Changeset 13

Timestamp:

07/19/15 15:05:42 (9 years ago)

Author:

tim

Message:

.

Location:

trunk

Files:

• bin/graph (modified) (8 diffs)
• bin/train (modified) (2 diffs)
• lib/nanownlib/__init__.py (modified) (5 diffs)
• lib/nanownlib/stats.py (modified) (3 diffs)
• lib/nanownlib/train.py (modified) (1 diff)

trunk/bin/graph

@@ -81 +81 @@
 diff_overtime.sort()
 
+print('packet_rtt diff mean: %f' % statistics.mean(diffs))
 print('packet_rtt diff median: %f' % statistics.median(diffs))
 print('packet_rtt diff midhinge: %f' % midsummary(diffs))
@@ -86 +87 @@
 print('packet_rtt diff quadsummary: %f' % quadsummary(diffs))
 print('packet_rtt diff ubersummary: %f' % ubersummary(diffs))
+print('packet_rtt diff septasummary: %f' % septasummary(diffs))
 print('packet_rtt diff MAD: %f' % mad(diffs))
 try:
@@ -91 +93 @@
     print('reported diff quadsummary: %f' % quadsummary(reported_diffs))
     print('reported diff ubersummary: %f' % ubersummary(reported_diffs))
+    print('reported diff septasummary: %f' % septasummary(reported_diffs))
     print('reported diff MAD: %f' % mad(reported_diffs))
 
-    import cProfile
-    start = time.time()
-    kresults = kfilter({},diffs)
+    #import cProfile
+    #start = time.time()
+    #kresults = kfilter({},diffs)
     #print('packet_rtt diff kfilter: ', numpy.mean(kresults['est']), kresults['var'])
-    print('packet_rtt diff kfilter: ', kresults['est'][-1], kresults['var'][-1])
-    kresults = kfilter({},reported_diffs)
+    #print('packet_rtt diff kfilter: ', kresults['est'][-1], kresults['var'][-1])
+    #kresults = kfilter({},reported_diffs)
     #print('reported diff kfilter: ', numpy.mean(kresults['est']), kresults['var'][-1])
-    print('reported diff kfilter: ', kresults['est'][-1], kresults['var'][-1])
-    print("kfilter time: %f" % (time.time()-start))
+    #print('reported diff kfilter: ', kresults['est'][-1], kresults['var'][-1])
+    #print("kfilter time: %f" % (time.time()-start))
 except:
     pass
@@ -111 +114 @@
 
 
+
+
+def testKalman4D(params=None):
+    from pykalman import KalmanFilter
+    train = db.subseries('train','long', offset=0)
+    test = db.subseries('test','long', offset=0)
+    null = db.subseries('train_null','long', offset=0)
+    measurements = numpy.asarray([(s['unusual_packet'],s['other_packet'],s['unusual_tsval'],s['other_tsval']) for s in (train+test)])
+    null_measurements = numpy.asarray([(s['unusual_packet'],s['other_packet'],s['unusual_tsval'],s['other_tsval']) for s in null])
+    
+    if params == None:
+        kf = KalmanFilter(n_dim_obs=4, n_dim_state=4,
+                          initial_state_mean=[quadsummary([s['unusual_packet'] for s in train]),
+                                              quadsummary([s['other_packet'] for s in train]),
+                                              numpy.mean([s['unusual_tsval'] for s in train]),
+                                              numpy.mean([s['other_tsval'] for s in train])])
+        kf = KalmanFilter(n_dim_obs=4, n_dim_state=4)
+        
+        start=time.time()
+        kf = kf.em(measurements[0:len(train)]+null_measurements[0:50000], n_iter=10,
+                   em_vars=('transition_matrices',
+                            'observation_matrices',
+                            'transition_offsets',
+                            'observation_offsets',
+                            'transition_covariance',
+                            'observation_covariance',
+                            'initial_state_mean',
+                            'initial_state_covariance'))
+        params = {'transition_matrices': kf.transition_matrices.tolist(),
+                  'observation_matrices': kf.observation_matrices.tolist(),
+                  'transition_offsets': kf.transition_offsets.tolist(),
+                  'observation_offsets': kf.observation_offsets.tolist(),
+                  'transition_covariance': kf.transition_covariance.tolist(),
+                  'observation_covariance': kf.observation_covariance.tolist(),
+                  'initial_state_mean': kf.initial_state_mean.tolist(),
+                  'initial_state_covariance': kf.initial_state_covariance.tolist()}
+        print("Learned Params:\n")
+        import pprint
+        pprint.pprint(params)
+        print("pykalman em time: %f" % (time.time()-start))
+        
+    #kf = KalmanFilter(n_dim_obs=2, n_dim_state=2, **params)
+
+    num_obs=5000
+    for offset in range(50000,100000+num_obs,num_obs):
+        start=time.time()
+        m = measurements[offset:offset+num_obs]
+        #params['initial_state_mean']=[quadsummary([s[0] for s in m]),
+        #                              quadsummary([s[1] for s in m]),
+        #                              numpy.mean([s[2] for s in m]),
+        #                              numpy.mean([s[3] for s in m])]
+        kf = KalmanFilter(n_dim_obs=4, n_dim_state=4, **params)
+        (smoothed_state_means, smoothed_state_covariances) = kf.smooth(m)
+        #print("pykalman smooth time: %f" % (time.time()-start))
+        up = numpy.mean([m[0] for m in smoothed_state_means])
+        op = numpy.mean([m[1] for m in smoothed_state_means])
+        #print("packet_rtt pykalman final:", smoothed_state_means[-1][0]-smoothed_state_means[-1][1])
+        print("packet_rtt pykalman mean:", up-op)
+        print("packet_rtt mean:", numpy.mean([s[0]-s[1] for s in m]))
+        #up = numpy.mean([m[2] for m in smoothed_state_means])
+        #op = numpy.mean([m[3] for m in smoothed_state_means])
+        #print("tsval_rtt pykalman final:", smoothed_state_means[-1][2]-smoothed_state_means[-1][3])
+        #print("tsval_rtt pykalman mean:", up-op)
+        #print("tsval_rtt mean:", numpy.mean([s[2]-s[3] for s in m]))
+
+    for offset in range(0,len(null_measurements)+num_obs,num_obs):
+        start=time.time()
+        m = null_measurements[offset:offset+num_obs]
+        #params['initial_state_mean']=[quadsummary([s[0] for s in m]),
+        #                              quadsummary([s[1] for s in m]),
+        #                              numpy.mean([s[2] for s in m]),
+        #                              numpy.mean([s[3] for s in m])]
+        kf = KalmanFilter(n_dim_obs=4, n_dim_state=4, **params)
+        (smoothed_state_means, smoothed_state_covariances) = kf.smooth(m)
+        up = numpy.mean([m[0] for m in smoothed_state_means])
+        op = numpy.mean([m[1] for m in smoothed_state_means])
+        #print("null packet_rtt pykalman final:", smoothed_state_means[-1][0]-smoothed_state_means[-1][1])
+        print("null packet_rtt pykalman mean:", up-op)
+        print("null packet_rtt mean:", numpy.mean([s[0]-s[1] for s in m]))
+        #up = numpy.mean([m[2] for m in smoothed_state_means])
+        #op = numpy.mean([m[3] for m in smoothed_state_means])
+        #print("null tsval_rtt pykalman final:", smoothed_state_means[-1][2]-smoothed_state_means[-1][3])
+        #print("null tsval_rtt pykalman mean:", up-op)
+        #print("null tsval_rtt mean:", numpy.mean([s[2]-s[3] for s in m]))
+
+        
+
+echo_vm_5k={'initial_state_covariance': [[33599047.5,
+                               -18251285.25,
+                               3242535690.59375,
+                               -8560730487.84375],
+                              [-18251285.25,
+                               9914252.3125,
+                               -1761372688.59375,
+                               4650260880.1875],
+                              [3242535690.59375,
+                               -1761372688.59375,
+                               312926663745.03125,
+                               -826168494791.7188],
+                              [-8560730487.84375,
+                               4650260880.1875,
+                               -826168494791.7188,
+                               2181195982530.4688]],
+ 'initial_state_mean': [12939012.5625,
+                        12934563.71875,
+                        13134751.608,
+                        13138990.9985],
+ 'observation_covariance': [[11960180434.411114,
+                             4760272534.795976,
+                             8797551081.431936,
+                             6908794128.927051],
+                            [4760272534.795962,
+                             12383598172.428213,
+                             5470747537.2599745,
+                             11252625555.297853],
+                            [8797551081.431955,
+                             5470747537.2601185,
+                             1466222848395.7058,
+                             72565713883.12643],
+                            [6908794128.927095,
+                             11252625555.297981,
+                             72565713883.12654,
+                             1519760903943.507]],
+ 'observation_matrices': [[1.4255288693095167,
+                           -0.4254638445329988,
+                           0.0003406844036817347,
+                           -0.0005475021956726778],
+                          [-0.46467270827589857,
+                           1.4654311778340343,
+                           -0.0003321330280128265,
+                           -0.0002853945703691352],
+                          [-0.2644570970067974,
+                           -0.33955835481495455,
+                           1.7494161615202275,
+                           -0.15394117603733548],
+                          [-0.3419097544041847,
+                           -0.23992883666045373,
+                           -0.15587790880447727,
+                           1.7292393175137022]],
+ 'observation_offsets': [165.2279084503762,
+                         157.76807691937614,
+                         168.4235495099334,
+                         225.33433430227353],
+ 'transition_covariance': [[2515479496.145993,
+                            -401423541.70620924,
+                            1409951418.1627903,
+                            255932902.74454522],
+                           [-401423541.706214,
+                            2744353887.676857,
+                            1162316.2019491254,
+                            1857251491.3987627],
+                           [1409951418.1628358,
+                            1162316.2020361447,
+                            543279068599.8229,
+                            -39399311190.5746],
+                           [255932902.74459982,
+                            1857251491.398838,
+                            -39399311190.574585,
+                            537826124257.5266]],
+ 'transition_matrices': [[0.52163952865412,
+                          0.47872618354122665,
+                          -0.0004322286766109684,
+                          0.00017293351811531466],
+                         [0.5167436693545113,
+                          0.48319044922845933,
+                          7.765428142114672e-05,
+                          -0.00021518950285326355],
+                         [0.2091705950622469,
+                          0.41051399729482796,
+                          0.19341113299389256,
+                          0.19562916616052917],
+                         [0.368592004009912,
+                          0.22263632461118732,
+                          0.20756792378812872,
+                          0.20977025833570906]],
+ 'transition_offsets': [592.5708159274,
+                        583.3804671015271,
+                        414.4187239098291,
+                        562.166786712371]}
+
+echo_vm_5k={'initial_state_covariance': [[0.375, 0.0, 0.0, 0.0],
+                              [0.0, 0.375, 0.0, 0.0],
+                              [0.0, 0.0, 0.375, 0.0],
+                              [0.0, 0.0, 0.0, 0.375]],
+ 'initial_state_mean': [15997944.198361743,
+                        16029825.435899183,
+                        17093077.26228404,
+                        17524263.088803563],
+ 'observation_covariance': [[36572556646.179054,
+                             21816054953.37006,
+                             31144379008.310543,
+                             19651005729.823025],
+                            [21816054953.372543,
+                             440428106325.20325,
+                             41103447776.740585,
+                             427146570672.51227],
+                            [31144379008.31037,
+                             41103447776.74027,
+                             3280009435458.6953,
+                             458734528073.65686],
+                            [19651005729.82234,
+                             427146570672.5109,
+                             458734528073.6557,
+                             3769493190697.773]],
+ 'observation_matrices': [[1.0248853427592337,
+                           -0.031198859962501047,
+                           0.001613706836380402,
+                           0.004720209443291878],
+                          [-0.8604422900368718,
+                           1.8583369609057172,
+                           -0.0022646214457040514,
+                           0.004437933935378169],
+                          [-0.5814771409524866,
+                           0.22228184387142846,
+                           1.6259599749174072,
+                           -0.271594798325566],
+                          [-0.5862601003257453,
+                           0.2598285939005791,
+                           -0.28286590143513024,
+                           1.604087079832425]],
+ 'observation_offsets': [1979.4518332096984,
+                         1889.3380163762793,
+                         2132.9112026744906,
+                         1750.7759421584785],
+ 'transition_covariance': [[6176492087.271547,
+                            762254719.4171592,
+                            4584288694.652873,
+                            3044796192.4357214],
+                           [762254719.4185101,
+                            173302376079.4761,
+                            5261303152.757347,
+                            167562483383.9925],
+                           [4584288694.651718,
+                            5261303152.755746,
+                            1056156956874.4131,
+                            -115859156952.07962],
+                           [3044796192.434162,
+                            167562483383.9901,
+                            -115859156952.08018,
+                            1225788436266.3086]],
+ 'transition_matrices': [[0.9673912485796876,
+                          0.03252962227543321,
+                          0.0006756067792537124,
+                          -0.0006566638567164773],
+                         [0.9548761966068113,
+                          0.03841774395880293,
+                          0.00426067282319309,
+                          0.002303362691861821],
+                         [0.6215040230859188,
+                          -0.2584476837756142,
+                          0.3176491193420503,
+                          0.3241682768126566],
+                         [0.6634028281470279,
+                          -0.33548335246018723,
+                          0.3298144902195048,
+                          0.3475836278392421]],
+ 'transition_offsets': [1751.3049487348183,
+                        1764.989515773476,
+                        1986.8405778425586,
+                        2232.830254345267]}
+#testKalman4D(echo_vm_5k)
+
+
+
+def testKalman(params=None):
+    from pykalman import AdditiveUnscentedKalmanFilter,KalmanFilter
+    train = db.subseries('train','long', offset=0)
+    test = db.subseries('test','long', offset=0)
+    measurements = numpy.asarray([(s['unusual_packet'],s['other_packet']) for s in (train+test)])
+
+    #kf = KalmanFilter(transition_matrices = [[1, 1], [0, 1]], observation_matrices = [[0.1, 0.5], [-0.3, 0.0]])
+    kf = KalmanFilter(n_dim_obs=2, n_dim_state=2,
+                      initial_state_mean=[quadsummary([s['unusual_packet'] for s in train]),
+                                          quadsummary([s['other_packet'] for s in train])])
+    #kf = AdditiveUnscentedKalmanFilter(n_dim_obs=2, n_dim_state=2)
+
+    if params == None:
+        start=time.time()
+        kf = kf.em(measurements[0:len(train)], n_iter=10,
+                   em_vars=('transition_matrices',
+                            'observation_matrices',
+                            'transition_offsets',
+                            'observation_offsets',
+                            'transition_covariance',
+                            'observation_covariance',
+                            'initial_state_covariance'))
+        params = {'transition_matrices': kf.transition_matrices.tolist(),
+                  'observation_matrices': kf.observation_matrices.tolist(),
+                  'transition_offsets': kf.transition_offsets.tolist(),
+                  'observation_offsets': kf.observation_offsets.tolist(),
+                  'transition_covariance': kf.transition_covariance.tolist(),
+                  'observation_covariance': kf.observation_covariance.tolist(),
+                  'initial_state_mean': kf.initial_state_mean.tolist(),
+                  'initial_state_covariance': kf.initial_state_covariance.tolist()}
+        print("Learned Params:\n")
+        import pprint
+        pprint.pprint(params)
+        print("pykalman em time: %f" % (time.time()-start))
+        
+    #kf = KalmanFilter(n_dim_obs=2, n_dim_state=2, **params)
+
+    num_obs=10000
+    for offset in range(50000,100000+num_obs,num_obs):
+        start=time.time()
+        kf = KalmanFilter(n_dim_obs=2, n_dim_state=2, **params)
+        m = measurements[offset:offset+num_obs]
+        (smoothed_state_means, smoothed_state_covariances) = kf.smooth(m)
+        print("pykalman smooth time: %f" % (time.time()-start))
+        up = numpy.mean([m[0] for m in smoothed_state_means])
+        op = numpy.mean([m[1] for m in smoothed_state_means])
+        print("packet_rtt pykalman final:", smoothed_state_means[-1][0]-smoothed_state_means[-1][1])
+        print("packet_rtt pykalman mean:", up-op)
+        print("packet_rtt mean:", numpy.mean([s[0]-s[1] for s in m]))
+
+
+five_iter = {'observation_offsets': [-54.53185823, -55.25219184],
+            'observation_covariance': [[  1.15059170e+10,   4.36743765e+09],
+                                       [  4.36743765e+09,   1.19410313e+10]],
+            'initial_state_mean': [ 12939012.5625 ,  12934563.71875],
+            'transition_covariance': [[  2.98594543e+09,   6.86355073e+07],
+                                      [  6.86355073e+07,   3.21368699e+09]],
+            'initial_state_covariance': [[  2.36836696e+09,   1.63195635e+09],
+                                         [  1.63195635e+09,   1.12452233e+09]],
+            'transition_offsets': [ 343.69740217,  338.5042467 ],
+            'observation_matrices': [[ 1.42539895, -0.4255261 ],
+                                     [-0.46280375,  1.46295189]],
+            'transition_matrices': [[ 0.56151623,  0.4385931 ],
+                                    [ 0.47309189,  0.52673508]]}
+ten_iter = {'initial_state_covariance': [[229936928.28125, 41172601.0],
+                                         [41172601.0, 7372383.46875]],
+            'initial_state_mean': [12939012.5625, 12934563.71875],
+            'observation_covariance': [[11958914107.88334, 4761048283.066559],
+                                       [4761048283.066557, 12388186543.42032]],
+            'observation_matrices': [[1.4258395826727792, -0.42598392357467674],
+                                     [-0.4647443890462455, 1.4648767294384015]],
+            'observation_offsets': [165.409715349344, 157.96206130876212],
+            'transition_covariance': [[2515594742.7187943, -401728959.41375697],
+                                      [-401728959.41375697, 2743831805.402682]],
+            'transition_matrices': [[0.521306461057975, 0.47879632652984583],
+                                    [0.5167881285851763, 0.483006520280469]],
+            'transition_offsets': [592.4419187566978, 583.2272403965366]}
+#testKalman(ten_iter)
+
+
 def getTCPTSPrecision():
     cursor = db.conn.cursor()
-    query="""SELECT tcpts_mean FROM meta;"""
+    query="""SELECT tcpts_mean FROM meta"""
     cursor.execute(query)
     row = cursor.fetchone()
@@ -170 +517 @@
     #plt.savefig('paper/graphs/dists-vs-dist-of-diffs2.svg')
 
-tsFilteredHistogram()
-sys.exit(0)
-
-
-from pykalman import KalmanFilter
-#kf = KalmanFilter(transition_matrices = [[1, 1], [0, 1]], observation_matrices = [[0.1, 0.5], [-0.3, 0.0]])
-kf = KalmanFilter(transition_matrices = [[1, 0], [0, 1]], n_dim_obs=2, observation_matrices = [[1.0, 0], [0, 1.0]])
-
-#delta = 1e-5
-#trans_cov = delta / (1 - delta) * np.eye(2)
-
-#kf = KalmanFilter(n_dim_obs=2, n_dim_state=2,
-#                  initial_state_mean=np.zeros(2),
-#                  initial_state_covariance=np.ones((2, 2)),
-#                  transition_matrices=np.eye(2),
-#                  observation_matrices=obs_mat,
-#                  observation_covariance=1.0,
-#                  transition_covariance=trans_cov)
-
-
-#measurements = numpy.asarray([[1,0], [0,0], [0,1]])  # 3 observations
-measurements = numpy.asarray([(s['unusual_packet'],s['other_packet']) for s in (db.subseries('train','long')+db.subseries('test','long'))])
-kf = kf.em(measurements, n_iter=5)
-#(filtered_state_means, filtered_state_covariances) = kf.filter(measurements)
-#print("packet_rtt pykalman:", filtered_state_means[-1][0]-filtered_state_means[-1][1])
-#print("packet_rtt pykalman:", filtered_state_means[-1])
-
-(smoothed_state_means, smoothed_state_covariances) = kf.smooth(measurements)
-#up = numpy.mean([m[0] for m in smoothed_state_means])
-#op = numpy.mean([m[1] for m in smoothed_state_means])
-print("packet_rtt pykalman:", smoothed_state_means[-1], smoothed_state_means[-1][0]-smoothed_state_means[-1][1])
-#print("packet_rtt pykalman:", up, op, up-op)
+#tsFilteredHistogram()
+
+
+
 
 
@@ -267 +586 @@
         classifiers.append(c[0])
 
+    max_obs = 0
     for classifier in classifiers:
         query="""
@@ -304 +624 @@
         performance = []
         for row in cursor:
+            max_obs = max(max_obs, row[0])
             num_obs.append(row[0])
             performance.append(row[1])
@@ -313 +634 @@
 
     plt.legend([l[1] for l in legend], [l[0] for l in legend], scatterpoints=1, fontsize='xx-small')
+    plt.plot([0, max_obs], [5.0, 5.0], "k--")
     plt.show()
         

trunk/bin/train

@@ -55 +55 @@
     trainer = classifiers[classifier]['train']
     threshold = 5.0 # in percent
-    num_obs = 1000
+    num_obs = 7
     max_obs = int(db.populationSize('train')/5)
     result = None
@@ -75 +75 @@
         classifiers[classifier]['train_results'].append(result)
 
-        if error < threshold:
+        if error < threshold and num_obs > 100:
             break
 

trunk/lib/nanownlib/__init__.py

@@ -176 +176 @@
     my_ip = getLocalIP(target_ip, target_port)
     my_iface = getIfaceForIP(my_ip)
-    return subprocess.Popen(['chrt', '-r', '99', './bin/csamp', my_iface, my_ip,
+    return subprocess.Popen(['chrt', '-r', '99', 'nanown-csamp', my_iface, my_ip,
                              target_ip, "%d" % target_port, output_file, '0'])
 
@@ -256 +256 @@
         suspect += 'R' # reordered received packets
     
-    packet_rtt = last_rcvd['observed'] - last_sent['observed']
-    if packet_rtt < 0:
-        sys.stderr.write("WARN: Negative packet_rtt. last_rcvd=%s,last_sent=%s\n" % (last_rcvd, last_sent))
-
     last_sent_ack = None
     try:
-        last_sent_ack = min(((p['observed'],p) for p in packets
-                             if p['sent']==0 and p['payload_len']+last_sent['tcpseq']==p['tcpack']))[1]
+        last_sent_ack = min(((p['tcpack'],p['observed'],p) for p in packets
+                             if p['sent']==0 and p['payload_len']+last_sent['tcpseq']>=p['tcpack']))[2]
         
     except Exception as e:
         sys.stderr.write("WARN: Could not find last_sent_ack.\n")
 
+    packet_rtt = last_rcvd['observed'] - last_sent['observed']
     tsval_rtt = None
     if None not in (timestamp_precision, last_sent_ack):
         tsval_rtt = int(round((last_rcvd['tsval'] - last_sent_ack['tsval'])*timestamp_precision))
 
+    if packet_rtt < 0 or (tsval_rtt != None and tsval_rtt < 0):
+        #sys.stderr.write("WARN: Negative packet or tsval RTT. last_rcvd=%s,last_sent=%s\n" % (last_rcvd, last_sent))
+        suspect += 'N'
+        
     return {'packet_rtt':packet_rtt,
             'tsval_rtt':tsval_rtt,
@@ -279 +280 @@
 
 
-# trimean and mad for each dist of differences
+# septasummary and mad for each dist of differences
 def evaluateTrim(db, unusual_case, strim, rtrim):
     cursor = db.conn.cursor()
@@ -292 +293 @@
       FROM (SELECT probes.sample s,packet_rtt FROM probes,trim_analysis WHERE sent_trimmed=:strim AND rcvd_trimmed=:rtrim AND trim_analysis.probe_id=probes.id AND probes.test_case=:unusual_case AND probes.type in ('train','test')) u
     """
-
+    #TODO: check for "N" in suspect field and return a flag
+    
     params = {"strim":strim,"rtrim":rtrim,"unusual_case":unusual_case}
     cursor.execute(query, params)
     differences = [row[0] for row in cursor]
     
-    return ubersummary(differences),mad(differences)
+    return septasummary(differences),mad(differences)
 
 
@@ -362 +364 @@
     for strim in range(0,num_sent):
         for rtrim in range(0,num_rcvd):
+            #print(strim,rtrim)
             if strim == 0 and rtrim == 0:
                 continue # no point in doing 0,0 again

trunk/lib/nanownlib/stats.py

@@ -166 +166 @@
     #return statistics.mean((l1,l2,l3,m,r3,r2,r1))
 
+    
+def septasummary(values, distance=25):
+    left2 = 50-distance
+    left3 = 50-(distance/2.0)
+    left1 = left2/2.0
+    right2 = 50+distance
+    right3 = 50+(distance/2.0)
+    right1 = (right2+100)/2.0
+    l1,l2,l3,m,r3,r2,r1 = numpy.percentile(values, (left1,left2,left3,50,right3,right2,right1))
+    return (l1+l2+l3+m+r3+r2+r1)/7.0
+
 
 def tsvalwmean(subseries):
@@ -254 +265 @@
 ubersummaryTest = functools.partial(summaryTest, ubersummary)
 quadsummaryTest = functools.partial(summaryTest, quadsummary)
+septasummaryTest = functools.partial(summaryTest, septasummary)
 
 def rmse(expected, measurements):
@@ -327 +339 @@
         else:
             return 0
+
+
+from pykalman import KalmanFilter
+def pyKalman4DTest(params, greater, samples):
+    kp = params['kparams']
+    #kp['initial_state_mean']=[quadsummary([s['unusual_packet'] for s in samples]),
+    #                          quadsummary([s['other_packet'] for s in samples]),
+    #                          numpy.mean([s['unusual_tsval'] for s in samples]),
+    #                          numpy.mean([s['other_tsval'] for s in samples])]
+    kf = KalmanFilter(n_dim_obs=4, n_dim_state=4, **kp)
+    smooth,covariance = kf.smooth([(s['unusual_packet'],s['other_packet'],s['unusual_tsval'],s['other_tsval'])
+                                   for s in samples])
+    m = numpy.mean(smooth)
+    if greater:
+        if m > params['threshold']:
+            return 1
+        else:
+            return 0
+    else:
+        if m < params['threshold']:
+            return 1
+        else:
+            return 0
+    

trunk/lib/nanownlib/train.py

@@ -300 +300 @@
 
 
+from pykalman import KalmanFilter
+_pykalman4d_params = None
+_pykalman4d_params = {'observation_covariance': [[11960180434.411114, 4760272534.795976, 8797551081.431936, 6908794128.927051], [4760272534.795962, 12383598172.428213, 5470747537.2599745, 11252625555.297853], [8797551081.431955, 5470747537.2601185, 1466222848395.7058, 72565713883.12643], [6908794128.927095, 11252625555.297981, 72565713883.12654, 1519760903943.507]], 'transition_offsets': [592.5708159274, 583.3804671015271, 414.4187239098291, 562.166786712371], 'observation_offsets': [165.2279084503762, 157.76807691937614, 168.4235495099334, 225.33433430227353], 'initial_state_covariance': [[33599047.5, -18251285.25, 3242535690.59375, -8560730487.84375], [-18251285.25, 9914252.3125, -1761372688.59375, 4650260880.1875], [3242535690.59375, -1761372688.59375, 312926663745.03125, -826168494791.7188], [-8560730487.84375, 4650260880.1875, -826168494791.7188, 2181195982530.4688]], 'initial_state_mean': [12939012.5625, 12934563.71875, 13134751.608, 13138990.9985], 'transition_covariance': [[2515479496.145993, -401423541.70620924, 1409951418.1627903, 255932902.74454522], [-401423541.706214, 2744353887.676857, 1162316.2019491254, 1857251491.3987627], [1409951418.1628358, 1162316.2020361447, 543279068599.8229, -39399311190.5746], [255932902.74459982, 1857251491.398838, -39399311190.574585, 537826124257.5266]], 'observation_matrices': [[1.4255288693095167, -0.4254638445329988, 0.0003406844036817347, -0.0005475021956726778], [-0.46467270827589857, 1.4654311778340343, -0.0003321330280128265, -0.0002853945703691352], [-0.2644570970067974, -0.33955835481495455, 1.7494161615202275, -0.15394117603733548], [-0.3419097544041847, -0.23992883666045373, -0.15587790880447727, 1.7292393175137022]], 'transition_matrices': [[0.52163952865412, 0.47872618354122665, -0.0004322286766109684, 0.00017293351811531466], [0.5167436693545113, 0.48319044922845933, 7.765428142114672e-05, -0.00021518950285326355], [0.2091705950622469, 0.41051399729482796, 0.19341113299389256, 0.19562916616052917], [0.368592004009912, 0.22263632461118732, 0.20756792378812872, 0.20977025833570906]]}
+_pykalman4d_good_threshold = 2009.25853272
+_pykalman4d_params = None
+
+_pykalman4d_params = {'observation_covariance': [[32932883342.63772, 18054300398.442295, 27538911550.824535, 17152378956.778696], [18054300398.446983, 436546443436.5115, 37327644533.69647, 424485386677.31274], [27538911550.838238, 37327644533.706024, 3276324705772.982, 456017515263.88715], [17152378956.788027, 424485386677.317, 456017515263.88245, 3767844180658.1724]], 'observation_matrices': [[1.025773112769464, -0.028755990114063934, 0.0003540921897382532, 0.0025748564713126143], [-0.8595457826320256, 1.8607522167556567, -0.003520779053701517, 0.002309145982167138], [-0.5806427858959466, 0.22466075141448982, 1.6247192012813798, -0.27363797512617793], [-0.5853369461874607, 0.262177909212312, -0.28415108658843735, 1.6020343138710018]], 'initial_state_mean': [0.0, 0.0, 0.0, 0.0], 'observation_offsets': [549.4498515668686, 484.2106453284049, 648.556719142234, 380.10978090584763], 'transition_covariance': [[4147844406.7768326, -1308763245.5992138, 2920744388.523955, 860096280.797968], [-1308763245.5998695, 171190325905.83395, 3557618712.218984, 165332873663.83142], [2920744388.532502, 3557618712.2283373, 1054894349089.0673, -117551209299.73402], [860096280.805706, 165332873663.83963, -117551209299.73474, 1223605046475.7324]], 'transition_offsets': [1156.9264087977374, 1150.752680207601, 1312.2595286459816, 1267.4069537452415], 'initial_state_covariance': [[667999273207241.0, 669330484615232.1, 713726904326576.2, 731731206363217.4], [669330484615390.9, 670664348906228.8, 715149243295271.9, 733189424910272.2], [713726904326843.4, 715149243295370.6, 762584802695960.9, 781821582244358.5], [731731206363417.0, 733189424910299.0, 781821582244278.6, 801543624134758.0]], 'transition_matrices': [[0.9680677036616316, 0.03260717171917804, 0.0005279411071512641, -0.0012363486571871363], [0.9555219601128613, 0.03851351491891819, 0.00411268796118236, 0.0017357967358293536], [0.622254432930994, -0.2583795512595657, 0.31745705251401546, 0.32357126976364725], [0.6644076824932768, -0.33545285094373867, 0.3295778964272671, 0.34682391469482354]]}
+_pykalman4d_good_threshold = -253.849393803
+def trainPyKalman4D(db, unusual_case, greater, num_observations):
+    global _pykalman4d_params
+    global _pykalman4d_good_threshold
+    db.resetOffsets()
+
+    if _pykalman4d_params == None:
+        train = db.subseries('train',unusual_case, offset=0)
+        null = db.subseries('train_null',unusual_case, offset=0)
+        train_array = numpy.asarray([(s['unusual_packet'],s['other_packet'],s['unusual_tsval'],s['other_tsval'])
+                                     for s in train])
+        null_array = numpy.asarray([(s['unusual_packet'],s['other_packet'],s['unusual_tsval'],s['other_tsval'])
+                                    for s in null])
+        kf = KalmanFilter(n_dim_obs=4, n_dim_state=4)
+        #initial_state_mean=[quadsummary([s['unusual_packet'] for s in train]),
+        #                                      quadsummary([s['other_packet'] for s in train]),
+        #                                      numpy.mean([s['unusual_tsval'] for s in train]),
+        #                                      numpy.mean([s['other_tsval'] for s in train])])
+
+        kf = kf.em(train_array+null_array[0:50000], n_iter=10,
+                   em_vars=('transition_matrices',
+                            'observation_matrices',
+                            'transition_offsets',
+                            'observation_offsets',
+                            'transition_covariance',
+                            'observation_covariance',
+                            'initial_state_covariance'))
+        _pykalman4d_params = {'transition_matrices': kf.transition_matrices.tolist(),
+                              'observation_matrices': kf.observation_matrices.tolist(),
+                              'transition_offsets': kf.transition_offsets.tolist(),
+                              'observation_offsets': kf.observation_offsets.tolist(),
+                              'transition_covariance': kf.transition_covariance.tolist(),
+                              'observation_covariance': kf.observation_covariance.tolist(),
+                              'initial_state_mean': kf.initial_state_mean.tolist(),
+                              'initial_state_covariance': kf.initial_state_covariance.tolist()}
+        print(_pykalman4d_params)
+    
+        kf = KalmanFilter(n_dim_obs=4, n_dim_state=4, **_pykalman4d_params)
+        smoothed,covariance = kf.smooth(train_array)
+        null_smoothed,covariance = kf.smooth(null_array)
+
+        kp = _pykalman4d_params.copy()
+        #kp['initial_state_mean']=[quadsummary([s['unusual_packet'] for s in train]),
+        #                          quadsummary([s['other_packet'] for s in train]),
+        #                          numpy.mean([s['unusual_tsval'] for s in train]),
+        #                          numpy.mean([s['other_tsval'] for s in train])]
+        #kf = KalmanFilter(n_dim_obs=4, n_dim_state=4, **kp)
+        #null_smoothed,covariance = kf.smooth(null_array)
+        
+        _pykalman4d_good_threshold = (numpy.mean([m[0]-m[1] for m in smoothed])+numpy.mean([m[0]-m[1] for m in null_smoothed]))/2.0
+        print(_pykalman4d_good_threshold)
+
+    
+    def trainAux(params, num_trials):
+        estimator = functools.partial(pyKalman4DTest, params, greater)
+        estimates = bootstrap3(estimator, db, 'train', unusual_case, num_observations, num_trials)
+        null_estimates = bootstrap3(estimator, db, 'train_null', unusual_case, num_observations, num_trials)
+        
+        bad_estimates = len([e for e in estimates if e != 1])
+        bad_null_estimates = len([e for e in null_estimates if e != 0])
+        
+        false_negatives = 100.0*bad_estimates/num_trials
+        false_positives = 100.0*bad_null_estimates/num_trials
+        return false_positives,false_negatives
+
+    params = {'threshold':_pykalman4d_good_threshold, 'kparams':_pykalman4d_params}
+
+    wt = WorkerThreads(2, trainAux)
+    num_trials = 50
+    performance = []
+    for t in range(-80,100,20):
+        thresh = _pykalman4d_good_threshold + abs(_pykalman4d_good_threshold)*(t/100.0)
+        params['threshold'] = thresh
+        wt.addJob(thresh, (params.copy(),num_trials))
+    wt.wait()
+    while not wt.resultq.empty():
+        job_id,errors = wt.resultq.get()
+        fp,fn = errors
+        #performance.append(((fp+fn)/2.0, job_id, fn, fp))
+        performance.append((abs(fp-fn), job_id, fn, fp))
+    performance.sort()
+    #pprint.pprint(performance)
+    best_threshold = performance[0][1]
+    #print("best_threshold:", best_threshold)
+    params['threshold']=best_threshold
+
+    wt.stop()
+    
+    return {'trial_type':"train",
+            'num_observations':num_observations,
+            'num_trials':num_trials,
+            'params':json.dumps(params, sort_keys=True),
+            'false_positives':performance[0][3],
+            'false_negatives':performance[0][2]}
+
+
+
 classifiers = {'boxtest':{'train':trainBoxTest, 'test':multiBoxTest, 'train_results':[]},
                'midsummary':{'train':functools.partial(trainSummary, midsummary), 'test':midsummaryTest, 'train_results':[]},
-               'ubersummary':{'train':functools.partial(trainSummary, ubersummary), 'test':ubersummaryTest, 'train_results':[]},
+               #'ubersummary':{'train':functools.partial(trainSummary, ubersummary), 'test':ubersummaryTest, 'train_results':[]},
                'quadsummary':{'train':functools.partial(trainSummary, quadsummary), 'test':quadsummaryTest, 'train_results':[]},
-               'tsvalwmean':{'train':trainTsval, 'test':tsvalwmeanTest, 'train_results':[]},
+               'septasummary':{'train':functools.partial(trainSummary, septasummary), 'test':septasummaryTest, 'train_results':[]},
+               #'pykalman4d':{'train':trainPyKalman4D, 'test':pyKalman4DTest, 'train_results':[]},
+               #'tsvalwmean':{'train':trainTsval, 'test':tsvalwmeanTest, 'train_results':[]},
                #'kalman':{'train':trainKalman, 'test':kalmanTest, 'train_results':[]},
                #'_trimean':{'train':None, 'test':trimeanTest, 'train_results':[]},