Default: %(default)s")
+    parser.add_argument(
+        '--log-every',
+        type=int,
+        metavar='N',
+        default=10,
+        help="Log on every N-th iteration. Default: %(default)s")
+    parser.add_argument(
+        '--save-every',
+        type=int,
+        metavar='N',
+        default=500,
+        help="Save on every N-th iteration. Default: %(default)s")
+    parser.add_argument(
+        '--calculate-every',
+        type=int,
+        metavar='N',
+        default=1000,
+        help="Calculate error rate every N-th iteration. Default: %(default)s")
+    #  imgdir = '/home/kb/build/github.com/tmbdev/clstm/book'
+    args = parser.parse_args()
+    trainer = OcrTrainer(
+        args.imgdir,
+        gtdir=args.gtdir,
+        model_name=args.model_name,
+        train_ratio=args.train_ratio,
+        max_iterations=args.max_iterations,
+        num_hidden=args.num_hidden,
+        learning_rate=args.learning_rate,
+        momentum=args.momentum)
+    trainer.train(
+        log_out=args.log_out,
+        log_every=args.log_every,
+        save_every=args.save_every,
+        calculate_every=args.calculate_every)
+
+if __name__ == '__main__':
+    main()
diff --git a/pydoc/Makefile b/pydoc/Makefile
new file mode 100644
index 0000000..4f4eedf
--- /dev/null
+++ b/pydoc/Makefile
@@ -0,0 +1,225 @@
+# Makefile for Sphinx documentation
+#
+
+# You can set these variables from the command line.
+SPHINXOPTS    =
+SPHINXBUILD   = sphinx-build
+PAPER         =
+BUILDDIR      = _build
+
+# Internal variables.
+PAPEROPT_a4     = -D latex_paper_size=a4
+PAPEROPT_letter = -D latex_paper_size=letter
+ALLSPHINXOPTS   = -d $(BUILDDIR)/doctrees $(PAPEROPT_$(PAPER)) $(SPHINXOPTS) .
+# the i18n builder cannot share the environment and doctrees with the others
+I18NSPHINXOPTS  = $(PAPEROPT_$(PAPER)) $(SPHINXOPTS) .
+
+.PHONY: help
+help:
+	@echo "Please use \`make <target>' where <target> is one of"
+	@echo "  html       to make standalone HTML files"
+	@echo "  dirhtml    to make HTML files named index.html in directories"
+	@echo "  singlehtml to make a single large HTML file"
+	@echo "  pickle     to make pickle files"
+	@echo "  json       to make JSON files"
+	@echo "  htmlhelp   to make HTML files and a HTML help project"
+	@echo "  qthelp     to make HTML files and a qthelp project"
+	@echo "  applehelp  to make an Apple Help Book"
+	@echo "  devhelp    to make HTML files and a Devhelp project"
+	@echo "  epub       to make an epub"
+	@echo "  epub3      to make an epub3"
+	@echo "  latex      to make LaTeX files, you can set PAPER=a4 or PAPER=letter"
+	@echo "  latexpdf   to make LaTeX files and run them through pdflatex"
+	@echo "  latexpdfja to make LaTeX files and run them through platex/dvipdfmx"
+	@echo "  text       to make text files"
+	@echo "  man        to make manual pages"
+	@echo "  texinfo    to make Texinfo files"
+	@echo "  info       to make Texinfo files and run them through makeinfo"
+	@echo "  gettext    to make PO message catalogs"
+	@echo "  changes    to make an overview of all changed/added/deprecated items"
+	@echo "  xml        to make Docutils-native XML files"
+	@echo "  pseudoxml  to make pseudoxml-XML files for display purposes"
+	@echo "  linkcheck  to check all external links for integrity"
+	@echo "  doctest    to run all doctests embedded in the documentation (if enabled)"
+	@echo "  coverage   to run coverage check of the documentation (if enabled)"
+	@echo "  dummy      to check syntax errors of document sources"
+
+.PHONY: clean
+clean:
+	rm -rf $(BUILDDIR)/*
+
+.PHONY: html
+html:
+	$(SPHINXBUILD) -b html $(ALLSPHINXOPTS) $(BUILDDIR)/html
+	@echo
+	@echo "Build finished. The HTML pages are in $(BUILDDIR)/html."
+
+.PHONY: dirhtml
+dirhtml:
+	$(SPHINXBUILD) -b dirhtml $(ALLSPHINXOPTS) $(BUILDDIR)/dirhtml
+	@echo
+	@echo "Build finished. The HTML pages are in $(BUILDDIR)/dirhtml."
+
+.PHONY: singlehtml
+singlehtml:
+	$(SPHINXBUILD) -b singlehtml $(ALLSPHINXOPTS) $(BUILDDIR)/singlehtml
+	@echo
+	@echo "Build finished. The HTML page is in $(BUILDDIR)/singlehtml."
+
+.PHONY: pickle
+pickle:
+	$(SPHINXBUILD) -b pickle $(ALLSPHINXOPTS) $(BUILDDIR)/pickle
+	@echo
+	@echo "Build finished; now you can process the pickle files."
+
+.PHONY: json
+json:
+	$(SPHINXBUILD) -b json $(ALLSPHINXOPTS) $(BUILDDIR)/json
+	@echo
+	@echo "Build finished; now you can process the JSON files."
+
+.PHONY: htmlhelp
+htmlhelp:
+	$(SPHINXBUILD) -b htmlhelp $(ALLSPHINXOPTS) $(BUILDDIR)/htmlhelp
+	@echo
+	@echo "Build finished; now you can run HTML Help Workshop with the" \
+	      ".hhp project file in $(BUILDDIR)/htmlhelp."
+
+.PHONY: qthelp
+qthelp:
+	$(SPHINXBUILD) -b qthelp $(ALLSPHINXOPTS) $(BUILDDIR)/qthelp
+	@echo
+	@echo "Build finished; now you can run "qcollectiongenerator" with the" \
+	      ".qhcp project file in $(BUILDDIR)/qthelp, like this:"
+	@echo "# qcollectiongenerator $(BUILDDIR)/qthelp/pyclstm.qhcp"
+	@echo "To view the help file:"
+	@echo "# assistant -collectionFile $(BUILDDIR)/qthelp/pyclstm.qhc"
+
+.PHONY: applehelp
+applehelp:
+	$(SPHINXBUILD) -b applehelp $(ALLSPHINXOPTS) $(BUILDDIR)/applehelp
+	@echo
+	@echo "Build finished. The help book is in $(BUILDDIR)/applehelp."
+	@echo "N.B. You won't be able to view it unless you put it in" \
+	      "~/Library/Documentation/Help or install it in your application" \
+	      "bundle."
+
+.PHONY: devhelp
+devhelp:
+	$(SPHINXBUILD) -b devhelp $(ALLSPHINXOPTS) $(BUILDDIR)/devhelp
+	@echo
+	@echo "Build finished."
+	@echo "To view the help file:"
+	@echo "# mkdir -p $$HOME/.local/share/devhelp/pyclstm"
+	@echo "# ln -s $(BUILDDIR)/devhelp $$HOME/.local/share/devhelp/pyclstm"
+	@echo "# devhelp"
+
+.PHONY: epub
+epub:
+	$(SPHINXBUILD) -b epub $(ALLSPHINXOPTS) $(BUILDDIR)/epub
+	@echo
+	@echo "Build finished. The epub file is in $(BUILDDIR)/epub."
+
+.PHONY: epub3
+epub3:
+	$(SPHINXBUILD) -b epub3 $(ALLSPHINXOPTS) $(BUILDDIR)/epub3
+	@echo
+	@echo "Build finished. The epub3 file is in $(BUILDDIR)/epub3."
+
+.PHONY: latex
+latex:
+	$(SPHINXBUILD) -b latex $(ALLSPHINXOPTS) $(BUILDDIR)/latex
+	@echo
+	@echo "Build finished; the LaTeX files are in $(BUILDDIR)/latex."
+	@echo "Run \`make' in that directory to run these through (pdf)latex" \
+	      "(use \`make latexpdf' here to do that automatically)."
+
+.PHONY: latexpdf
+latexpdf:
+	$(SPHINXBUILD) -b latex $(ALLSPHINXOPTS) $(BUILDDIR)/latex
+	@echo "Running LaTeX files through pdflatex..."
+	$(MAKE) -C $(BUILDDIR)/latex all-pdf
+	@echo "pdflatex finished; the PDF files are in $(BUILDDIR)/latex."
+
+.PHONY: latexpdfja
+latexpdfja:
+	$(SPHINXBUILD) -b latex $(ALLSPHINXOPTS) $(BUILDDIR)/latex
+	@echo "Running LaTeX files through platex and dvipdfmx..."
+	$(MAKE) -C $(BUILDDIR)/latex all-pdf-ja
+	@echo "pdflatex finished; the PDF files are in $(BUILDDIR)/latex."
+
+.PHONY: text
+text:
+	$(SPHINXBUILD) -b text $(ALLSPHINXOPTS) $(BUILDDIR)/text
+	@echo
+	@echo "Build finished. The text files are in $(BUILDDIR)/text."
+
+.PHONY: man
+man:
+	$(SPHINXBUILD) -b man $(ALLSPHINXOPTS) $(BUILDDIR)/man
+	@echo
+	@echo "Build finished. The manual pages are in $(BUILDDIR)/man."
+
+.PHONY: texinfo
+texinfo:
+	$(SPHINXBUILD) -b texinfo $(ALLSPHINXOPTS) $(BUILDDIR)/texinfo
+	@echo
+	@echo "Build finished. The Texinfo files are in $(BUILDDIR)/texinfo."
+	@echo "Run \`make' in that directory to run these through makeinfo" \
+	      "(use \`make info' here to do that automatically)."
+
+.PHONY: info
+info:
+	$(SPHINXBUILD) -b texinfo $(ALLSPHINXOPTS) $(BUILDDIR)/texinfo
+	@echo "Running Texinfo files through makeinfo..."
+	make -C $(BUILDDIR)/texinfo info
+	@echo "makeinfo finished; the Info files are in $(BUILDDIR)/texinfo."
+
+.PHONY: gettext
+gettext:
+	$(SPHINXBUILD) -b gettext $(I18NSPHINXOPTS) $(BUILDDIR)/locale
+	@echo
+	@echo "Build finished. The message catalogs are in $(BUILDDIR)/locale."
+
+.PHONY: changes
+changes:
+	$(SPHINXBUILD) -b changes $(ALLSPHINXOPTS) $(BUILDDIR)/changes
+	@echo
+	@echo "The overview file is in $(BUILDDIR)/changes."
+
+.PHONY: linkcheck
+linkcheck:
+	$(SPHINXBUILD) -b linkcheck $(ALLSPHINXOPTS) $(BUILDDIR)/linkcheck
+	@echo
+	@echo "Link check complete; look for any errors in the above output " \
+	      "or in $(BUILDDIR)/linkcheck/output.txt."
+
+.PHONY: doctest
+doctest:
+	$(SPHINXBUILD) -b doctest $(ALLSPHINXOPTS) $(BUILDDIR)/doctest
+	@echo "Testing of doctests in the sources finished, look at the " \
+	      "results in $(BUILDDIR)/doctest/output.txt."
+
+.PHONY: coverage
+coverage:
+	$(SPHINXBUILD) -b coverage $(ALLSPHINXOPTS) $(BUILDDIR)/coverage
+	@echo "Testing of coverage in the sources finished, look at the " \
+	      "results in $(BUILDDIR)/coverage/python.txt."
+
+.PHONY: xml
+xml:
+	$(SPHINXBUILD) -b xml $(ALLSPHINXOPTS) $(BUILDDIR)/xml
+	@echo
+	@echo "Build finished. The XML files are in $(BUILDDIR)/xml."
+
+.PHONY: pseudoxml
+pseudoxml:
+	$(SPHINXBUILD) -b pseudoxml $(ALLSPHINXOPTS) $(BUILDDIR)/pseudoxml
+	@echo
+	@echo "Build finished. The pseudo-XML files are in $(BUILDDIR)/pseudoxml."
+
+.PHONY: dummy
+dummy:
+	$(SPHINXBUILD) -b dummy $(ALLSPHINXOPTS) $(BUILDDIR)/dummy
+	@echo
+	@echo "Build finished. Dummy builder generates no files."
diff --git a/pydoc/conf.py b/pydoc/conf.py
new file mode 100644
index 0000000..3379c6a
--- /dev/null
+++ b/pydoc/conf.py
@@ -0,0 +1,348 @@
+# -*- coding: utf-8 -*-
+#
+# pyclstm documentation build configuration file, created by
+# sphinx-quickstart on Thu Oct  6 12:57:42 2016.
+#
+# This file is execfile()d with the current directory set to its
+# containing dir.
+#
+# Note that not all possible configuration values are present in this
+# autogenerated file.
+#
+# All configuration values have a default; values that are commented out
+# serve to show the default.
+
+# If extensions (or modules to document with autodoc) are in another directory,
+# add these directories to sys.path here. If the directory is relative to the
+# documentation root, use os.path.abspath to make it absolute, like shown here.
+#
+# import os
+# import sys
+# sys.path.insert(0, os.path.abspath('.'))
+import sphinx_rtd_theme
+
+# -- General configuration ------------------------------------------------
+
+# If your documentation needs a minimal Sphinx version, state it here.
+#
+# needs_sphinx = '1.0'
+
+# Add any Sphinx extension module names here, as strings. They can be
+# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
+# ones.
+extensions = [
+    'sphinx.ext.autodoc',
+    'sphinx.ext.intersphinx',
+]
+
+# Add any paths that contain templates here, relative to this directory.
+templates_path = ['_templates']
+
+# The suffix(es) of source filenames.
+# You can specify multiple suffix as a list of string:
+#
+# source_suffix = ['.rst', '.md']
+source_suffix = '.rst'
+
+# The encoding of source files.
+#
+# source_encoding = 'utf-8-sig'
+
+# The master toctree document.
+master_doc = 'index'
+
+# General information about the project.
+project = u'pyclstm'
+copyright = u'2016, Thomas Breuel, Johannes Baiter'
+author = u'Thomas Breuel, Johannes Baiter'
+
+# The version info for the project you're documenting, acts as replacement for
+# |version| and |release|, also used in various other places throughout the
+# built documents.
+#
+# The short X.Y version.
+version = u'0.1'
+# The full version, including alpha/beta/rc tags.
+release = u'0.1'
+
+# The language for content autogenerated by Sphinx. Refer to documentation
+# for a list of supported languages.
+#
+# This is also used if you do content translation via gettext catalogs.
+# Usually you set "language" from the command line for these cases.
+language = None
+
+# There are two options for replacing |today|: either, you set today to some
+# non-false value, then it is used:
+#
+# today = ''
+#
+# Else, today_fmt is used as the format for a strftime call.
+#
+# today_fmt = '%B %d, %Y'
+
+# List of patterns, relative to source directory, that match files and
+# directories to ignore when looking for source files.
+# This patterns also effect to html_static_path and html_extra_path
+exclude_patterns = ['_build', 'Thumbs.db', '.DS_Store']
+
+# The reST default role (used for this markup: `text`) to use for all
+# documents.
+#
+# default_role = None
+
+# If true, '()' will be appended to :func: etc. cross-reference text.
+#
+# add_function_parentheses = True
+
+# If true, the current module name will be prepended to all description
+# unit titles (such as .. function::).
+#
+# add_module_names = True
+
+# If true, sectionauthor and moduleauthor directives will be shown in the
+# output. They are ignored by default.
+#
+# show_authors = False
+
+# The name of the Pygments (syntax highlighting) style to use.
+pygments_style = 'sphinx'
+
+# A list of ignored prefixes for module index sorting.
+# modindex_common_prefix = []
+
+# If true, keep warnings as "system message" paragraphs in the built documents.
+# keep_warnings = False
+
+# If true, `todo` and `todoList` produce output, else they produce nothing.
+todo_include_todos = False
+
+
+# -- Options for HTML output ----------------------------------------------
+
+# The theme to use for HTML and HTML Help pages.  See the documentation for
+# a list of builtin themes.
+#
+html_theme = 'sphinx_rtd_theme'
+
+# Theme options are theme-specific and customize the look and feel of a theme
+# further.  For a list of options available for each theme, see the
+# documentation.
+#
+# html_theme_options = {}
+
+# Add any paths that contain custom themes here, relative to this directory.
+html_theme_path = [sphinx_rtd_theme.get_html_theme_path()]
+
+# The name for this set of Sphinx documents.
+# "<project> v<release> documentation" by default.
+#
+# html_title = u'pyclstm v0.1'
+
+# A shorter title for the navigation bar.  Default is the same as html_title.
+#
+# html_short_title = None
+
+# The name of an image file (relative to this directory) to place at the top
+# of the sidebar.
+#
+# html_logo = None
+
+# The name of an image file (relative to this directory) to use as a favicon of
+# the docs.  This file should be a Windows icon file (.ico) being 16x16 or 32x32
+# pixels large.
+#
+# html_favicon = None
+
+# Add any paths that contain custom static files (such as style sheets) here,
+# relative to this directory. They are copied after the builtin static files,
+# so a file named "default.css" will overwrite the builtin "default.css".
+html_static_path = ['_static']
+
+# Add any extra paths that contain custom files (such as robots.txt or
+# .htaccess) here, relative to this directory. These files are copied
+# directly to the root of the documentation.
+#
+# html_extra_path = []
+
+# If not None, a 'Last updated on:' timestamp is inserted at every page
+# bottom, using the given strftime format.
+# The empty string is equivalent to '%b %d, %Y'.
+#
+# html_last_updated_fmt = None
+
+# If true, SmartyPants will be used to convert quotes and dashes to
+# typographically correct entities.
+#
+# html_use_smartypants = True
+
+# Custom sidebar templates, maps document names to template names.
+#
+# html_sidebars = {}
+
+# Additional templates that should be rendered to pages, maps page names to
+# template names.
+#
+# html_additional_pages = {}
+
+# If false, no module index is generated.
+#
+# html_domain_indices = True
+
+# If false, no index is generated.
+#
+# html_use_index = True
+
+# If true, the index is split into individual pages for each letter.
+#
+# html_split_index = False
+
+# If true, links to the reST sources are added to the pages.
+#
+# html_show_sourcelink = True
+
+# If true, "Created using Sphinx" is shown in the HTML footer. Default is True.
+#
+# html_show_sphinx = True
+
+# If true, "(C) Copyright ..." is shown in the HTML footer. Default is True.
+#
+# html_show_copyright = True
+
+# If true, an OpenSearch description file will be output, and all pages will
+# contain a <link> tag referring to it.  The value of this option must be the
+# base URL from which the finished HTML is served.
+#
+# html_use_opensearch = ''
+
+# This is the file name suffix for HTML files (e.g. ".xhtml").
+# html_file_suffix = None
+
+# Language to be used for generating the HTML full-text search index.
+# Sphinx supports the following languages:
+#   'da', 'de', 'en', 'es', 'fi', 'fr', 'hu', 'it', 'ja'
+#   'nl', 'no', 'pt', 'ro', 'ru', 'sv', 'tr', 'zh'
+#
+# html_search_language = 'en'
+
+# A dictionary with options for the search language support, empty by default.
+# 'ja' uses this config value.
+# 'zh' user can custom change `jieba` dictionary path.
+#
+# html_search_options = {'type': 'default'}
+
+# The name of a javascript file (relative to the configuration directory) that
+# implements a search results scorer. If empty, the default will be used.
+#
+# html_search_scorer = 'scorer.js'
+
+# Output file base name for HTML help builder.
+htmlhelp_basename = 'pyclstmdoc'
+
+# -- Options for LaTeX output ---------------------------------------------
+
+latex_elements = {
+     # The paper size ('letterpaper' or 'a4paper').
+     #
+     # 'papersize': 'letterpaper',
+
+     # The font size ('10pt', '11pt' or '12pt').
+     #
+     # 'pointsize': '10pt',
+
+     # Additional stuff for the LaTeX preamble.
+     #
+     # 'preamble': '',
+
+     # Latex figure (float) alignment
+     #
+     # 'figure_align': 'htbp',
+}
+
+# Grouping the document tree into LaTeX files. List of tuples
+# (source start file, target name, title,
+#  author, documentclass [howto, manual, or own class]).
+latex_documents = [
+    (master_doc, 'pyclstm.tex', u'pyclstm Documentation',
+     u'Thomas Breuel, Johannes Baiter', 'manual'),
+]
+
+# The name of an image file (relative to this directory) to place at the top of
+# the title page.
+#
+# latex_logo = None
+
+# For "manual" documents, if this is true, then toplevel headings are parts,
+# not chapters.
+#
+# latex_use_parts = False
+
+# If true, show page references after internal links.
+#
+# latex_show_pagerefs = False
+
+# If true, show URL addresses after external links.
+#
+# latex_show_urls = False
+
+# Documents to append as an appendix to all manuals.
+#
+# latex_appendices = []
+
+# It false, will not define \strong, \code, 	itleref, \crossref ... but only
+# \sphinxstrong, ..., \sphinxtitleref, ... To help avoid clash with user added
+# packages.
+#
+# latex_keep_old_macro_names = True
+
+# If false, no module index is generated.
+#
+# latex_domain_indices = True
+
+
+# -- Options for manual page output ---------------------------------------
+
+# One entry per manual page. List of tuples
+# (source start file, name, description, authors, manual section).
+man_pages = [
+    (master_doc, 'pyclstm', u'pyclstm Documentation',
+     [author], 1)
+]
+
+# If true, show URL addresses after external links.
+#
+# man_show_urls = False
+
+
+# -- Options for Texinfo output -------------------------------------------
+
+# Grouping the document tree into Texinfo files. List of tuples
+# (source start file, target name, title, author,
+#  dir menu entry, description, category)
+texinfo_documents = [
+    (master_doc, 'pyclstm', u'pyclstm Documentation',
+     author, 'pyclstm', 'One line description of project.',
+     'Miscellaneous'),
+]
+
+# Documents to append as an appendix to all manuals.
+#
+# texinfo_appendices = []
+
+# If false, no module index is generated.
+#
+# texinfo_domain_indices = True
+
+# How to display URL addresses: 'footnote', 'no', or 'inline'.
+#
+# texinfo_show_urls = 'footnote'
+
+# If true, do not generate a @detailmenu in the "Top" node's menu.
+#
+# texinfo_no_detailmenu = False
+
+
+# Example configuration for intersphinx: refer to the Python standard library.
+intersphinx_mapping = {'https://docs.python.org/': None,
+                       'http://pillow.readthedocs.io/en/3.2.x/': None,
+                       'http://docs.scipy.org/doc/numpy/': None}
diff --git a/pydoc/index.rst b/pydoc/index.rst
new file mode 100644
index 0000000..7096414
--- /dev/null
+++ b/pydoc/index.rst
@@ -0,0 +1,69 @@
+.. pyclstm documentation master file, created by
+   sphinx-quickstart on Thu Oct  6 12:57:42 2016.
+   You can adapt this file completely to your liking, but it should at least
+   contain the root `toctree` directive.
+
+pyclstm
+=======
+
+Installation
+------------
+
+**Requirements:**
+    - A recent version of Eigen (>= 3.3) with development headers
+    - A C++ compiler (g++ is recommended)
+    - Cython (>= 0.23)
+
+**Installation:**
+
+.. code::
+
+    $ pip install git+https://github.com/jbaiter/clstm.git@cython
+
+
+Example Usage
+-------------
+
+**Training:**
+
+Refer to ``run_uw3_500.py`` in the root directory for a more comprehensive
+example.
+
+.. code::
+
+    import pyclstm
+    ocr = pyclstm.ClstmOcr()
+    ocr.prepare_training(
+        graphemes=graphemes,  # A list of characters the engine is supposed to recognize
+    )
+
+    # line_img can be an image loaded with PIL/Pillow or a numpy array
+    for line_img, ground_truth in training_data:
+        ocr.train(line_img, ground_truth)
+    ocr.save("my_model.clstm")
+
+
+**Recognition:**
+
+.. code::
+
+    import pyclstm
+    ocr = pyclstm.ClstmOcr()
+    ocr.load("my_model.clstm")
+    text = ocr.recognize(line_img)
+
+
+API Reference
+-------------
+
+.. automodule:: pyclstm
+    :members:
+    :undoc-members:
+
+Indices and tables
+==================
+
+* :ref:`genindex`
+* :ref:`modindex`
+* :ref:`search`
+
diff --git a/pydoc/make.bat b/pydoc/make.bat
new file mode 100644
index 0000000..62fe0f0
--- /dev/null
+++ b/pydoc/make.bat
@@ -0,0 +1,281 @@
+@ECHO OFF
+
+REM Command file for Sphinx documentation
+
+if "%SPHINXBUILD%" == "" (
+	set SPHINXBUILD=sphinx-build
+)
+set BUILDDIR=_build
+set ALLSPHINXOPTS=-d %BUILDDIR%/doctrees %SPHINXOPTS% .
+set I18NSPHINXOPTS=%SPHINXOPTS% .
+if NOT "%PAPER%" == "" (
+	set ALLSPHINXOPTS=-D latex_paper_size=%PAPER% %ALLSPHINXOPTS%
+	set I18NSPHINXOPTS=-D latex_paper_size=%PAPER% %I18NSPHINXOPTS%
+)
+
+if "%1" == "" goto help
+
+if "%1" == "help" (
+	:help
+	echo.Please use `make ^<target^>` where ^<target^> is one of
+	echo.  html       to make standalone HTML files
+	echo.  dirhtml    to make HTML files named index.html in directories
+	echo.  singlehtml to make a single large HTML file
+	echo.  pickle     to make pickle files
+	echo.  json       to make JSON files
+	echo.  htmlhelp   to make HTML files and a HTML help project
+	echo.  qthelp     to make HTML files and a qthelp project
+	echo.  devhelp    to make HTML files and a Devhelp project
+	echo.  epub       to make an epub
+	echo.  epub3      to make an epub3
+	echo.  latex      to make LaTeX files, you can set PAPER=a4 or PAPER=letter
+	echo.  text       to make text files
+	echo.  man        to make manual pages
+	echo.  texinfo    to make Texinfo files
+	echo.  gettext    to make PO message catalogs
+	echo.  changes    to make an overview over all changed/added/deprecated items
+	echo.  xml        to make Docutils-native XML files
+	echo.  pseudoxml  to make pseudoxml-XML files for display purposes
+	echo.  linkcheck  to check all external links for integrity
+	echo.  doctest    to run all doctests embedded in the documentation if enabled
+	echo.  coverage   to run coverage check of the documentation if enabled
+	echo.  dummy      to check syntax errors of document sources
+	goto end
+)
+
+if "%1" == "clean" (
+	for /d %%i in (%BUILDDIR%\*) do rmdir /q /s %%i
+	del /q /s %BUILDDIR%\*
+	goto end
+)
+
+
+REM Check if sphinx-build is available and fallback to Python version if any
+%SPHINXBUILD% 1>NUL 2>NUL
+if errorlevel 9009 goto sphinx_python
+goto sphinx_ok
+
+:sphinx_python
+
+set SPHINXBUILD=python -m sphinx.__init__
+%SPHINXBUILD% 2> nul
+if errorlevel 9009 (
+	echo.
+	echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
+	echo.installed, then set the SPHINXBUILD environment variable to point
+	echo.to the full path of the 'sphinx-build' executable. Alternatively you
+	echo.may add the Sphinx directory to PATH.
+	echo.
+	echo.If you don't have Sphinx installed, grab it from
+	echo.http://sphinx-doc.org/
+	exit /b 1
+)
+
+:sphinx_ok
+
+
+if "%1" == "html" (
+	%SPHINXBUILD% -b html %ALLSPHINXOPTS% %BUILDDIR%/html
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished. The HTML pages are in %BUILDDIR%/html.
+	goto end
+)
+
+if "%1" == "dirhtml" (
+	%SPHINXBUILD% -b dirhtml %ALLSPHINXOPTS% %BUILDDIR%/dirhtml
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished. The HTML pages are in %BUILDDIR%/dirhtml.
+	goto end
+)
+
+if "%1" == "singlehtml" (
+	%SPHINXBUILD% -b singlehtml %ALLSPHINXOPTS% %BUILDDIR%/singlehtml
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished. The HTML pages are in %BUILDDIR%/singlehtml.
+	goto end
+)
+
+if "%1" == "pickle" (
+	%SPHINXBUILD% -b pickle %ALLSPHINXOPTS% %BUILDDIR%/pickle
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished; now you can process the pickle files.
+	goto end
+)
+
+if "%1" == "json" (
+	%SPHINXBUILD% -b json %ALLSPHINXOPTS% %BUILDDIR%/json
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished; now you can process the JSON files.
+	goto end
+)
+
+if "%1" == "htmlhelp" (
+	%SPHINXBUILD% -b htmlhelp %ALLSPHINXOPTS% %BUILDDIR%/htmlhelp
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished; now you can run HTML Help Workshop with the ^
+.hhp project file in %BUILDDIR%/htmlhelp.
+	goto end
+)
+
+if "%1" == "qthelp" (
+	%SPHINXBUILD% -b qthelp %ALLSPHINXOPTS% %BUILDDIR%/qthelp
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished; now you can run "qcollectiongenerator" with the ^
+.qhcp project file in %BUILDDIR%/qthelp, like this:
+	echo.^> qcollectiongenerator %BUILDDIR%\qthelp\pyclstm.qhcp
+	echo.To view the help file:
+	echo.^> assistant -collectionFile %BUILDDIR%\qthelp\pyclstm.ghc
+	goto end
+)
+
+if "%1" == "devhelp" (
+	%SPHINXBUILD% -b devhelp %ALLSPHINXOPTS% %BUILDDIR%/devhelp
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished.
+	goto end
+)
+
+if "%1" == "epub" (
+	%SPHINXBUILD% -b epub %ALLSPHINXOPTS% %BUILDDIR%/epub
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished. The epub file is in %BUILDDIR%/epub.
+	goto end
+)
+
+if "%1" == "epub3" (
+	%SPHINXBUILD% -b epub3 %ALLSPHINXOPTS% %BUILDDIR%/epub3
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished. The epub3 file is in %BUILDDIR%/epub3.
+	goto end
+)
+
+if "%1" == "latex" (
+	%SPHINXBUILD% -b latex %ALLSPHINXOPTS% %BUILDDIR%/latex
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished; the LaTeX files are in %BUILDDIR%/latex.
+	goto end
+)
+
+if "%1" == "latexpdf" (
+	%SPHINXBUILD% -b latex %ALLSPHINXOPTS% %BUILDDIR%/latex
+	cd %BUILDDIR%/latex
+	make all-pdf
+	cd %~dp0
+	echo.
+	echo.Build finished; the PDF files are in %BUILDDIR%/latex.
+	goto end
+)
+
+if "%1" == "latexpdfja" (
+	%SPHINXBUILD% -b latex %ALLSPHINXOPTS% %BUILDDIR%/latex
+	cd %BUILDDIR%/latex
+	make all-pdf-ja
+	cd %~dp0
+	echo.
+	echo.Build finished; the PDF files are in %BUILDDIR%/latex.
+	goto end
+)
+
+if "%1" == "text" (
+	%SPHINXBUILD% -b text %ALLSPHINXOPTS% %BUILDDIR%/text
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished. The text files are in %BUILDDIR%/text.
+	goto end
+)
+
+if "%1" == "man" (
+	%SPHINXBUILD% -b man %ALLSPHINXOPTS% %BUILDDIR%/man
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished. The manual pages are in %BUILDDIR%/man.
+	goto end
+)
+
+if "%1" == "texinfo" (
+	%SPHINXBUILD% -b texinfo %ALLSPHINXOPTS% %BUILDDIR%/texinfo
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished. The Texinfo files are in %BUILDDIR%/texinfo.
+	goto end
+)
+
+if "%1" == "gettext" (
+	%SPHINXBUILD% -b gettext %I18NSPHINXOPTS% %BUILDDIR%/locale
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished. The message catalogs are in %BUILDDIR%/locale.
+	goto end
+)
+
+if "%1" == "changes" (
+	%SPHINXBUILD% -b changes %ALLSPHINXOPTS% %BUILDDIR%/changes
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.The overview file is in %BUILDDIR%/changes.
+	goto end
+)
+
+if "%1" == "linkcheck" (
+	%SPHINXBUILD% -b linkcheck %ALLSPHINXOPTS% %BUILDDIR%/linkcheck
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Link check complete; look for any errors in the above output ^
+or in %BUILDDIR%/linkcheck/output.txt.
+	goto end
+)
+
+if "%1" == "doctest" (
+	%SPHINXBUILD% -b doctest %ALLSPHINXOPTS% %BUILDDIR%/doctest
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Testing of doctests in the sources finished, look at the ^
+results in %BUILDDIR%/doctest/output.txt.
+	goto end
+)
+
+if "%1" == "coverage" (
+	%SPHINXBUILD% -b coverage %ALLSPHINXOPTS% %BUILDDIR%/coverage
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Testing of coverage in the sources finished, look at the ^
+results in %BUILDDIR%/coverage/python.txt.
+	goto end
+)
+
+if "%1" == "xml" (
+	%SPHINXBUILD% -b xml %ALLSPHINXOPTS% %BUILDDIR%/xml
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished. The XML files are in %BUILDDIR%/xml.
+	goto end
+)
+
+if "%1" == "pseudoxml" (
+	%SPHINXBUILD% -b pseudoxml %ALLSPHINXOPTS% %BUILDDIR%/pseudoxml
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished. The pseudo-XML files are in %BUILDDIR%/pseudoxml.
+	goto end
+)
+
+if "%1" == "dummy" (
+	%SPHINXBUILD% -b dummy %ALLSPHINXOPTS% %BUILDDIR%/dummy
+	if errorlevel 1 exit /b 1
+	echo.
+	echo.Build finished. Dummy builder generates no files.
+	goto end
+)
+
+:end
diff --git a/pyextra_defs.h b/pyextra_defs.h
new file mode 100644
index 0000000..0b5ed86
--- /dev/null
+++ b/pyextra_defs.h
@@ -0,0 +1,18 @@
+#include <Python.h>
+
+/** This extra header is needed to make the code Python 3 compatible.
+ * The reason for this is that the signature of `PyUnicode_AsWideChar` changed
+ * due to the removal of the `PyUnicodeObject` (`unicode`) type in Python 3.
+ *
+ * Unfortunately Cython doesn't support conditional compilation out of the
+ * box, so we need to use this workaround.
+ **/
+#if PY_MAJOR_VERSION >= 3
+typedef PyObject UnicodeObject;
+#else
+typedef PyUnicodeObject UnicodeObject;
+#endif
+
+Py_ssize_t Unicode_AsWideChar(PyObject* str, Py_ssize_t length, wchar_t *wchars) {
+    return PyUnicode_AsWideChar((UnicodeObject*) str, wchars, length);
+}
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 0000000..8ccc63c
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1 @@
+Cython>=0.23
diff --git a/run_uw3_500.py b/run_uw3_500.py
new file mode 100644
index 0000000..ee126b2
--- /dev/null
+++ b/run_uw3_500.py
@@ -0,0 +1,53 @@
+from __future__ import division
+
+import glob
+import os
+import random
+import sys
+from itertools import chain
+
+import pyclstm
+from PIL import Image
+
+if not os.path.isdir('./book'):
+    print("Please download the sample data:\n\t"
+          "curl -L http://tmbdev.net/ocrdata/uw3-500.tgz|tar xf")
+    sys.exit(1)
+
+all_imgs = [Image.open(p) for p in sorted(glob.glob("./book/*/*.png"))]
+all_texts = [open(p).read().strip()
+             for p in sorted(glob.glob("./book/*/*.gt.txt"))]
+if sys.version_info <= (3,):
+    all_texts = [t.decode('utf8') for t in all_texts]
+
+all_data = list(zip(all_imgs, all_texts))
+
+train_data = all_data[:400]
+test_data = all_data[400:]
+
+ocr = pyclstm.ClstmOcr()
+graphemes = set(chain.from_iterable(all_texts))
+ocr.prepare_training(graphemes)
+
+for i in range(2000000):
+    best_error = 1.
+    img, txt = random.choice(train_data)
+    out = ocr.train(img, txt)
+    if not i % 10:
+        aligned = ocr.aligned()
+        print("Truth:   {}".format(txt))
+        print("Aligned: {}".format(aligned))
+        print("Output:  {}".format(out))
+    if not i % 1000:
+        errors = 0
+        chars = 0
+        for img, txt in test_data:
+            out = ocr.recognize(img)
+            errors += pyclstm.levenshtein(txt, out)
+            chars += len(txt)
+        error = errors / chars
+        print ("=== Test set error after {} iterations: {:.2f}"
+               .format(i, error))
+        if error < best_error:
+            print("=== New best error, saving model to model.clstm")
+            ocr.save("./model.clstm")
diff --git a/setup.py b/setup.py
index 80085f3..6adb75a 100644
--- a/setup.py
+++ b/setup.py
@@ -1,51 +1,46 @@
-#!/usr/bin/env python
+from __future__ import print_function
 
-import os, os.path
-from distutils.core import setup, Extension
-from numpy.distutils.misc_util import get_numpy_include_dirs
-from os.path import getctime
+import os
 
-import distutils.sysconfig
-config = distutils.sysconfig.get_config_vars()
-# OPT=-DNDEBUG -g -fwrapv -O2 -Wall -Wstrict-prototypes
-# CFLAGS
-for k,v in config.items():
-  if type(v)==str and "-W" in v:
-    print k,v
-def remove_warnings(opt):
-  opt = opt.split()
-  opt = [s for s in opt if not s.startswith("-W")]
-  return " ".join(opt)
-config["OPT"] = remove_warnings(config["OPT"])
-config["CFLAGS"] = remove_warnings(config["CFLAGS"])
-config["CONFIGURE_CFLAGS"] = remove_warnings(config["CONFIGURE_CFLAGS"])
-config["LDSHARED"] = remove_warnings(config["LDSHARED"])
+from distutils.core import setup, Extension
+from Cython.Build import cythonize
 
-# hgversion = os.popen("hg -q id").read().strip()
-hgversion = "unknown"
 
-include_dirs = ['/usr/include/eigen3', '/usr/local/include/eigen3', '/usr/local/include', '/usr/include/hdf5/serial'] + get_numpy_include_dirs()
-swig_opts = ["-c++"] + ["-I" + d for d in include_dirs]
-swiglib = os.popen("swig -swiglib").read()[:-1]
+def ensure_protobuf():
+    exists = os.path.exists("clstm.pb.cc")
+    stale = (exists and
+             os.path.getctime("clstm.pb.cc") < os.path.getctime("clstm.proto"))
+    if not exists or stale:
+        print("Generating proto file")
+        os.system("protoc clstm.proto --cpp_out=.")
 
-if not os.path.exists("clstm.pb.cc") or \
-    getctime("clstm.pb.cc") < getctime("clstm.proto"):
-  print "making proto file"
-  os.system("protoc clstm.proto --cpp_out=.")
 
-clstm = Extension('_clstm',
-        libraries = ['png','protobuf'],
-        swig_opts = swig_opts,
-        include_dirs = include_dirs,
-        extra_compile_args = ['-std=c++11','-Wno-sign-compare',
-            '-Dadd_raw=add','-DNODISPLAY=1','-DTHROW=throw',
-            '-DHGVERSION="\\"'+hgversion+'\\""'],
-        sources=['clstm.i','clstm.cc','clstm_prefab.cc','extras.cc',
-                 'ctc.cc','clstm_proto.cc','clstm.pb.cc'])
+ext = Extension(
+    "pyclstm",
+    sources=['pyclstm.pyx', 'clstm.cc', 'clstm_prefab.cc', 'extras.cc',
+             'batches.cc', 'ctc.cc', 'clstm_proto.cc', 'clstm.pb.cc',
+             'clstm_compute.cc', 'tensor.cc'],
+    include_dirs=['/usr/include/eigen3', '/usr/local/include/eigen3',
+                  '/usr/local/include', '/usr/include',
+                  '/usr/include/hdf5/serial', '/usr/include/hdf5'],
+    libraries=['protobuf', 'png'],
+    language='c++',
+    extra_compile_args=['-w', '--std=c++11', '-Wno-unused-result', '-g',
+                        '-DNODISPLAY=1', '-DTHROW=throw', '-DNDEBUG', '-Ofast',
+                        '-DEIGEN_NO_DEBUG', '-finline', '-ffast-math',
+                        '-fno-signaling-nans', '-funsafe-math-optimizations',
+                        '-ffinite-math-only', '-march=native'])
 
-setup (name = 'clstm',
-       version = '0.0',
-       author      = "Thomas Breuel",
-       description = """clstm library bindings""",
-       ext_modules = [clstm],
-       py_modules = ["clstm"])
+ensure_protobuf()
+setup(
+    name='clstm',
+    version='0.1',
+    packages = ['pyclstm_cli'],
+    author="Thomas Breuel, Johannes Baiter",
+    description="CLSTM Python bindings",
+    entry_points = {
+        'console_scripts':[
+            'pyclstm-train=pyclstm_cli.train:main'
+        ]
+    },
+    ext_modules=cythonize([ext], compiler_directives={'embedsignature': True}))