annif.backend.nn_ensemble.NNEnsembleBackend._fit_model() - Code Metrics - Inspection of "Issue690 use python type hints" - NatLibFi/Annif - Measure and Improve Code Quality continuously with Scrutinizer

Passed

Pull Request — main (#708)

by Juho

created 2023-05-25 11:26 UTC

NNEnsembleBackend._fit_model() B

↳ Parent: annif.backend.nn_ensemble

Complexity

Conditions

Size

Total Lines	25
Code Lines	20

Duplication

Lines	0
Ratio	0 %

Importance

Changes

Metric	Value
cc	5
eloc	20
nop	5
dl	0
loc	25
rs	8.9332
c	0
b	0
f	0

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

import os.path
import shutil
from io import BytesIO
from typing import TYPE_CHECKING, Any, Dict, List, Tuple, Union

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 SuggestionBatch, vector_to_suggestions

from . import backend, ensemble

if TYPE_CHECKING:
    from tensorflow.python.framework.ops import EagerTensor

    from annif.corpus.document import DocumentCorpus


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


def key_to_idx(key: Union[memoryview, bytes]) -> int:
    """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: np.ndarray, targets: np.ndarray) -> None:
        # 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: int) -> Tuple[np.ndarray, np.ndarray]:
        """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) -> int:
        """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: EagerTensor) -> EagerTensor:
        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) -> Dict[str, Any]:
        params = backend.AnnifBackend.DEFAULT_PARAMETERS.copy()
        params.update(self.DEFAULT_PARAMETERS)
        return params

    def initialize(self, parallel: bool = False) -> None:
        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 _merge_source_batches(
        self,
        batch_by_source: Dict[str, SuggestionBatch],
        sources: List[Tuple[str, float]],
        params: Dict[str, Any],
    ) -> SuggestionBatch:
        src_weight = dict(sources)
        score_vectors = np.array(
            [
                [
                    np.sqrt(suggestions.as_vector())
                    * src_weight[project_id]
                    * len(batch_by_source)
                    for suggestions in batch
                ]
                for project_id, batch in batch_by_source.items()
            ],
            dtype=np.float32,
        ).transpose(1, 2, 0)
        prediction = self._model(score_vectors).numpy()
        return SuggestionBatch.from_sequence(
            [
                vector_to_suggestions(row, limit=int(params["limit"]))
                for row in prediction
            ],
            self.project.subjects,
        )

    def _create_model(self, sources: List[Tuple[str, float]]) -> None:
        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: DocumentCorpus,
        params: Dict[str, Any],
        jobs: int = 0,
    ) -> None:
        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: DocumentCorpus,
        seq: LMDBSequence,
        n_jobs: int,
    ) -> None:
        # 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()
                    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: DocumentCorpus,
        epochs: int,
        lmdb_map_size: int,
        n_jobs: int = 1,
    ) -> None:
        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: DocumentCorpus,
        params: Dict[str, Any],
    ) -> None:
        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			from __future__ import annotations
4
5			import os.path
6			import shutil
7			from io import BytesIO
8			from typing import TYPE_CHECKING, Any, Dict, List, Tuple, Union
9
10			import joblib
11			import lmdb
12			import numpy as np
13			import tensorflow.keras.backend as K
14			from scipy.sparse import csc_matrix, csr_matrix
15			from tensorflow.keras.layers import Add, Dense, Dropout, Flatten, Input, Layer
16			from tensorflow.keras.models import Model, load_model
17			from tensorflow.keras.utils import Sequence
18
19			import annif.corpus
20			import annif.parallel
21			import annif.util
22			from annif.exception import NotInitializedException, NotSupportedException
23			from annif.suggestion import SuggestionBatch, vector_to_suggestions
24
25			from . import backend, ensemble
26
27			if TYPE_CHECKING:
28			from tensorflow.python.framework.ops import EagerTensor
29
30			from annif.corpus.document import DocumentCorpus
31
32
33			def idx_to_key(idx: int) -> bytes:
34			"""convert an integer index to a binary key for use in LMDB"""
35			return b"%08d" % idx
36
37
38			def key_to_idx(key: Union[memoryview, bytes]) -> int:
39			"""convert a binary LMDB key to an integer index"""
40			return int(key)
41
42
43			class LMDBSequence(Sequence):
44			"""A sequence of samples stored in a LMDB database."""
45
46			def __init__(self, txn, batch_size):
47			self._txn = txn
48			cursor = txn.cursor()
49			if cursor.last():
50			# Counter holds the number of samples in the database
51			self._counter = key_to_idx(cursor.key()) + 1
52			else: # empty database
53			self._counter = 0
54			self._batch_size = batch_size
55
56			def add_sample(self, inputs: np.ndarray, targets: np.ndarray) -> None:
57			# use zero-padded 8-digit key
58			key = idx_to_key(self._counter)
59			self._counter += 1
60			# convert the sample into a sparse matrix and serialize it as bytes
61			sample = (csc_matrix(inputs), csr_matrix(targets))
62			buf = BytesIO()
63			joblib.dump(sample, buf)
64			buf.seek(0)
65			self._txn.put(key, buf.read())
66
67			def __getitem__(self, idx: int) -> Tuple[np.ndarray, np.ndarray]:
68			"""get a particular batch of samples"""
69			cursor = self._txn.cursor()
70			first_key = idx * self._batch_size
71			cursor.set_key(idx_to_key(first_key))
72			input_arrays = []
73			target_arrays = []
74			for key, value in cursor.iternext():
75			if key_to_idx(key) >= (first_key + self._batch_size):
76			break
77			input_csr, target_csr = joblib.load(BytesIO(value))
78			input_arrays.append(input_csr.toarray())
79			target_arrays.append(target_csr.toarray().flatten())
80			return np.array(input_arrays), np.array(target_arrays)
81
82			def __len__(self) -> int:
83			"""return the number of available batches"""
84			return int(np.ceil(self._counter / self._batch_size))
85
86
87			class MeanLayer(Layer):
88			"""Custom Keras layer that calculates mean values along the 2nd axis."""
89
90			def call(self, inputs: EagerTensor) -> EagerTensor:
91			return K.mean(inputs, axis=2)
92
93
94			class NNEnsembleBackend(backend.AnnifLearningBackend, ensemble.BaseEnsembleBackend):
95			"""Neural network ensemble backend that combines results from multiple
96			projects"""
97
98			name = "nn_ensemble"
99
100			MODEL_FILE = "nn-model.h5"
101			LMDB_FILE = "nn-train.mdb"
102
103			DEFAULT_PARAMETERS = {
104			"nodes": 100,
105			"dropout_rate": 0.2,
106			"optimizer": "adam",
107			"epochs": 10,
108			"learn-epochs": 1,
109			"lmdb_map_size": 1024 * 1024 * 1024,
110			}
111
112			# defaults for uninitialized instances
113			_model = None
114
115			def default_params(self) -> Dict[str, Any]:
116			params = backend.AnnifBackend.DEFAULT_PARAMETERS.copy()
117			params.update(self.DEFAULT_PARAMETERS)
118			return params
119
120			def initialize(self, parallel: bool = False) -> None:
121			super().initialize(parallel)
122			if self._model is not None:
123			return # already initialized
124			if parallel:
125			# Don't load TF model just before parallel execution,
126			# since it won't work after forking worker processes
127			return
128			model_filename = os.path.join(self.datadir, self.MODEL_FILE)
129			if not os.path.exists(model_filename):
130			raise NotInitializedException(
131			"model file {} not found".format(model_filename),
132			backend_id=self.backend_id,
133			)
134			self.debug("loading Keras model from {}".format(model_filename))
135			self._model = load_model(
136			model_filename, custom_objects={"MeanLayer": MeanLayer}
137			)
138
139			def _merge_source_batches(
140			self,
141			batch_by_source: Dict[str, SuggestionBatch],
142			sources: List[Tuple[str, float]],
143			params: Dict[str, Any],
144			) -> SuggestionBatch:
145			src_weight = dict(sources)
146			score_vectors = np.array(
147			[
148			[
149			np.sqrt(suggestions.as_vector())
150			* src_weight[project_id]
151			* len(batch_by_source)
152			for suggestions in batch
153			]
154			for project_id, batch in batch_by_source.items()
155			],
156			dtype=np.float32,
157			).transpose(1, 2, 0)
158			prediction = self._model(score_vectors).numpy()
159			return SuggestionBatch.from_sequence(
160			[
161			vector_to_suggestions(row, limit=int(params["limit"]))
162			for row in prediction
163			],
164			self.project.subjects,
165			)
166
167			def _create_model(self, sources: List[Tuple[str, float]]) -> None:
168			self.info("creating NN ensemble model")
169
170			inputs = Input(shape=(len(self.project.subjects), len(sources)))
171
172			flat_input = Flatten()(inputs)
173			drop_input = Dropout(rate=float(self.params["dropout_rate"]))(flat_input)
174			hidden = Dense(int(self.params["nodes"]), activation="relu")(drop_input)
175			drop_hidden = Dropout(rate=float(self.params["dropout_rate"]))(hidden)
176			delta = Dense(
177			len(self.project.subjects),
178			kernel_initializer="zeros",
179			bias_initializer="zeros",
180			)(drop_hidden)
181
182			mean = MeanLayer()(inputs)
183
184			predictions = Add()([mean, delta])
185
186			self._model = Model(inputs=inputs, outputs=predictions)
187			self._model.compile(
188			optimizer=self.params["optimizer"],
189			loss="binary_crossentropy",
190			metrics=["top_k_categorical_accuracy"],
191			)
192			if "lr" in self.params:
193			self._model.optimizer.learning_rate.assign(float(self.params["lr"]))
194
195			summary = []
196			self._model.summary(print_fn=summary.append)
197			self.debug("Created model: \n" + "\n".join(summary))
198
199			def _train(
200			self,
201			corpus: DocumentCorpus,
202			params: Dict[str, Any],
203			jobs: int = 0,
204			) -> None:
205			sources = annif.util.parse_sources(self.params["sources"])
206			self._create_model(sources)
207			self._fit_model(
208			corpus,
209			epochs=int(params["epochs"]),
210			lmdb_map_size=int(params["lmdb_map_size"]),
211			n_jobs=jobs,
212			)
213
214			def _corpus_to_vectors(
215			self,
216			corpus: DocumentCorpus,
217			seq: LMDBSequence,
218			n_jobs: int,
219			) -> None:
220			# pass corpus through all source projects
221			sources = dict(annif.util.parse_sources(self.params["sources"]))
222
223			# initialize the source projects before forking, to save memory
224			self.info(f"Initializing source projects: {', '.join(sources.keys())}")
225			for project_id in sources.keys():
226			project = self.project.registry.get_project(project_id)
227			project.initialize(parallel=True)
228
229			psmap = annif.parallel.ProjectSuggestMap(
230			self.project.registry,
231			list(sources.keys()),
232			backend_params=None,
233			limit=None,
234			threshold=0.0,
235			)
236
237			jobs, pool_class = annif.parallel.get_pool(n_jobs)
238
239			self.info("Processing training documents...")
240			with pool_class(jobs) as pool:
241			for hits, subject_set in pool.imap_unordered(
242			psmap.suggest, corpus.documents
243			):
244			doc_scores = []
245			for project_id, p_hits in hits.items():
246			vector = p_hits.as_vector()
247			doc_scores.append(
248			np.sqrt(vector) * sources[project_id] * len(sources)
249			)
250			score_vector = np.array(doc_scores, dtype=np.float32).transpose()
251			true_vector = subject_set.as_vector(len(self.project.subjects))
252			seq.add_sample(score_vector, true_vector)
253
254			def _open_lmdb(self, cached, lmdb_map_size):
255			lmdb_path = os.path.join(self.datadir, self.LMDB_FILE)
256			if not cached and os.path.exists(lmdb_path):
257			shutil.rmtree(lmdb_path)
258			return lmdb.open(lmdb_path, map_size=lmdb_map_size, writemap=True)
259
260			def _fit_model(
261			self,
262			corpus: DocumentCorpus,
263			epochs: int,
264			lmdb_map_size: int,
265			n_jobs: int = 1,
266			) -> None:
267			env = self._open_lmdb(corpus == "cached", lmdb_map_size)
268			if corpus != "cached":
269			if corpus.is_empty():
270			raise NotSupportedException(
271			"Cannot train nn_ensemble project with no documents"
272			)
273			with env.begin(write=True, buffers=True) as txn:
274			seq = LMDBSequence(txn, batch_size=32)
275			self._corpus_to_vectors(corpus, seq, n_jobs)
276			else:
277			self.info("Reusing cached training data from previous run.")
278			# fit the model using a read-only view of the LMDB
279			self.info("Training neural network model...")
280			with env.begin(buffers=True) as txn:
281			seq = LMDBSequence(txn, batch_size=32)
282			self._model.fit(seq, verbose=True, epochs=epochs)
283
284			annif.util.atomic_save(self._model, self.datadir, self.MODEL_FILE)
285
286			def _learn(
287			self,
288			corpus: DocumentCorpus,
289			params: Dict[str, Any],
290			) -> None:
291			self.initialize()
292			self._fit_model(
293			corpus, int(params["learn-epochs"]), int(params["lmdb_map_size"])
294			)
295

NatLibFi / Annif

Pull Request — main (#708)

NNEnsembleBackend._fit_model() B

Complexity

Size

Duplication

Importance

Duplication Side-by-Side

Filter issues like