annif.backend.nn_ensemble.NNEnsembleBackend._suggest_batch_with_sources() - Code Metrics - Inspection of "Batch suggest in NN ensemble" - NatLibFi/Annif - Measure and Improve Code Quality continuously with Scrutinizer

Passed

Pull Request — master (#677)

by Juho

created 2023-03-03 10:38 UTC

NNEnsembleBackend._suggest_batch_with_sources() A

↳ Parent: annif.backend.nn_ensemble

Complexity

Conditions

Size

Total Lines	17
Code Lines	13

Duplication

Lines	0
Ratio	0 %

Importance

Changes

Metric	Value
cc	2
eloc	13
nop	3
dl	0
loc	17
rs	9.75
c	0
b	0
f	0

"""Neural network based ensemble backend that combines results from multiple
projects."""


import os.path
import shutil
from io import BytesIO

import joblib
import lmdb
import numpy as np
import tensorflow.keras.backend as K
from scipy.sparse import csc_matrix, csr_matrix
from tensorflow.keras.layers import Add, Dense, Dropout, Flatten, Input, Layer
from tensorflow.keras.models import Model, load_model
from tensorflow.keras.utils import Sequence

import annif.corpus
import annif.parallel
import annif.util
from annif.exception import NotInitializedException, NotSupportedException
from annif.suggestion import VectorSuggestionResult

from . import backend, ensemble


def idx_to_key(idx):
    """convert an integer index to a binary key for use in LMDB"""
    return b"%08d" % idx


def key_to_idx(key):
    """convert a binary LMDB key to an integer index"""
    return int(key)


class LMDBSequence(Sequence):
    """A sequence of samples stored in a LMDB database."""

    def __init__(self, txn, batch_size):
        self._txn = txn
        cursor = txn.cursor()
        if cursor.last():
            # Counter holds the number of samples in the database
            self._counter = key_to_idx(cursor.key()) + 1
        else:  # empty database
            self._counter = 0
        self._batch_size = batch_size

    def add_sample(self, inputs, targets):
        # use zero-padded 8-digit key
        key = idx_to_key(self._counter)
        self._counter += 1
        # convert the sample into a sparse matrix and serialize it as bytes
        sample = (csc_matrix(inputs), csr_matrix(targets))
        buf = BytesIO()
        joblib.dump(sample, buf)
        buf.seek(0)
        self._txn.put(key, buf.read())

    def __getitem__(self, idx):
        """get a particular batch of samples"""
        cursor = self._txn.cursor()
        first_key = idx * self._batch_size
        cursor.set_key(idx_to_key(first_key))
        input_arrays = []
        target_arrays = []
        for key, value in cursor.iternext():
            if key_to_idx(key) >= (first_key + self._batch_size):
                break
            input_csr, target_csr = joblib.load(BytesIO(value))
            input_arrays.append(input_csr.toarray())
            target_arrays.append(target_csr.toarray().flatten())
        return np.array(input_arrays), np.array(target_arrays)

    def __len__(self):
        """return the number of available batches"""
        return int(np.ceil(self._counter / self._batch_size))


class MeanLayer(Layer):
    """Custom Keras layer that calculates mean values along the 2nd axis."""

    def call(self, inputs):
        return K.mean(inputs, axis=2)


class NNEnsembleBackend(backend.AnnifLearningBackend, ensemble.BaseEnsembleBackend):
    """Neural network ensemble backend that combines results from multiple
    projects"""

    name = "nn_ensemble"

    MODEL_FILE = "nn-model.h5"
    LMDB_FILE = "nn-train.mdb"

    DEFAULT_PARAMETERS = {
        "nodes": 100,
        "dropout_rate": 0.2,
        "optimizer": "adam",
        "epochs": 10,
        "learn-epochs": 1,
        "lmdb_map_size": 1024 * 1024 * 1024,
    }

    # defaults for uninitialized instances
    _model = None

    def default_params(self):
        params = backend.AnnifBackend.DEFAULT_PARAMETERS.copy()
        params.update(self.DEFAULT_PARAMETERS)
        return params

    def initialize(self, parallel=False):
        super().initialize(parallel)
        if self._model is not None:
            return  # already initialized
        if parallel:
            # Don't load TF model just before parallel execution,
            # since it won't work after forking worker processes
            return
        model_filename = os.path.join(self.datadir, self.MODEL_FILE)
        if not os.path.exists(model_filename):
            raise NotInitializedException(
                "model file {} not found".format(model_filename),
                backend_id=self.backend_id,
            )
        self.debug("loading Keras model from {}".format(model_filename))
        self._model = load_model(
            model_filename, custom_objects={"MeanLayer": MeanLayer}
        )

    def _suggest_batch_with_sources(self, texts, sources):
        hit_sets_from_sources = []
        for project_id, weight in sources:
            source_project = self.project.registry.get_project(project_id)
            hit_sets = source_project.suggest(texts)
            norm_hit_sets = [
                self._normalize_hits(hits, source_project) for hits in hit_sets
            ]
            hit_sets_from_sources.append(
                [
                    annif.suggestion.WeightedSuggestion(
                        hits=norm_hits, weight=weight, subjects=source_project.subjects
                    )
                    for norm_hits in norm_hit_sets
                ]
            )
        return hit_sets_from_sources

    def _merge_hits_from_sources(self, hit_sets_from_sources, params):
        score_vectors = np.array(
            [
                [
                    np.sqrt(hits.as_vector(len(subjects)))
                    * weight
                    * len(hit_sets_from_sources)
                    for hits, weight, subjects in hits_from_sources
                ]
                for hits_from_sources in hit_sets_from_sources
            ],
            dtype=np.float32,
        ).transpose(1, 2, 0)
        results = self._model(score_vectors).numpy()
        return [VectorSuggestionResult(res) for res in results]

    def _suggest_batch(self, texts, params):
        sources = annif.util.parse_sources(params["sources"])
        hit_sets_from_sources = self._suggest_batch_with_sources(texts, sources)
        merged_hits = self._merge_hits_from_sources(hit_sets_from_sources, params)
        return merged_hits

    def _create_model(self, sources):
        self.info("creating NN ensemble model")

        inputs = Input(shape=(len(self.project.subjects), len(sources)))

        flat_input = Flatten()(inputs)
        drop_input = Dropout(rate=float(self.params["dropout_rate"]))(flat_input)
        hidden = Dense(int(self.params["nodes"]), activation="relu")(drop_input)
        drop_hidden = Dropout(rate=float(self.params["dropout_rate"]))(hidden)
        delta = Dense(
            len(self.project.subjects),
            kernel_initializer="zeros",
            bias_initializer="zeros",
        )(drop_hidden)

        mean = MeanLayer()(inputs)

        predictions = Add()([mean, delta])

        self._model = Model(inputs=inputs, outputs=predictions)
        self._model.compile(
            optimizer=self.params["optimizer"],
            loss="binary_crossentropy",
            metrics=["top_k_categorical_accuracy"],
        )
        if "lr" in self.params:
            self._model.optimizer.learning_rate.assign(float(self.params["lr"]))

        summary = []
        self._model.summary(print_fn=summary.append)
        self.debug("Created model: \n" + "\n".join(summary))

    def _train(self, corpus, params, jobs=0):
        sources = annif.util.parse_sources(self.params["sources"])
        self._create_model(sources)
        self._fit_model(
            corpus,
            epochs=int(params["epochs"]),
            lmdb_map_size=int(params["lmdb_map_size"]),
            n_jobs=jobs,
        )

    def _corpus_to_vectors(self, corpus, seq, n_jobs):
        # pass corpus through all source projects
        sources = dict(annif.util.parse_sources(self.params["sources"]))

        # initialize the source projects before forking, to save memory
        self.info(f"Initializing source projects: {', '.join(sources.keys())}")
        for project_id in sources.keys():
            project = self.project.registry.get_project(project_id)
            project.initialize(parallel=True)

        psmap = annif.parallel.ProjectSuggestMap(
            self.project.registry,
            list(sources.keys()),
            backend_params=None,
            limit=None,
            threshold=0.0,
        )

        jobs, pool_class = annif.parallel.get_pool(n_jobs)

        self.info("Processing training documents...")
        with pool_class(jobs) as pool:
            for hits, subject_set in pool.imap_unordered(
                psmap.suggest, corpus.documents
            ):
                doc_scores = []
                for project_id, p_hits in hits.items():
                    vector = p_hits.as_vector(len(self.project.subjects))
                    doc_scores.append(
                        np.sqrt(vector) * sources[project_id] * len(sources)
                    )
                score_vector = np.array(doc_scores, dtype=np.float32).transpose()
                true_vector = subject_set.as_vector(len(self.project.subjects))
                seq.add_sample(score_vector, true_vector)

    def _open_lmdb(self, cached, lmdb_map_size):
        lmdb_path = os.path.join(self.datadir, self.LMDB_FILE)
        if not cached and os.path.exists(lmdb_path):
            shutil.rmtree(lmdb_path)
        return lmdb.open(lmdb_path, map_size=lmdb_map_size, writemap=True)

    def _fit_model(self, corpus, epochs, lmdb_map_size, n_jobs=1):
        env = self._open_lmdb(corpus == "cached", lmdb_map_size)
        if corpus != "cached":
            if corpus.is_empty():
                raise NotSupportedException(
                    "Cannot train nn_ensemble project with no documents"
                )
            with env.begin(write=True, buffers=True) as txn:
                seq = LMDBSequence(txn, batch_size=32)
                self._corpus_to_vectors(corpus, seq, n_jobs)
        else:
            self.info("Reusing cached training data from previous run.")
        # fit the model using a read-only view of the LMDB
        self.info("Training neural network model...")
        with env.begin(buffers=True) as txn:
            seq = LMDBSequence(txn, batch_size=32)
            self._model.fit(seq, verbose=True, epochs=epochs)

        annif.util.atomic_save(self._model, self.datadir, self.MODEL_FILE)

    def _learn(self, corpus, params):
        self.initialize()
        self._fit_model(
            corpus, int(params["learn-epochs"]), int(params["lmdb_map_size"])
        )


1			"""Neural network based ensemble backend that combines results from multiple
2			projects."""
3
4
5			import os.path
6			import shutil
7			from io import BytesIO
8
9			import joblib
10			import lmdb
11			import numpy as np
12			import tensorflow.keras.backend as K
13			from scipy.sparse import csc_matrix, csr_matrix
14			from tensorflow.keras.layers import Add, Dense, Dropout, Flatten, Input, Layer
15			from tensorflow.keras.models import Model, load_model
16			from tensorflow.keras.utils import Sequence
17
18			import annif.corpus
19			import annif.parallel
20			import annif.util
21			from annif.exception import NotInitializedException, NotSupportedException
22			from annif.suggestion import VectorSuggestionResult
23
24			from . import backend, ensemble
25
26
27			def idx_to_key(idx):
28			"""convert an integer index to a binary key for use in LMDB"""
29			return b"%08d" % idx
30
31
32			def key_to_idx(key):
33			"""convert a binary LMDB key to an integer index"""
34			return int(key)
35
36
37			class LMDBSequence(Sequence):
38			"""A sequence of samples stored in a LMDB database."""
39
40			def __init__(self, txn, batch_size):
41			self._txn = txn
42			cursor = txn.cursor()
43			if cursor.last():
44			# Counter holds the number of samples in the database
45			self._counter = key_to_idx(cursor.key()) + 1
46			else: # empty database
47			self._counter = 0
48			self._batch_size = batch_size
49
50			def add_sample(self, inputs, targets):
51			# use zero-padded 8-digit key
52			key = idx_to_key(self._counter)
53			self._counter += 1
54			# convert the sample into a sparse matrix and serialize it as bytes
55			sample = (csc_matrix(inputs), csr_matrix(targets))
56			buf = BytesIO()
57			joblib.dump(sample, buf)
58			buf.seek(0)
59			self._txn.put(key, buf.read())
60
61			def __getitem__(self, idx):
62			"""get a particular batch of samples"""
63			cursor = self._txn.cursor()
64			first_key = idx * self._batch_size
65			cursor.set_key(idx_to_key(first_key))
66			input_arrays = []
67			target_arrays = []
68			for key, value in cursor.iternext():
69			if key_to_idx(key) >= (first_key + self._batch_size):
70			break
71			input_csr, target_csr = joblib.load(BytesIO(value))
72			input_arrays.append(input_csr.toarray())
73			target_arrays.append(target_csr.toarray().flatten())
74			return np.array(input_arrays), np.array(target_arrays)
75
76			def __len__(self):
77			"""return the number of available batches"""
78			return int(np.ceil(self._counter / self._batch_size))
79
80
81			class MeanLayer(Layer):
82			"""Custom Keras layer that calculates mean values along the 2nd axis."""
83
84			def call(self, inputs):
85			return K.mean(inputs, axis=2)
86
87
88			class NNEnsembleBackend(backend.AnnifLearningBackend, ensemble.BaseEnsembleBackend):
89			"""Neural network ensemble backend that combines results from multiple
90			projects"""
91
92			name = "nn_ensemble"
93
94			MODEL_FILE = "nn-model.h5"
95			LMDB_FILE = "nn-train.mdb"
96
97			DEFAULT_PARAMETERS = {
98			"nodes": 100,
99			"dropout_rate": 0.2,
100			"optimizer": "adam",
101			"epochs": 10,
102			"learn-epochs": 1,
103			"lmdb_map_size": 1024 * 1024 * 1024,
104			}
105
106			# defaults for uninitialized instances
107			_model = None
108
109			def default_params(self):
110			params = backend.AnnifBackend.DEFAULT_PARAMETERS.copy()
111			params.update(self.DEFAULT_PARAMETERS)
112			return params
113
114			def initialize(self, parallel=False):
115			super().initialize(parallel)
116			if self._model is not None:
117			return # already initialized
118			if parallel:
119			# Don't load TF model just before parallel execution,
120			# since it won't work after forking worker processes
121			return
122			model_filename = os.path.join(self.datadir, self.MODEL_FILE)
123			if not os.path.exists(model_filename):
124			raise NotInitializedException(
125			"model file {} not found".format(model_filename),
126			backend_id=self.backend_id,
127			)
128			self.debug("loading Keras model from {}".format(model_filename))
129			self._model = load_model(
130			model_filename, custom_objects={"MeanLayer": MeanLayer}
131			)
132
133			def _suggest_batch_with_sources(self, texts, sources):
134			hit_sets_from_sources = []
135			for project_id, weight in sources:
136			source_project = self.project.registry.get_project(project_id)
137			hit_sets = source_project.suggest(texts)
138			norm_hit_sets = [
139			self._normalize_hits(hits, source_project) for hits in hit_sets
140			]
141			hit_sets_from_sources.append(
142			[
143			annif.suggestion.WeightedSuggestion(
144			hits=norm_hits, weight=weight, subjects=source_project.subjects
145			)
146			for norm_hits in norm_hit_sets
147			]
148			)
149			return hit_sets_from_sources
150
151			def _merge_hits_from_sources(self, hit_sets_from_sources, params):
152			score_vectors = np.array(
153			[
154			[
155			np.sqrt(hits.as_vector(len(subjects)))
156			* weight
157			* len(hit_sets_from_sources)
158			for hits, weight, subjects in hits_from_sources
159			]
160			for hits_from_sources in hit_sets_from_sources
161			],
162			dtype=np.float32,
163			).transpose(1, 2, 0)
164			results = self._model(score_vectors).numpy()
165			return [VectorSuggestionResult(res) for res in results]
166
167			def _suggest_batch(self, texts, params):
168			sources = annif.util.parse_sources(params["sources"])
169			hit_sets_from_sources = self._suggest_batch_with_sources(texts, sources)
170			merged_hits = self._merge_hits_from_sources(hit_sets_from_sources, params)
171			return merged_hits
172
173			def _create_model(self, sources):
174			self.info("creating NN ensemble model")
175
176			inputs = Input(shape=(len(self.project.subjects), len(sources)))
177
178			flat_input = Flatten()(inputs)
179			drop_input = Dropout(rate=float(self.params["dropout_rate"]))(flat_input)
180			hidden = Dense(int(self.params["nodes"]), activation="relu")(drop_input)
181			drop_hidden = Dropout(rate=float(self.params["dropout_rate"]))(hidden)
182			delta = Dense(
183			len(self.project.subjects),
184			kernel_initializer="zeros",
185			bias_initializer="zeros",
186			)(drop_hidden)
187
188			mean = MeanLayer()(inputs)
189
190			predictions = Add()([mean, delta])
191
192			self._model = Model(inputs=inputs, outputs=predictions)
193			self._model.compile(
194			optimizer=self.params["optimizer"],
195			loss="binary_crossentropy",
196			metrics=["top_k_categorical_accuracy"],
197			)
198			if "lr" in self.params:
199			self._model.optimizer.learning_rate.assign(float(self.params["lr"]))
200
201			summary = []
202			self._model.summary(print_fn=summary.append)
203			self.debug("Created model: \n" + "\n".join(summary))
204
205			def _train(self, corpus, params, jobs=0):
206			sources = annif.util.parse_sources(self.params["sources"])
207			self._create_model(sources)
208			self._fit_model(
209			corpus,
210			epochs=int(params["epochs"]),
211			lmdb_map_size=int(params["lmdb_map_size"]),
212			n_jobs=jobs,
213			)
214
215			def _corpus_to_vectors(self, corpus, seq, n_jobs):
216			# pass corpus through all source projects
217			sources = dict(annif.util.parse_sources(self.params["sources"]))
218
219			# initialize the source projects before forking, to save memory
220			self.info(f"Initializing source projects: {', '.join(sources.keys())}")
221			for project_id in sources.keys():
222			project = self.project.registry.get_project(project_id)
223			project.initialize(parallel=True)
224
225			psmap = annif.parallel.ProjectSuggestMap(
226			self.project.registry,
227			list(sources.keys()),
228			backend_params=None,
229			limit=None,
230			threshold=0.0,
231			)
232
233			jobs, pool_class = annif.parallel.get_pool(n_jobs)
234
235			self.info("Processing training documents...")
236			with pool_class(jobs) as pool:
237			for hits, subject_set in pool.imap_unordered(
238			psmap.suggest, corpus.documents
239			):
240			doc_scores = []
241			for project_id, p_hits in hits.items():
242			vector = p_hits.as_vector(len(self.project.subjects))
243			doc_scores.append(
244			np.sqrt(vector) * sources[project_id] * len(sources)
245			)
246			score_vector = np.array(doc_scores, dtype=np.float32).transpose()
247			true_vector = subject_set.as_vector(len(self.project.subjects))
248			seq.add_sample(score_vector, true_vector)
249
250			def _open_lmdb(self, cached, lmdb_map_size):
251			lmdb_path = os.path.join(self.datadir, self.LMDB_FILE)
252			if not cached and os.path.exists(lmdb_path):
253			shutil.rmtree(lmdb_path)
254			return lmdb.open(lmdb_path, map_size=lmdb_map_size, writemap=True)
255
256			def _fit_model(self, corpus, epochs, lmdb_map_size, n_jobs=1):
257			env = self._open_lmdb(corpus == "cached", lmdb_map_size)
258			if corpus != "cached":
259			if corpus.is_empty():
260			raise NotSupportedException(
261			"Cannot train nn_ensemble project with no documents"
262			)
263			with env.begin(write=True, buffers=True) as txn:
264			seq = LMDBSequence(txn, batch_size=32)
265			self._corpus_to_vectors(corpus, seq, n_jobs)
266			else:
267			self.info("Reusing cached training data from previous run.")
268			# fit the model using a read-only view of the LMDB
269			self.info("Training neural network model...")
270			with env.begin(buffers=True) as txn:
271			seq = LMDBSequence(txn, batch_size=32)
272			self._model.fit(seq, verbose=True, epochs=epochs)
273
274			annif.util.atomic_save(self._model, self.datadir, self.MODEL_FILE)
275
276			def _learn(self, corpus, params):
277			self.initialize()
278			self._fit_model(
279			corpus, int(params["learn-epochs"]), int(params["lmdb_map_size"])
280			)
281

NatLibFi / Annif

Pull Request — master (#677)

NNEnsembleBackend._suggest_batch_with_sources() A

Complexity

Size

Duplication

Importance

Duplication Side-by-Side

Filter issues like