syn

dataset.py

@@ -48,7 +48,7 @@ def dsphere(n=100, d=2, r=1, noise=None, ambient=None):
 
     return data
 
-def create_sphere_dataset10000(n_samples=500, d=100, n_spheres=11, r=5, seed=42):
+def create_sphere_dataset10000(n_samples=1000, d=100, n_spheres=11, r=5, seed=42):
     np.random.seed(seed)
 
     # it seemed that rescaling the shift variance by sqrt of d lets big sphere stay around the inner spheres
@@ -78,9 +78,9 @@ def create_sphere_dataset10000(n_samples=500, d=100, n_spheres=11, r=5, seed=42)
         labels[label_index:label_index + n_sphere_samples] = index
         label_index += n_sphere_samples
 
-    index_seed = np.linspace(0, 10000, num=20, dtype='int16', endpoint=False)
+    index_seed = np.linspace(0, 20000, num=20, dtype='int16', endpoint=False)
     arr = np.array([], dtype='int16')
-    for i in range(500):
+    for i in range(n_samples):
         arr = np.concatenate((arr, index_seed+int(i)))
     dataset = dataset[arr]
     labels = labels[arr]

main.py

@@ -246,7 +246,7 @@ def SetParam():
     parser.add_argument("-ND", "--N_Dataset", default=800, type=int)   # The data number used for training
     parser.add_argument("-GC", "--GradualChanging", default=[500, 1000], type=int, nargs='+')   # Range for the gradual changing of push-away Loss
     parser.add_argument("-R", "--ratio", default=[0.2, 1.0, 0.0, 1.0], type=float, nargs='+')   # The weight ratio for loss_ae/loss_iso/loss_angle/loss_push-away
-    parser.add_argument("-EPS", "--Epsilon", default=0.23, type=float)   # The boundary parameters used to determine the neighborhood
+    parser.add_argument("-EPS", "--Epsilon", default=0.30, type=float)   # The boundary parameters used to determine the neighborhood
     parser.add_argument("-MK", "--MAEK", default=15, type=int)
     parser.add_argument("-E", "--EPOCHS", default=10000, type=int)
     parser.add_argument("-P", "--PlotForloop", default=1000, type=int)   # Save data and plot every 1000 epochs
@@ -265,8 +265,8 @@ def SetParam():
         args.ParamPath = './param/spheres5500_enc.json'
     if args.DATASET == 'Spheres5500' and args.Mode == 'ML-AE':
         args.ParamPath = './param/spheres5500_ae.json'
-    if args.DATASET == 'Spheres10000':
-        args.ParamPath = './param/spheres10000_ae.json'
+    if args.DATASET == 'Spheres10000' and args.Mode == 'ML-Enc':
+        args.ParamPath = './param/spheres10000_enc.json'
     if args.ParamPath is not 'None':
         jsontxt = open(args.ParamPath, 'r').read()
         param = json.loads(jsontxt)
@@ -282,6 +282,7 @@ def SetParam():
     else:
         path = "./pic/{}".format(param['name'])
 
+    path = path+str(time.time())[:20]
     # Save parameters
     if not os.path.exists(path):
         os.makedirs(path)
@@ -353,8 +354,8 @@ def Train_MultiRun():
             train_data = train_data[:-5500]
             train_label = train_label[:-5500]
         if param['DATASET'] == 'Spheres10000':
-            test_data = train_data[9000:]
-            test_label = train_label[9000:]
+            test_data = train_data[18500:]
+            test_label = train_label[18500:]
             train_data = train_data[:7500]
             train_label = train_label[:7500]
 

param/spheres10000_enc.json

@@ -0,0 +1,35 @@
+{
+  "name": "MLDL_Spheres5500_N11000_Enc",
+  "ParamPath": "None",
+  "Mode": "ML-Enc",
+  "DATASET": "Spheres10000",
+  "LEARNINGRATE": 0.001,
+  "BATCHSIZE": 10000,
+  "RegularB": 3,
+  "N_Dataset": 20000,
+  "GradualChanging": [
+    200,
+    400
+  ],
+  "ratio": [
+    0.0,
+    1.0,
+    0.0,
+    0.0
+  ],
+  "Epsilon": 2,
+  "MAEK": 15,
+  "EPOCHS": 10000,
+  "PlotForloop": 1000,
+  "SEED": 42,
+  "NetworkStructure": [
+    101,
+    90,
+    80,
+    70,
+    2
+  ],
+  "Noise": 0.0,
+  "Train_MultiRun": false,
+  "Visualization": false
+}

utils.py

@@ -57,7 +57,7 @@ def __init__(self, args, model):
 
 
     # Plotting the results of a single layer in the network based on the input data
-    def PlotOtherLayer(self, fig, data, label, title='', fig_position0=1, fig_position1=1, fig_position2=1, s=10, dataset='None'):
+    def PlotOtherLayer(self, fig, data, label, title='', fig_position0=1, fig_position1=1, fig_position2=1, s=1, dataset='None'):
 
         color_list = []
         for i in range(label.shape[0]):