Mergery

bin/parser/train_ud.py

@@ -18,6 +18,8 @@ import spacy.attrs
 import io
 from thinc.neural.ops import CupyOps
 from thinc.neural import Model
+from spacy.es import Spanish
+from spacy.attrs import POS
 
 try:
     import cupy
@@ -156,20 +158,15 @@ def main(lang_name, train_loc, dev_loc, model_dir, clusters_loc=None):
                 for tag in tags:
                     vocab.morphology.tag_map[tag] = {POS: tag.split('__', 1)[0]}
     tagger = Tagger(vocab)
-    encoder = TokenVectorEncoder(vocab)
+    encoder = TokenVectorEncoder(vocab, width=64)
     parser = DependencyParser(vocab, actions=actions, features=features, L1=0.0)
 
     Xs, ys = organize_data(vocab, train_sents)
     dev_Xs, dev_ys = organize_data(vocab, dev_sents)
-    #Xs = Xs[:1000]
-    #ys = ys[:1000]
-    #dev_Xs = dev_Xs[:1000]
-    #dev_ys = dev_ys[:1000]
     with encoder.model.begin_training(Xs[:100], ys[:100]) as (trainer, optimizer):
         docs = list(Xs)
         for doc in docs:
             encoder(doc)
-        parser.begin_training(docs, ys)
         nn_loss = [0.]
         def track_progress():
             with encoder.tagger.use_params(optimizer.averages):
@@ -191,11 +188,23 @@ def main(lang_name, train_loc, dev_loc, model_dir, clusters_loc=None):
             upd_tokvecs(d_tokvecs, sgd=optimizer)
             encoder.update(docs, golds, sgd=optimizer)
             nn_loss[-1] += loss
-    nlp = LangClass(vocab=vocab, tagger=tagger, parser=parser)
-    #nlp.end_training(model_dir)
-    scorer = score_model(vocab, tagger, parser, read_conllx(dev_loc))
+    nlp = LangClass(vocab=vocab, parser=parser)
+    scorer = score_model(vocab, encoder, parser, read_conllx(dev_loc))
     print('%d:\t%.3f\t%.3f\t%.3f' % (itn, scorer.uas, scorer.las, scorer.tags_acc))
+    #nlp.end_training(model_dir)
+    #scorer = score_model(vocab, tagger, parser, read_conllx(dev_loc))
+    #print('%d:\t%.3f\t%.3f\t%.3f' % (itn, scorer.uas, scorer.las, scorer.tags_acc))
 
 
 if __name__ == '__main__':
+    import cProfile
+    import pstats
+    if 1:
+        plac.call(main)
+    else:
+        cProfile.runctx("plac.call(main)", globals(), locals(), "Profile.prof")
+    s = pstats.Stats("Profile.prof")
+    s.strip_dirs().sort_stats("time").print_stats()
+
+
     plac.call(main)

spacy/_ml.py

@@ -7,8 +7,125 @@ from thinc.neural._classes.static_vectors import StaticVectors
 from thinc.neural._classes.batchnorm import BatchNorm
 from thinc.neural._classes.resnet import Residual
 
+from thinc import describe
+from thinc.describe import Dimension, Synapses, Biases, Gradient
+from thinc.neural._classes.affine import _set_dimensions_if_needed
 from .attrs import ID, LOWER, PREFIX, SUFFIX, SHAPE, TAG, DEP
 
+import numpy
+
+
+@describe.on_data(_set_dimensions_if_needed)
+@describe.attributes(
+    nI=Dimension("Input size"),
+    nF=Dimension("Number of features"),
+    nO=Dimension("Output size"),
+    W=Synapses("Weights matrix",
+        lambda obj: (obj.nO, obj.nF, obj.nI),
+        lambda W, ops: ops.xavier_uniform_init(W)),
+    b=Biases("Bias vector",
+        lambda obj: (obj.nO,)),
+    d_W=Gradient("W"),
+    d_b=Gradient("b")
+)
+class PrecomputableAffine(Model):
+    def __init__(self, nO=None, nI=None, nF=None, **kwargs):
+        Model.__init__(self, **kwargs)
+        self.nO = nO
+        self.nI = nI
+        self.nF = nF
+
+    def begin_update(self, X, drop=0.):
+        # X: (b, i)
+        # Xf: (b, f, i)
+        # dY: (b, o)
+        # dYf: (b, f, o)
+        #Yf = numpy.einsum('bi,ofi->bfo', X, self.W)
+        Yf = self.ops.xp.tensordot(
+                X, self.W, axes=[[1], [2]]).transpose((0, 2, 1))
+        Yf += self.b
+        def backward(dY_ids, sgd=None):
+            dY, ids = dY_ids
+            Xf = X[ids]
+
+            #dW = numpy.einsum('bo,bfi->ofi', dY, Xf)
+            dW = self.ops.xp.tensordot(dY, Xf, axes=[[0], [0]])
+            db = dY.sum(axis=0)
+            #dXf = numpy.einsum('bo,ofi->bfi', dY, self.W)
+            dXf = self.ops.xp.tensordot(dY, self.W, axes=[[1], [0]])
+
+            self.d_W += dW
+            self.d_b += db
+
+            if sgd is not None:
+                sgd(self._mem.weights, self._mem.gradient, key=self.id)
+            return dXf
+        return Yf, backward
+
+
+@describe.on_data(_set_dimensions_if_needed)
+@describe.attributes(
+    nI=Dimension("Input size"),
+    nF=Dimension("Number of features"),
+    nP=Dimension("Number of pieces"),
+    nO=Dimension("Output size"),
+    W=Synapses("Weights matrix",
+        lambda obj: (obj.nF, obj.nO, obj.nP, obj.nI),
+        lambda W, ops: ops.xavier_uniform_init(W)),
+    b=Biases("Bias vector",
+        lambda obj: (obj.nO, obj.nP)),
+    d_W=Gradient("W"),
+    d_b=Gradient("b")
+)
+class PrecomputableMaxouts(Model):
+    def __init__(self, nO=None, nI=None, nF=None, pieces=2, **kwargs):
+        Model.__init__(self, **kwargs)
+        self.nO = nO
+        self.nP = pieces
+        self.nI = nI
+        self.nF = nF
+
+    def begin_update(self, X, drop=0.):
+        # X: (b, i)
+        # Yfp: (f, b, o, p)
+        # Yf: (f, b, o)
+        # Xf: (b, f, i)
+        # dY: (b, o)
+        # dYp: (b, o, p)
+        # W: (f, o, p, i)
+        # b: (o, p)
+
+        # Yfp = numpy.einsum('bi,fopi->fbop', X, self.W)
+        Yfp = self.ops.xp.tensordot(X, self.W,
+                axes=[[1], [3]]).transpose((1, 0, 2, 3))
+        Yfp = self.ops.xp.ascontiguousarray(Yfp)
+        Yfp += self.b
+        Yf = self.ops.allocate((self.nF, X.shape[0], self.nO))
+        which = self.ops.allocate((self.nF, X.shape[0], self.nO), dtype='i')
+        for i in range(self.nF):
+            Yf[i], which[i] = self.ops.maxout(Yfp[i])
+        def backward(dY_ids, sgd=None):
+            dY, ids = dY_ids
+            Xf = X[ids]
+            dYp = self.ops.allocate((dY.shape[0], self.nO, self.nP))
+            for i in range(self.nF):
+                dYp += self.ops.backprop_maxout(dY, which[i], self.nP)
+
+            #dXf = numpy.einsum('bop,fopi->bfi', dYp, self.W)
+            dXf = self.ops.xp.tensordot(dYp, self.W, axes=[[1,2], [1,2]])
+            #dW = numpy.einsum('bfi,bop->fopi', Xf, dYp)
+            dW = self.ops.xp.tensordot(Xf, dYp, axes=[[0], [0]])
+            dW = dW.transpose((0, 2, 3, 1))
+            db = dYp.sum(axis=0)
+
+            self.d_W += dW
+            self.d_b += db
+
+            if sgd is not None:
+                sgd(self._mem.weights, self._mem.gradient, key=self.id)
+            return dXf
+        return Yf, backward
+
 
 def get_col(idx):
     def forward(X, drop=0.):
@@ -22,55 +139,36 @@ def get_col(idx):
     return layerize(forward)
 
 
-def build_model(state2vec, width, depth, nr_class):
-    with Model.define_operators({'>>': chain, '**': clone}):
-        model = (
-            state2vec
-            >> Maxout(width, 1344)
-            >> Maxout(width, width)
-            >> Affine(nr_class, width)
+def build_tok2vec(lang, width, depth=2, embed_size=1000):
+    cols = [ID, LOWER, PREFIX, SUFFIX, SHAPE]
+    with Model.define_operators({'>>': chain, '|': concatenate, '**': clone}):
+        #static = get_col(cols.index(ID))     >> StaticVectors(lang, width)
+        lower = get_col(cols.index(LOWER))     >> HashEmbed(width, embed_size)
+        prefix = get_col(cols.index(PREFIX)) >> HashEmbed(width//4, embed_size)
+        suffix = get_col(cols.index(SUFFIX)) >> HashEmbed(width//4, embed_size)
+        shape = get_col(cols.index(SHAPE))   >> HashEmbed(width//4, embed_size)
+        tok2vec = (
+            doc2feats(cols)
+            >> with_flatten(
+                #(static | prefix | suffix | shape)
+                (lower | prefix | suffix | shape)
+                >> Maxout(width)
+                >> (ExtractWindow(nW=1) >> Maxout(width, width*3))
+                >> (ExtractWindow(nW=1) >> Maxout(width, width*3))
+            )
         )
-    return model
+    return tok2vec
 
 
-def build_debug_model(state2vec, width, depth, nr_class):
-    with Model.define_operators({'>>': chain, '**': clone}):
-        model = (
-            state2vec
-            #>> Maxout(width)
-            >> Maxout(nr_class)
-        )
-    return model
-
-
-def build_debug_state2vec(width, nr_vector=1000, nF=1, nB=0, nS=1, nL=2, nR=2):
-    ops = Model.ops
-    def forward(tokens_attrs_vectors, drop=0.):
-        tokens, attr_vals, tokvecs = tokens_attrs_vectors
-
-        orig_tokvecs_shape = tokvecs.shape
-        tokvecs = tokvecs.reshape((tokvecs.shape[0], tokvecs.shape[1] *
-                                   tokvecs.shape[2]))
-
-        vector = tokvecs
-
-        def backward(d_vector, sgd=None):
-            d_tokvecs = vector.reshape(orig_tokvecs_shape)
-            return (tokens, d_tokvecs)
-        return vector, backward
+def doc2feats(cols):
+    def forward(docs, drop=0.):
+        feats = [doc.to_array(cols) for doc in docs]
+        feats = [model.ops.asarray(f, dtype='uint64') for f in feats]
+        return feats, None
     model = layerize(forward)
     return model
 
 
-def build_state2vec(nr_context_tokens, width, nr_vector=1000):
-    ops = Model.ops
-    with Model.define_operators({'|': concatenate, '+': add, '>>': chain}):
-        #hiddens = [get_col(i) >> Maxout(width) for i in range(nr_context_tokens)]
-        features = [get_col(i) for i in range(nr_context_tokens)]
-        model = get_token_vectors >> concatenate(*features) >> ReLu(width)
-    return model
-
-
 def print_shape(prefix):
     def forward(X, drop=0.):
         return X, lambda dX, **kwargs: dX
@@ -86,87 +184,6 @@ def get_token_vectors(tokens_attrs_vectors, drop=0.):
     return vectors, backward
 
 
-def build_parser_state2vec(width, nr_vector=1000, nF=1, nB=0, nS=1, nL=2, nR=2):
-    embed_tags = _reshape(chain(get_col(0), HashEmbed(16, nr_vector)))
-    embed_deps = _reshape(chain(get_col(1), HashEmbed(16, nr_vector)))
-    ops = embed_tags.ops
-    def forward(tokens_attrs_vectors, drop=0.):
-        tokens, attr_vals, tokvecs = tokens_attrs_vectors
-        tagvecs, bp_tagvecs = embed_deps.begin_update(attr_vals, drop=drop)
-        depvecs, bp_depvecs = embed_tags.begin_update(attr_vals, drop=drop)
-        orig_tokvecs_shape = tokvecs.shape
-        tokvecs = tokvecs.reshape((tokvecs.shape[0], tokvecs.shape[1] *
-                                   tokvecs.shape[2]))
-
-        shapes = (tagvecs.shape, depvecs.shape, tokvecs.shape)
-        assert tagvecs.shape[0] == depvecs.shape[0] == tokvecs.shape[0], shapes
-        vector = ops.xp.hstack((tagvecs, depvecs, tokvecs))
-
-        def backward(d_vector, sgd=None):
-            d_tagvecs, d_depvecs, d_tokvecs = backprop_concatenate(d_vector, shapes)
-            assert d_tagvecs.shape == shapes[0], (d_tagvecs.shape, shapes)
-            assert d_depvecs.shape == shapes[1], (d_depvecs.shape, shapes)
-            assert d_tokvecs.shape == shapes[2], (d_tokvecs.shape, shapes)
-            bp_tagvecs(d_tagvecs)
-            bp_depvecs(d_depvecs)
-            d_tokvecs = d_tokvecs.reshape(orig_tokvecs_shape)
-
-            return (tokens, d_tokvecs)
-        return vector, backward
-    model = layerize(forward)
-    model._layers = [embed_tags, embed_deps]
-    return model
-
-
-def backprop_concatenate(gradient, shapes):
-    grads = []
-    start = 0
-    for shape in shapes:
-        end = start + shape[1]
-        grads.append(gradient[:, start : end])
-        start = end
-    return grads
-
-
-def _reshape(layer):
-    '''Transforms input with shape
-      (states, tokens, features)
-    into input with shape:
-      (states * tokens, features)
-    So that it can be used with a token-wise feature extraction layer, e.g.
-    an embedding layer. The embedding layer outputs:
-      (states * tokens, ndim)
-    But we want to concatenate the vectors for the tokens, so we produce:
-      (states, tokens * ndim)
-    We then need to reverse the transforms to do the backward pass. Recall
-    the simple rule here: each layer is a map:
-        inputs -> (outputs, (d_outputs->d_inputs))
-    So the shapes must match like this:
-        shape of forward input == shape of backward output
-        shape of backward input == shape of forward output
-    '''
-    def forward(X__bfm, drop=0.):
-        b, f, m = X__bfm.shape
-        B = b*f
-        M = f*m
-        X__Bm = X__bfm.reshape((B, m))
-        y__Bn, bp_yBn = layer.begin_update(X__Bm, drop=drop)
-        n = y__Bn.shape[1]
-        N = f * n
-        y__bN = y__Bn.reshape((b, N))
-        def backward(dy__bN, sgd=None):
-            dy__Bn = dy__bN.reshape((B, n))
-            dX__Bm = bp_yBn(dy__Bn, sgd)
-            if dX__Bm is None:
-                return None
-            else:
-                return dX__Bm.reshape((b, f, m))
-        return y__bN, backward
-    model = layerize(forward)
-    model._layers.append(layer)
-    return model
-
-
 @layerize
 def flatten(seqs, drop=0.):
     ops = Model.ops
@@ -177,32 +194,44 @@ def flatten(seqs, drop=0.):
     return X, finish_update
 
 
-def build_tok2vec(lang, width, depth=2, embed_size=1000):
-    cols = [ID, LOWER, PREFIX, SUFFIX, SHAPE, TAG]
-    with Model.define_operators({'>>': chain, '|': concatenate, '**': clone}):
-        #static = get_col(cols.index(ID))     >> StaticVectors(lang, width)
-        lower = get_col(cols.index(LOWER))     >> HashEmbed(width, embed_size)
-        prefix = get_col(cols.index(PREFIX)) >> HashEmbed(width, embed_size)
-        suffix = get_col(cols.index(SUFFIX)) >> HashEmbed(width, embed_size)
-        shape = get_col(cols.index(SHAPE))   >> HashEmbed(width, embed_size)
-        tok2vec = (
-            doc2feats(cols)
-            >> with_flatten(
-                #(static | prefix | suffix | shape)
-                (lower | prefix | suffix | shape)
-                >> Maxout(width, width*4)
-                >> Residual((ExtractWindow(nW=1) >> Maxout(width, width*3)))
-                >> Residual((ExtractWindow(nW=1) >> Maxout(width, width*3)))
-                >> Residual((ExtractWindow(nW=1) >> Maxout(width, width*3)))
-            )
-        )
-    return tok2vec
+#def build_feature_precomputer(model, feat_maps):
+#    '''Allow a model to be "primed" by pre-computing input features in bulk.
+#
+#    This is used for the parser, where we want to take a batch of documents,
+#    and compute vectors for each (token, position) pair. These vectors can then
+#    be reused, especially for beam-search.
+#
+#    Let's say we're using 12 features for each state, e.g. word at start of
+#    buffer, three words on stack, their children, etc. In the normal arc-eager
+#    system, a document of length N is processed in 2*N states. This means we'll
+#    create 2*N*12 feature vectors --- but if we pre-compute, we only need
+#    N*12 vector computations. The saving for beam-search is much better:
+#    if we have a beam of k, we'll normally make 2*N*12*K computations --
+#    so we can save the factor k. This also gives a nice CPU/GPU division:
+#    we can do all our hard maths up front, packed into large multiplications,
+#    and do the hard-to-program parsing on the CPU.
+#    '''
+#    def precompute(input_vectors):
+#        cached, backprops = zip(*[lyr.begin_update(input_vectors)
+#                                for lyr in feat_maps)
+#        def forward(batch_token_ids, drop=0.):
+#            output = ops.allocate((batch_size, output_width))
+#            # i: batch index
+#            # j: position index (i.e. N0, S0, etc
+#            # tok_i: Index of the token within its document
+#            for i, token_ids in enumerate(batch_token_ids):
+#                for j, tok_i in enumerate(token_ids):
+#                    output[i] += cached[j][tok_i]
+#            def backward(d_vector, sgd=None):
+#                d_inputs = ops.allocate((batch_size, n_feat, vec_width))
+#                for i, token_ids in enumerate(batch_token_ids):
+#                    for j in range(len(token_ids)):
+#                        d_inputs[i][j] = backprops[j](d_vector, sgd)
+#                # Return the IDs, so caller can associate to correct token
+#                return (batch_token_ids, d_inputs)
+#            return vector, backward
+#        return chain(layerize(forward), model)
+#    return precompute
+#
+#
 
-
-def doc2feats(cols):
-    def forward(docs, drop=0.):
-        feats = [doc.to_array(cols) for doc in docs]
-        feats = [model.ops.asarray(f, dtype='uint64') for f in feats]
-        return feats, None
-    model = layerize(forward)
-    return model

spacy/pipeline.pyx

@@ -23,7 +23,7 @@ class TokenVectorEncoder(object):
     '''Assign position-sensitive vectors to tokens, using a CNN or RNN.'''
     def __init__(self, vocab, **cfg):
         self.vocab = vocab
-        self.model = build_tok2vec(vocab.lang, 64, **cfg)
+        self.model = build_tok2vec(vocab.lang, **cfg)
         self.tagger = chain(
                         self.model,
                         flatten,

spacy/syntax/parser.pxd

@@ -13,5 +13,6 @@ cdef class Parser:
     cdef readonly object model
     cdef readonly TransitionSystem moves
     cdef readonly object cfg
+    cdef public object feature_maps
 
     #cdef int parseC(self, TokenC* tokens, int length, int nr_feat) nogil

spacy/syntax/parser.pyx

@@ -28,8 +28,11 @@ from murmurhash.mrmr cimport hash64
 from preshed.maps cimport MapStruct
 from preshed.maps cimport map_get
 
-from numpy import exp
 
+from thinc.api import layerize, chain
+from thinc.neural import Model, Maxout
+
+from .._ml import PrecomputableAffine, PrecomputableMaxouts
 from . import _parse_features
 from ._parse_features cimport CONTEXT_SIZE
 from ._parse_features cimport fill_context
@@ -44,10 +47,9 @@ from ..strings cimport StringStore
 from ..gold cimport GoldParse
 from ..attrs cimport TAG, DEP
 
-from .._ml import build_parser_state2vec, build_model
-from .._ml import build_state2vec, build_model
-from .._ml import build_debug_state2vec, build_debug_model
 
+def get_templates(*args, **kwargs):
+    return []
 
 USE_FTRL = True
 DEBUG = False
@@ -56,8 +58,109 @@ def set_debug(val):
     DEBUG = val
 
 
-def get_templates(*args, **kwargs):
-    return []
+def get_greedy_model_for_batch(tokvecs, TransitionSystem moves, upper_model, lower_model):
+    cdef int[:, :] is_valid_
+    cdef float[:, :] costs_
+    lengths = [len(t) for t in tokvecs]
+    tokvecs = upper_model.ops.flatten(tokvecs)
+    is_valid = upper_model.ops.allocate((len(tokvecs), moves.n_moves), dtype='i')
+    costs = upper_model.ops.allocate((len(tokvecs), moves.n_moves), dtype='f')
+    token_ids = upper_model.ops.allocate((len(tokvecs), lower_model.nF), dtype='i')
+
+    cached, bp_features = lower_model.begin_update(tokvecs, drop=0.)
+
+    is_valid_ = is_valid
+    costs_ = costs
+
+    def forward(states_offsets, drop=0.):
+        nonlocal is_valid, costs, token_ids, moves
+        states, offsets = states_offsets
+        assert len(states) != 0
+        is_valid = is_valid[:len(states)]
+        costs = costs[:len(states)]
+        token_ids = token_ids[:len(states)]
+        is_valid = is_valid[:len(states)]
+        cdef StateClass state
+        cdef int i
+        for i, (offset, state) in enumerate(zip(offsets, states)):
+            state.set_context_tokens(token_ids[i])
+            moves.set_valid(&is_valid_[i, 0], state.c)
+        adjusted_ids = token_ids.copy()
+        for i, offset in enumerate(offsets):
+            adjusted_ids[i] *= token_ids[i] >= 0
+            adjusted_ids[i] += offset
+        features = upper_model.ops.allocate((len(states), lower_model.nO), dtype='f')
+        for i in range(len(states)):
+            for j, tok_i in enumerate(adjusted_ids[i]):
+                if tok_i >= 0:
+                    features[i] += cached[j, tok_i]
+
+        scores, bp_scores = upper_model.begin_update(features, drop=drop)
+        scores = upper_model.ops.relu(scores)
+        softmaxed = upper_model.ops.softmax(scores)
+        # Renormalize for invalid actions
+        softmaxed *= is_valid
+        totals = softmaxed.sum(axis=1)
+        for total in totals:
+            assert total > 0, (totals, scores, softmaxed)
+            assert total <= 1.1, totals
+        softmaxed /= softmaxed.sum(axis=1).reshape((softmaxed.shape[0], 1))
+
+        def backward(golds, sgd=None):
+            nonlocal costs_, is_valid_, moves
+            cdef int i
+            for i, (state, gold) in enumerate(zip(states, golds)):
+                moves.set_costs(&is_valid_[i, 0], &costs_[i, 0],
+                    state, gold)
+            d_scores = scores.copy()
+            d_scores.fill(0)
+            set_log_loss(upper_model.ops, d_scores,
+                scores, is_valid, costs)
+            upper_model.ops.backprop_relu(d_scores, scores, inplace=True)
+            d_features = bp_scores(d_scores, sgd)
+            d_tokens = bp_features((d_features, adjusted_ids), sgd)
+            return (token_ids, d_tokens)
+
+        return softmaxed, backward
+
+    return layerize(forward)
+
+
+def set_log_loss(ops, gradients, scores, is_valid, costs):
+    """Do multi-label log loss"""
+    n = gradients.shape[0]
+    scores = scores * is_valid
+    g_scores = scores * is_valid * (costs <= 0.)
+    exps = ops.xp.exp(scores - scores.max(axis=1).reshape((n, 1)))
+    exps *= is_valid
+    g_exps = ops.xp.exp(g_scores - g_scores.max(axis=1).reshape((n, 1)))
+    g_exps *= costs <= 0.
+    g_exps *= is_valid
+    gradients[:] = exps / exps.sum(axis=1).reshape((n, 1))
+    gradients -= g_exps / g_exps.sum(axis=1).reshape((n, 1))
+
+
+def transition_batch(TransitionSystem moves, states, scores):
+    cdef StateClass state
+    cdef int guess
+    for state, guess in zip(states, scores.argmax(axis=1)):
+        action = moves.c[guess]
+        action.do(state.c, action.label)
+
+
+def init_states(TransitionSystem moves, docs):
+    cdef Doc doc
+    cdef StateClass state
+    offsets = []
+    states = []
+    offset = 0
+    for i, doc in enumerate(docs):
+        state = StateClass.init(doc.c, doc.length)
+        moves.initialize_state(state.c)
+        states.append(state)
+        offsets.append(offset)
+        offset += len(doc)
+    return states, offsets
 
 
 cdef class Parser:
@@ -107,8 +210,9 @@ cdef class Parser:
         cfg['actions'] = TransitionSystem.get_actions(**cfg)
         self.moves = TransitionSystem(vocab.strings, cfg['actions'])
         if model is None:
-            model = self.build_model(**cfg)
-        self.model = model
+            self.model, self.feature_maps = self.build_model(**cfg)
+        else:
+            self.model, self.feature_maps = model
         self.cfg = cfg
 
     def __reduce__(self):
@@ -116,10 +220,10 @@ cdef class Parser:
 
     def build_model(self, width=128, nr_vector=1000, nF=1, nB=1, nS=1, nL=1, nR=1, **_):
         nr_context_tokens = StateClass.nr_context_tokens(nF, nB, nS, nL, nR)
-        state2vec = build_state2vec(nr_context_tokens, width, nr_vector)
-        #state2vec = build_debug_state2vec(width, nr_vector)
-        model = build_debug_model(state2vec, width*2, 2, self.moves.n_moves)
-        return model
+
+        upper = chain(Maxout(width, width), Maxout(self.moves.n_moves, width))
+        lower = PrecomputableMaxouts(width, nF=nr_context_tokens, nI=width*2)
+        return upper, lower
 
     def __call__(self, Doc tokens):
         """
@@ -131,7 +235,6 @@ cdef class Parser:
             None
         """
         self.parse_batch([tokens])
-        self.moves.finalize_doc(tokens)
 
     def pipe(self, stream, int batch_size=1000, int n_threads=2):
         """
@@ -167,169 +270,53 @@ cdef class Parser:
                 yield doc
 
     def parse_batch(self, docs):
-        states = self._init_states(docs)
-        nr_class = self.moves.n_moves
         cdef Doc doc
         cdef StateClass state
-        cdef int guess
-        tokvecs = [d.tensor for d in docs]
+        model = get_greedy_model_for_batch([d.tensor for d in docs],
+                    self.moves, self.model, self.feature_maps)
+        states, offsets = init_states(self.moves, docs)
         all_states = list(states)
-        todo = zip(states, tokvecs)
+        todo = list(zip(states, offsets))
         while todo:
             todo = filter(lambda sp: not sp[0].py_is_final(), todo)
             if not todo:
                 break
-            states, tokvecs = zip(*todo)
-            scores, _ = self._begin_update(states, tokvecs)
-            self._transition_batch(states, docs, scores)
+            states, offsets = zip(*todo)
+            scores = model((states, offsets))
+            transition_batch(self.moves, states, scores)
+            todo = [st for st in todo if not st[0].py_is_final()]
         for state, doc in zip(all_states, docs):
             self.moves.finalize_state(state.c)
             for i in range(doc.length):
                 doc.c[i] = state.c._sent[i]
-
-    def begin_training(self, docs, golds):
-        for gold in golds:
-            self.moves.preprocess_gold(gold)
-        states = self._init_states(docs)
-        tokvecs = [d.tensor for d in docs]
-        d_tokens = [self.model.ops.allocate(d.tensor.shape) for d in docs]
-        nr_class = self.moves.n_moves
-        costs = self.model.ops.allocate((len(docs), nr_class), dtype='f')
-        gradients = self.model.ops.allocate((len(docs), nr_class), dtype='f')
-        is_valid = self.model.ops.allocate((len(docs), nr_class), dtype='i')
-        attr_names = numpy.zeros((2,), dtype='i')
-        attr_names[0] = TAG
-        attr_names[1] = DEP
-
-        features = self._get_features(states, tokvecs, attr_names)
-        self.model.begin_training(features)
+        for doc in docs:
+            self.moves.finalize_doc(doc)
 
     def update(self, docs, golds, drop=0., sgd=None):
         if isinstance(docs, Doc) and isinstance(golds, GoldParse):
             return self.update([docs], [golds], drop=drop)
         for gold in golds:
             self.moves.preprocess_gold(gold)
-        states = self._init_states(docs)
-        tokvecs = [d.tensor for d in docs]
+
+        model = get_greedy_model_for_batch([d.tensor for d in docs],
+                    self.moves, self.model, self.feature_maps)
+        states, offsets = init_states(self.moves, docs)
+
         d_tokens = [self.model.ops.allocate(d.tensor.shape) for d in docs]
-        nr_class = self.moves.n_moves
         output = list(d_tokens)
-        todo = zip(states, tokvecs, golds, d_tokens)
-        assert len(states) == len(todo)
-        losses = []
+        todo = zip(states, offsets, golds, d_tokens)
         while todo:
             # Get unfinished states (and their matching gold and token gradients)
             todo = filter(lambda sp: not sp[0].py_is_final(), todo)
             if not todo:
                 break
-            states, tokvecs, golds, d_tokens = zip(*todo)
-            scores, finish_update = self._begin_update(states, tokvecs)
-            token_ids, batch_token_grads = finish_update(golds, sgd=sgd, losses=losses,
-                                                         force_gold=False)
-            batch_token_grads *= (token_ids >= 0).reshape((token_ids.shape[0], token_ids.shape[1], 1))
-            token_ids *= token_ids >= 0
-            if hasattr(self.model.ops.xp, 'scatter_add'):
-                for i, tok_ids in enumerate(token_ids):
-                    self.model.ops.xp.scatter_add(d_tokens[i],
-                        tok_ids, batch_token_grads[i])
-            else:
-                for i, tok_ids in enumerate(token_ids):
-                    self.model.ops.xp.add.at(d_tokens[i],
-                        tok_ids, batch_token_grads[i])
-            self._transition_batch(states, docs, scores)
-        return output, sum(losses)
-
-    def _begin_update(self, states, tokvecs, drop=0.):
-        nr_class = self.moves.n_moves
-        attr_names = numpy.zeros((2,), dtype='i')
-        attr_names[0] = TAG
-        attr_names[1] = DEP
-
-        features = self._get_features(states, tokvecs, attr_names)
-        scores, finish_update = self.model.begin_update(features, drop=drop)
-        assert scores.shape[0] == len(states), (len(states), scores.shape)
-        assert len(scores.shape) == 2
-        is_valid = self.model.ops.allocate((len(states), nr_class), dtype='i')
-        self._validate_batch(is_valid, states)
-        softmaxed = self.model.ops.softmax(scores)
-        softmaxed *= is_valid
-        softmaxed /= softmaxed.sum(axis=1).reshape((softmaxed.shape[0], 1))
-        def backward(golds, sgd=None, losses=[], force_gold=False):
-            nonlocal softmaxed
-            costs = self.model.ops.allocate((len(states), nr_class), dtype='f')
-            d_scores = self.model.ops.allocate((len(states), nr_class), dtype='f')
-
-            self._cost_batch(costs, is_valid, states, golds)
-            self._set_gradient(d_scores, scores, is_valid, costs)
-            losses.append(self.model.ops.xp.abs(d_scores).sum())
-            if force_gold:
-                softmaxed *= costs <= 0
-            return finish_update(d_scores, sgd=sgd)
-        return softmaxed, backward
-
-    def _init_states(self, docs):
-        states = []
-        cdef Doc doc
-        cdef StateClass state
-        for i, doc in enumerate(docs):
-            state = StateClass.init(doc.c, doc.length)
-            self.moves.initialize_state(state.c)
-            states.append(state)
-        return states
-
-    def _get_features(self, states, all_tokvecs, attr_names,
-            nF=1, nB=0, nS=2, nL=2, nR=2):
-        n_tokens = states[0].nr_context_tokens(nF, nB, nS, nL, nR)
-        vector_length = all_tokvecs[0].shape[1]
-        cpu_tokens = numpy.zeros((len(states), n_tokens), dtype='int32')
-        features = numpy.zeros((len(states), n_tokens, attr_names.shape[0]), dtype='uint64')
-        tokvecs = self.model.ops.allocate((len(states), n_tokens, vector_length), dtype='f')
-        for i, state in enumerate(states):
-            state.set_context_tokens(cpu_tokens[i], nF, nB, nS, nL, nR)
-        for i in range(len(states)):
-            for j, tok_i in enumerate(cpu_tokens[i]):
-                if tok_i >= 0:
-                    tokvecs[i, j] = all_tokvecs[i][tok_i]
-        return (cpu_tokens, self.model.ops.asarray(features), tokvecs)
-
-    def _validate_batch(self, int[:, ::1] is_valid, states):
-        cdef StateClass state
-        cdef int i
-        for i, state in enumerate(states):
-            self.moves.set_valid(&is_valid[i, 0], state.c)
-
-    def _cost_batch(self, float[:, ::1] costs, int[:, ::1] is_valid,
-            states, golds):
-        cdef int i
-        cdef StateClass state
-        cdef GoldParse gold
-        for i, (state, gold) in enumerate(zip(states, golds)):
-            self.moves.set_costs(&is_valid[i, 0], &costs[i, 0], state, gold)
-
-    def _transition_batch(self, states, docs, scores):
-        cdef StateClass state
-        cdef int guess
-        for state, doc, guess in zip(states, docs, scores.argmax(axis=1)):
-            action = self.moves.c[guess]
-            orths = [t.lex.orth for t in state.c._sent[:state.c.length]]
-            words = [doc.vocab.strings[w] for w in orths]
-            if not action.is_valid(state.c, action.label):
-                ValueError("Invalid action", scores)
-            action.do(state.c, action.label)
-
-    def _set_gradient(self, gradients, scores, is_valid, costs):
-        """Do multi-label log loss"""
-        cdef double Z, gZ, max_, g_max
-        n = gradients.shape[0]
-        scores = scores * is_valid
-        g_scores = scores * is_valid * (costs <= 0.)
-        exps = self.model.ops.xp.exp(scores - scores.max(axis=1).reshape((n, 1)))
-        exps *= is_valid
-        g_exps = self.model.ops.xp.exp(g_scores - g_scores.max(axis=1).reshape((n, 1)))
-        g_exps *= costs <= 0.
-        g_exps *= is_valid
-        gradients[:] = exps / exps.sum(axis=1).reshape((n, 1))
-        gradients -= g_exps / g_exps.sum(axis=1).reshape((n, 1))
+            states, offsets, golds, d_tokens = zip(*todo)
+            scores, finish_update = model.begin_update((states, offsets))
+            (token_ids, d_state_features) = finish_update(golds, sgd=sgd)
+            for i, token_ids in enumerate(token_ids):
+                d_tokens[i][token_ids] += d_state_features[i]
+            transition_batch(self.moves, states, scores)
+        return output
 
     def step_through(self, Doc doc, GoldParse gold=None):
         """
@@ -366,6 +353,50 @@ cdef class Parser:
                 self.cfg.setdefault('extra_labels', []).append(label)
 
 
+def _begin_update(self, model, states, tokvecs, drop=0.):
+    nr_class = self.moves.n_moves
+    attr_names = self.model.ops.allocate((2,), dtype='i')
+    attr_names[0] = TAG
+    attr_names[1] = DEP
+
+    features = self._get_features(states, tokvecs, attr_names)
+    scores, finish_update = self.model.begin_update(features, drop=drop)
+    assert scores.shape[0] == len(states), (len(states), scores.shape)
+    assert len(scores.shape) == 2
+    is_valid = self.model.ops.allocate((len(states), nr_class), dtype='i')
+    self._validate_batch(is_valid, states)
+    softmaxed = self.model.ops.softmax(scores)
+    softmaxed *= is_valid
+    softmaxed /= softmaxed.sum(axis=1).reshape((softmaxed.shape[0], 1))
+    def backward(golds, sgd=None, losses=[], force_gold=False):
+        nonlocal softmaxed
+        costs = self.model.ops.allocate((len(states), nr_class), dtype='f')
+        d_scores = self.model.ops.allocate((len(states), nr_class), dtype='f')
+
+        self._cost_batch(costs, is_valid, states, golds)
+        self._set_gradient(d_scores, scores, is_valid, costs)
+        losses.append(numpy.abs(d_scores).sum())
+        if force_gold:
+            softmaxed *= costs <= 0
+        return finish_update(d_scores, sgd=sgd)
+    return softmaxed, backward
+
+
+def _get_features(self, states, all_tokvecs, attr_names,
+        nF=1, nB=0, nS=2, nL=2, nR=2):
+    n_tokens = states[0].nr_context_tokens(nF, nB, nS, nL, nR)
+    vector_length = all_tokvecs[0].shape[1]
+    tokens = self.model.ops.allocate((len(states), n_tokens), dtype='int32')
+    features = self.model.ops.allocate((len(states), n_tokens, attr_names.shape[0]), dtype='uint64')
+    tokvecs = self.model.ops.allocate((len(states), n_tokens, vector_length), dtype='f')
+    for i, state in enumerate(states):
+        state.set_context_tokens(tokens[i], nF, nB, nS, nL, nR)
+        state.set_attributes(features[i], tokens[i], attr_names)
+        state.set_token_vectors(tokvecs[i], all_tokvecs[i], tokens[i])
+    return (tokens, features, tokvecs)
+
+
+
 cdef int dropout(FeatureC* feats, int nr_feat, float prob) except -1:
     if prob <= 0 or prob >= 1.:
         return 0