Followed contributions guidelines.

keras_app/custom_layers/capsule_layer.py

@@ -3,58 +3,59 @@
 import keras
 
 
-
-def squash_activation(vectors,axis=-1):
+def squash_activation(vectors, axis=-1):
     s_squared_norm = K.sum(K.square(vectors), axis, keepdims=True)
-    scale = s_squared_norm / (1 + s_squared_norm) / K.sqrt(s_squared_norm +1e-7)
+    scale = s_squared_norm / (1 + s_squared_norm) / K.sqrt(s_squared_norm + 1e-7)
     return scale * vectors
 
+
 class CapsuleLayer(Layer):
-    def __init__(self,num_capsule,dim_capsule,num_routing=3,**kwargs):
-        super(CapsuleLayer,self).__init__(**kwargs)
-        self.num_capsule=num_capsule
-        self.dim_capsule=dim_capsule
-        self.num_routing=num_routing
-        self.kernel_initializer=keras.initializers.random_uniform(-1, 1)
-        self.bias_initializer=keras.initializers.Zeros()
-        super(CapsuleLayer,self).__init__(**kwargs)
-
-    def build(self,input_shape):
-        assert len(input_shape)>=3
-        self.W=self.add_weight(shape=[input_shape[1],self.num_capsule,input_shape[2],self.dim_capsule],
-                              initializer=self.kernel_initializer,name='W')
-        self.b=self.add_weight(shape=[input_shape[1],self.num_capsule],initializer=self.bias_initializer,name='b')
-
-        super(CapsuleLayer,self).build(input_shape)
-
-    def call(self,inputs,training=None):
-        inputs_expand=K.expand_dims(inputs,2)
-        inputs_tiled=K.repeat_elements(inputs_expand,self.num_capsule,axis=2)
-        inputs_hat=K.map_fn(lambda x:K.batch_dot(x,self.W,[2,2]),inputs_tiled)
-
-        input_shape=K.shape(inputs_hat)
-        b=self.b
-        b=K.expand_dims(b,axis=0)
-        assert self.num_routing>0
+    def __init__(self, num_capsule, dim_capsule, num_routing=3, **kwargs):
+        super(CapsuleLayer, self).__init__(**kwargs)
+        self.num_capsule = num_capsule
+        self.dim_capsule = dim_capsule
+        self.num_routing = num_routing
+        self.kernel_initializer = keras.initializers.random_uniform(-1, 1)
+        self.bias_initializer = keras.initializers.Zeros()
+        super(CapsuleLayer, self).__init__(**kwargs)
+
+    def build(self, input_shape):
+        assert len(input_shape) >= 3
+        self.W = self.add_weight(shape=[input_shape[1], self.num_capsule, input_shape[2], self.dim_capsule], 
+                                 initializer=self.kernel_initializer,
+                                 name='W')
+        self.b = self.add_weight(shape=[input_shape[1], self.num_capsule], 
+                                 initializer=self.bias_initializer,
+                                 name ='b')
+        super(CapsuleLayer, self).build(input_shape)
+
+    def call(self, inputs, training=None):
+        inputs_expand = K.expand_dims(inputs, 2)
+        inputs_tiled = K.repeat_elements(inputs_expand, self.num_capsule, axis=2)
+        inputs_hat = K.map_fn(lambda x: K.batch_dot(x, self.W, [2, 2]), inputs_tiled)
+        input_shape = K.shape(inputs_hat)
+        b = self.b
+        b = K.expand_dims(b, axis=0)
+        assert self.num_routing > 0
         for i in range(self.num_routing):
-            c=K.softmax(b)
-            c=K.expand_dims(c,axis=-1)
-            c=K.repeat_elements(c,rep=self.dim_capsule,axis=-1)
-            S=K.sum(c*inputs_hat,axis=1)
-            V=squash_activation(S)
-            if i!=self.num_routing-1:
-                V_expanded=K.expand_dims(V,axis=1)
-                V_expanded=K.tile(V_expanded,[1,input_shape[1],1,1])
-                b=b+K.sum(inputs_hat*V_expanded,axis=-1)
-                
-        return V    
-    def compute_output_shape(self,input_shape):
-        return tuple([None,self.num_capsule,self.dim_capsule])
-    
+            c = K.softmax(b)
+            c = K.expand_dims(c, axis=-1)
+            c = K.repeat_elements(c, rep=self.dim_capsule, axis=-1)
+            S = K.sum(c * inputs_hat, axis=1)
+            V = squash_activation(S)
+            if i != self.num_routing-1:
+                V_expanded = K.expand_dims(V, axis=1)
+                V_expanded = K.tile(V_expanded, [1, input_shape[1], 1, 1])
+                b = b + K.sum(inputs_hat * V_expanded, axis=-1)
+        return V
+
+    def compute_output_shape(self, input_shape):
+        return tuple([None, self.num_capsule, self.dim_capsule])
+
     def get_config(self):
-        base_config=super(CapsuleLayer,self).get_config()
-        base_config['num_capsule']=self.num_capsule
-        base_config['num_routing']=self.num_routing
-        base_config['dim_capsule']=self.dim_capsule
+        base_config = super(CapsuleLayer, self).get_config()
+        base_config['num_capsule'] = self.num_capsule
+        base_config['num_routing'] = self.num_routing
+        base_config['dim_capsule'] = self.dim_capsule
         return base_config
-
+        

keras_app/custom_layers/config.py

@@ -6,19 +6,19 @@
         'url': '/media/lrn.py'
     },
     'CapsuleLayer': {
-    	'filename': 'capsule_layer.py',
-    	'url': '/media/capsule_layer.py'
+        'filename': 'capsule_layer.py',
+        'url': '/media/capsule_layer.py'
     },
     'Length': {
-    	'filename': 'length.py',
-    	'url': 'media/length.py'
+        'filename': 'length.py',
+        'url': 'media/length.py'
     },
     'MaskCapsule': {
-    	'filename': 'mask_capsule.py',
-    	'url':'media/mask_capsule.py'
+        'filename': 'mask_capsule.py',
+        'url': 'media/mask_capsule.py'
     },
     'Squash': {
-    	'filename': 'squash.py',
-    	'url':'media/squash.py'
+        'filename': 'squash.py',
+        'url': 'media/squash.py'
     }
-}
+}

keras_app/custom_layers/length.py

@@ -2,20 +2,20 @@
 from keras import backend as K
 
 
-
 class Length(Layer):
-    def __init__(self,**kwargs):
-        super(Length,self).__init__(**kwargs)
-    def build(self,input_shape):
-        super(Length,self).build(input_shape)
-
-    def call(self,input):
-        return K.sqrt(K.sum(K.square(input),-1))
-
-    def compute_output_shape(self,input_shape):
+    def __init__(self, **kwargs):
+        super(Length, self).__init__(**kwargs)
+
+    def build(self, input_shape):
+        super(Length, self).build(input_shape)
+
+    def call(self, input):
+        return K.sqrt(K.sum(K.square(input), -1))
+
+    def compute_output_shape(self, input_shape):
         return input_shape[:-1]
-    
+
     def get_config(self):
-        base_config=super(Length,self).get_config()
+        base_config = super(Length, self).get_config()
         return base_config
 

keras_app/custom_layers/mask_capsule.py

@@ -3,28 +3,28 @@
 
 
 class MaskCapsule(Layer):
-    def __init__(self,**kwargs):
-        super(MaskCapsule,self).__init__(**kwargs)
-        
-    def build(self,input_shape):
-        super(MaskCapsule,self).build(input_shape)
-    
-    def call(self,inputs):
-        if type(inputs)==list:
-            assert len(inputs)==2
-            inputs,mask=inputs
+    def __init__(self, **kwargs):
+        super(MaskCapsule, self).__init__(**kwargs)
+
+    def build(self, input_shape):
+        super(MaskCapsule, self).build(input_shape)
+
+    def call(self, inputs):
+        if type(inputs) == list:
+            assert len(inputs) == 2
+            inputs, mask = inputs
         else:
-            x=K.sqrt(K.sum(K.square(inputs),-1))
-            mask=K.one_hot(indices=K.argmax(x,1),num_classes=x.get_shape().as_list()[1])
-        masked=K.batch_flatten(inputs*K.expand_dims(mask,-1))
+            x = K.sqrt(K.sum(K.square(inputs), -1))
+            mask = K.one_hot(indices=K.argmax(x, 1), num_classes=x.get_shape().as_list()[1])
+        masked = K.batch_flatten(inputs*K.expand_dims(mask, -1))
         return masked
-    
-    def compute_output_shape(self,input_shape):
+
+    def compute_output_shape(self, input_shape):
         if type(input_shape[0]) is tuple:
-            return tuple([None,input_shape[0][1]*input_shape[0][2]])
+            return tuple([None, input_shape[0][1]*input_shape[0][2]])
         else:
-            return tuple([None ,input_shape[1]*input_shape[2]])
-    
+            return tuple([None, input_shape[1]*input_shape[2]])
+
     def get_config(self):
-        return super(MaskCapsule,self).get_config()
-            
+        return super(MaskCapsule, self).get_config()
+

keras_app/custom_layers/squash.py

@@ -1,19 +1,25 @@
 from keras.layers.core import Layer
 from keras import backend as K
 
+
 class Squash(Layer):
-    def __init__(self,axis=-1,**kwargs):
-        self.axis=axis
-        super(Squash,self).__init__(**kwargs)
-    def build(self,input_shape):
-        super(Squash,self).build(input_shape)
-    def call(self,inputs):
+    def __init__(self, axis=-1, **kwargs):
+        self.axis = axis
+        super(Squash, self).__init__(**kwargs)
+
+    def build(self, input_shape):
+        super(Squash, self).build(input_shape)
+
+    def call(self, inputs):
         s_squared_norm = K.sum(K.square(inputs), self.axis, keepdims=True)
-        scale = s_squared_norm / (1 + s_squared_norm) / K.sqrt(s_squared_norm +1e-7)
+        scale = s_squared_norm / (1 + s_squared_norm) / K.sqrt(s_squared_norm + 1e-7)
         return scale * inputs
-    def compute_output_shape(self,input_shape):
+
+    def compute_output_shape(self, input_shape):
         return input_shape
+
     def get_config(self):
-        base_config=super(Squash,self).get_config()
-        base_config['axis']=self.axis
-        return base_config
+        base_config = super(Squash, self).get_config()
+        base_config['axis'] = self.axis
+        return base_config
+

keras_app/views/export_json.py

@@ -10,7 +10,7 @@
 from layers_export import data, convolution, deconvolution, pooling, dense, dropout, embed,\
     recurrent, batch_norm, activation, flatten, reshape, eltwise, concat, upsample, locally_connected,\
     permute, repeat_vector, regularization, masking, gaussian_noise, gaussian_dropout, alpha_dropout, \
-    bidirectional, time_distributed, lrn, depthwiseConv,capsule_layer,length,mask_capsule,squash
+    bidirectional, time_distributed, lrn, depthwiseConv, capsule_layer, length, mask_capsule, squash
 from ..custom_layers import config as custom_layers_config
 
 
@@ -82,10 +82,10 @@ def export_json(request, is_tf=False):
 
         custom_layers_map = {
             'LRN': lrn,
-            'CapsuleLayer':capsule_layer,
-            'Length':length,
-            'MaskCapsule':mask_capsule,
-            'Squash':squash
+            'CapsuleLayer': capsule_layer,
+            'Length': length,
+            'MaskCapsule': mask_capsule,
+            'Squash': squash
         }
 
         # Remove any duplicate activation layers (timedistributed and bidirectional layers)

keras_app/views/import_json.py

@@ -10,11 +10,11 @@
     Recurrent, BatchNorm, Activation, LeakyReLU, PReLU, ELU, Scale, Flatten, Reshape, Concat, \
     Eltwise, Padding, Upsample, LocallyConnected, ThresholdedReLU, Permute, RepeatVector,\
     ActivityRegularization, Masking, GaussianNoise, GaussianDropout, AlphaDropout, \
-    TimeDistributed, Bidirectional, DepthwiseConv, lrn ,capsule_layer,length,mask_capsule,squash
+    TimeDistributed, Bidirectional, DepthwiseConv, lrn, capsule_layer, length, mask_capsule, squash
 from keras.models import model_from_json, Sequential
 from keras.layers import deserialize
 from ..custom_layers.lrn import LRN
-from ..custom_layers.capsule_layer import *
+from ..custom_layers.capsule_layer import CapsuleLayer
 from ..custom_layers.length import Length
 from ..custom_layers.mask_capsule import MaskCapsule
 from ..custom_layers.squash import Squash
@@ -53,7 +53,9 @@ def import_json(request):
         except Exception:
             return JsonResponse({'result': 'error', 'error': 'Invalid JSON'})
 
-    model = model_from_json(json.dumps(model), custom_objects={'LRN': LRN,'Length':Length,'MaskCapsule':MaskCapsule,'Squash':Squash,'CapsuleLayer':CapsuleLayer})
+    model = model_from_json(json.dumps(model),
+                             custom_objects={'LRN': LRN, 'Length': Length, 'MaskCapsule': MaskCapsule, 
+                                             'Squash': Squash, 'CapsuleLayer': CapsuleLayer})
     layer_map = {
         'InputLayer': Input,
         'Dense': Dense,
@@ -119,9 +121,9 @@ def import_json(request):
         'Bidirectional': Bidirectional,
         'LRN': lrn,
         'CapsuleLayer': capsule_layer,
-        'Length':length,
-        'MaskCapsule':mask_capsule,
-        'Squash':squash
+        'Length': length,
+        'MaskCapsule': mask_capsule,
+        'Squash': squash
     }
 
     hasActivation = ['Conv1D', 'Conv2D', 'Conv3D', 'Conv2DTranspose', 'Dense', 'LocallyConnected1D',

keras_app/views/layers_export.py

@@ -18,7 +18,7 @@
 from keras.layers import TimeDistributed, Bidirectional
 from keras import regularizers
 from ..custom_layers.lrn import LRN
-from ..custom_layers.capsule_layer import *
+from ..custom_layers.capsule_layer import CapsuleLayer
 from ..custom_layers.length import Length
 from ..custom_layers.mask_capsule import MaskCapsule
 from ..custom_layers.squash import Squash
@@ -132,9 +132,9 @@ def activation(layer, layer_in, layerId, tensor=True):
     return out
 
 
-def squash(layer,layer_in,layerId):
-    axis=layer['params']['axis']
-    out={}
+def squash(layer, layer_in, layerId):
+    axis = layer['params']['axis']
+    out = {}
     out[layerId] = Squash(axis=axis)(*layer_in)
 
 
@@ -189,15 +189,19 @@ def masking(layer, layer_in, layerId, tensor=True):
         out[layerId] = out[layerId](*layer_in)
     return out
 
-def length(layer,layer_in,layerId):
-    out={}
-    out[layerId]=Length()(*layer_in)
+
+def length(layer, layer_in, layerId):
+    out = {}
+    out[layerId] = Length()(*layer_in)
     return out
 
-def mask_capsule(layer,layer_in,layerId):
-    out={}
-    out[layerId]=MaskCapsule()(*layer_in)
+
+def mask_capsule(layer, layer_in, layerId):
+    out = {}
+    out[layerId] = MaskCapsule()(*layer_in)
     return out
+
+
 # ********** Convolution Layers **********
 def convolution(layer, layer_in, layerId, tensor=True):
     convMap = {
@@ -383,14 +387,17 @@ def upsample(layer, layer_in, layerId, tensor=True):
         out[layerId] = out[layerId](*layer_in)
     return out
 
+
 # ********** Capsule Layers **********
-def capsule_layer(layer,layer_in,layerId):
-    num_capsule=layer['params']['num_capsule']
-    dim_capsule=layer['params']['dim_capsule']
-    num_routing=layer['params']['num_routing']
+def capsule_layer(layer, layer_in, layerId):
+    num_capsule = layer['params']['num_capsule']
+    dim_capsule = layer['params']['dim_capsule']
+    num_routing = layer['params']['num_routing']
     out = {}
-    out[layerId]=CapsuleLayer(num_capsule=num_capsule,dim_capsule=dim_capsule,num_routing=num_routing)(*layer_in)
+    out[layerId] = CapsuleLayer(num_capsule=num_capsule, dim_capsule=dim_capsule, num_routing=num_routing)(*layer_in)
     return out
+
+
 # ********** Pooling Layers **********
 def pooling(layer, layer_in, layerId, tensor=True):
     poolMap = {