def classify(self, dataset, output_frequency='per_row'): """ Return a classification, for each ``prediction_window`` examples in the ``dataset``, using the trained activity classification model. The output SFrame contains predictions as both class labels as well as probabilities that the predicted value is the associated label. Parameters ---------- dataset : SFrame Dataset of new observations. Must include columns with the same names as the features and session id used for model training, but does not require a target column. Additional columns are ignored. output_frequency : {'per_row', 'per_window'}, optional The frequency of the predictions which is one of: - 'per_row': Each prediction is returned ``prediction_window`` times. - 'per_window': Return a single prediction for each ``prediction_window`` rows in ``dataset`` per ``session_id``. Returns ------- out : SFrame An SFrame with model predictions i.e class labels and probabilities. See Also ---------- create, evaluate, predict Examples ---------- >>> classes = model.classify(data) """ _tkutl._check_categorical_option_type('output_frequency', output_frequency, ['per_window', 'per_row']) id_target_map = self._id_target_map preds = self.predict(dataset, output_type='probability_vector', output_frequency=output_frequency) if output_frequency == 'per_row': return _SFrame({ 'class': preds.apply(lambda p: id_target_map[_np.argmax(p)]), 'probability': preds.apply(_np.max) }) elif output_frequency == 'per_window': preds['class'] = preds['probability_vector'].apply( lambda p: id_target_map[_np.argmax(p)]) preds['probability'] = preds['probability_vector'].apply(_np.max) preds = preds.remove_column('probability_vector') return preds
def evaluate(self, dataset, metric='auto', batch_size=64): """ Evaluate the model by making predictions of target values and comparing these to actual values. Parameters ---------- dataset : SFrame Dataset to use for evaluation, must include a column with the same name as the features used for model training. Additional columns are ignored. metric : str, optional Name of the evaluation metric. Possible values are: - 'auto' : Returns all available metrics. - 'accuracy' : Classification accuracy (micro average). - 'auc' : Area under the ROC curve (macro average) - 'precision' : Precision score (macro average) - 'recall' : Recall score (macro average) - 'f1_score' : F1 score (macro average) - 'log_loss' : Log loss - 'confusion_matrix' : An SFrame with counts of possible prediction/true label combinations. - 'roc_curve' : An SFrame containing information needed for an ROC curve batch_size : int, optional If you are getting memory errors, try decreasing this value. If you have a powerful computer, increasing this value may improve performance. Returns ------- out : dict Dictionary of evaluation results where the key is the name of the evaluation metric (e.g. `accuracy`) and the value is the evaluation score. See Also ---------- classify, predict Examples ---------- .. sourcecode:: python >>> results = model.evaluate(data) >>> print results['accuracy'] """ from turicreate.toolkits import evaluation # parameter checking if not isinstance(dataset, _tc.SFrame): raise TypeError('\'dataset\' parameter must be an SFrame') if(batch_size < 1): raise ValueError('\'batch_size\' must be greater than or equal to 1') avail_metrics = ['accuracy', 'auc', 'precision', 'recall', 'f1_score', 'log_loss', 'confusion_matrix', 'roc_curve'] _tk_utils._check_categorical_option_type( 'metric', metric, avail_metrics + ['auto']) if metric == 'auto': metrics = avail_metrics else: metrics = [metric] if any([m in metrics for m in ('roc_curve', 'log_loss', 'auc')]): probs = self.predict(dataset, output_type='probability_vector', batch_size=batch_size) if any([m in metrics for m in ('accuracy', 'precision', 'recall', 'f1_score', 'confusion_matrix')]): classes = self.predict(dataset, output_type='class', batch_size=batch_size) ret = {} if 'accuracy' in metrics: ret['accuracy'] = evaluation.accuracy(dataset[self.target], classes) if 'auc' in metrics: ret['auc'] = evaluation.auc(dataset[self.target], probs, index_map=self._class_label_to_id) if 'precision' in metrics: ret['precision'] = evaluation.precision(dataset[self.target], classes) if 'recall' in metrics: ret['recall'] = evaluation.recall(dataset[self.target], classes) if 'f1_score' in metrics: ret['f1_score'] = evaluation.f1_score(dataset[self.target], classes) if 'log_loss' in metrics: ret['log_loss'] = evaluation.log_loss(dataset[self.target], probs, index_map=self._class_label_to_id) if 'confusion_matrix' in metrics: ret['confusion_matrix'] = evaluation.confusion_matrix(dataset[self.target], classes) if 'roc_curve' in metrics: ret['roc_curve'] = evaluation.roc_curve(dataset[self.target], probs, index_map=self._class_label_to_id) return ret
def create( dataset, target, feature=None, model='resnet-50', l2_penalty=0.01, l1_penalty=0.0, solver='auto', feature_rescaling=True, convergence_threshold=_DEFAULT_SOLVER_OPTIONS['convergence_threshold'], step_size=_DEFAULT_SOLVER_OPTIONS['step_size'], lbfgs_memory_level=_DEFAULT_SOLVER_OPTIONS['lbfgs_memory_level'], max_iterations=_DEFAULT_SOLVER_OPTIONS['max_iterations'], class_weights=None, validation_set='auto', verbose=True, seed=None, batch_size=64): """ Create a :class:`ImageClassifier` model. Parameters ---------- dataset : SFrame Input data. The column named by the 'feature' parameter will be extracted for modeling. target : string, or int Name of the column containing the target variable. The values in this column must be of string or integer type. String target variables are automatically mapped to integers in the order in which they are provided. For example, a target variable with 'cat' and 'dog' as possible values is mapped to 0 and 1 respectively with 0 being the base class and 1 being the reference class. Use `model.classes` to retrieve the order in which the classes are mapped. feature : string, optional indicates that the SFrame has only column of Image type and that will Name of the column containing the input images. 'None' (the default) indicates the only image column in `dataset` should be used as the feature. l2_penalty : float, optional Weight on l2 regularization of the model. The larger this weight, the more the model coefficients shrink toward 0. This introduces bias into the model but decreases variance, potentially leading to better predictions. The default value is 0.01; setting this parameter to 0 corresponds to unregularized logistic regression. See the ridge regression reference for more detail. l1_penalty : float, optional Weight on l1 regularization of the model. Like the l2 penalty, the higher the l1 penalty, the more the estimated coefficients shrink toward 0. The l1 penalty, however, completely zeros out sufficiently small coefficients, automatically indicating features that are not useful for the model. The default weight of 0 prevents any features from being discarded. See the LASSO regression reference for more detail. solver : string, optional Name of the solver to be used to solve the regression. See the references for more detail on each solver. Available solvers are: - *auto (default)*: automatically chooses the best solver for the data and model parameters. - *newton*: Newton-Raphson - *lbfgs*: limited memory BFGS - *fista*: accelerated gradient descent For this model, the Newton-Raphson method is equivalent to the iteratively re-weighted least squares algorithm. If the l1_penalty is greater than 0, use the 'fista' solver. The model is trained using a carefully engineered collection of methods that are automatically picked based on the input data. The ``newton`` method works best for datasets with plenty of examples and few features (long datasets). Limited memory BFGS (``lbfgs``) is a robust solver for wide datasets (i.e datasets with many coefficients). ``fista`` is the default solver for l1-regularized linear regression. The solvers are all automatically tuned and the default options should function well. See the solver options guide for setting additional parameters for each of the solvers. See the user guide for additional details on how the solver is chosen. (see `here <https://apple.github.io/turicreate/docs/userguide/supervised-learning/linear-regression.html>`_) feature_rescaling : boolean, optional Feature rescaling is an important pre-processing step that ensures that all features are on the same scale. An l2-norm rescaling is performed to make sure that all features are of the same norm. Categorical features are also rescaled by rescaling the dummy variables that are used to represent them. The coefficients are returned in original scale of the problem. This process is particularly useful when features vary widely in their ranges. convergence_threshold : float, optional Convergence is tested using variation in the training objective. The variation in the training objective is calculated using the difference between the objective values between two steps. Consider reducing this below the default value (0.01) for a more accurately trained model. Beware of overfitting (i.e a model that works well only on the training data) if this parameter is set to a very low value. lbfgs_memory_level : float, optional The L-BFGS algorithm keeps track of gradient information from the previous ``lbfgs_memory_level`` iterations. The storage requirement for each of these gradients is the ``num_coefficients`` in the problem. Increasing the ``lbfgs_memory_level ``can help improve the quality of the model trained. Setting this to more than ``max_iterations`` has the same effect as setting it to ``max_iterations``. model : string optional Uses a pretrained model to bootstrap an image classifier: - "resnet-50" : Uses a pretrained resnet model. Exported Core ML model will be ~90M. - "squeezenet_v1.1" : Uses a pretrained squeezenet model. Exported Core ML model will be ~4.7M. - "VisionFeaturePrint_Scene": Uses an OS internal feature extractor. Only on available on iOS 12.0+, macOS 10.14+ and tvOS 12.0+. Exported Core ML model will be ~41K. Models are downloaded from the internet if not available locally. Once downloaded, the models are cached for future use. step_size : float, optional The starting step size to use for the ``fista`` solver. The default is set to 1.0, this is an aggressive setting. If the first iteration takes a considerable amount of time, reducing this parameter may speed up model training. class_weights : {dict, `auto`}, optional Weights the examples in the training data according to the given class weights. If set to `None`, all classes are supposed to have weight one. The `auto` mode set the class weight to be inversely proportional to number of examples in the training data with the given class. validation_set : SFrame, optional A dataset for monitoring the model's generalization performance. The format of this SFrame must be the same as the training set. By default this argument is set to 'auto' and a validation set is automatically sampled and used for progress printing. If validation_set is set to None, then no additional metrics are computed. The default value is 'auto'. max_iterations : int, optional The maximum number of allowed passes through the data. More passes over the data can result in a more accurately trained model. Consider increasing this (the default value is 10) if the training accuracy is low and the *Grad-Norm* in the display is large. verbose : bool, optional If True, prints progress updates and model details. seed : int, optional Seed for random number generation. Set this value to ensure that the same model is created every time. batch_size : int, optional If you are getting memory errors, try decreasing this value. If you have a powerful computer, increasing this value may improve performance. Returns ------- out : ImageClassifier A trained :class:`ImageClassifier` model. Examples -------- .. sourcecode:: python >>> model = turicreate.image_classifier.create(data, target='is_expensive') # Make predictions (in various forms) >>> predictions = model.predict(data) # predictions >>> predictions = model.classify(data) # predictions with confidence >>> predictions = model.predict_topk(data) # Top-5 predictions (multiclass) # Evaluate the model with ground truth data >>> results = model.evaluate(data) See Also -------- ImageClassifier """ start_time = _time.time() # Check model parameter allowed_models = list(_pre_trained_models.MODELS.keys()) if _mac_ver() >= (10, 14): allowed_models.append('VisionFeaturePrint_Scene') # Also, to make sure existing code doesn't break, replace incorrect name # with the correct name version if model == "VisionFeaturePrint_Screen": print( "WARNING: Correct spelling of model name is VisionFeaturePrint_Scene; VisionFeaturePrint_Screen will be removed in subsequent versions." ) model = "VisionFeaturePrint_Scene" _tkutl._check_categorical_option_type('model', model, allowed_models) # Check dataset parameter if len(dataset) == 0: raise _ToolkitError('Unable to train on empty dataset') if (feature is not None) and (feature not in dataset.column_names()): raise _ToolkitError("Image feature column '%s' does not exist" % feature) if target not in dataset.column_names(): raise _ToolkitError("Target column '%s' does not exist" % target) if (batch_size < 1): raise ValueError("'batch_size' must be greater than or equal to 1") if not (isinstance(validation_set, _tc.SFrame) or validation_set == 'auto' or validation_set is None): raise TypeError("Unrecognized value for 'validation_set'.") if feature is None: feature = _tkutl._find_only_image_column(dataset) feature_extractor = _image_feature_extractor._create_feature_extractor( model) # Extract features extracted_features = _tc.SFrame({ target: dataset[target], '__image_features__': feature_extractor.extract_features(dataset, feature, verbose=verbose, batch_size=batch_size), }) if isinstance(validation_set, _tc.SFrame): extracted_features_validation = _tc.SFrame({ target: validation_set[target], '__image_features__': feature_extractor.extract_features(validation_set, feature, verbose=verbose, batch_size=batch_size), }) else: extracted_features_validation = validation_set # Train a classifier using the extracted features extracted_features[target] = dataset[target] lr_model = _tc.logistic_classifier.create( extracted_features, features=['__image_features__'], target=target, max_iterations=max_iterations, validation_set=extracted_features_validation, seed=seed, verbose=verbose, l2_penalty=l2_penalty, l1_penalty=l1_penalty, solver=solver, feature_rescaling=feature_rescaling, convergence_threshold=convergence_threshold, step_size=step_size, lbfgs_memory_level=lbfgs_memory_level, class_weights=class_weights) # set input image shape if model in _pre_trained_models.MODELS: input_image_shape = _pre_trained_models.MODELS[model].input_image_shape else: # model == VisionFeaturePrint_Scene input_image_shape = (3, 299, 299) # Save the model state = { 'classifier': lr_model, 'model': model, 'max_iterations': max_iterations, 'feature_extractor': feature_extractor, 'input_image_shape': input_image_shape, 'target': target, 'feature': feature, 'num_features': 1, 'num_classes': lr_model.num_classes, 'classes': lr_model.classes, 'num_examples': lr_model.num_examples, 'training_time': _time.time() - start_time, 'training_loss': lr_model.training_loss, } return ImageClassifier(state)
def get_topics(self, topic_ids=None, num_words=5, cdf_cutoff=1.0, output_type='topic_probabilities'): """ Get the words associated with a given topic. The score column is the probability of choosing that word given that you have chosen a particular topic. Parameters ---------- topic_ids : list of int, optional The topics to retrieve words. Topic ids are zero-based. Throws an error if greater than or equal to m['num_topics'], or if the requested topic name is not present. num_words : int, optional The number of words to show. cdf_cutoff : float, optional Allows one to only show the most probable words whose cumulative probability is below this cutoff. For example if there exist three words where .. math:: p(word_1 | topic_k) = .1 p(word_2 | topic_k) = .2 p(word_3 | topic_k) = .05 then setting :math:`cdf_{cutoff}=.3` would return only :math:`word_1` and :math:`word_2` since :math:`p(word_1 | topic_k) + p(word_2 | topic_k) <= cdf_{cutoff}` output_type : {'topic_probabilities' | 'topic_words'}, optional Determine the type of desired output. See below. Returns ------- out : SFrame If output_type is 'topic_probabilities', then the returned value is an SFrame with a column of words ranked by a column of scores for each topic. Otherwise, the returned value is a SArray where each element is a list of the most probable words for each topic. Examples -------- Get the highest ranked words for all topics. >>> docs = turicreate.SArray('https://static.turi.com/datasets/nips-text') >>> m = turicreate.topic_model.create(docs, num_iterations=50) >>> m.get_topics() +-------+----------+-----------------+ | topic | word | score | +-------+----------+-----------------+ | 0 | cell | 0.028974400831 | | 0 | input | 0.0259470208503 | | 0 | image | 0.0215721599763 | | 0 | visual | 0.0173635081992 | | 0 | object | 0.0172447874156 | | 1 | function | 0.0482834508265 | | 1 | input | 0.0456270024091 | | 1 | point | 0.0302662839454 | | 1 | result | 0.0239474934631 | | 1 | problem | 0.0231750116011 | | ... | ... | ... | +-------+----------+-----------------+ Get the highest ranked words for topics 0 and 1 and show 15 words per topic. >>> m.get_topics([0, 1], num_words=15) +-------+----------+------------------+ | topic | word | score | +-------+----------+------------------+ | 0 | cell | 0.028974400831 | | 0 | input | 0.0259470208503 | | 0 | image | 0.0215721599763 | | 0 | visual | 0.0173635081992 | | 0 | object | 0.0172447874156 | | 0 | response | 0.0139740298286 | | 0 | layer | 0.0122585145062 | | 0 | features | 0.0115343177265 | | 0 | feature | 0.0103530459301 | | 0 | spatial | 0.00823387994361 | | ... | ... | ... | +-------+----------+------------------+ If one wants to instead just get the top words per topic, one may change the format of the output as follows. >>> topics = m.get_topics(output_type='topic_words') dtype: list Rows: 10 [['cell', 'image', 'input', 'object', 'visual'], ['algorithm', 'data', 'learning', 'method', 'set'], ['function', 'input', 'point', 'problem', 'result'], ['model', 'output', 'pattern', 'set', 'unit'], ['action', 'learning', 'net', 'problem', 'system'], ['error', 'function', 'network', 'parameter', 'weight'], ['information', 'level', 'neural', 'threshold', 'weight'], ['control', 'field', 'model', 'network', 'neuron'], ['hidden', 'layer', 'system', 'training', 'vector'], ['component', 'distribution', 'local', 'model', 'optimal']] """ _check_categorical_option_type('output_type', output_type, ['topic_probabilities', 'topic_words']) if topic_ids is None: topic_ids = list(range(self._get('num_topics'))) assert isinstance(topic_ids, list), \ "The provided topic_ids is not a list." if any([type(x) == str for x in topic_ids]): raise ValueError( "Only integer topic_ids can be used at this point in time.") if not all([x >= 0 and x < self.num_topics for x in topic_ids]): raise ValueError("Topic id values must be non-negative and less than the " + \ "number of topics used to fit the model.") opts = { 'model': self.__proxy__, 'topic_ids': topic_ids, 'num_words': num_words, 'cdf_cutoff': cdf_cutoff } response = _turicreate.extensions._text.topicmodel_get_topic(opts) ret = response['top_words'] def sort_wordlist_by_prob(z): words = sorted(z.items(), key=_operator.itemgetter(1), reverse=True) return [word for (word, prob) in words] if output_type != 'topic_probabilities': ret = ret.groupby( 'topic', {'word': _turicreate.aggregate.CONCAT('word', 'score')}) words = ret.sort('topic')['word'].apply(sort_wordlist_by_prob) ret = _SFrame({'words': words}) return ret
def predict_topk(self, dataset, output_type='probability', k=3, output_frequency='per_row'): """ Return top-k predictions for the ``dataset``, using the trained model. Predictions are returned as an SFrame with three columns: `prediction_id`, `class`, and `probability`, or `rank`, depending on the ``output_type`` parameter. Parameters ---------- dataset : SFrame Dataset of new observations. Must include columns with the same names as the features and session id used for model training, but does not require a target column. Additional columns are ignored. output_type : {'probability', 'rank'}, optional Choose the return type of the prediction: - `probability`: Probability associated with each label in the prediction. - `rank` : Rank associated with each label in the prediction. k : int, optional Number of classes to return for each input example. output_frequency : {'per_row', 'per_window'}, optional The frequency of the predictions which is one of: - 'per_row': Each prediction is returned ``prediction_window`` times. - 'per_window': Return a single prediction for each ``prediction_window`` rows in ``dataset`` per ``session_id``. Returns ------- out : SFrame An SFrame with model predictions. See Also -------- predict, classify, evaluate Examples -------- >>> pred = m.predict_topk(validation_data, k=3) >>> pred +---------------+-------+-------------------+ | row_id | class | probability | +---------------+-------+-------------------+ | 0 | 4 | 0.995623886585 | | 0 | 9 | 0.0038311756216 | | 0 | 7 | 0.000301006948575 | | 1 | 1 | 0.928708016872 | | 1 | 3 | 0.0440889261663 | | 1 | 2 | 0.0176190119237 | | 2 | 3 | 0.996967732906 | | 2 | 2 | 0.00151345680933 | | 2 | 7 | 0.000637513934635 | | 3 | 1 | 0.998070061207 | | ... | ... | ... | +---------------+-------+-------------------+ """ _tkutl._check_categorical_option_type('output_type', output_type, ['probability', 'rank']) id_target_map = self._id_target_map preds = self.predict( dataset, output_type='probability_vector', output_frequency=output_frequency) if output_frequency == 'per_row': probs = preds elif output_frequency == 'per_window': probs = preds['probability_vector'] if output_type == 'rank': probs = probs.apply(lambda p: [ {'class': id_target_map[i], 'rank': i} for i in reversed(_np.argsort(p)[-k:])] ) elif output_type == 'probability': probs = probs.apply(lambda p: [ {'class': id_target_map[i], 'probability': p[i]} for i in reversed(_np.argsort(p)[-k:])] ) if output_frequency == 'per_row': output = _SFrame({'probs': probs}) output = output.add_row_number(column_name='row_id') elif output_frequency == 'per_window': output = _SFrame({ 'probs': probs, self.session_id: preds[self.session_id], 'prediction_id': preds['prediction_id'] }) output = output.stack('probs', new_column_name='probs') output = output.unpack('probs', column_name_prefix='') return output
def predict(self, dataset, output_type='class', output_frequency='per_row'): """ Return predictions for ``dataset``, using the trained activity classifier. Predictions can be generated as class labels, or as a probability vector with probabilities for each class. The activity classifier generates a single prediction for each ``prediction_window`` rows in ``dataset``, per ``session_id``. Thus the number of predictions is smaller than the length of ``dataset``. By default each prediction is replicated by ``prediction_window`` to return a prediction for each row of ``dataset``. Use ``output_frequency`` to get the unreplicated predictions. Parameters ---------- dataset : SFrame Dataset of new observations. Must include columns with the same names as the features used for model training, but does not require a target column. Additional columns are ignored. output_type : {'class', 'probability_vector'}, optional Form of each prediction which is one of: - 'probability_vector': Prediction probability associated with each class as a vector. The probability of the first class (sorted alphanumerically by name of the class in the training set) is in position 0 of the vector, the second in position 1 and so on. - 'class': Class prediction. This returns the class with maximum probability. output_frequency : {'per_row', 'per_window'}, optional The frequency of the predictions which is one of: - 'per_window': Return a single prediction for each ``prediction_window`` rows in ``dataset`` per ``session_id``. - 'per_row': Convenience option to make sure the number of predictions match the number of rows in the dataset. Each prediction from the model is repeated ``prediction_window`` times during that window. Returns ------- out : SArray | SFrame If ``output_frequency`` is 'per_row' return an SArray with predictions for each row in ``dataset``. If ``output_frequency`` is 'per_window' return an SFrame with predictions for ``prediction_window`` rows in ``dataset``. See Also ---------- create, evaluate, classify Examples -------- .. sourcecode:: python # One prediction per row >>> probability_predictions = model.predict( ... data, output_type='probability_vector', output_frequency='per_row')[:4] >>> probability_predictions dtype: array Rows: 4 [array('d', [0.01857384294271469, 0.0348394550383091, 0.026018327102065086]), array('d', [0.01857384294271469, 0.0348394550383091, 0.026018327102065086]), array('d', [0.01857384294271469, 0.0348394550383091, 0.026018327102065086]), array('d', [0.01857384294271469, 0.0348394550383091, 0.026018327102065086])] # One prediction per window >>> class_predictions = model.predict( ... data, output_type='class', output_frequency='per_window') >>> class_predictions +---------------+------------+-----+ | prediction_id | session_id |class| +---------------+------------+-----+ | 0 | 3 | 5 | | 1 | 3 | 5 | | 2 | 3 | 5 | | 3 | 3 | 5 | | 4 | 3 | 5 | | 5 | 3 | 5 | | 6 | 3 | 5 | | 7 | 3 | 4 | | 8 | 3 | 4 | | 9 | 3 | 4 | | ... | ... | ... | +---------------+------------+-----+ """ _tkutl._raise_error_if_not_sframe(dataset, 'dataset') _tkutl._check_categorical_option_type( 'output_frequency', output_frequency, ['per_window', 'per_row']) _tkutl._check_categorical_option_type( 'output_type', output_type, ['probability_vector', 'class']) from ._sframe_sequence_iterator import SFrameSequenceIter as _SFrameSequenceIter from ._sframe_sequence_iterator import prep_data as _prep_data from ._sframe_sequence_iterator import _ceil_dev prediction_window = self.prediction_window chunked_dataset, _ = _prep_data(dataset, self.features, self.session_id, prediction_window, self._predictions_in_chunk, verbose=False) data_iter = _SFrameSequenceIter(chunked_dataset, len(self.features), prediction_window, self._predictions_in_chunk, self._recalibrated_batch_size, use_pad=True) chunked_data = data_iter.dataset preds = self._pred_model.predict(data_iter).asnumpy() if output_frequency == 'per_row': # Replicate each prediction times prediction_window preds = preds.repeat(prediction_window, axis=1) # Remove predictions for padded rows unpadded_len = chunked_data['chunk_len'].to_numpy() preds = [p[:unpadded_len[i]] for i, p in enumerate(preds)] # Reshape from (num_of_chunks, chunk_size, num_of_classes) # to (ceil(length / prediction_window), num_of_classes) # chunk_size is DIFFERENT between chunks - since padding was removed. out = _np.concatenate(preds) out = out.reshape((-1, len(self._target_id_map))) out = _SArray(out) if output_type == 'class': id_target_map = self._id_target_map out = out.apply(lambda c: id_target_map[_np.argmax(c)]) elif output_frequency == 'per_window': # Calculate the number of expected predictions and # remove predictions for padded data unpadded_len = chunked_data['chunk_len'].apply( lambda l: _ceil_dev(l, prediction_window)).to_numpy() preds = [p[:unpadded_len[i]] for i, p in enumerate(preds)] out = _SFrame({ self.session_id: chunked_data['session_id'], 'preds': _SArray(preds, dtype=list) }).stack('preds', new_column_name='probability_vector') # Calculate the prediction index per session out = out.add_row_number(column_name='prediction_id') start_sess_idx = out.groupby( self.session_id, {'start_idx': _agg.MIN('prediction_id')}) start_sess_idx = start_sess_idx.unstack( [self.session_id, 'start_idx'], new_column_name='idx')['idx'][0] if output_type == 'class': id_target_map = self._id_target_map out['probability_vector'] = out['probability_vector'].apply( lambda c: id_target_map[_np.argmax(c)]) out = out.rename({'probability_vector': 'class'}) return out
def predict(self, dataset, output_type="class", missing_value_action="auto"): """ Return predictions for ``dataset``, using the trained logistic regression model. Predictions can be generated as class labels (0 or 1), or margins (i.e. the distance of the observations from the hyperplane separating the classes). By default, the predict method returns class labels. For each new example in ``dataset``, the margin---also known as the linear predictor---is the inner product of the example and the model coefficients plus the intercept term. Predicted classes are obtained by thresholding the margins at 0. Parameters ---------- dataset : SFrame | dict Dataset of new observations. Must include columns with the same names as the features used for model training, but does not require a target column. Additional columns are ignored. output_type : {'margin', 'class'}, optional Form of the predictions which are one of: - 'margin': Distance of the observations from the hyperplane separating the classes. - 'class': Class prediction. missing_value_action : str, optional Action to perform when missing values are encountered. This can be one of: - 'auto': Default to 'impute' - 'impute': Proceed with evaluation by filling in the missing values with the mean of the training data. Missing values are also imputed if an entire column of data is missing during evaluation. - 'error' : Do not proceed with prediction and terminate with an error message. Returns ------- out : SArray An SArray with model predictions. See Also ---------- create, evaluate, classify Examples ---------- >>> data = turicreate.SFrame('https://static.turi.com/datasets/regression/houses.csv') >>> data['is_expensive'] = data['price'] > 30000 >>> model = turicreate.svm_classifier.create(data, target='is_expensive', features=['bath', 'bedroom', 'size']) >>> class_predictions = model.progressredict(data) >>> margin_predictions = model.progressredict(data, output_type='margin') """ _check_categorical_option_type("output_type", output_type, ["class", "margin"]) return super(_Classifier, self).predict( dataset, output_type=output_type, missing_value_action=missing_value_action )
def predict(self, dataset, output_type='class', missing_value_action='auto'): """ A flexible and advanced prediction API. The target column is provided during :func:`~turicreate.random_forest.create`. If the target column is in the `dataset` it will be ignored. Parameters ---------- dataset : SFrame A dataset that has the same columns that were used during training. If the target column exists in ``dataset`` it will be ignored while making predictions. output_type : {'probability', 'margin', 'class', 'probability_vector'}, optional. Form of the predictions which are one of: - 'probability': Prediction probability associated with the True class (not applicable for multi-class classification) - 'margin': Margin associated with the prediction (not applicable for multi-class classification) - 'probability_vector': Prediction probability associated with each class as a vector. The probability of the first class (sorted alphanumerically by name of the class in the training set) is in position 0 of the vector, the second in position 1 and so on. - 'class': Class prediction. For multi-class classification, this returns the class with maximum probability. missing_value_action : str, optional Action to perform when missing values are encountered. Can be one of: - 'auto': By default the model will treat missing value as is. - 'impute': Proceed with evaluation by filling in the missing values with the mean of the training data. Missing values are also imputed if an entire column of data is missing during evaluation. - 'error': Do not proceed with evaluation and terminate with an error message. Returns ------- out : SArray Predicted target value for each example (i.e. row) in the dataset. See Also ---------- create, evaluate, classify Examples -------- >>> m.predict(testdata) >>> m.predict(testdata, output_type='probability') >>> m.predict(testdata, output_type='margin') """ _check_categorical_option_type( 'output_type', output_type, ['class', 'margin', 'probability', 'probability_vector']) return super(_Classifier, self).predict(dataset, output_type=output_type, missing_value_action=missing_value_action)
def create(dataset, annotations=None, feature=None, model="darknet-yolo", classes=None, batch_size=0, max_iterations=0, verbose=True, grid_shape=[13, 13], **kwargs): """ Create a :class:`ObjectDetector` model. Parameters ---------- dataset : SFrame Input data. The columns named by the ``feature`` and ``annotations`` parameters will be extracted for training the detector. annotations : string Name of the column containing the object detection annotations. This column should be a list of dictionaries (or a single dictionary), with each dictionary representing a bounding box of an object instance. Here is an example of the annotations for a single image with two object instances:: [{'label': 'dog', 'type': 'rectangle', 'coordinates': {'x': 223, 'y': 198, 'width': 130, 'height': 230}}, {'label': 'cat', 'type': 'rectangle', 'coordinates': {'x': 40, 'y': 73, 'width': 80, 'height': 123}}] The value for `x` is the horizontal center of the box paired with `width` and `y` is the vertical center of the box paired with `height`. 'None' (the default) indicates the only list column in `dataset` should be used for the annotations. feature : string Name of the column containing the input images. 'None' (the default) indicates the only image column in `dataset` should be used as the feature. model : string optional Object detection model to use: - "darknet-yolo" : Fast and medium-sized model grid_shape : array optional Shape of the grid used for object detection. Higher values increase precision for small objects, but at a higher computational cost - [13, 13] : Default grid value for a Fast and medium-sized model classes : list optional List of strings containing the names of the classes of objects. Inferred from the data if not provided. batch_size: int The number of images per training iteration. If 0, then it will be automatically determined based on resource availability. max_iterations : int The number of training iterations. If 0, then it will be automatically be determined based on the amount of data you provide. verbose : bool, optional If True, print progress updates and model details. Returns ------- out : ObjectDetector A trained :class:`ObjectDetector` model. See Also -------- ObjectDetector Examples -------- .. sourcecode:: python # Train an object detector model >>> model = turicreate.object_detector.create(data) # Make predictions on the training set and as column to the SFrame >>> data['predictions'] = model.predict(data) # Visualize predictions by generating a new column of marked up images >>> data['image_pred'] = turicreate.object_detector.util.draw_bounding_boxes(data['image'], data['predictions']) """ _raise_error_if_not_sframe(dataset, "dataset") if len(dataset) == 0: raise _ToolkitError("Unable to train on empty dataset") _numeric_param_check_range("max_iterations", max_iterations, 0, _six.MAXSIZE) start_time = _time.time() supported_detectors = ["darknet-yolo"] if feature is None: feature = _tkutl._find_only_image_column(dataset) if verbose: print("Using '%s' as feature column" % feature) if annotations is None: annotations = _tkutl._find_only_column_of_type( dataset, target_type=[list, dict], type_name="list", col_name="annotations") if verbose: print("Using '%s' as annotations column" % annotations) _raise_error_if_not_detection_sframe(dataset, feature, annotations, require_annotations=True) _tkutl._handle_missing_values(dataset, feature, "dataset") _tkutl._check_categorical_option_type("model", model, supported_detectors) base_model = model.split("-", 1)[0] ref_model = _pre_trained_models.OBJECT_DETECTION_BASE_MODELS[base_model]() pretrained_model = _pre_trained_models.OBJECT_DETECTION_BASE_MODELS[ "darknet_mlmodel"]() pretrained_model_path = pretrained_model.get_model_path() params = { "anchors": [ (1.0, 2.0), (1.0, 1.0), (2.0, 1.0), (2.0, 4.0), (2.0, 2.0), (4.0, 2.0), (4.0, 8.0), (4.0, 4.0), (8.0, 4.0), (8.0, 16.0), (8.0, 8.0), (16.0, 8.0), (16.0, 32.0), (16.0, 16.0), (32.0, 16.0), ], "grid_shape": grid_shape, "aug_resize": 0, "aug_rand_crop": 0.9, "aug_rand_pad": 0.9, "aug_rand_gray": 0.0, "aug_aspect_ratio": 1.25, "aug_hue": 0.05, "aug_brightness": 0.05, "aug_saturation": 0.05, "aug_contrast": 0.05, "aug_horizontal_flip": True, "aug_min_object_covered": 0, "aug_min_eject_coverage": 0.5, "aug_area_range": (0.15, 2), "aug_pca_noise": 0.0, "aug_max_attempts": 20, "aug_inter_method": 2, "lmb_coord_xy": 10.0, "lmb_coord_wh": 10.0, "lmb_obj": 100.0, "lmb_noobj": 5.0, "lmb_class": 2.0, "non_maximum_suppression_threshold": 0.45, "rescore": True, "clip_gradients": 0.025, "weight_decay": 0.0005, "sgd_momentum": 0.9, "learning_rate": 1.0e-3, "shuffle": True, "mps_loss_mult": 8, # This large buffer size (8 batches) is an attempt to mitigate against # the SFrame shuffle operation that can occur after each epoch. "io_thread_buffer_size": 8, "mlmodel_path": pretrained_model_path, } # create tensorflow model here import turicreate.toolkits.libtctensorflow if classes == None: classes = [] _raise_error_if_not_iterable(classes) _raise_error_if_not_iterable(grid_shape) grid_shape = [int(x) for x in grid_shape] assert len(grid_shape) == 2 tf_config = { "grid_height": params["grid_shape"][0], "grid_width": params["grid_shape"][1], "mlmodel_path": params["mlmodel_path"], "classes": classes, "compute_final_metrics": False, "verbose": verbose, "model": "darknet-yolo", } # If batch_size or max_iterations = 0, they will be automatically # generated in C++. if batch_size > 0: tf_config["batch_size"] = batch_size if max_iterations > 0: tf_config["max_iterations"] = max_iterations model = _tc.extensions.object_detector() model.train( data=dataset, annotations_column_name=annotations, image_column_name=feature, options=tf_config, ) return ObjectDetector(model_proxy=model, name="object_detector")
def roc_curve(targets, predictions, average=None, index_map=None): r""" Compute an ROC curve for the given targets and predictions. Currently, only binary classification is supported. Parameters ---------- targets : SArray An SArray containing the observed values. For binary classification, the alpha-numerically first category is considered the reference category. predictions : SArray The prediction that corresponds to each target value. This vector must have the same length as ``targets``. Target scores, can either be probability estimates of the positive class, confidence values, or binary decisions. average : string, [None (default)] Metric averaging strategies for multiclass classification. Averaging strategies can be one of the following: - None: No averaging is performed and a single metric is returned for each class. index_map : dict[int], [None (default)] For binary classification, a dictionary mapping the two target labels to either 0 (negative) or 1 (positive). For multi-class classification, a dictionary mapping potential target labels to the associated index into the vectors in ``predictions``. Returns ------- out : SFrame Each row represents the predictive performance when using a given cutoff threshold, where all predictions above that cutoff are considered "positive". Four columns are used to describe the performance: - tpr : True positive rate, the number of true positives divided by the number of positives. - fpr : False positive rate, the number of false positives divided by the number of negatives. - p : Total number of positive values. - n : Total number of negative values. - class : Reference class for this ROC curve. See Also -------- confusion_matrix, auc References ---------- `An introduction to ROC analysis. Tom Fawcett. <https://ccrma.stanford.edu/workshops/mir2009/references/ROCintro.pdf>`_ Notes ----- - For binary classification, when the target label is of type "string", then the labels are sorted alphanumerically and the largest label is chosen as the "positive" label. For example, if the classifier labels are {"cat", "dog"}, then "dog" is chosen as the positive label for the binary classification case. This behavior can be overridden by providing an explicit ``index_map``. - For multi-class classification, when the target label is of type "string", then the probability vector is assumed to be a vector of probabilities of classes as sorted alphanumerically. Hence, for the probability vector [0.1, 0.2, 0.7] for a dataset with classes "cat", "dog", and "rat"; the 0.1 corresponds to "cat", the 0.2 to "dog" and the 0.7 to "rat". This behavior can be overridden by providing an explicit ``index_map``. - The ROC curve is computed using a binning approximation with 1M bins and is hence accurate only to the 5th decimal. Examples -------- .. sourcecode:: python >>> targets = turicreate.SArray([0, 1, 1, 0]) >>> predictions = turicreate.SArray([0.1, 0.35, 0.7, 0.99]) # Calculate the roc-curve. >>> roc_curve = turicreate.evaluation.roc_curve(targets, predictions) +-------------------+-----+-----+---+---+ | threshold | fpr | tpr | p | n | +-------------------+-----+-----+---+---+ | 0.0 | 1.0 | 1.0 | 2 | 2 | | 9.99999974738e-06 | 1.0 | 1.0 | 2 | 2 | | 1.99999994948e-05 | 1.0 | 1.0 | 2 | 2 | | 2.99999992421e-05 | 1.0 | 1.0 | 2 | 2 | | 3.99999989895e-05 | 1.0 | 1.0 | 2 | 2 | | 4.99999987369e-05 | 1.0 | 1.0 | 2 | 2 | | 5.99999984843e-05 | 1.0 | 1.0 | 2 | 2 | | 7.00000018696e-05 | 1.0 | 1.0 | 2 | 2 | | 7.9999997979e-05 | 1.0 | 1.0 | 2 | 2 | | 9.00000013644e-05 | 1.0 | 1.0 | 2 | 2 | +-------------------+-----+-----+---+---+ [100001 rows x 5 columns] For the multi-class setting, an ROC curve is returned for each class. .. sourcecode:: python # Targets and Predictions >>> targets = turicreate.SArray([0, 1, 2, 3, 0, 1, 2, 3]) >>> predictions = turicreate.SArray([1, 0, 2, 1, 3, 1, 2, 1]) # Micro average of the recall scores for each class. >>> turicreate.evaluation.recall(targets, predictions, ... average = 'micro') 0.375 # Macro average of the recall scores for each class. >>> turicreate.evaluation.recall(targets, predictions, ... average = 'macro') 0.375 # Recall score for each class. >>> turicreate.evaluation.recall(targets, predictions, ... average = None) {0: 0.0, 1: 0.5, 2: 1.0, 3: 0.0} This metric also works in the multi-class setting. .. sourcecode:: python # Targets and Predictions >>> targets = turicreate.SArray([ 1, 0, 2, 1]) >>> predictions = turicreate.SArray([[.1, .8, 0.1], ... [.9, .1, 0.0], ... [.8, .1, 0.1], ... [.3, .6, 0.1]]) # Compute the ROC curve. >>> roc_curve = turicreate.evaluation.roc_curve(targets, predictions) +-----------+-----+-----+---+---+-------+ | threshold | fpr | tpr | p | n | class | +-----------+-----+-----+---+---+-------+ | 0.0 | 1.0 | 1.0 | 1 | 3 | 0 | | 1e-05 | 1.0 | 1.0 | 1 | 3 | 0 | | 2e-05 | 1.0 | 1.0 | 1 | 3 | 0 | | 3e-05 | 1.0 | 1.0 | 1 | 3 | 0 | | 4e-05 | 1.0 | 1.0 | 1 | 3 | 0 | | 5e-05 | 1.0 | 1.0 | 1 | 3 | 0 | | 6e-05 | 1.0 | 1.0 | 1 | 3 | 0 | | 7e-05 | 1.0 | 1.0 | 1 | 3 | 0 | | 8e-05 | 1.0 | 1.0 | 1 | 3 | 0 | | 9e-05 | 1.0 | 1.0 | 1 | 3 | 0 | +-----------+-----+-----+---+---+-------+ [300003 rows x 6 columns] This metric also works for string classes. .. sourcecode:: python # Targets and Predictions >>> targets = turicreate.SArray(["cat", "dog", "foosa", "dog"]) >>> predictions = turicreate.SArray([[.1, .8, 0.1], ... [.9, .1, 0.0], ... [.8, .1, 0.1], ... [.3, .6, 0.1]]) # Compute the ROC curve. >>> roc_curve = turicreate.evaluation.roc_curve(targets, predictions) +-----------+-----+-----+---+---+-------+ | threshold | fpr | tpr | p | n | class | +-----------+-----+-----+---+---+-------+ | 0.0 | 1.0 | 1.0 | 1 | 3 | cat | | 1e-05 | 1.0 | 1.0 | 1 | 3 | cat | | 2e-05 | 1.0 | 1.0 | 1 | 3 | cat | | 3e-05 | 1.0 | 1.0 | 1 | 3 | cat | | 4e-05 | 1.0 | 1.0 | 1 | 3 | cat | | 5e-05 | 1.0 | 1.0 | 1 | 3 | cat | | 6e-05 | 1.0 | 1.0 | 1 | 3 | cat | | 7e-05 | 1.0 | 1.0 | 1 | 3 | cat | | 8e-05 | 1.0 | 1.0 | 1 | 3 | cat | | 9e-05 | 1.0 | 1.0 | 1 | 3 | cat | +-----------+-----+-----+---+---+-------+ [300003 rows x 6 columns] """ _supervised_evaluation_error_checking(targets, predictions) _check_categorical_option_type('average', average, [None]) _check_prob_and_prob_vector(predictions) _check_target_not_float(targets) _check_index_map(index_map) opts = {"average": average, "binary": predictions.dtype in [int, float]} if index_map is not None: opts['index_map'] = index_map return _turicreate.extensions._supervised_streaming_evaluator( targets, predictions, "roc_curve", opts)
def recall(targets, predictions, average='macro'): r""" Compute the recall score for classification tasks. The recall score quantifies the ability of a classifier to predict `positive` examples. Recall can be interpreted as the probability that a randomly selected `positive` example is correctly identified by the classifier. The score is in the range [0,1] with 0 being the worst, and 1 being perfect. The recall score is defined as the ratio: .. math:: \frac{tp}{tp + fn} where `tp` is the number of true positives and `fn` the number of false negatives. Parameters ---------- targets : SArray Ground truth class labels. The SArray can be of any type. predictions : SArray The prediction that corresponds to each target value. This SArray must have the same length as ``targets`` and must be of the same type as the ``targets`` SArray. average : string, [None, 'macro' (default), 'micro'] Metric averaging strategies for multiclass classification. Averaging strategies can be one of the following: - None: No averaging is performed and a single metric is returned for each class. - 'micro': Calculate metrics globally by counting the total true positives, false negatives, and false positives. - 'macro': Calculate metrics for each label and find their unweighted mean. This does not take label imbalance into account. Returns ------- out : float (for binary classification) or dict[float] Score for the positive class (for binary classification) or an average score for each class for multi-class classification. If `average=None`, then a dictionary is returned where the key is the class label and the value is the score for the corresponding class label. Notes ----- - For binary classification, when the target label is of type "string", then the labels are sorted alphanumerically and the largest label is chosen as the "positive" label. For example, if the classifier labels are {"cat", "dog"}, then "dog" is chosen as the positive label for the binary classification case. See Also -------- confusion_matrix, accuracy, precision, f1_score Examples -------- .. sourcecode:: python # Targets and Predictions >>> targets = turicreate.SArray([0, 1, 2, 3, 0, 1, 2, 3]) >>> predictions = turicreate.SArray([1, 0, 2, 1, 3, 1, 2, 1]) # Micro average of the recall scores for each class. >>> turicreate.evaluation.recall(targets, predictions, ... average = 'micro') 0.375 # Macro average of the recall scores for each class. >>> turicreate.evaluation.recall(targets, predictions, ... average = 'macro') 0.375 # Recall score for each class. >>> turicreate.evaluation.recall(targets, predictions, ... average = None) {0: 0.0, 1: 0.5, 2: 1.0, 3: 0.0} This metric also works for string classes. .. sourcecode:: python # Targets and Predictions >>> targets = turicreate.SArray( ... ["cat", "dog", "foosa", "snake", "cat", "dog", "foosa", "snake"]) >>> predictions = turicreate.SArray( ... ["dog", "cat", "foosa", "dog", "snake", "dog", "cat", "dog"]) # Micro average of the recall scores for each class. >>> turicreate.evaluation.recall(targets, predictions, ... average = 'micro') 0.375 # Macro average of the recall scores for each class. >>> turicreate.evaluation.recall(targets, predictions, ... average = 'macro') 0.375 # Recall score for each class. >>> turicreate.evaluation.recall(targets, predictions, ... average = None) {0: 0.0, 1: 0.5, 2: 1.0, 3: 0.0} """ _supervised_evaluation_error_checking(targets, predictions) _check_categorical_option_type('average', average, ['micro', 'macro', None]) _check_same_type_not_float(targets, predictions) opts = {"average": average} return _turicreate.extensions._supervised_streaming_evaluator( targets, predictions, "recall", opts)
def fbeta_score(targets, predictions, beta=1.0, average='macro'): r""" Compute the F-beta score. The F-beta score is the weighted harmonic mean of precision and recall. The score lies in the range [0,1] with 1 being ideal and 0 being the worst. The `beta` value is the weight given to `precision` vs `recall` in the combined score. `beta=0` considers only precision, as `beta` increases, more weight is given to recall with `beta > 1` favoring recall over precision. The F-beta score is defined as: .. math:: f_{\beta} = (1 + \beta^2) \times \frac{(p \times r)}{(\beta^2 p + r)} Where :math:`p` is the precision and :math:`r` is the recall. Parameters ---------- targets : SArray An SArray of ground truth class labels. Can be of any type except float. predictions : SArray The prediction that corresponds to each target value. This SArray must have the same length as ``targets`` and must be of the same type as the ``targets`` SArray. beta: float Weight of the `precision` term in the harmonic mean. average : string, [None, 'macro' (default), 'micro'] Metric averaging strategies for multiclass classification. Averaging strategies can be one of the following: - None: No averaging is performed and a single metric is returned for each class. - 'micro': Calculate metrics globally by counting the total true positives, false negatives and false positives. - 'macro': Calculate metrics for each label, and find their unweighted mean. This does not take label imbalance into account. For a more precise definition of `micro` and `macro` averaging refer to [1] below. Returns ------- out : float (for binary classification) or dict[float] (for multi-class, average=None) Score for the positive class (for binary classification) or an average score for each class for multi-class classification. If `average=None`, then a dictionary is returned where the key is the class label and the value is the score for the corresponding class label. Notes ----- - For binary classification, if the target label is of type "string", then the labels are sorted alphanumerically and the largest label is chosen as the "positive" label. For example, if the classifier labels are {"cat", "dog"}, then "dog" is chosen as the positive label for the binary classification case. See Also -------- confusion_matrix, accuracy, precision, recall, f1_score Examples -------- .. sourcecode:: python # Targets and Predictions >>> targets = turicreate.SArray([0, 1, 2, 3, 0, 1, 2, 3]) >>> predictions = turicreate.SArray([1, 0, 2, 1, 3, 1, 0, 1]) # Micro average of the F-Beta score >>> turicreate.evaluation.fbeta_score(targets, predictions, ... beta=2.0, average = 'micro') 0.25 # Macro average of the F-Beta score >>> turicreate.evaluation.fbeta_score(targets, predictions, ... beta=2.0, average = 'macro') 0.24305555555555558 # F-Beta score for each class. >>> turicreate.evaluation.fbeta_score(targets, predictions, ... beta=2.0, average = None) {0: 0.0, 1: 0.4166666666666667, 2: 0.5555555555555556, 3: 0.0} This metric also works when the targets are of type `str` .. sourcecode:: python # Targets and Predictions >>> targets = turicreate.SArray( ... ["cat", "dog", "foosa", "snake", "cat", "dog", "foosa", "snake"]) >>> predictions = turicreate.SArray( ... ["dog", "cat", "foosa", "dog", "snake", "dog", "cat", "dog"]) # Micro average of the F-Beta score >>> turicreate.evaluation.fbeta_score(targets, predictions, ... beta=2.0, average = 'micro') 0.25 # Macro average of the F-Beta score >>> turicreate.evaluation.fbeta_score(targets, predictions, ... beta=2.0, average = 'macro') 0.24305555555555558 # F-Beta score for each class. >>> turicreate.evaluation.fbeta_score(targets, predictions, ... beta=2.0, average = None) {'cat': 0.0, 'dog': 0.4166666666666667, 'foosa': 0.5555555555555556, 'snake': 0.0} References ---------- - [1] Sokolova, Marina, and Guy Lapalme. "A systematic analysis of performance measures for classification tasks." Information Processing & Management 45.4 (2009): 427-437. """ _supervised_evaluation_error_checking(targets, predictions) _check_categorical_option_type('average', average, ['micro', 'macro', None]) _check_same_type_not_float(targets, predictions) opts = {"beta": beta, "average": average} return _turicreate.extensions._supervised_streaming_evaluator( targets, predictions, "fbeta_score", opts)
def auc(targets, predictions, average='macro', index_map=None): r""" Compute the area under the ROC curve for the given targets and predictions. Parameters ---------- targets : SArray An SArray containing the observed values. For binary classification, the alpha-numerically first category is considered the reference category. predictions : SArray Prediction probability that corresponds to each target value. This must be of same length as ``targets``. average : string, [None, 'macro' (default)] Metric averaging strategies for multiclass classification. Averaging strategies can be one of the following: - None: No averaging is performed and a single metric is returned for each class. - 'macro': Calculate metrics for each label, and find their unweighted mean. This does not take label imbalance into account. index_map : dict[int], [None (default)] For binary classification, a dictionary mapping the two target labels to either 0 (negative) or 1 (positive). For multi-class classification, a dictionary mapping potential target labels to the associated index into the vectors in ``predictions``. Returns ------- out : float (for binary classification) or dict[float] Score for the positive class (for binary classification) or an average score for each class for multi-class classification. If `average=None`, then a dictionary is returned where the key is the class label and the value is the score for the corresponding class label. See Also -------- roc_curve, confusion_matrix Examples -------- .. sourcecode:: python >>> targets = turicreate.SArray([0, 1, 1, 0]) >>> predictions = turicreate.SArray([0.1, 0.35, 0.7, 0.99]) # Calculate the auc-score >>> auc = turicreate.evaluation.auc(targets, predictions) 0.5 This metric also works when the targets are strings (Here "cat" is chosen as the reference class). .. sourcecode:: python >>> targets = turicreate.SArray(["cat", "dog", "dog", "cat"]) >>> predictions = turicreate.SArray([0.1, 0.35, 0.7, 0.99]) # Calculate the auc-score >>> auc = turicreate.evaluation.auc(targets, predictions) 0.5 For the multi-class setting, the auc-score can be averaged. .. sourcecode:: python # Targets and Predictions >>> targets = turicreate.SArray([ 1, 0, 2, 1]) >>> predictions = turicreate.SArray([[.1, .8, 0.1], ... [.9, .1, 0.0], ... [.8, .1, 0.1], ... [.3, .6, 0.1]]) # Macro average of the scores for each class. >>> turicreate.evaluation.auc(targets, predictions, average = 'macro') 0.8888888888888888 # Scores for each class. >>> turicreate.evaluation.auc(targets, predictions, average = None) {0: 1.0, 1: 1.0, 2: 0.6666666666666666} This metric also works for "string" targets in the multi-class setting .. sourcecode:: python # Targets and Predictions >>> targets = turicreate.SArray([ "dog", "cat", "foosa", "dog"]) >>> predictions = turicreate.SArray([[.1, .8, 0.1], [.9, .1, 0.0], [.8, .1, 0.1], [.3, .6, 0.1]]) # Macro average. >>> auc = turicreate.evaluation.auc(targets, predictions) 0.8888888888888888 # Score for each class. >>> auc = turicreate.evaluation.auc(targets, predictions, average=None) {'cat': 1.0, 'dog': 1.0, 'foosa': 0.6666666666666666} """ _supervised_evaluation_error_checking(targets, predictions) _check_categorical_option_type('average', average, ['macro', None]) _check_prob_and_prob_vector(predictions) _check_target_not_float(targets) _check_index_map(index_map) opts = {"average": average, "binary": predictions.dtype in [int, float]} if index_map is not None: opts['index_map'] = index_map return _turicreate.extensions._supervised_streaming_evaluator( targets, predictions, "auc", opts)
def evaluate(self, dataset, metric = 'auto', verbose = True): """ Evaluate the model by making predictions of target values and comparing these to actual values. Parameters ---------- dataset : SFrame Dataset of new observations. Must include columns with the same names as the feature and target columns used for model training. Additional columns are ignored. metric : str optional Name of the evaluation metric. Possible values are: - 'auto' : Returns all available metrics. - 'accuracy' : Classification accuracy (micro average). - 'auc' : Area under the ROC curve (macro average) - 'precision' : Precision score (macro average) - 'recall' : Recall score (macro average) - 'f1_score' : F1 score (macro average) - 'confusion_matrix' : An SFrame with counts of possible prediction/true label combinations. - 'roc_curve' : An SFrame containing information needed for an ROC curve verbose : bool optional If True, prints prediction progress. Returns ------- out : dict Dictionary of evaluation results where the key is the name of the evaluation metric (e.g. `accuracy`) and the value is the evaluation score. See Also ---------- create, predict Examples ---------- .. sourcecode:: python >>> results = model.evaluate(data) >>> print(results['accuracy']) """ if self.target not in dataset.column_names(): raise _ToolkitError("Dataset provided to evaluate does not have " + "ground truth in the " + self.target + " column.") predicted = self._predict_with_probabilities(dataset, verbose) avail_metrics = ['accuracy', 'auc', 'precision', 'recall', 'f1_score', 'confusion_matrix', 'roc_curve'] _tkutl._check_categorical_option_type( 'metric', metric, avail_metrics + ['auto']) metrics = avail_metrics if metric == 'auto' else [metric] ret = {} if 'accuracy' in metrics: ret['accuracy'] = _evaluation.accuracy( dataset[self.target], predicted[self.target]) if 'auc' in metrics: ret['auc'] = _evaluation.auc( dataset[self.target], predicted['probability'], index_map=self._class_to_index) if 'precision' in metrics: ret['precision'] = _evaluation.precision( dataset[self.target], predicted[self.target]) if 'recall' in metrics: ret['recall'] = _evaluation.recall( dataset[self.target], predicted[self.target]) if 'f1_score' in metrics: ret['f1_score'] = _evaluation.f1_score( dataset[self.target], predicted[self.target]) if 'confusion_matrix' in metrics: ret['confusion_matrix'] = _evaluation.confusion_matrix( dataset[self.target], predicted[self.target]) if 'roc_curve' in metrics: ret['roc_curve'] = _evaluation.roc_curve( dataset[self.target], predicted['probability'], index_map=self._class_to_index) return ret
def predict_topk(self, dataset, output_type='probability', k=3, batch_size=64): """ Return top-k predictions for the ``dataset``. Predictions are returned as an SFrame with three columns: `id`, `class`, and `probability` or `rank` depending on the ``output_type`` parameter. Parameters ---------- dataset : SFrame Dataset to classify. Must include columns with the same names as the features. Additional columns are ignored. output_type : {'probability', 'rank'}, optional Choose the return type of the prediction: - `probability`: Probability associated with each label in the prediction. - `rank` : Rank associated with each label in the prediction. k : int, optional Number of classes to return for each input example. batch_size : int, optional If you are getting memory errors, try decreasing this value. If you have a powerful computer, increasing this value may improve performance. Returns ------- out : SFrame An SFrame with model predictions. See Also -------- predict, classify, evaluate Examples -------- >>> pred = m.predict_topk(validation_data, k=3) >>> pred +------+-------+-------------------+ | id | class | probability | +------+-------+-------------------+ | 0 | 4 | 0.995623886585 | | 0 | 9 | 0.0038311756216 | | 0 | 7 | 0.000301006948575 | | 1 | 1 | 0.928708016872 | | 1 | 3 | 0.0440889261663 | | 1 | 2 | 0.0176190119237 | | 2 | 3 | 0.996967732906 | | 2 | 2 | 0.00151345680933 | | 2 | 7 | 0.000637513934635 | | 3 | 1 | 0.998070061207 | | ... | ... | ... | +------+-------+-------------------+ """ # parameter checking if not isinstance(dataset, _tc.SFrame): raise TypeError('\'dataset\' parameter must be an SFrame') _tk_utils._check_categorical_option_type('output_type', output_type, ['probability', 'rank']) if(batch_size < 1): raise ValueError('\'batch_size\' must be greater than or equal to 1') prob_vector = self.predict(dataset, output_type='probability_vector', batch_size=64) id_to_label = self._id_to_class_label if output_type == 'probability': results = prob_vector.apply(lambda p: [ {'class': id_to_label[i], 'probability': p[i]} for i in reversed(_np.argsort(p)[-k:])] ) else: assert(output_type == 'rank') results = prob_vector.apply(lambda p: [ {'class': id_to_label[i], 'rank': rank} for rank, i in enumerate(reversed(_np.argsort(p)[-k:]))] ) results = _tc.SFrame({'X': results}) results = results.add_row_number() results = results.stack('X', new_column_name='X') results = results.unpack('X', column_name_prefix='') return results
def evaluate(self, dataset, metric='auto', batch_size=256, verbose=True): """ Evaluate the model by making predictions of target values and comparing these to actual values. Parameters ---------- dataset : SFrame Dataset of new observations. Must include columns with the same names as the feature and target columns used for model training. Additional columns are ignored. metric : str, optional Name of the evaluation metric. Possible values are: - 'auto' : Returns all available metrics. - 'accuracy' : Classification accuracy (micro average). - 'auc' : Area under the ROC curve (macro average) - 'precision' : Precision score (macro average) - 'recall' : Recall score (macro average) - 'f1_score' : F1 score (macro average) - 'log_loss' : Log loss - 'confusion_matrix' : An SFrame with counts of possible prediction/true label combinations. - 'roc_curve' : An SFrame containing information needed for an ROC curve batch_size : int, optional If you are getting memory errors, try decreasing this value. If you have a powerful computer, increasing this value may improve performance. verbose : bool, optional If True, prints prediction progress. Returns ------- out : dict Dictionary of evaluation results where the key is the name of the evaluation metric (e.g. `accuracy`) and the value is the evaluation score. See Also ---------- create, predict Examples ---------- .. sourcecode:: python >>> results = model.evaluate(data) >>> print(results['accuracy']) """ import os, json, math if self.target not in dataset.column_names(): raise _ToolkitError("Must provide ground truth column, '" + self.target + "' in the evaluation dataset.") predicted = self._predict_with_probabilities(dataset, batch_size, verbose) avail_metrics = ['accuracy', 'auc', 'precision', 'recall', 'f1_score', 'confusion_matrix', 'roc_curve', 'log_loss'] _tkutl._check_categorical_option_type( 'metric', metric, avail_metrics + ['auto']) metrics = avail_metrics if metric == 'auto' else [metric] labels = self.classes ret = {} if 'accuracy' in metrics: ret['accuracy'] = _evaluation.accuracy( dataset[self.target], predicted[self.target]) if 'auc' in metrics: ret['auc'] = _evaluation.auc( dataset[self.target], predicted['probability'], index_map=self._class_to_index) if 'precision' in metrics: ret['precision'] = _evaluation.precision( dataset[self.target], predicted[self.target]) if 'recall' in metrics: ret['recall'] = _evaluation.recall( dataset[self.target], predicted[self.target]) if 'f1_score' in metrics: ret['f1_score'] = _evaluation.f1_score( dataset[self.target], predicted[self.target]) if 'confusion_matrix' in metrics: ret['confusion_matrix'] = _evaluation.confusion_matrix( dataset[self.target], predicted[self.target]) if 'roc_curve' in metrics: ret['roc_curve'] = _evaluation.roc_curve( dataset[self.target], predicted['probability'], index_map=self._class_to_index) if 'log_loss' in metrics: ret['log_loss'] = _evaluation.log_loss( dataset[self.target], predicted['probability'], index_map=self._class_to_index) from .._evaluate_utils import ( entropy, confidence, relative_confidence, get_confusion_matrix, hclusterSort, l2Dist ) evaluation_result = {k: ret[k] for k in metrics} evaluation_result['num_test_examples'] = len(dataset) for k in ['num_classes', 'num_examples', 'training_loss', 'training_time', 'max_iterations']: evaluation_result[k] = getattr(self, k) #evaluation_result['input_image_shape'] = getattr(self, 'input_image_shape') evaluation_result["model_name"] = "Drawing Classifier" extended_test = dataset.add_column(predicted["probability"], 'probs') extended_test['label'] = dataset[self.target] extended_test = extended_test.add_columns( [extended_test.apply(lambda d: labels[d['probs'].index(confidence(d['probs']))]), extended_test.apply(lambda d: entropy(d['probs'])), extended_test.apply(lambda d: confidence(d['probs'])), extended_test.apply(lambda d: relative_confidence(d['probs']))], ['predicted_label', 'entropy', 'confidence', 'relative_confidence']) extended_test = extended_test.add_column(extended_test.apply(lambda d: d['label'] == d['predicted_label']), 'correct') sf_conf_mat = get_confusion_matrix(extended_test, labels) confidence_threshold = 0.5 hesitant_threshold = 0.2 evaluation_result['confidence_threshold'] = confidence_threshold evaluation_result['hesitant_threshold'] = hesitant_threshold evaluation_result['confidence_metric_for_threshold'] = 'relative_confidence' evaluation_result['conf_mat'] = list(sf_conf_mat) vectors = map(lambda l: {'name': l, 'pos':list(sf_conf_mat[sf_conf_mat['target_label']==l].sort('predicted_label')['norm_prob'])}, labels) evaluation_result['sorted_labels'] = hclusterSort(vectors, l2Dist)[0]['name'].split("|") per_l = extended_test.groupby(['label'], {'count': _tc.aggregate.COUNT, 'correct_count': _tc.aggregate.SUM('correct') }) per_l['recall'] = per_l.apply(lambda l: l['correct_count']*1.0 / l['count']) per_pl = extended_test.groupby(['predicted_label'], {'predicted_count': _tc.aggregate.COUNT, 'correct_count': _tc.aggregate.SUM('correct') }) per_pl['precision'] = per_pl.apply(lambda l: l['correct_count']*1.0 / l['predicted_count']) per_pl = per_pl.rename({'predicted_label': 'label'}) evaluation_result['label_metrics'] = list(per_l.join(per_pl, on='label', how='outer').select_columns(['label', 'count', 'correct_count', 'predicted_count', 'recall', 'precision'])) evaluation_result['labels'] = labels extended_test = extended_test.add_row_number('__idx').rename({'label': 'target_label'}) evaluation_result['test_data'] = extended_test evaluation_result['feature'] = self.feature return _Evaluation(evaluation_result)
def evaluate(self, dataset, metric='auto', verbose=True, batch_size=64): """ Evaluate the model by making predictions of target values and comparing these to actual values. Parameters ---------- dataset : SFrame Dataset to use for evaluation, must include a column with the same name as the features used for model training. Additional columns are ignored. metric : str, optional Name of the evaluation metric. Possible values are: - 'auto' : Returns all available metrics. - 'accuracy' : Classification accuracy (micro average). - 'auc' : Area under the ROC curve (macro average) - 'precision' : Precision score (macro average) - 'recall' : Recall score (macro average) - 'f1_score' : F1 score (macro average) - 'log_loss' : Log loss - 'confusion_matrix' : An SFrame with counts of possible prediction/true label combinations. - 'roc_curve' : An SFrame containing information needed for an ROC curve verbose : bool, optional If True, prints progress updates and model details. batch_size : int, optional If you are getting memory errors, try decreasing this value. If you have a powerful computer, increasing this value may improve performance. Returns ------- out : dict Dictionary of evaluation results where the key is the name of the evaluation metric (e.g. `accuracy`) and the value is the evaluation score. See Also ---------- classify, predict Examples ---------- .. sourcecode:: python >>> results = model.evaluate(data) >>> print results['accuracy'] """ from turicreate.toolkits import evaluation # parameter checking if not isinstance(dataset, _tc.SFrame): raise TypeError('\'dataset\' parameter must be an SFrame') avail_metrics = ['accuracy', 'auc', 'precision', 'recall', 'f1_score', 'log_loss', 'confusion_matrix', 'roc_curve'] _tk_utils._check_categorical_option_type( 'metric', metric, avail_metrics + ['auto']) if metric == 'auto': metrics = avail_metrics else: metrics = [metric] if _is_deep_feature_sarray(dataset[self.feature]): deep_features = dataset[self.feature] else: deep_features = get_deep_features(dataset[self.feature], verbose=verbose) data = _tc.SFrame({'deep features': deep_features}) data = data.add_row_number() missing_ids = data.filter_by([[]], 'deep features')['id'] if len(missing_ids) > 0: data = data.filter_by([[]], 'deep features', exclude=True) # Remove the labels for entries without deep features _logging.warning("Dropping %d examples which are less than 975ms in length." % len(missing_ids)) labels = dataset[[self.target]].add_row_number() labels = data.join(labels, how='left')[self.target] else: labels = dataset[self.target] assert(len(labels) == len(data)) if any([m in metrics for m in ('roc_curve', 'log_loss', 'auc')]): probs = self.predict(data['deep features'], output_type='probability_vector', verbose=verbose, batch_size=batch_size) if any([m in metrics for m in ('accuracy', 'precision', 'recall', 'f1_score', 'confusion_matrix')]): classes = self.predict(data['deep features'], output_type='class', verbose=verbose, batch_size=batch_size) ret = {} if 'accuracy' in metrics: ret['accuracy'] = evaluation.accuracy(labels, classes) if 'auc' in metrics: ret['auc'] = evaluation.auc(labels, probs, index_map=self._class_label_to_id) if 'precision' in metrics: ret['precision'] = evaluation.precision(labels, classes) if 'recall' in metrics: ret['recall'] = evaluation.recall(labels, classes) if 'f1_score' in metrics: ret['f1_score'] = evaluation.f1_score(labels, classes) if 'log_loss' in metrics: ret['log_loss'] = evaluation.log_loss(labels, probs, index_map=self._class_label_to_id) if 'confusion_matrix' in metrics: ret['confusion_matrix'] = evaluation.confusion_matrix(labels, classes) if 'roc_curve' in metrics: ret['roc_curve'] = evaluation.roc_curve(labels, probs, index_map=self._class_label_to_id) return ret
def predict_topk(self, dataset, output_type="probability", k=3, batch_size=256): """ Return top-k predictions for the ``dataset``, using the trained model. Predictions are returned as an SFrame with three columns: `id`, `class`, and `probability` or `rank`, depending on the ``output_type`` parameter. Parameters ---------- dataset : SFrame | SArray | turicreate.Image Drawings to be classified. If dataset is an SFrame, it must include columns with the same names as the features used for model training, but does not require a target column. Additional columns are ignored. output_type : {'probability', 'rank'}, optional Choose the return type of the prediction: - `probability`: Probability associated with each label in the prediction. - `rank` : Rank associated with each label in the prediction. k : int, optional Number of classes to return for each input example. batch_size : int, optional If you are getting memory errors, try decreasing this value. If you have a powerful computer, increasing this value may improve performance. Returns ------- out : SFrame An SFrame with model predictions. See Also -------- predict, evaluate Examples -------- >>> pred = m.predict_topk(validation_data, k=3) >>> print(pred) +----+-------+-------------------+ | id | class | probability | +----+-------+-------------------+ | 0 | 4 | 0.995623886585 | | 0 | 9 | 0.0038311756216 | | 0 | 7 | 0.000301006948575 | | 1 | 1 | 0.928708016872 | | 1 | 3 | 0.0440889261663 | | 1 | 2 | 0.0176190119237 | | 2 | 3 | 0.996967732906 | | 2 | 2 | 0.00151345680933 | | 2 | 7 | 0.000637513934635 | | 3 | 1 | 0.998070061207 | | .. | ... | ... | +----+-------+-------------------+ [35688 rows x 3 columns] """ _tkutl._check_categorical_option_type("output_type", output_type, ["probability", "rank"]) if not isinstance(k, int): raise TypeError("'k' must be an integer >= 1") if k <= 0: raise ValueError("'k' must be >= 1") if batch_size is not None and not isinstance(batch_size, int): raise TypeError("'batch_size' must be an integer >= 1") if batch_size is not None and batch_size < 1: raise ValueError("'batch_size' must be >= 1") prob_vector = self.predict( dataset, output_type='probability_vector', batch_size=batch_size) classes = self.classes if output_type == 'probability': results = prob_vector.apply(lambda p: [ {'class': classes[i], 'probability': p[i]} for i in reversed(_np.argsort(p)[-k:])] ) else: assert(output_type == 'rank') results = prob_vector.apply(lambda p: [ {'class': classes[index], 'rank': rank} for rank, index in enumerate(reversed(_np.argsort(p)[-k:]))] ) results = _tc.SFrame({'X': results}) results = results.add_row_number() results = results.stack('X', new_column_name='X') results = results.unpack('X', column_name_prefix='') return results
def predict_topk(self, dataset, output_type="probability", k=3, missing_value_action='auto'): """ Return top-k predictions for the ``dataset``, using the trained model. Predictions are returned as an SFrame with three columns: `id`, `class`, and `probability`, `margin`, or `rank`, depending on the ``output_type`` parameter. Input dataset size must be the same as for training of the model. Parameters ---------- dataset : SFrame A dataset that has the same columns that were used during training. If the target column exists in ``dataset`` it will be ignored while making predictions. output_type : {'probability', 'rank', 'margin'}, optional Choose the return type of the prediction: - `probability`: Probability associated with each label in the prediction. - `rank` : Rank associated with each label in the prediction. - `margin` : Margin associated with each label in the prediction. k : int, optional Number of classes to return for each input example. missing_value_action : str, optional Action to perform when missing values are encountered. Can be one of: - 'auto': By default the model will treat missing value as is. - 'impute': Proceed with evaluation by filling in the missing values with the mean of the training data. Missing values are also imputed if an entire column of data is missing during evaluation. - 'error': Do not proceed with evaluation and terminate with an error message. Returns ------- out : SFrame An SFrame with model predictions. See Also -------- predict, classify, evaluate Examples -------- >>> pred = m.predict_topk(validation_data, k=3) >>> pred +--------+-------+-------------------+ | id | class | probability | +--------+-------+-------------------+ | 0 | 4 | 0.995623886585 | | 0 | 9 | 0.0038311756216 | | 0 | 7 | 0.000301006948575 | | 1 | 1 | 0.928708016872 | | 1 | 3 | 0.0440889261663 | | 1 | 2 | 0.0176190119237 | | 2 | 3 | 0.996967732906 | | 2 | 2 | 0.00151345680933 | | 2 | 7 | 0.000637513934635 | | 3 | 1 | 0.998070061207 | | ... | ... | ... | +--------+-------+-------------------+ [35688 rows x 3 columns] """ _check_categorical_option_type('output_type', output_type, ['rank', 'margin', 'probability']) if missing_value_action == 'auto': missing_value_action = _sl.select_default_missing_value_policy( self, 'predict') # Low latency path if isinstance(dataset, list): return _turicreate.extensions._fast_predict_topk( self.__proxy__, dataset, output_type, missing_value_action, k) if isinstance(dataset, dict): return _turicreate.extensions._fast_predict_topk( self.__proxy__, [dataset], output_type, missing_value_action, k) options = dict() options.update({ 'model': self.__proxy__, 'model_name': self.__name__, 'dataset': dataset, 'output_type': output_type, 'topk': k, 'missing_value_action': missing_value_action }) target = _turicreate.toolkits._main.run( 'supervised_learning_predict_topk', options) return _map_unity_proxy_to_object(target['predicted'])
def predict(self, data, output_type='class', batch_size=256, verbose=True): """ Predict on an SFrame or SArray of drawings, or on a single drawing. Parameters ---------- data : SFrame | SArray | tc.Image The drawing(s) on which to perform drawing classification. If dataset is an SFrame, it must have a column with the same name as the feature column during training. Additional columns are ignored. If the data is a single drawing, it can be either of type tc.Image, in which case it is a bitmap-based drawing input, or of type list, in which case it is a stroke-based drawing input. output_type : {'probability', 'class', 'probability_vector'}, optional Form of the predictions which are one of: - 'class': Class prediction. For multi-class classification, this returns the class with maximum probability. - 'probability': Prediction probability associated with the True class (not applicable for multi-class classification) - 'probability_vector': Prediction probability associated with each class as a vector. Label ordering is dictated by the ``classes`` member variable. batch_size : int, optional If you are getting memory errors, try decreasing this value. If you have a powerful computer, increasing this value may improve performance. verbose : bool, optional If True, prints prediction progress. Returns ------- out : SArray An SArray with model predictions. Each element corresponds to a drawing and contains a single value corresponding to the predicted label. Each prediction will have type integer or string depending on the type of the classes the model was trained on. If `data` is a single drawing, the return value will be a single prediction. See Also -------- evaluate Examples -------- .. sourcecode:: python # Make predictions >>> pred = model.predict(data) # Print predictions, for a better overview >>> print(pred) dtype: int Rows: 10 [3, 4, 3, 3, 4, 5, 8, 8, 8, 4] """ _tkutl._check_categorical_option_type("output_type", output_type, ["probability", "class", "probability_vector"]) if isinstance(data, _tc.SArray): predicted = self._predict_with_probabilities( _tc.SFrame({ self.feature: data }), batch_size, verbose ) elif isinstance(data, _tc.SFrame): predicted = self._predict_with_probabilities(data, batch_size, verbose) else: # single input predicted = self._predict_with_probabilities( _tc.SFrame({ self.feature: [data] }), batch_size, verbose ) if output_type == "class": return predicted[self.target] elif output_type == "probability": if len(self.classes) <= 2: _class_to_index = self._class_to_index target = self.target return predicted.apply( lambda row: row["probability"][_class_to_index[row[target]]]) else: raise _ToolkitError("Use probability vector in case of multi-class classification") else: assert (output_type == "probability_vector") return predicted["probability"]
def get_deep_features(images, model_name, batch_size=64, verbose=True): """ Extracts features from images from a specific model. Parameters ---------- images : SArray Input data. model_name : string string optional Uses a pretrained model to bootstrap an image classifier: - "resnet-50" : Uses a pretrained resnet model. Exported Core ML model will be ~90M. - "squeezenet_v1.1" : Uses a pretrained squeezenet model. Exported Core ML model will be ~4.7M. - "VisionFeaturePrint_Scene": Uses an OS internal feature extractor. Only on available on iOS 12.0+, macOS 10.14+ and tvOS 12.0+. Exported Core ML model will be ~41K. Models are downloaded from the internet if not available locally. Once downloaded, the models are cached for future use. Returns ------- out : SArray Returns an SArray with all the extracted features. See Also -------- turicreate.image_classifier.create turicreate.image_similarity.create Examples -------- >>> url = 'https://static.turi.com/datasets/images/nested' >>> image_sframe = turicreate.load_images(url) >>> image_sarray = image_sframe["image"] >>> deep_features_sframe = turicreate.image_analysis.get_deep_features(image_sarray, model_name="resnet-50") """ # Check model parameter allowed_models = list(_pre_trained_models.IMAGE_MODELS.keys()) if _mac_ver() >= (10, 14): allowed_models.append("VisionFeaturePrint_Scene") _tkutl._check_categorical_option_type("model", model_name, allowed_models) # Check images parameter if not isinstance(images, _tc.SArray): raise TypeError( "Unrecognized type for 'images'. An SArray is expected.") if len(images) == 0: raise _ToolkitError( "Unable to extract features on an empty SArray object") if batch_size < 1: raise ValueError("'batch_size' must be greater than or equal to 1") # Extract features feature_extractor = _image_feature_extractor._create_feature_extractor( model_name) images_sf = _tc.SFrame({"image": images}) return feature_extractor.extract_features(images_sf, "image", verbose=verbose, batch_size=batch_size)
def predict(self, dataset, output_type="class", output_frequency="per_row"): """ Return predictions for ``dataset``, using the trained activity classifier. Predictions can be generated as class labels, or as a probability vector with probabilities for each class. The activity classifier generates a single prediction for each ``prediction_window`` rows in ``dataset``, per ``session_id``. The number of these predictions is smaller than the length of ``dataset``. By default, when ``output_frequency='per_row'``, each prediction is repeated ``prediction_window`` to return a prediction for each row of ``dataset``. Use ``output_frequency=per_window`` to get the unreplicated predictions. Parameters ---------- dataset : SFrame Dataset of new observations. Must include columns with the same names as the features used for model training, but does not require a target column. Additional columns are ignored. output_type : {'class', 'probability_vector'}, optional Form of each prediction which is one of: - 'probability_vector': Prediction probability associated with each class as a vector. The probability of the first class (sorted alphanumerically by name of the class in the training set) is in position 0 of the vector, the second in position 1 and so on. - 'class': Class prediction. This returns the class with maximum probability. output_frequency : {'per_row', 'per_window'}, optional The frequency of the predictions which is one of: - 'per_window': Return a single prediction for each ``prediction_window`` rows in ``dataset`` per ``session_id``. - 'per_row': Convenience option to make sure the number of predictions match the number of rows in the dataset. Each prediction from the model is repeated ``prediction_window`` times during that window. Returns ------- out : SArray | SFrame If ``output_frequency`` is 'per_row' return an SArray with predictions for each row in ``dataset``. If ``output_frequency`` is 'per_window' return an SFrame with predictions for ``prediction_window`` rows in ``dataset``. See Also ---------- create, evaluate, classify Examples -------- .. sourcecode:: python # One prediction per row >>> probability_predictions = model.predict( ... data, output_type='probability_vector', output_frequency='per_row')[:4] >>> probability_predictions dtype: array Rows: 4 [array('d', [0.01857384294271469, 0.0348394550383091, 0.026018327102065086]), array('d', [0.01857384294271469, 0.0348394550383091, 0.026018327102065086]), array('d', [0.01857384294271469, 0.0348394550383091, 0.026018327102065086]), array('d', [0.01857384294271469, 0.0348394550383091, 0.026018327102065086])] # One prediction per window >>> class_predictions = model.predict( ... data, output_type='class', output_frequency='per_window') >>> class_predictions +---------------+------------+-----+ | prediction_id | session_id |class| +---------------+------------+-----+ | 0 | 3 | 5 | | 1 | 3 | 5 | | 2 | 3 | 5 | | 3 | 3 | 5 | | 4 | 3 | 5 | | 5 | 3 | 5 | | 6 | 3 | 5 | | 7 | 3 | 4 | | 8 | 3 | 4 | | 9 | 3 | 4 | | ... | ... | ... | +---------------+------------+-----+ """ _tkutl._check_categorical_option_type( "output_frequency", output_frequency, ["per_window", "per_row"] ) if output_frequency == "per_row": return self.__proxy__.predict(dataset, output_type) elif output_frequency == "per_window": return self.__proxy__.predict_per_window(dataset, output_type)
def evaluate(self, dataset, metric='auto'): """ Evaluate the model by making predictions of target values and comparing these to actual values. Parameters ---------- dataset : SFrame Dataset of new observations. Must include columns with the same names as the session_id, target and features used for model training. Additional columns are ignored. metric : str, optional Name of the evaluation metric. Possible values are: - 'auto' : Returns all available metrics. - 'accuracy' : Classification accuracy (micro average). - 'auc' : Area under the ROC curve (macro average) - 'precision' : Precision score (macro average) - 'recall' : Recall score (macro average) - 'f1_score' : F1 score (macro average) - 'log_loss' : Log loss - 'confusion_matrix' : An SFrame with counts of possible prediction/true label combinations. - 'roc_curve' : An SFrame containing information needed for an ROC curve Returns ------- out : dict Dictionary of evaluation results where the key is the name of the evaluation metric (e.g. `accuracy`) and the value is the evaluation score. See Also ---------- create, predict Examples ---------- .. sourcecode:: python >>> results = model.evaluate(data) >>> print results['accuracy'] """ avail_metrics = ['accuracy', 'auc', 'precision', 'recall', 'f1_score', 'log_loss', 'confusion_matrix', 'roc_curve'] _tkutl._check_categorical_option_type( 'metric', metric, avail_metrics + ['auto']) if metric == 'auto': metrics = avail_metrics else: metrics = [metric] probs = self.predict(dataset, output_type='probability_vector') classes = self.predict(dataset, output_type='class') ret = {} if 'accuracy' in metrics: ret['accuracy'] = _evaluation.accuracy(dataset[self.target], classes) if 'auc' in metrics: ret['auc'] = _evaluation.auc(dataset[self.target], probs) if 'precision' in metrics: ret['precision'] = _evaluation.precision(dataset[self.target], classes) if 'recall' in metrics: ret['recall'] = _evaluation.recall(dataset[self.target], classes) if 'f1_score' in metrics: ret['f1_score'] = _evaluation.f1_score(dataset[self.target], classes) if 'log_loss' in metrics: ret['log_loss'] = _evaluation.log_loss(dataset[self.target], probs) if 'confusion_matrix' in metrics: ret['confusion_matrix'] = _evaluation.confusion_matrix(dataset[self.target], classes) if 'roc_curve' in metrics: ret['roc_curve'] = _evaluation.roc_curve(dataset[self.target], probs) return ret
def create(dataset, label=None, feature=None, model='resnet-50', verbose=True): """ Create a :class:`ImageSimilarityModel` model. Parameters ---------- dataset : SFrame Input data. The column named by the 'feature' parameter will be extracted for modeling. label : string Name of the SFrame column with row labels to be used as uuid's to identify the data. If 'label' is set to None, row numbers are used to identify reference dataset rows when the model is queried. feature : string indicates that the SFrame has only column of Image type and that will Name of the column containing the input images. 'None' (the default) be used for similarity. model: string, optional Uses a pretrained model to bootstrap an image similarity model - "resnet-50" : Uses a pretrained resnet model. Models are downloaded from the internet if not available locally. Once downloaded, the models are cached for future use. verbose : bool, optional If True, print progress updates and model details. Returns ------- out : ImageSimilarityModel A trained :class:`ImageSimilarityModel` model. See Also -------- ImageSimilarityModel Examples -------- .. sourcecode:: python # Train an image similarity model >>> model = turicreate.image_similarity.create(data) # Query the model for similar images >>> similar_images = model.query(data) +-------------+-----------------+-------------------+------+ | query_label | reference_label | distance | rank | +-------------+-----------------+-------------------+------+ | 0 | 0 | 0.0 | 1 | | 0 | 519 | 12.5319706301 | 2 | | 0 | 1619 | 12.5563764596 | 3 | | 0 | 186 | 12.6132604915 | 4 | | 0 | 1809 | 12.9180964745 | 5 | | 1 | 1 | 2.02304872852e-06 | 1 | | 1 | 1579 | 11.4288186151 | 2 | | 1 | 1237 | 12.3764325949 | 3 | | 1 | 80 | 12.7264363676 | 4 | | 1 | 58 | 12.7675058558 | 5 | +-------------+-----------------+-------------------+------+ [500 rows x 4 columns] """ start_time = _time.time() # Check parameters _tkutl._check_categorical_option_type('model', model, _pre_trained_models.MODELS.keys()) if len(dataset) == 0: raise _ToolkitError('Unable to train on empty dataset') if (label is not None) and (label not in dataset.column_names()): raise _ToolkitError("Row label column '%s' does not exist" % label) if (feature is not None) and (feature not in dataset.column_names()): raise _ToolkitError("Image feature column '%s' does not exist" % feature) # Set defaults if feature is None: feature = _tkutl._find_only_image_column(dataset) # Load pre-trained model & feature extractor ptModel = _pre_trained_models.MODELS[model]() feature_extractor = _image_feature_extractor.MXFeatureExtractor(ptModel) # Extract features extracted_features = _tc.SFrame({ '__image_features__': feature_extractor.extract_features(dataset, feature, verbose=verbose), }) # Train a similarity model using the extracted features if label is not None: extracted_features[label] = dataset[label] nn_model = _tc.nearest_neighbors.create(extracted_features, label=label, features=['__image_features__'], verbose=verbose) # Save the model state = { 'similarity_model': nn_model, 'model': model, 'feature_extractor': feature_extractor, 'input_image_shape': ptModel.input_image_shape, 'label': label, 'feature': feature, 'num_features': 1, 'num_examples': nn_model.num_examples, 'training_time': _time.time() - start_time, } return ImageSimilarityModel(state)
def create(dataset, num_topics=10, initial_topics=None, alpha=None, beta=.1, num_iterations=10, num_burnin=5, associations=None, verbose=False, print_interval=10, validation_set=None, method='auto'): """ Create a topic model from the given data set. A topic model assumes each document is a mixture of a set of topics, where for each topic some words are more likely than others. One statistical approach to do this is called a "topic model". This method learns a topic model for the given document collection. Parameters ---------- dataset : SArray of type dict or SFrame with a single column of type dict A bag of words representation of a document corpus. Each element is a dictionary representing a single document, where the keys are words and the values are the number of times that word occurs in that document. num_topics : int, optional The number of topics to learn. initial_topics : SFrame, optional An SFrame with a column of unique words representing the vocabulary and a column of dense vectors representing probability of that word given each topic. When provided, these values are used to initialize the algorithm. alpha : float, optional Hyperparameter that controls the diversity of topics in a document. Smaller values encourage fewer topics per document. Provided value must be positive. Default value is 50/num_topics. beta : float, optional Hyperparameter that controls the diversity of words in a topic. Smaller values encourage fewer words per topic. Provided value must be positive. num_iterations : int, optional The number of iterations to perform. num_burnin : int, optional The number of iterations to perform when inferring the topics for documents at prediction time. verbose : bool, optional When True, print most probable words for each topic while printing progress. print_interval : int, optional The number of iterations to wait between progress reports. associations : SFrame, optional An SFrame with two columns named "word" and "topic" containing words and the topic id that the word should be associated with. These words are not considered during learning. validation_set : SArray of type dict or SFrame with a single column A bag of words representation of a document corpus, similar to the format required for `dataset`. This will be used to monitor model performance during training. Each document in the provided validation set is randomly split: the first portion is used estimate which topic each document belongs to, and the second portion is used to estimate the model's performance at predicting the unseen words in the test data. method : {'cgs', 'alias'}, optional The algorithm used for learning the model. - *cgs:* Collapsed Gibbs sampling - *alias:* AliasLDA method. Returns ------- out : TopicModel A fitted topic model. This can be used with :py:func:`~TopicModel.get_topics()` and :py:func:`~TopicModel.predict()`. While fitting is in progress, several metrics are shown, including: +------------------+---------------------------------------------------+ | Field | Description | +==================+===================================================+ | Elapsed Time | The number of elapsed seconds. | +------------------+---------------------------------------------------+ | Tokens/second | The number of unique words processed per second | +------------------+---------------------------------------------------+ | Est. Perplexity | An estimate of the model's ability to model the | | | training data. See the documentation on evaluate. | +------------------+---------------------------------------------------+ See Also -------- TopicModel, TopicModel.get_topics, TopicModel.predict, turicreate.SArray.dict_trim_by_keys, TopicModel.evaluate References ---------- - `Wikipedia - Latent Dirichlet allocation <http://en.wikipedia.org/wiki/Latent_Dirichlet_allocation>`_ - Alias method: Li, A. et al. (2014) `Reducing the Sampling Complexity of Topic Models. <http://www.sravi.org/pubs/fastlda-kdd2014.pdf>`_. KDD 2014. Examples -------- The following example includes an SArray of documents, where each element represents a document in "bag of words" representation -- a dictionary with word keys and whose values are the number of times that word occurred in the document: >>> docs = turicreate.SArray('https://static.turi.com/datasets/nytimes') Once in this form, it is straightforward to learn a topic model. >>> m = turicreate.topic_model.create(docs) It is also easy to create a new topic model from an old one -- whether it was created using Turi Create or another package. >>> m2 = turicreate.topic_model.create(docs, initial_topics=m['topics']) To manually fix several words to always be assigned to a topic, use the `associations` argument. The following will ensure that topic 0 has the most probability for each of the provided words: >>> from turicreate import SFrame >>> associations = SFrame({'word':['hurricane', 'wind', 'storm'], 'topic': [0, 0, 0]}) >>> m = turicreate.topic_model.create(docs, associations=associations) More advanced usage allows you to control aspects of the model and the learning method. >>> import turicreate as tc >>> m = tc.topic_model.create(docs, num_topics=20, # number of topics num_iterations=10, # algorithm parameters alpha=.01, beta=.1) # hyperparameters To evaluate the model's ability to generalize, we can create a train/test split where a portion of the words in each document are held out from training. >>> train, test = tc.text_analytics.random_split(.8) >>> m = tc.topic_model.create(train) >>> results = m.evaluate(test) >>> print results['perplexity'] """ dataset = _check_input(dataset) _check_categorical_option_type("method", method, ['auto', 'cgs', 'alias']) if method == 'cgs' or method == 'auto': model_name = 'cgs_topic_model' else: model_name = 'alias_topic_model' # If associations are provided, check they are in the proper format if associations is None: associations = _turicreate.SFrame({'word': [], 'topic': []}) if isinstance(associations, _turicreate.SFrame) and \ associations.num_rows() > 0: assert set(associations.column_names()) == set(['word', 'topic']), \ "Provided associations must be an SFrame containing a word column\ and a topic column." assert associations['word'].dtype == str, \ "Words must be strings." assert associations['topic'].dtype == int, \ "Topic ids must be of int type." if alpha is None: alpha = float(50) / num_topics if validation_set is not None: _check_input(validation_set) # Must be a single column if isinstance(validation_set, _turicreate.SFrame): column_name = validation_set.column_names()[0] validation_set = validation_set[column_name] (validation_train, validation_test) = _random_split(validation_set) else: validation_train = _SArray() validation_test = _SArray() opts = { 'model_name': model_name, 'data': dataset, 'num_topics': num_topics, 'num_iterations': num_iterations, 'print_interval': print_interval, 'alpha': alpha, 'beta': beta, 'num_burnin': num_burnin, 'associations': associations } # Initialize the model with basic parameters response = _turicreate.extensions._text.topicmodel_init(opts) m = TopicModel(response['model']) # If initial_topics provided, load it into the model if isinstance(initial_topics, _turicreate.SFrame): assert set(['vocabulary', 'topic_probabilities']) == \ set(initial_topics.column_names()), \ "The provided initial_topics does not have the proper format, \ e.g. wrong column names." observed_topics = initial_topics['topic_probabilities'].apply( lambda x: len(x)) assert all(observed_topics == num_topics), \ "Provided num_topics value does not match the number of provided initial_topics." # Rough estimate of total number of words weight = len(dataset) * 1000 opts = { 'model': m.__proxy__, 'topics': initial_topics['topic_probabilities'], 'vocabulary': initial_topics['vocabulary'], 'weight': weight } response = _turicreate.extensions._text.topicmodel_set_topics(opts) m = TopicModel(response['model']) # Train the model on the given data set and retrieve predictions opts = { 'model': m.__proxy__, 'data': dataset, 'verbose': verbose, 'validation_train': validation_train, 'validation_test': validation_test } response = _turicreate.extensions._text.topicmodel_train(opts) m = TopicModel(response['model']) return m
def create(dataset, target, feature=None, model='resnet-50', max_iterations=10, verbose=True, seed=None): """ Create a :class:`ImageClassifier` model. Parameters ---------- dataset : SFrame Input data. The column named by the 'feature' parameter will be extracted for modeling. target : string, or int Name of the column containing the target variable. The values in this column must be of string or integer type. String target variables are automatically mapped to integers in the order in which they are provided. For example, a target variable with 'cat' and 'dog' as possible values is mapped to 0 and 1 respectively with 0 being the base class and 1 being the reference class. Use `model.classes` to retrieve the order in which the classes are mapped. feature : string, optional indicates that the SFrame has only column of Image type and that will Name of the column containing the input images. 'None' (the default) indicates the only image column in `dataset` should be used as the feature. model : string optional Uses a pretrained model to bootstrap an image classifier - "resnet-50" : Uses a pretrained resnet model. - "squeezenet_v1.1" : Uses a pretrained squeezenet model. Models are downloaded from the internet if not available locally. Once downloaded, the models are cached for future use. max_iterations : float, optional The maximum number of allowed passes through the data. More passes over the data can result in a more accurately trained model. Consider increasing this (the default value is 10) if the training accuracy is low and the *Grad-Norm* in the display is large. verbose : bool, optional If True, prints progress updates and model details. seed : int, optional Seed for random number generation. Set this value to ensure that the same model is created every time. Returns ------- out : ImageClassifier A trained :class:`ImageClassifier` model. Examples -------- .. sourcecode:: python >>> model = turicreate.image_classifier.create(data, target='is_expensive') # Make predictions (in various forms) >>> predictions = model.predict(data) # predictions >>> predictions = model.classify(data) # predictions with confidence >>> predictions = model.predict_topk(data) # Top-5 predictions (multiclass) # Evaluate the model with ground truth data >>> results = model.evaluate(data) See Also -------- ImageClassifier """ start_time = _time.time() # Check parameters _tkutl._check_categorical_option_type('model', model, _pre_trained_models.MODELS.keys()) if len(dataset) == 0: raise _ToolkitError('Unable to train on empty dataset') if (feature is not None) and (feature not in dataset.column_names()): raise _ToolkitError("Image feature column '%s' does not exist" % feature) if target not in dataset.column_names(): raise _ToolkitError("Target column '%s' does not exist" % target) if feature is None: feature = _tkutl._find_only_image_column(dataset) # Load pre-trained model & feature extractor ptModel = _pre_trained_models.MODELS[model]() feature_extractor = _image_feature_extractor.MXFeatureExtractor(ptModel) # Extract features extracted_features = _tc.SFrame({ target: dataset[target], '__image_features__': feature_extractor.extract_features(dataset, feature, verbose=verbose), }) # Train a classifier using the extracted features extracted_features[target] = dataset[target] lr_model = _tc.logistic_classifier.create(extracted_features, features=['__image_features__'], target=target, max_iterations=max_iterations, seed=seed, verbose=verbose) # Save the model state = { 'classifier': lr_model, 'model': model, 'max_iterations': max_iterations, 'feature_extractor': feature_extractor, 'input_image_shape': ptModel.input_image_shape, 'target': target, 'feature': feature, 'num_features': 1, 'num_classes': lr_model.num_classes, 'classes': lr_model.classes, 'num_examples': lr_model.num_examples, 'training_time': _time.time() - start_time, 'training_loss': lr_model.training_loss, } return ImageClassifier(state)
def create(dataset, label=None, feature=None, model="resnet-50", verbose=True, batch_size=64): """ Create a :class:`ImageSimilarityModel` model. Parameters ---------- dataset : SFrame Input data. The column named by the 'feature' parameter will be extracted for modeling. label : string Name of the SFrame column with row labels to be used as uuid's to identify the data. If 'label' is set to None, row numbers are used to identify reference dataset rows when the model is queried. feature : string Name of the column containing the input images. 'None' (the default) indicates that the SFrame has only one column of Image type and that will be used for similarity. model: string, optional Uses a pretrained model to bootstrap an image similarity model - "resnet-50" : Uses a pretrained resnet model. - "squeezenet_v1.1" : Uses a pretrained squeezenet model. - "VisionFeaturePrint_Scene": Uses an OS internal feature extractor. Only on available on iOS 12.0+, macOS 10.14+ and tvOS 12.0+. Models are downloaded from the internet if not available locally. Once downloaded, the models are cached for future use. verbose : bool, optional If True, print progress updates and model details. batch_size : int, optional If you are getting memory errors, try decreasing this value. If you have a powerful computer, increasing this value may improve performance. Returns ------- out : ImageSimilarityModel A trained :class:`ImageSimilarityModel` model. See Also -------- ImageSimilarityModel Examples -------- .. sourcecode:: python # Train an image similarity model >>> model = turicreate.image_similarity.create(data) # Query the model for similar images >>> similar_images = model.query(data) +-------------+-----------------+-------------------+------+ | query_label | reference_label | distance | rank | +-------------+-----------------+-------------------+------+ | 0 | 0 | 0.0 | 1 | | 0 | 519 | 12.5319706301 | 2 | | 0 | 1619 | 12.5563764596 | 3 | | 0 | 186 | 12.6132604915 | 4 | | 0 | 1809 | 12.9180964745 | 5 | | 1 | 1 | 2.02304872852e-06 | 1 | | 1 | 1579 | 11.4288186151 | 2 | | 1 | 1237 | 12.3764325949 | 3 | | 1 | 80 | 12.7264363676 | 4 | | 1 | 58 | 12.7675058558 | 5 | +-------------+-----------------+-------------------+------+ [500 rows x 4 columns] """ start_time = _time.time() if not isinstance(dataset, _tc.SFrame): raise TypeError("'dataset' must be of type SFrame.") # Check parameters allowed_models = list(_pre_trained_models.IMAGE_MODELS.keys()) if _mac_ver() >= (10, 14): allowed_models.append("VisionFeaturePrint_Scene") # Also, to make sure existing code doesn't break, replace incorrect name # with the correct name version if model == "VisionFeaturePrint_Screen": print( "WARNING: Correct spelling of model name is VisionFeaturePrint_Scene. VisionFeaturePrint_Screen will be removed in future releases." ) model = "VisionFeaturePrint_Scene" _tkutl._check_categorical_option_type("model", model, allowed_models) if len(dataset) == 0: raise _ToolkitError("Unable to train on empty dataset") if (label is not None) and (label not in dataset.column_names()): raise _ToolkitError("Row label column '%s' does not exist" % label) if (feature is not None) and (feature not in dataset.column_names()): raise _ToolkitError("Image feature column '%s' does not exist" % feature) if batch_size < 1: raise ValueError("'batch_size' must be greater than or equal to 1") # Set defaults if feature is None: feature = _tkutl._find_only_image_column(dataset) feature_extractor = _image_feature_extractor._create_feature_extractor( model) # Extract features extracted_features = _tc.SFrame({ "__image_features__": feature_extractor.extract_features(dataset, feature, verbose=verbose, batch_size=batch_size), }) # Train a similarity model using the extracted features if label is not None: extracted_features[label] = dataset[label] nn_model = _tc.nearest_neighbors.create( extracted_features, label=label, features=["__image_features__"], verbose=verbose, ) # set input image shape if model in _pre_trained_models.IMAGE_MODELS: input_image_shape = _pre_trained_models.IMAGE_MODELS[ model].input_image_shape else: # model == VisionFeaturePrint_Scene input_image_shape = (3, 299, 299) # Save the model state = { "similarity_model": nn_model, "model": model, "feature_extractor": feature_extractor, "input_image_shape": input_image_shape, "label": label, "feature": feature, "num_features": 1, "num_examples": nn_model.num_examples, "training_time": _time.time() - start_time, } return ImageSimilarityModel(state)
def create(dataset, target, feature = None, model = 'resnet-50', validation_set='auto', max_iterations = 10, verbose = True, seed = None, batch_size=64): """ Create a :class:`ImageClassifier` model. Parameters ---------- dataset : SFrame Input data. The column named by the 'feature' parameter will be extracted for modeling. target : string, or int Name of the column containing the target variable. The values in this column must be of string or integer type. String target variables are automatically mapped to integers in the order in which they are provided. For example, a target variable with 'cat' and 'dog' as possible values is mapped to 0 and 1 respectively with 0 being the base class and 1 being the reference class. Use `model.classes` to retrieve the order in which the classes are mapped. feature : string, optional indicates that the SFrame has only column of Image type and that will Name of the column containing the input images. 'None' (the default) indicates the only image column in `dataset` should be used as the feature. model : string optional Uses a pretrained model to bootstrap an image classifier: - "resnet-50" : Uses a pretrained resnet model. Exported Core ML model will be ~90M. - "squeezenet_v1.1" : Uses a pretrained squeezenet model. Exported Core ML model will be ~4.7M. - "VisionFeaturePrint_Screen": Uses an OS internal feature extractor. Only on available on iOS 12.0+, macOS 10.14+ and tvOS 12.0+. Exported Core ML model will be ~41K. Models are downloaded from the internet if not available locally. Once downloaded, the models are cached for future use. validation_set : SFrame, optional A dataset for monitoring the model's generalization performance. The format of this SFrame must be the same as the training set. By default this argument is set to 'auto' and a validation set is automatically sampled and used for progress printing. If validation_set is set to None, then no additional metrics are computed. The default value is 'auto'. max_iterations : float, optional The maximum number of allowed passes through the data. More passes over the data can result in a more accurately trained model. Consider increasing this (the default value is 10) if the training accuracy is low and the *Grad-Norm* in the display is large. verbose : bool, optional If True, prints progress updates and model details. seed : int, optional Seed for random number generation. Set this value to ensure that the same model is created every time. batch_size : int, optional If you are getting memory errors, try decreasing this value. If you have a powerful computer, increasing this value may improve performance. Returns ------- out : ImageClassifier A trained :class:`ImageClassifier` model. Examples -------- .. sourcecode:: python >>> model = turicreate.image_classifier.create(data, target='is_expensive') # Make predictions (in various forms) >>> predictions = model.predict(data) # predictions >>> predictions = model.classify(data) # predictions with confidence >>> predictions = model.predict_topk(data) # Top-5 predictions (multiclass) # Evaluate the model with ground truth data >>> results = model.evaluate(data) See Also -------- ImageClassifier """ start_time = _time.time() # Check model parameter allowed_models = list(_pre_trained_models.MODELS.keys()) if _mac_ver() >= (10,14): allowed_models.append('VisionFeaturePrint_Screen') _tkutl._check_categorical_option_type('model', model, allowed_models) # Check dataset parameter if len(dataset) == 0: raise _ToolkitError('Unable to train on empty dataset') if (feature is not None) and (feature not in dataset.column_names()): raise _ToolkitError("Image feature column '%s' does not exist" % feature) if target not in dataset.column_names(): raise _ToolkitError("Target column '%s' does not exist" % target) if(batch_size < 1): raise ValueError("'batch_size' must be greater than or equal to 1") if not (isinstance(validation_set, _tc.SFrame) or validation_set == 'auto' or validation_set is None): raise TypeError("Unrecognized value for 'validation_set'.") if feature is None: feature = _tkutl._find_only_image_column(dataset) feature_extractor = _image_feature_extractor._create_feature_extractor(model) # Extract features extracted_features = _tc.SFrame({ target: dataset[target], '__image_features__': feature_extractor.extract_features(dataset, feature, verbose=verbose, batch_size=batch_size), }) if isinstance(validation_set, _tc.SFrame): extracted_features_validation = _tc.SFrame({ target: validation_set[target], '__image_features__': feature_extractor.extract_features(validation_set, feature, verbose=verbose, batch_size=batch_size), }) else: extracted_features_validation = validation_set # Train a classifier using the extracted features extracted_features[target] = dataset[target] lr_model = _tc.logistic_classifier.create(extracted_features, features=['__image_features__'], target=target, max_iterations=max_iterations, validation_set=extracted_features_validation, seed=seed, verbose=verbose) # set input image shape if model in _pre_trained_models.MODELS: input_image_shape = _pre_trained_models.MODELS[model].input_image_shape else: # model == VisionFeaturePrint_Screen input_image_shape = (3, 299, 299) # Save the model state = { 'classifier': lr_model, 'model': model, 'max_iterations': max_iterations, 'feature_extractor': feature_extractor, 'input_image_shape': input_image_shape, 'target': target, 'feature': feature, 'num_features': 1, 'num_classes': lr_model.num_classes, 'classes': lr_model.classes, 'num_examples': lr_model.num_examples, 'training_time': _time.time() - start_time, 'training_loss': lr_model.training_loss, } return ImageClassifier(state)
def create(dataset, annotations=None, feature=None, model='darknet-yolo', classes=None, max_iterations=0, verbose=True, **kwargs): """ Create a :class:`ObjectDetector` model. Parameters ---------- dataset : SFrame Input data. The columns named by the ``feature`` and ``annotations`` parameters will be extracted for training the detector. annotations : string Name of the column containing the object detection annotations. This column should be a list of dictionaries, with each dictionary representing a bounding box of an object instance. Here is an example of the annotations for a single image with two object instances:: [{'label': 'dog', 'type': 'rectangle', 'coordinates': {'x': 223, 'y': 198, 'width': 130, 'height': 230}}, {'label': 'cat', 'type': 'rectangle', 'coordinates': {'x': 40, 'y': 73, 'width': 80, 'height': 123}}] The value for `x` is the horizontal center of the box paired with `width` and `y` is the vertical center of the box paired with `height`. 'None' (the default) indicates the only list column in `dataset` should be used for the annotations. feature : string Name of the column containing the input images. 'None' (the default) indicates the only image column in `dataset` should be used as the feature. model : string optional Object detection model to use: - "darknet-yolo" : Fast and medium-sized model classes : list optional List of strings containing the names of the classes of objects. Inferred from the data if not provided. max_iterations : int The number of training iterations. If 0, then it will be automatically be determined based on the amount of data you provide. verbose : bool, optional If True, print progress updates and model details. Returns ------- out : ObjectDetector A trained :class:`ObjectDetector` model. See Also -------- ObjectDetector Examples -------- .. sourcecode:: python # Train an object detector model >>> model = turicreate.object_detector.create(data) # Make predictions on the training set and as column to the SFrame >>> data['predictions'] = model.predict(data) # Visualize predictions by generating a new column of marked up images >>> data['image_pred'] = turicreate.object_detector.util.draw_bounding_boxes(data['image'], data['predictions']) """ _raise_error_if_not_sframe(dataset, "dataset") from ._mx_detector import YOLOLoss as _YOLOLoss from ._model import tiny_darknet as _tiny_darknet from ._sframe_loader import SFrameDetectionIter as _SFrameDetectionIter from ._manual_scheduler import ManualScheduler as _ManualScheduler import mxnet as _mx if len(dataset) == 0: raise _ToolkitError('Unable to train on empty dataset') _numeric_param_check_range('max_iterations', max_iterations, 0, _six.MAXSIZE) start_time = _time.time() supported_detectors = ['darknet-yolo'] if feature is None: feature = _tkutl._find_only_image_column(dataset) if verbose: print("Using '%s' as feature column" % feature) if annotations is None: annotations = _tkutl._find_only_column_of_type(dataset, target_type=list, type_name='list', col_name='annotations') if verbose: print("Using '%s' as annotations column" % annotations) _raise_error_if_not_detection_sframe(dataset, feature, annotations, require_annotations=True) _tkutl._check_categorical_option_type('model', model, supported_detectors) base_model = model.split('-', 1)[0] ref_model = _pre_trained_models.OBJECT_DETECTION_BASE_MODELS[base_model]() params = { 'anchors': [ (1.0, 2.0), (1.0, 1.0), (2.0, 1.0), (2.0, 4.0), (2.0, 2.0), (4.0, 2.0), (4.0, 8.0), (4.0, 4.0), (8.0, 4.0), (8.0, 16.0), (8.0, 8.0), (16.0, 8.0), (16.0, 32.0), (16.0, 16.0), (32.0, 16.0), ], 'grid_shape': [13, 13], 'batch_size': 32, 'aug_resize': 0, 'aug_rand_crop': 0.9, 'aug_rand_pad': 0.9, 'aug_rand_gray': 0.0, 'aug_aspect_ratio': 1.25, 'aug_hue': 0.05, 'aug_brightness': 0.05, 'aug_saturation': 0.05, 'aug_contrast': 0.05, 'aug_horizontal_flip': True, 'aug_min_object_covered': 0, 'aug_min_eject_coverage': 0.5, 'aug_area_range': (.15, 2), 'aug_pca_noise': 0.0, 'aug_max_attempts': 20, 'aug_inter_method': 2, 'lmb_coord_xy': 10.0, 'lmb_coord_wh': 10.0, 'lmb_obj': 100.0, 'lmb_noobj': 5.0, 'lmb_class': 2.0, 'non_maximum_suppression_threshold': 0.45, 'rescore': True, 'clip_gradients': 0.025, 'learning_rate': 1.0e-3, 'shuffle': True, } if '_advanced_parameters' in kwargs: # Make sure no additional parameters are provided new_keys = set(kwargs['_advanced_parameters'].keys()) set_keys = set(params.keys()) unsupported = new_keys - set_keys if unsupported: raise _ToolkitError('Unknown advanced parameters: {}'.format(unsupported)) params.update(kwargs['_advanced_parameters']) anchors = params['anchors'] num_anchors = len(anchors) num_gpus = _mxnet_utils.get_num_gpus_in_use(max_devices=params['batch_size']) batch_size_each = params['batch_size'] // max(num_gpus, 1) # Note, this may slightly alter the batch size to fit evenly on the GPUs batch_size = max(num_gpus, 1) * batch_size_each grid_shape = params['grid_shape'] input_image_shape = (3, grid_shape[0] * ref_model.spatial_reduction, grid_shape[1] * ref_model.spatial_reduction) try: instances = (dataset.stack(annotations, new_column_name='_bbox', drop_na=True) .unpack('_bbox', limit=['label'])) except (TypeError, RuntimeError): # If this fails, the annotation format isinvalid at the coarsest level raise _ToolkitError("Annotations format is invalid. Must be a list of " "dictionaries containing 'label' and 'coordinates'.") num_images = len(dataset) num_instances = len(instances) if classes is None: classes = instances['_bbox.label'].unique() classes = sorted(classes) # Make a class-to-index look-up table class_to_index = {name: index for index, name in enumerate(classes)} num_classes = len(classes) # Create data loader loader = _SFrameDetectionIter(dataset, batch_size=batch_size, input_shape=input_image_shape[1:], output_shape=grid_shape, anchors=anchors, class_to_index=class_to_index, aug_params=params, shuffle=params['shuffle'], loader_type='augmented', feature_column=feature, annotations_column=annotations) # Predictions per anchor box: x/y + w/h + object confidence + class probs preds_per_box = 5 + num_classes output_size = preds_per_box * num_anchors ymap_shape = (batch_size_each,) + tuple(grid_shape) + (num_anchors, preds_per_box) net = _tiny_darknet(output_size=output_size) loss = _YOLOLoss(input_shape=input_image_shape[1:], output_shape=grid_shape, batch_size=batch_size_each, num_classes=num_classes, anchors=anchors, parameters=params) base_lr = params['learning_rate'] if max_iterations == 0: # Set number of iterations through a heuristic num_iterations_raw = 5000 * _np.sqrt(num_instances) / batch_size num_iterations = 1000 * max(1, int(round(num_iterations_raw / 1000))) else: num_iterations = max_iterations steps = [num_iterations // 2, 3 * num_iterations // 4, num_iterations] steps_and_factors = [(step, 10**(-i)) for i, step in enumerate(steps)] steps, factors = zip(*steps_and_factors) lr_scheduler = _ManualScheduler(step=steps, factor=factors) ctx = _mxnet_utils.get_mxnet_context(max_devices=batch_size) net_params = net.collect_params() net_params.initialize(_mx.init.Xavier(), ctx=ctx) net_params['conv7_weight'].initialize(_mx.init.Xavier(factor_type='avg'), ctx=ctx, force_reinit=True) net_params['conv8_weight'].initialize(_mx.init.Uniform(0.00005), ctx=ctx, force_reinit=True) # Initialize object confidence low, preventing an unnecessary adjustment # period toward conservative estimates bias = _np.zeros(output_size, dtype=_np.float32) bias[4::preds_per_box] -= 6 from ._mx_detector import ConstantArray net_params['conv8_bias'].initialize(ConstantArray(bias), ctx, force_reinit=True) # Take a subset and then load the rest of the parameters. It is possible to # do allow_missing=True directly on net_params. However, this will more # easily hide bugs caused by names getting out of sync. ref_model.available_parameters_subset(net_params).load(ref_model.model_path, ctx) options = {'learning_rate': base_lr, 'lr_scheduler': lr_scheduler, 'momentum': 0.9, 'wd': 0.00005, 'rescale_grad': 1.0} clip_grad = params.get('clip_gradients') if clip_grad: options['clip_gradient'] = clip_grad trainer = _mx.gluon.Trainer(net.collect_params(), 'sgd', options) iteration = 0 smoothed_loss = None last_time = 0 while iteration < num_iterations: loader.reset() for batch in loader: data = _mx.gluon.utils.split_and_load(batch.data[0], ctx_list=ctx, batch_axis=0) label = _mx.gluon.utils.split_and_load(batch.label[0], ctx_list=ctx, batch_axis=0) Ls = [] with _mx.autograd.record(): for x, y in zip(data, label): z = net(x) z0 = _mx.nd.transpose(z, [0, 2, 3, 1]).reshape(ymap_shape) L = loss(z0, y) Ls.append(L) for L in Ls: L.backward() cur_loss = _np.mean([L.asnumpy()[0] for L in Ls]) if smoothed_loss is None: smoothed_loss = cur_loss else: smoothed_loss = 0.9 * smoothed_loss + 0.1 * cur_loss trainer.step(1) iteration += 1 cur_time = _time.time() if verbose and cur_time > last_time + 10: print('{now:%Y-%m-%d %H:%M:%S} Training {cur_iter:{width}d}/{num_iterations:{width}d} Loss {loss:6.3f}'.format( now=_datetime.now(), cur_iter=iteration, num_iterations=num_iterations, loss=smoothed_loss, width=len(str(num_iterations)))) last_time = cur_time if iteration == num_iterations: break training_time = _time.time() - start_time # Save the model state = { '_model': net, '_class_to_index': class_to_index, '_training_time_as_string': _seconds_as_string(training_time), '_grid_shape': grid_shape, 'anchors': anchors, 'model': model, 'classes': classes, 'batch_size': batch_size, 'input_image_shape': input_image_shape, 'feature': feature, 'non_maximum_suppression_threshold': params['non_maximum_suppression_threshold'], 'annotations': annotations, 'num_classes': num_classes, 'num_examples': num_images, 'num_bounding_boxes': num_instances, 'training_time': training_time, 'training_epochs': loader.cur_epoch, 'training_iterations': iteration, 'max_iterations': max_iterations, 'training_loss': smoothed_loss, } return ObjectDetector(state)
def predict_topk(self, dataset, output_type="probability", k=3, missing_value_action="auto"): """ Return top-k predictions for the ``dataset``, using the trained model. Predictions are returned as an SFrame with three columns: `id`, `class`, and `probability`, `margin`, or `rank`, depending on the ``output_type`` parameter. Input dataset size must be the same as for training of the model. Parameters ---------- dataset : SFrame A dataset that has the same columns that were used during training. If the target column exists in ``dataset`` it will be ignored while making predictions. output_type : {'probability', 'rank', 'margin'}, optional Choose the return type of the prediction: - `probability`: Probability associated with each label in the prediction. - `rank` : Rank associated with each label in the prediction. - `margin` : Margin associated with each label in the prediction. k : int, optional Number of classes to return for each input example. missing_value_action : str, optional Action to perform when missing values are encountered. Can be one of: - 'auto': Default to 'impute' - 'impute': Proceed with evaluation by filling in the missing values with the mean of the training data. Missing values are also imputed if an entire column of data is missing during evaluation. - 'error': Do not proceed with evaluation and terminate with an error message. Returns ------- out : SFrame An SFrame with model predictions. See Also -------- predict, classify, evaluate Examples -------- >>> pred = m.predict_topk(validation_data, k=3) >>> pred +--------+-------+-------------------+ | id | class | probability | +--------+-------+-------------------+ | 0 | 4 | 0.995623886585 | | 0 | 9 | 0.0038311756216 | | 0 | 7 | 0.000301006948575 | | 1 | 1 | 0.928708016872 | | 1 | 3 | 0.0440889261663 | | 1 | 2 | 0.0176190119237 | | 2 | 3 | 0.996967732906 | | 2 | 2 | 0.00151345680933 | | 2 | 7 | 0.000637513934635 | | 3 | 1 | 0.998070061207 | | ... | ... | ... | +--------+-------+-------------------+ [35688 rows x 3 columns] """ _check_categorical_option_type("output_type", output_type, ["rank", "margin", "probability"]) _check_categorical_option_type("missing_value_action", missing_value_action, ["auto", "impute", "error"]) if missing_value_action == "auto": missing_value_action = "impute" # Low latency path if isinstance(dataset, list): return self.__proxy__.fast_predict_topk(dataset, missing_value_action, output_type, k) if isinstance(dataset, dict): return self.__proxy__.fast_predict_topk([dataset], missing_value_action, output_type, k) # Fast path _raise_error_if_not_sframe(dataset, "dataset") if missing_value_action == "auto": missing_value_action = _sl.select_default_missing_value_policy( self, "predict") return self.__proxy__.predict_topk(dataset, missing_value_action, output_type, k)