From d9f5534b840e7b87cdccbe68eadb3d8209c9ab07 Mon Sep 17 00:00:00 2001
From: Tiago Freitas Pereira <tiagofrepereira@gmail.com>
Date: Fri, 22 May 2020 22:00:52 +0200
Subject: [PATCH] Polishing ZT-Norm
---
.../pipelines/vanilla_biometrics/zt_norm.py | 62 ++++---------------
bob/bio/base/script/vanilla_biometrics.py | 15 ++---
.../test_vanilla_biometrics_score_norm.py | 20 ++----
3 files changed, 27 insertions(+), 70 deletions(-)
diff --git a/bob/bio/base/pipelines/vanilla_biometrics/zt_norm.py b/bob/bio/base/pipelines/vanilla_biometrics/zt_norm.py
index 0de212ec..0b58fb88 100644
--- a/bob/bio/base/pipelines/vanilla_biometrics/zt_norm.py
+++ b/bob/bio/base/pipelines/vanilla_biometrics/zt_norm.py
@@ -10,7 +10,7 @@ from bob.pipelines import DelayedSample, Sample, SampleSet
import numpy as np
import dask
import functools
-import pickle
+import cloudpickle
import os
from .score_writers import FourColumnsScoreWriter
import logging
@@ -78,6 +78,7 @@ class ZTNormPipeline(object):
allow_scoring_with_all_biometric_references=False,
):
+
self.transformer = self.train_background_model(background_model_samples)
# Create biometric samples
@@ -128,11 +129,6 @@ class ZTNormPipeline(object):
allow_scoring_with_all_biometric_references,
)
-
- # TODO: Do the score write
- #if self.vanilla_biometrics_pipeline.score_writer is not None:
- # return self.write_scores(scores)
-
return raw_scores, z_normed_scores, t_normed_scores, zt_normed_scores
def train_background_model(self, background_model_samples):
@@ -166,8 +162,6 @@ class ZTNormPipeline(object):
allow_scoring_with_all_biometric_references=False,
):
- # zprobe_samples = self._inject_references(zprobe_samples, biometric_references)
-
z_scores, z_probe_features = self.compute_scores(
zprobe_samples, biometric_references
)
@@ -214,10 +208,6 @@ class ZTNormPipeline(object):
allow_scoring_with_all_biometric_references=False,
):
- # z_probe_features = self._inject_references(
- # z_probe_features, t_biometric_references
- # )
-
# Reusing the zprobe_features and t_biometric_references
zt_scores = self.vanilla_biometrics_pipeline.biometric_algorithm.score_samples(
z_probe_features,
@@ -258,20 +248,6 @@ class ZTNorm(object):
# Axi1 1=TNORM
return (score - mu) / std
- """
- if axis == 1:
- return (
- score
- - np.tile(mu.reshape(N, 1), (1, score.shape[1]))
- ) / np.tile(std.reshape(N, 1), (1, score.shape[1]))
- else:
- return (
- score
- - np.tile(mu.reshape(1, N), (score.shape[0], 1))
- ) / np.tile(std.reshape(1, N), (score.shape[0], 1))
- """
-
-
def _compute_std(self, mu, norm_base_scores, axis=1):
# Reference: https://gitlab.idiap.ch/bob/bob.learn.em/-/blob/master/bob/learn/em/test/test_ztnorm.py
# Axis 0=ZNORM
@@ -317,16 +293,10 @@ class ZTNorm(object):
axis=0 computes CORRECTLY the statistics for ZNorm
axis=1 computes CORRECTLY the statistics for TNorm
"""
-
# Dumping all scores
- if isinstance(sampleset_for_norm[0][0], DelayedSample):
- score_floats = np.array(
- [f.data for sset in sampleset_for_norm for s in sset for f in s.data]
- )
- else:
- score_floats = np.array(
- [s.data for sset in sampleset_for_norm for s in sset]
- )
+ score_floats = np.array(
+ [s.data for sset in sampleset_for_norm for s in sset]
+ )
# Reshaping in PROBE vs BIOMETRIC_REFERENCES
n_probes = len(sampleset_for_norm)
@@ -334,23 +304,18 @@ class ZTNorm(object):
score_floats = score_floats.reshape((n_probes, n_references))
# AXIS ON THE MODELS
- big_mu = np.mean(score_floats, axis=axis)
- #big_std = np.std(score_floats, axis=axis)
+ big_mu = np.mean(score_floats, axis=axis)
big_std = self._compute_std(big_mu, score_floats, axis=axis)
# Creating statistics structure with subject id as the key
stats = {}
if axis == 0:
- # TODO: NEED TO SOLVE THIS FETCHING.
- # IT SHOULD BE TRANSPARENT
- if isinstance(sampleset_for_norm[0][0], DelayedSample):
- sset = sampleset_for_norm[0].samples[0].data
- else:
- sset = sampleset_for_norm[0]
-
- for mu, std, s in zip(big_mu, big_std, sset):
- stats[s.subject] = {"big_mu": mu, "big_std": std}
+ # Z-Norm is one statistic per biometric references
+ biometric_reference_subjects = [br.subject for br in sampleset_for_norm[0]]
+ for mu, std, s in zip(big_mu, big_std, biometric_reference_subjects):
+ stats[s] = {"big_mu": mu, "big_std": std}
else:
+ # T-Norm is one statistic per probe
for mu, std, sset in zip(big_mu, big_std, sampleset_for_norm):
stats[sset.subject] = {"big_mu": mu, "big_std": std}
@@ -359,7 +324,6 @@ class ZTNorm(object):
def _znorm_samplesets(self, probe_scores, stats):
# Normalizing
# TODO: THIS TENDS TO BE EXTREMLY SLOW
-
z_normed_score_samples = []
for probe_score in probe_scores:
z_normed_score_samples.append(self._apply_znorm(probe_score, stats))
@@ -497,10 +461,10 @@ class ZTNormCheckpointWrapper(object):
def _write_scores(self, samples, path):
os.makedirs(os.path.dirname(path), exist_ok=True)
- open(path, "wb").write(pickle.dumps(samples))
+ open(path, "wb").write(cloudpickle.dumps(samples))
def _load(self, path):
- return pickle.loads(open(path, "rb").read())
+ return cloudpickle.loads(open(path, "rb").read())
def _apply_znorm(self, probe_score, stats):
diff --git a/bob/bio/base/script/vanilla_biometrics.py b/bob/bio/base/script/vanilla_biometrics.py
index 89516f79..f54d3f64 100644
--- a/bob/bio/base/script/vanilla_biometrics.py
+++ b/bob/bio/base/script/vanilla_biometrics.py
@@ -157,7 +157,7 @@ def vanilla_biometrics(
"""
def _compute_scores(result, dask_client):
- if isinstance(result, Delayed):
+ if isinstance(result, Delayed) or isinstance(result, dask.bag.Bag):
if dask_client is not None:
result = result.compute(scheduler=dask_client)
else:
@@ -254,11 +254,12 @@ def vanilla_biometrics(
def _build_filename(score_file_name, suffix):
return os.path.join(score_file_name, suffix)
- # Running RAW_SCORES
+ # Running RAW_SCORES
raw_scores = _post_process_scores(
pipeline, raw_scores, _build_filename(score_file_name, "raw_scores")
)
- _compute_scores(raw_scores, dask_client)
+
+ _ = _compute_scores(raw_scores, dask_client)
# Z-SCORES
z_normed_scores = _post_process_scores(
@@ -266,7 +267,7 @@ def vanilla_biometrics(
z_normed_scores,
_build_filename(score_file_name, "z_normed_scores"),
)
- _compute_scores(z_normed_scores, dask_client)
+ _ = _compute_scores(z_normed_scores, dask_client)
# T-SCORES
t_normed_scores = _post_process_scores(
@@ -274,7 +275,7 @@ def vanilla_biometrics(
t_normed_scores,
_build_filename(score_file_name, "t_normed_scores"),
)
- _compute_scores(t_normed_scores, dask_client)
+ _ = _compute_scores(t_normed_scores, dask_client)
# ZT-SCORES
zt_normed_scores = _post_process_scores(
@@ -282,7 +283,7 @@ def vanilla_biometrics(
zt_normed_scores,
_build_filename(score_file_name, "zt_normed_scores"),
)
- _compute_scores(zt_normed_scores, dask_client)
+ _ = _compute_scores(zt_normed_scores, dask_client)
else:
@@ -296,7 +297,7 @@ def vanilla_biometrics(
post_processed_scores = _post_process_scores(
pipeline, result, score_file_name
)
- _compute_scores(post_processed_scores, dask_client)
+ _ = _compute_scores(post_processed_scores, dask_client)
if dask_client is not None:
dask_client.shutdown()
diff --git a/bob/bio/base/test/test_vanilla_biometrics_score_norm.py b/bob/bio/base/test/test_vanilla_biometrics_score_norm.py
index 99b57dba..eb9f1593 100644
--- a/bob/bio/base/test/test_vanilla_biometrics_score_norm.py
+++ b/bob/bio/base/test/test_vanilla_biometrics_score_norm.py
@@ -41,9 +41,9 @@ import copy
def zt_norm_stubs(references, probes, t_references, z_probes):
def _norm(scores, norm_base_scores, axis=1):
mu = np.mean(norm_base_scores, axis=axis)
-
- #old = True
- #if old:
+
+ # old = True
+ # if old:
# std = np.std(norm_base_scores, axis=axis)
# if axis == 1:
# return ((scores.T - mu) / std).T
@@ -96,7 +96,6 @@ def zt_norm_stubs(references, probes, t_references, z_probes):
)
) / np.tile(std.reshape(1, norm_base_scores.shape[1]), (scores.shape[0], 1))
-
n_reference = references.shape[0]
n_probes = probes.shape[0]
n_t_references = t_references.shape[0]
@@ -159,14 +158,7 @@ def test_norm_mechanics():
def _dump_scores_from_samples(scores, shape):
# We have to transpose because the tests are BIOMETRIC_REFERENCES vs PROBES
# and bob.bio.base is PROBES vs BIOMETRIC_REFERENCES
- if isinstance(scores[0][0], DelayedSample):
- return (
- np.array([f.data for sset in scores for s in sset for f in s.data])
- .reshape(shape)
- .T
- )
- else:
- return np.array([s.data for sset in scores for s in sset]).reshape(shape).T
+ return np.array([s.data for sset in scores for s in sset]).reshape(shape).T
with tempfile.TemporaryDirectory() as dir_name:
@@ -394,8 +386,8 @@ def test_norm_mechanics():
# With checkpoing
run(False, with_checkpoint=True)
run(False, with_checkpoint=True)
- # shutil.rmtree(dir_name) # Deleting the cache so it runs again from scratch
- # os.makedirs(dir_name, exist_ok=True)
+ shutil.rmtree(dir_name) # Deleting the cache so it runs again from scratch
+ os.makedirs(dir_name, exist_ok=True)
# With dask
run(True) # On memory
--
GitLab