def build_model(nr_class, width, depth, conv_depth, **kwargs): with Model.define_operators({"|": concatenate, ">>": chain, "**": clone}): embed = (HashEmbed(width, 5000, column=1) | StaticVectors("spacy_pretrained_vectors", width, column=5) | HashEmbed(width // 2, 750, column=2) | HashEmbed(width // 2, 750, column=3) | HashEmbed(width // 2, 750, column=4)) >> LN(Maxout(width)) sent2vec = (flatten_add_lengths >> with_getitem( 0, embed >> Residual(ExtractWindow(nW=1) >> LN(Maxout(width)))** conv_depth, ) >> ParametricAttention(width) >> Pooling(sum_pool) >> Residual( LN(Maxout(width)))**depth) model = ( foreach(sent2vec, drop_factor=2.0) >> flatten_add_lengths # This block would allow the model to learn some cross-sentence # features. It's not useful on this problem. It might make more # sense to use a BiLSTM here, following Liang et al (2016). # >> with_getitem(0, # Residual(ExtractWindow(nW=1) >> LN(Maxout(width))) ** conv_depth # ) >> ParametricAttention(width, hard=False) >> Pooling(sum_pool) >> Residual(LN(Maxout(width)))**depth >> Softmax(nr_class)) model.lsuv = False return model
def build_text_classifier(nr_class, width=64, **cfg): nr_vector = cfg.get('nr_vector', 5000) pretrained_dims = cfg.get('pretrained_dims', 0) with Model.define_operators({ '>>': chain, '+': add, '|': concatenate, '**': clone }): if cfg.get('low_data') and pretrained_dims: model = (SpacyVectors >> flatten_add_lengths >> with_getitem( 0, Affine(width, pretrained_dims)) >> ParametricAttention(width) >> Pooling(sum_pool) >> Residual(ReLu(width, width))**2 >> zero_init( Affine(nr_class, width, drop_factor=0.0)) >> logistic) return model lower = HashEmbed(width, nr_vector, column=1) prefix = HashEmbed(width // 2, nr_vector, column=2) suffix = HashEmbed(width // 2, nr_vector, column=3) shape = HashEmbed(width // 2, nr_vector, column=4) trained_vectors = (FeatureExtracter( [ORTH, LOWER, PREFIX, SUFFIX, SHAPE, ID]) >> with_flatten( uniqued((lower | prefix | suffix | shape) >> LN( Maxout(width, width + (width // 2) * 3)), column=0))) if pretrained_dims: static_vectors = ( SpacyVectors >> with_flatten(Affine(width, pretrained_dims))) # TODO Make concatenate support lists vectors = concatenate_lists(trained_vectors, static_vectors) vectors_width = width * 2 else: vectors = trained_vectors vectors_width = width static_vectors = None cnn_model = ( vectors >> with_flatten( LN(Maxout(width, vectors_width)) >> Residual( (ExtractWindow(nW=1) >> LN(Maxout(width, width * 3))))**2, pad=2) >> flatten_add_lengths >> ParametricAttention(width) >> Pooling(sum_pool) >> Residual(zero_init(Maxout(width, width))) >> zero_init(Affine(nr_class, width, drop_factor=0.0))) linear_model = ( _preprocess_doc >> LinearModel(nr_class, drop_factor=0.)) model = ((linear_model | cnn_model) >> zero_init( Affine(nr_class, nr_class * 2, drop_factor=0.0)) >> logistic) model.nO = nr_class model.lsuv = False return model
def build_nel_encoder(embed_width, hidden_width, ner_types, **cfg): if "entity_width" not in cfg: raise ValueError(Errors.E144.format(param="entity_width")) conv_depth = cfg.get("conv_depth", 2) cnn_maxout_pieces = cfg.get("cnn_maxout_pieces", 3) pretrained_vectors = cfg.get("pretrained_vectors", None) context_width = cfg.get("entity_width") with Model.define_operators({">>": chain, "**": clone}): # context encoder tok2vec = Tok2Vec( width=hidden_width, embed_size=embed_width, pretrained_vectors=pretrained_vectors, cnn_maxout_pieces=cnn_maxout_pieces, subword_features=True, conv_depth=conv_depth, bilstm_depth=0, ) model = ( tok2vec >> flatten_add_lengths >> Pooling(mean_pool) >> Residual( zero_init(Maxout(hidden_width, hidden_width))) >> zero_init( Affine(context_width, hidden_width, drop_factor=0.0))) model.tok2vec = tok2vec model.nO = context_width return model
def build_model(nr_class, width, **kwargs): with Model.define_operators({'|': concatenate, '>>': chain, '**': clone}): model = (FeatureExtracter([ORTH]) >> flatten_add_lengths >> with_getitem(0, uniqued(HashEmbed(width, 10000, column=0))) >> Pooling(mean_pool) >> Softmax(nr_class)) model.lsuv = False return model
def build_model(nr_class, width, depth, conv_depth, **kwargs): with Model.define_operators({'|': concatenate, '>>': chain, '**': clone}): embed = ((HashEmbed(width, 5000, column=1) | StaticVectors('spacy_pretrained_vectors', width, column=5) | HashEmbed(width // 2, 750, column=2) | HashEmbed(width // 2, 750, column=3) | HashEmbed(width // 2, 750, column=4)) >> LN(Maxout(width))) sent2vec = (flatten_add_lengths >> with_getitem( 0, embed >> Residual(ExtractWindow(nW=1) >> LN(Maxout(width)))** conv_depth) >> ParametricAttention(width) >> Pooling(sum_pool) >> Residual(LN(Maxout(width)))**depth) model = (foreach(sent2vec, drop_factor=2.0) >> flatten_add_lengths >> ParametricAttention(width, hard=False) >> Pooling(sum_pool) >> Residual(LN(Maxout(width)))**depth >> Softmax(nr_class)) model.lsuv = False return model
def softmax_last_hidden(nr_class, *, exclusive_classes=True, **cfg): """Select features from the class-vectors from the last hidden state, mean-pool them, and softmax to produce one vector per document. The gradients of the class vectors are incremented in the backward pass, to allow fine-tuning. """ width = cfg["token_vector_width"] return chain(get_pytt_last_hidden, flatten_add_lengths, Pooling(mean_pool), Softmax(nr_class, width))
def sigmoid_last_hidden(nr_class, *, exclusive_classes=False, **cfg): width = cfg["token_vector_width"] return chain( get_last_hidden, flatten_add_lengths, Pooling(mean_pool), zero_init(Affine(nr_class, width, drop_factor=0.0)), logistic, )
def build_textcat_model(tok2vec, nr_class, width): from thinc.v2v import Model, Softmax from thinc.api import flatten_add_lengths, chain from thinc.t2v import Pooling, mean_pool with Model.define_operators({">>": chain}): model = (tok2vec >> flatten_add_lengths >> Pooling(mean_pool) >> Softmax(nr_class, width)) model.tok2vec = chain(tok2vec, flatten) return model
def fine_tune_pooler_output(nr_class, *, exclusive_classes=True, **cfg): """Select features from the class-vectors from the last hidden state, softmax them, and then mean-pool them to produce one feature per vector. The gradients of the class vectors are incremented in the backward pass, to allow fine-tuning. """ return chain( get_pytt_pooler_output, flatten_add_lengths, with_getitem(0, Softmax(nr_class, cfg["token_vector_width"])), Pooling(mean_pool), )
def softmax_pooler_output(nr_class, *, exclusive_classes=True, **cfg): """Select features from the pooler output, (if necessary) mean-pool them to produce one vector per item, and then softmax them. The gradients of the class vectors are incremented in the backward pass, to allow fine-tuning. """ return chain( get_pooler_output, flatten_add_lengths, with_getitem(0, Softmax(nr_class, cfg["token_vector_width"])), Pooling(mean_pool), )
def build_textcat_model(tok2vec, nr_class, width): from thinc.v2v import Model, Softmax, Maxout from thinc.api import flatten_add_lengths, chain from thinc.t2v import Pooling, sum_pool, mean_pool, max_pool from thinc.misc import Residual, LayerNorm from spacy._ml import logistic, zero_init with Model.define_operators({">>": chain}): model = (tok2vec >> flatten_add_lengths >> Pooling(mean_pool) >> Softmax(nr_class, width)) model.tok2vec = tok2vec return model
def build_model(nr_class, width, depth, conv_depth, vectors_name, **kwargs): with Model.define_operators({"|": concatenate, ">>": chain, "**": clone}): embed = (HashEmbed(width, 5000, column=1) | StaticVectors(vectors_name, width, column=5) | HashEmbed(width // 2, 750, column=2) | HashEmbed(width // 2, 750, column=3) | HashEmbed(width // 2, 750, column=4)) >> LN(Maxout(width)) sent2vec = (with_flatten(embed) >> Residual( prepare_self_attention(Affine(width * 3, width), nM=width, nH=4) >> MultiHeadedAttention() >> with_flatten( Maxout(width, width, pieces=3))) >> flatten_add_lengths >> ParametricAttention(width, hard=False) >> Pooling(mean_pool) >> Residual(LN(Maxout(width)))) model = (foreach(sent2vec, drop_factor=2.0) >> Residual( prepare_self_attention(Affine(width * 3, width), nM=width, nH=4) >> MultiHeadedAttention() >> with_flatten(LN(Affine(width, width)))) >> flatten_add_lengths >> ParametricAttention( width, hard=False) >> Pooling(mean_pool) >> Residual( LN(Maxout(width)))**2 >> Softmax(nr_class)) model.lsuv = False return model
def softmax_tanh_class_vector(nr_class, *, exclusive_classes=True, **cfg): """Select features from the class-vectors from the last hidden state, mean-pool them, apply a tanh-activated hidden layer, and then softmax-activated output layer to produce one vector per document. The gradients of the class vectors are incremented in the backward pass, to allow fine-tuning. """ width = cfg["token_vector_width"] return chain( get_class_tokens, flatten_add_lengths, with_getitem(0, chain(Affine(width, width), tanh)), Pooling(mean_pool), Softmax(nr_class, width), )
def build_simple_cnn_text_classifier(tok2vec, nr_class, exclusive_classes=False, **cfg): """ Build a simple CNN text classifier, given a token-to-vector model as inputs. If exclusive_classes=True, a softmax non-linearity is applied, so that the outputs sum to 1. If exclusive_classes=False, a logistic non-linearity is applied instead, so that outputs are in the range [0, 1]. """ with Model.define_operators({">>": chain}): if exclusive_classes: output_layer = Softmax(nr_class, tok2vec.nO) else: output_layer = zero_init(Affine(nr_class, tok2vec.nO, drop_factor=0.0)) >> logistic model = tok2vec >> flatten_add_lengths >> Pooling(mean_pool) >> output_layer model.tok2vec = chain(tok2vec, flatten) model.nO = nr_class return model
def build_nel_encoder(embed_width, hidden_width, ner_types, **cfg): # TODO proper error if "entity_width" not in cfg: raise ValueError("entity_width not found") if "context_width" not in cfg: raise ValueError("context_width not found") conv_depth = cfg.get("conv_depth", 2) cnn_maxout_pieces = cfg.get("cnn_maxout_pieces", 3) pretrained_vectors = cfg.get( "pretrained_vectors") # self.nlp.vocab.vectors.name context_width = cfg.get("context_width") entity_width = cfg.get("entity_width") with Model.define_operators({">>": chain, "**": clone}): model = ( Affine(entity_width, entity_width + context_width + 1 + ner_types) >> Affine(1, entity_width, drop_factor=0.0) >> logistic) # context encoder tok2vec = (Tok2Vec( width=hidden_width, embed_size=embed_width, pretrained_vectors=pretrained_vectors, cnn_maxout_pieces=cnn_maxout_pieces, subword_features=True, conv_depth=conv_depth, bilstm_depth=0, ) >> flatten_add_lengths >> Pooling(mean_pool) >> Residual( zero_init(Maxout(hidden_width, hidden_width))) >> zero_init( Affine(context_width, hidden_width))) model.tok2vec = tok2vec model.tok2vec = tok2vec model.tok2vec.nO = context_width model.nO = 1 return model
def build_text_classifier(nr_class, width=64, **cfg): depth = cfg.get("depth", 2) nr_vector = cfg.get("nr_vector", 5000) pretrained_dims = cfg.get("pretrained_dims", 0) with Model.define_operators({ ">>": chain, "+": add, "|": concatenate, "**": clone }): if cfg.get("low_data") and pretrained_dims: model = (SpacyVectors >> flatten_add_lengths >> with_getitem( 0, Affine(width, pretrained_dims)) >> ParametricAttention(width) >> Pooling(sum_pool) >> Residual(ReLu(width, width))**2 >> zero_init( Affine(nr_class, width, drop_factor=0.0)) >> logistic) return model lower = HashEmbed(width, nr_vector, column=1) prefix = HashEmbed(width // 2, nr_vector, column=2) suffix = HashEmbed(width // 2, nr_vector, column=3) shape = HashEmbed(width // 2, nr_vector, column=4) trained_vectors = FeatureExtracter( [ORTH, LOWER, PREFIX, SUFFIX, SHAPE, ID]) >> with_flatten( uniqued( (lower | prefix | suffix | shape) >> LN( Maxout(width, width + (width // 2) * 3)), column=0, )) if pretrained_dims: static_vectors = SpacyVectors >> with_flatten( Affine(width, pretrained_dims)) # TODO Make concatenate support lists vectors = concatenate_lists(trained_vectors, static_vectors) vectors_width = width * 2 else: vectors = trained_vectors vectors_width = width static_vectors = None tok2vec = vectors >> with_flatten( LN(Maxout(width, vectors_width)) >> Residual( (ExtractWindow(nW=1) >> LN(Maxout(width, width * 3))))**depth, pad=depth, ) cnn_model = ( tok2vec >> flatten_add_lengths >> ParametricAttention(width) >> Pooling(sum_pool) >> Residual(zero_init(Maxout(width, width))) >> zero_init(Affine(nr_class, width, drop_factor=0.0))) linear_model = build_bow_text_classifier(nr_class, ngram_size=cfg.get( "ngram_size", 1), exclusive_classes=False) if cfg.get("exclusive_classes"): output_layer = Softmax(nr_class, nr_class * 2) else: output_layer = (zero_init( Affine(nr_class, nr_class * 2, drop_factor=0.0)) >> logistic) model = (linear_model | cnn_model) >> output_layer model.tok2vec = chain(tok2vec, flatten) model.nO = nr_class model.lsuv = False return model
def main(dataset='quora', width=50, depth=2, min_batch_size=1, max_batch_size=512, dropout=0.2, dropout_decay=0.0, pooling="mean+max", nb_epoch=5, pieces=3, L2=0.0, use_gpu=False, out_loc=None, quiet=False, job_id=None, ws_api_url=None, rest_api_url=None): global CTX if job_id is not None: CTX = neptune.Context() width = CTX.params.width L2 = CTX.params.L2 nb_epoch = CTX.params.nb_epoch depth = CTX.params.depth max_batch_size = CTX.params.max_batch_size cfg = dict(locals()) if out_loc: out_loc = Path(out_loc) if not out_loc.parent.exists(): raise IOError("Can't open output location: %s" % out_loc) print(cfg) if pooling == 'mean+max': pool_layer = Pooling(mean_pool, max_pool) elif pooling == "mean": pool_layer = mean_pool elif pooling == "max": pool_layer = max_pool else: raise ValueError("Unrecognised pooling", pooling) print("Load spaCy") nlp = get_spacy('en') if use_gpu: Model.ops = CupyOps() print("Construct model") # Bind operators for the scope of the block: # * chain (>>): Compose models in a 'feed forward' style, # i.e. chain(f, g)(x) -> g(f(x)) # * clone (**): Create n copies of a model, and chain them, i.e. # (f ** 3)(x) -> f''(f'(f(x))), where f, f' and f'' have distinct weights. # * concatenate (|): Merge the outputs of two models into a single vector, # i.e. (f|g)(x) -> hstack(f(x), g(x)) Model.lsuv = True #Model.ops = CupyOps() with Model.define_operators({ '>>': chain, '**': clone, '|': concatenate, '+': add }): mwe_encode = ExtractWindow(nW=1) >> BN( Maxout(width, drop_factor=0.0, pieces=pieces)) sent2vec = ( # List[spacy.token.Doc]{B} flatten_add_lengths # : (ids{T}, lengths{B}) >> with_getitem( 0, #(StaticVectors('en', width) HashEmbed(width, 3000) #+ HashEmbed(width, 3000)) #>> Residual(mwe_encode ** 2) ) # : word_ids{T} >> Pooling(mean_pool, max_pool) #>> Residual(BN(Maxout(width*2, pieces=pieces), nO=width*2)**2) >> Maxout(width * 2, pieces=pieces, drop_factor=0.0) >> logistic) model = Siamese(sent2vec, CauchySimilarity(width * 2)) print("Read and parse data: %s" % dataset) if dataset == 'quora': train, dev = datasets.quora_questions() elif dataset == 'snli': train, dev = datasets.snli() elif dataset == 'stackxc': train, dev = datasets.stack_exchange() elif dataset in ('quora+snli', 'snli+quora'): train, dev = datasets.quora_questions() train2, dev2 = datasets.snli() train.extend(train2) dev.extend(dev2) else: raise ValueError("Unknown dataset: %s" % dataset) get_ids = get_word_ids(Model.ops) train_X, train_y = preprocess(model.ops, nlp, train, get_ids) dev_X, dev_y = preprocess(model.ops, nlp, dev, get_ids) with model.begin_training(train_X[:10000], train_y[:10000], **cfg) as (trainer, optimizer): # Pass a callback to print progress. Give it all the local scope, # because why not? trainer.each_epoch.append(track_progress(**locals())) trainer.batch_size = min_batch_size batch_size = float(min_batch_size) print("Accuracy before training", model.evaluate_logloss(dev_X, dev_y)) print("Train") global epoch_train_acc n_iter = 0 for X, y in trainer.iterate(train_X, train_y, progress_bar=not quiet): # Slightly useful trick: Decay the dropout as training proceeds. yh, backprop = model.begin_update(X, drop=trainer.dropout) assert yh.shape == y.shape, (yh.shape, y.shape) assert (yh >= 0.).all(), yh train_acc = ((yh >= 0.5) == (y >= 0.5)).sum() loss = model.ops.xp.abs(yh - y).mean() epoch_train_acc += train_acc backprop(yh - y, optimizer) n_iter += 1 # Slightly useful trick: start with low batch size, accelerate. trainer.batch_size = min(int(batch_size), max_batch_size) batch_size *= 1.001 if out_loc: out_loc = Path(out_loc) print('Saving to', out_loc) with out_loc.open('wb') as file_: pickle.dump(model, file_, -1)
def Model(cls, length): return Siamese(Pooling(max_pool, mean_pool), CauchySimilarity(length))