add a first solution for fine-tuning models #2

dcase_models/model/container.py

@@ -4,8 +4,10 @@
 
 from keras.optimizers import Adam
 from keras.callbacks import ModelCheckpoint,CSVLogger
-from keras.models import model_from_json
+from keras.models import model_from_json, Model
 import keras.backend as K
+from keras.layers import Dense, Input
+
 
 from ..utils.files import save_json
 from ..utils.metrics import evaluate_metrics
@@ -214,3 +216,52 @@ def get_numer_of_parameters(self):
         trainable_count = int(np.sum([K.count_params(p) for p in set(models.trainable_weights)]))
         return trainable_count
 
+    def fine_tuning(self, layer_where_to_cut, new_number_of_classes=10, new_activation='softmax', 
+                    freeze_source_model=True, new_model=None):
+        """
+        Create a new model for fine-tuning. Cut the model in the layer_where_to_cut layer 
+        and add a new fully-connected layer.
+
+        Parameters
+        ----------
+        layer_where_to_cut : str or int
+            Name (str) of index (int) of the layer where cut the model. This layer
+            is included in the new model.
+
+        new_number_of_classes : int
+            Number of units in the new fully-connected layer (number of classes)
+
+        new_activation : str
+            Activitation of the new fully-connected layer
+
+        freeze_source_model : bool
+            If True, the source model is set to not be trainable
+
+        new_model : Keras Model
+            If is not None, this model is add after the cut model. This is useful if you
+            want add more than a fully-connected layer. 
+        """
+        # cut last layer
+        if type(layer_where_to_cut) == str:
+            last_layer = self.model.get_layer(layer_where_to_cut)
+        elif type(layer_where_to_cut) == int:
+            last_layer = self.model.layers[layer_where_to_cut]
+        else:
+            raise AttributeError("layer_where_to_cut has to be str or int type") 
+        model_without_last_layer = Model(self.model.input, last_layer.output, name='source_model')
+
+        # add a new fully connected layer
+        input_shape = self.model.layers[0].output_shape[1:]
+        x = Input(shape=input_shape, dtype='float32', name='input')
+        y = model_without_last_layer(x)
+
+        if new_model is None:
+            y = Dense(new_number_of_classes, activation=new_activation, name='new_dense_layer')(y)
+        else:
+            y = new_model(y)
+
+        # change self.model with fine_tuned model
+        self.model = Model(x, y)
+
+        # freeze the source model if freeze_source_model is True
+        self.model.get_layer('source_model').trainable = not freeze_source_model

dcase_models/model/models.py

@@ -65,7 +65,7 @@ def __init__(self, model=None, folder=None, metrics=['accuracy'], n_classes=10,
 
             # CONV 2
             y = Conv2D(48, filter_size_cnn, padding='valid', activation='relu', name='conv2')(y)
-            y = MaxPooling2D(pool_size=(2,2), strides=None, padding='valid', name='maxpool2')(y)
+            y = MaxPooling2D(pool_size=(4,2), strides=None, padding='valid', name='maxpool2')(y)
             y = BatchNormalization(name='batchnorm2')(y)
 
             # CONV 3

tests/SB_CNN/ESC50/fold1/model.json

@@ -0,0 +1 @@
+"{\"class_name\": \"Model\", \"config\": {\"name\": \"model_1\", \"layers\": [{\"name\": \"input\", \"class_name\": \"InputLayer\", \"config\": {\"batch_input_shape\": [null, 64, 128], \"dtype\": \"float32\", \"sparse\": false, \"name\": \"input\"}, \"inbound_nodes\": []}, {\"name\": \"source_model\", \"class_name\": \"Model\", \"config\": {\"name\": \"source_model\", \"layers\": [{\"name\": \"input\", \"class_name\": \"InputLayer\", \"config\": {\"batch_input_shape\": [null, 64, 128], \"dtype\": \"float32\", \"sparse\": false, \"name\": \"input\"}, \"inbound_nodes\": []}, {\"name\": \"lambda\", \"class_name\": \"Lambda\", \"config\": {\"name\": \"lambda\", \"trainable\": true, \"function\": [\"4wEAAAAAAAAAAQAAAAQAAABTAAAAcwwAAAB0AKABfABkAaECUwApAk7p/////ykC2gFL2gtleHBh\\nbmRfZGltcykB2gF4qQByBQAAAPofLi4vZGNhc2VfbW9kZWxzL21vZGVsL21vZGVscy5wedoIPGxh\\nbWJkYT44AAAA8wAAAAA=\\n\", null, null], \"function_type\": \"lambda\", \"output_shape\": null, \"output_shape_type\": \"raw\", \"arguments\": {}}, \"inbound_nodes\": [[[\"input\", 0, 0, {}]]]}, {\"name\": \"conv1\", \"class_name\": \"Conv2D\", \"config\": {\"name\": \"conv1\", \"trainable\": true, \"filters\": 24, \"kernel_size\": [5, 5], \"strides\": [1, 1], \"padding\": \"valid\", \"data_format\": \"channels_last\", \"dilation_rate\": [1, 1], \"activation\": \"relu\", \"use_bias\": true, \"kernel_initializer\": {\"class_name\": \"VarianceScaling\", \"config\": {\"scale\": 1.0, \"mode\": \"fan_avg\", \"distribution\": \"uniform\", \"seed\": null}}, \"bias_initializer\": {\"class_name\": \"Zeros\", \"config\": {}}, \"kernel_regularizer\": null, \"bias_regularizer\": null, \"activity_regularizer\": null, \"kernel_constraint\": null, \"bias_constraint\": null}, \"inbound_nodes\": [[[\"lambda\", 0, 0, {}]]]}, {\"name\": \"maxpool1\", \"class_name\": \"MaxPooling2D\", \"config\": {\"name\": \"maxpool1\", \"trainable\": true, \"pool_size\": [2, 2], \"padding\": \"valid\", \"strides\": [2, 2], \"data_format\": \"channels_last\"}, \"inbound_nodes\": [[[\"conv1\", 0, 0, {}]]]}, {\"name\": \"batchnorm1\", \"class_name\": \"BatchNormalization\", \"config\": {\"name\": \"batchnorm1\", \"trainable\": true, \"axis\": -1, \"momentum\": 0.99, \"epsilon\": 0.001, \"center\": true, \"scale\": true, \"beta_initializer\": {\"class_name\": \"Zeros\", \"config\": {}}, \"gamma_initializer\": {\"class_name\": \"Ones\", \"config\": {}}, \"moving_mean_initializer\": {\"class_name\": \"Zeros\", \"config\": {}}, \"moving_variance_initializer\": {\"class_name\": \"Ones\", \"config\": {}}, \"beta_regularizer\": null, \"gamma_regularizer\": null, \"beta_constraint\": null, \"gamma_constraint\": null}, \"inbound_nodes\": [[[\"maxpool1\", 0, 0, {}]]]}, {\"name\": \"conv2\", \"class_name\": \"Conv2D\", \"config\": {\"name\": \"conv2\", \"trainable\": true, \"filters\": 48, \"kernel_size\": [5, 5], \"strides\": [1, 1], \"padding\": \"valid\", \"data_format\": \"channels_last\", \"dilation_rate\": [1, 1], \"activation\": \"relu\", \"use_bias\": true, \"kernel_initializer\": {\"class_name\": \"VarianceScaling\", \"config\": {\"scale\": 1.0, \"mode\": \"fan_avg\", \"distribution\": \"uniform\", \"seed\": null}}, \"bias_initializer\": {\"class_name\": \"Zeros\", \"config\": {}}, \"kernel_regularizer\": null, \"bias_regularizer\": null, \"activity_regularizer\": null, \"kernel_constraint\": null, \"bias_constraint\": null}, \"inbound_nodes\": [[[\"batchnorm1\", 0, 0, {}]]]}, {\"name\": \"maxpool2\", \"class_name\": \"MaxPooling2D\", \"config\": {\"name\": \"maxpool2\", \"trainable\": true, \"pool_size\": [4, 2], \"padding\": \"valid\", \"strides\": [4, 2], \"data_format\": \"channels_last\"}, \"inbound_nodes\": [[[\"conv2\", 0, 0, {}]]]}, {\"name\": \"batchnorm2\", \"class_name\": \"BatchNormalization\", \"config\": {\"name\": \"batchnorm2\", \"trainable\": true, \"axis\": -1, \"momentum\": 0.99, \"epsilon\": 0.001, \"center\": true, \"scale\": true, \"beta_initializer\": {\"class_name\": \"Zeros\", \"config\": {}}, \"gamma_initializer\": {\"class_name\": \"Ones\", \"config\": {}}, \"moving_mean_initializer\": {\"class_name\": \"Zeros\", \"config\": {}}, \"moving_variance_initializer\": {\"class_name\": \"Ones\", \"config\": {}}, \"beta_regularizer\": null, \"gamma_regularizer\": null, \"beta_constraint\": null, \"gamma_constraint\": null}, \"inbound_nodes\": [[[\"maxpool2\", 0, 0, {}]]]}, {\"name\": \"conv3\", \"class_name\": \"Conv2D\", \"config\": {\"name\": \"conv3\", \"trainable\": true, \"filters\": 48, \"kernel_size\": [5, 5], \"strides\": [1, 1], \"padding\": \"valid\", \"data_format\": \"channels_last\", \"dilation_rate\": [1, 1], \"activation\": \"relu\", \"use_bias\": true, \"kernel_initializer\": {\"class_name\": \"VarianceScaling\", \"config\": {\"scale\": 1.0, \"mode\": \"fan_avg\", \"distribution\": \"uniform\", \"seed\": null}}, \"bias_initializer\": {\"class_name\": \"Zeros\", \"config\": {}}, \"kernel_regularizer\": null, \"bias_regularizer\": null, \"activity_regularizer\": null, \"kernel_constraint\": null, \"bias_constraint\": null}, \"inbound_nodes\": [[[\"batchnorm2\", 0, 0, {}]]]}, {\"name\": \"batchnorm3\", \"class_name\": \"BatchNormalization\", \"config\": {\"name\": \"batchnorm3\", \"trainable\": true, \"axis\": -1, \"momentum\": 0.99, \"epsilon\": 0.001, \"center\": true, \"scale\": true, \"beta_initializer\": {\"class_name\": \"Zeros\", \"config\": {}}, \"gamma_initializer\": {\"class_name\": \"Ones\", \"config\": {}}, \"moving_mean_initializer\": {\"class_name\": \"Zeros\", \"config\": {}}, \"moving_variance_initializer\": {\"class_name\": \"Ones\", \"config\": {}}, \"beta_regularizer\": null, \"gamma_regularizer\": null, \"beta_constraint\": null, \"gamma_constraint\": null}, \"inbound_nodes\": [[[\"conv3\", 0, 0, {}]]]}, {\"name\": \"flatten\", \"class_name\": \"Flatten\", \"config\": {\"name\": \"flatten\", \"trainable\": true, \"data_format\": \"channels_last\"}, \"inbound_nodes\": [[[\"batchnorm3\", 0, 0, {}]]]}, {\"name\": \"dropout1\", \"class_name\": \"Dropout\", \"config\": {\"name\": \"dropout1\", \"trainable\": true, \"rate\": 0.5, \"noise_shape\": null, \"seed\": null}, \"inbound_nodes\": [[[\"flatten\", 0, 0, {}]]]}, {\"name\": \"dense1\", \"class_name\": \"Dense\", \"config\": {\"name\": \"dense1\", \"trainable\": true, \"units\": 64, \"activation\": \"relu\", \"use_bias\": true, \"kernel_initializer\": {\"class_name\": \"VarianceScaling\", \"config\": {\"scale\": 1.0, \"mode\": \"fan_avg\", \"distribution\": \"uniform\", \"seed\": null}}, \"bias_initializer\": {\"class_name\": \"Zeros\", \"config\": {}}, \"kernel_regularizer\": {\"class_name\": \"L1L2\", \"config\": {\"l1\": 0.0, \"l2\": 0.0010000000474974513}}, \"bias_regularizer\": {\"class_name\": \"L1L2\", \"config\": {\"l1\": 0.0, \"l2\": 0.0010000000474974513}}, \"activity_regularizer\": null, \"kernel_constraint\": null, \"bias_constraint\": null}, \"inbound_nodes\": [[[\"dropout1\", 0, 0, {}]]]}, {\"name\": \"dropout2\", \"class_name\": \"Dropout\", \"config\": {\"name\": \"dropout2\", \"trainable\": true, \"rate\": 0.5, \"noise_shape\": null, \"seed\": null}, \"inbound_nodes\": [[[\"dense1\", 0, 0, {}]]]}], \"input_layers\": [[\"input\", 0, 0]], \"output_layers\": [[\"dropout2\", 0, 0]]}, \"inbound_nodes\": [[[\"input\", 0, 0, {}]]]}, {\"name\": \"new_dense_layer\", \"class_name\": \"Dense\", \"config\": {\"name\": \"new_dense_layer\", \"trainable\": true, \"units\": 50, \"activation\": \"softmax\", \"use_bias\": true, \"kernel_initializer\": {\"class_name\": \"VarianceScaling\", \"config\": {\"scale\": 1.0, \"mode\": \"fan_avg\", \"distribution\": \"uniform\", \"seed\": null}}, \"bias_initializer\": {\"class_name\": \"Zeros\", \"config\": {}}, \"kernel_regularizer\": null, \"bias_regularizer\": null, \"activity_regularizer\": null, \"kernel_constraint\": null, \"bias_constraint\": null}, \"inbound_nodes\": [[[\"source_model\", 1, 0, {}]]]}], \"input_layers\": [[\"input\", 0, 0]], \"output_layers\": [[\"new_dense_layer\", 0, 0]]}, \"keras_version\": \"2.2.4\", \"backend\": \"tensorflow\"}"

tests/SB_CNN/ESC50/fold1/training.log

@@ -0,0 +1,51 @@
+epoch,Acc,loss
+0,0.03,7.502168061392648
+1,0.0625,6.382974977039155
+2,0.0925,5.609923143840971
+3,0.1225,4.946143490019299
+4,0.1275,4.51800890604655
+5,0.14,4.206934458414714
+6,0.17,3.9105675583794004
+7,0.1825,3.74058198928833
+8,0.185,3.5537690730321976
+9,0.1975,3.4097133327665783
+10,0.1975,3.3246297645568847
+11,0.2025,3.1981849284399124
+12,0.205,3.157173556373233
+13,0.2075,3.0660635684785387
+14,0.2125,3.041565242948986
+15,0.2225,2.971737453823998
+16,0.22,2.9743683887663344
+17,0.2275,2.918078219095866
+18,0.2225,2.9116986974080405
+19,0.22,2.8902962466648647
+20,0.23,2.8917558279491606
+21,0.225,2.869637153262184
+22,0.2325,2.869133186340332
+23,0.2425,2.8532540253230505
+24,0.2375,2.8584180604843867
+25,0.235,2.833264215106056
+26,0.235,2.8339362516857327
+27,0.2325,2.8220349543435232
+28,0.23,2.8212475331624347
+29,0.2225,2.8232532464890254
+30,0.2225,2.841775826045445
+31,0.225,2.803494078318278
+32,0.23,2.81758319627671
+33,0.23,2.8031206031072706
+34,0.2375,2.84076034228007
+35,0.2375,2.8026808257330034
+36,0.2375,2.798538010006859
+37,0.2275,2.8218897683279853
+38,0.2225,2.8118302486056375
+39,0.2325,2.812475911095029
+40,0.2325,2.794565626780192
+41,0.2325,2.832269555046445
+42,0.2225,2.8263108544122604
+43,0.22,2.812968007950556
+44,0.2175,2.815446391786848
+45,0.23,2.7967070988246374
+46,0.225,2.801809174673898
+47,0.2325,2.8028654884156725
+48,0.2375,2.802789167676653
+49,0.2325,2.8161884684789746