Bob 2.0-based training of the binary Logistic Regression model
Algorithms have at least one input and one output. All algorithm endpoints are organized in groups. Groups are used by the platform to indicate which inputs and outputs are synchronized together. The first group is automatically synchronized with the channel defined by the block in which the algorithm is deployed.
| Endpoint Name | Data Format | Nature |
|---|---|---|
| asv_real_scores | elie_khoury/string_probe_scores/1 | Input |
| asv_probe_id | system/text/1 | Input |
| classifier | tutorial/linear_machine/1 | Output |
| Endpoint Name | Data Format | Nature |
|---|---|---|
| pad_scores_class | system/text/1 | Input |
| pad_file_id | system/text/1 | Input |
| pad_scores | system/float/1 | Input |
| Endpoint Name | Data Format | Nature |
|---|---|---|
| asv_attack_scores | elie_khoury/string_probe_scores/1 | Input |
| asv_attack_id | system/text/1 | Input |
xxxxxxxxxximport bobimport bob.learn.linearimport numpyimport mathdef zeromean_unitvar_norm(data, mean, std): """ Normalize the data with zero mean and unit variance. Mean and variance are in numpy.ndarray format""" return numpy.divide(data-mean,std)class Algorithm: def __init__(self): self.all_pad_scores = [] # scores from pad self.all_asv_scores = [] # scores form asv self.train_real = [] self.train_zimp = [] self.train_attack = None self.pad_real_id_matched = None self.pad_attack_id_matched = None def process(self, inputs, outputs): if self.pad_real_id_matched is None and self.pad_attack_id_matched is None: self.pad_real_id_matched = {} self.pad_attack_id_matched = {} # accumulates the scores from PAD while inputs['pad_scores'].hasMoreData(): inputs['pad_scores'].next() inputs['pad_scores_class'].next() inputs['pad_file_id'].next() score_pad = inputs['pad_scores'].data.value label_pad = inputs['pad_scores_class'].data.text file_id = inputs['pad_file_id'].data.text if label_pad == 'real': self.pad_real_id_matched[file_id] = score_pad # add real score matching file id else: self.pad_attack_id_matched[file_id] = score_pad # add attack score matching file id self.all_pad_scores.append(score_pad) # add score to the array with all if self.train_attack is None: self.train_attack = [] while inputs['asv_attack_id'].hasMoreData(): inputs['asv_attack_id'].next() inputs['asv_attack_scores'].next() asv_attack_id = inputs['asv_attack_id'].data.text asv_attack_scores = inputs['asv_attack_scores'].data corresponding_pad_attack_score = self.pad_attack_id_matched[asv_attack_id] for score in asv_attack_scores.scores: if score.template_identity == asv_attack_scores.client_identity: self.train_attack.append((score.score, corresponding_pad_attack_score)) self.all_asv_scores.append(score.score) # accumulate the scores from ASV asv_probe_id = inputs['asv_probe_id'].data.text print ("asv_probe_id: %s" % asv_probe_id) corresponding_pad_real_score = self.pad_real_id_matched[asv_probe_id] data_asv = inputs['asv_real_scores'].data for score in data_asv.scores: # if this is real score if score.template_identity == data_asv.client_identity: self.train_real.append((score.score, corresponding_pad_real_score)) # if this is zero-impostor score else: self.train_zimp.append((score.score, corresponding_pad_real_score)) self.all_asv_scores.append(score.score) if not(inputs.hasMoreData()): self.all_pad_scores = numpy.asarray(self.all_pad_scores, dtype=numpy.float64) self.all_asv_scores = numpy.asarray(self.all_asv_scores, dtype=numpy.float64) # compute a combined (asv,pad) mean and std pad_mean = numpy.average(self.all_pad_scores, axis=0) pad_std = numpy.std(self.all_pad_scores, axis=0) asv_mean = numpy.average(self.all_asv_scores, axis=0) asv_std = numpy.std(self.all_asv_scores, axis=0) scores_mean = numpy.asarray([asv_mean, pad_mean], dtype=numpy.float64) print ("scores_mean ", scores_mean) scores_std = numpy.asarray([asv_std, pad_std], dtype=numpy.float64) print ("scores_std ", scores_std) # put together and normalize real scores from pad and asv self.train_real = zeromean_unitvar_norm(self.train_real, scores_mean, scores_std) # put together and normalize zimp scores from pad and asv self.train_zimp = zeromean_unitvar_norm(self.train_zimp, scores_mean, scores_std) # put together and normalize attack scores from pad and asv self.train_attack = zeromean_unitvar_norm(self.train_attack, scores_mean, scores_std) print("self.train_real len: %d, self.train_zimp len: %d, self.train_attack len: %d" %(len(self.train_real), len(self.train_zimp), len(self.train_attack))) # trains the Logistic Regression classifier positives = self.train_real negatives = numpy.concatenate((self.train_zimp, self.train_attack), axis=0) trainer = bob.learn.linear.CGLogRegTrainer() machine = trainer.train(negatives, positives) if scores_mean is not None and scores_std is not None: machine.input_subtract = scores_mean machine.input_divide = scores_std # outputs data outputs["classifier"].write({ 'input_subtract': machine.input_subtract, 'input_divide': machine.input_divide, 'weights': machine.weights, 'biases': machine.biases, }) return TrueThe code for this algorithm in Python
The ruler at 80 columns indicate suggested POSIX line breaks (for readability).
The editor will automatically enlarge to accomodate the entirety of your input
Use keyboard shortcuts for search/replace and faster editing. For example, use Ctrl-F (PC) or Cmd-F (Mac) to search through this box
This algorithm will run a Logistic Regression model [LR] for a binary classification problem using features as inputs.
The inputs take feature vectors as input and a text flag indicating if the data is a hit (it should be 'real') or a miss.
| [LR] | https://en.wikipedia.org/wiki/Logistic_regression |
| Updated | Name | Databases/Protocols | Analyzers | |||
|---|---|---|---|---|---|---|
| pkorshunov/pkorshunov/isv-asv-pad-fusion-complete/1/asv_isv-pad_gmm-fusion_lr-pa | avspoof/2@physicalaccess_verification,avspoof/2@physicalaccess_verify_train,avspoof/2@physicalaccess_verify_train_spoof,avspoof/2@physicalaccess_antispoofing,avspoof/2@physicalaccess_verification_spoof | pkorshunov/spoof-score-fusion-roc_hist/1 |
This table shows the number of times this algorithm has been successfully run using the given environment. Note this does not provide sufficient information to evaluate if the algorithm will run when submitted to different conditions.