Add L1 regularization

chatbot_constructor.py

@@ -29,7 +29,7 @@ def todset(text: str):
 def hash_str(data: str):
     return hashlib.md5(data.encode('utf-8')).hexdigest()
 
-def train(message: str = "", dropout: float = 0.5, learning_rate: float = 0.001, epochs: int = 16, emb_size: int = 128, input_len: int = 16, kernels_count: int = 8, kernel_size: int = 8, data: str = ""):
+def train(message: str = "", regularization: float = 0.0001, dropout: float = 0.1, learning_rate: float = 0.001, epochs: int = 16, emb_size: int = 128, input_len: int = 16, kernels_count: int = 8, kernel_size: int = 8, data: str = ""):
     data_hash = None
     if "→" not in data or "\n" not in data:
             if data in os.listdir("cache"):
@@ -44,7 +44,7 @@ def train(message: str = "", dropout: float = 0.5, learning_rate: float = 0.001,
     vocab_size = len(tokenizer.word_index) + 1
     inp_len = input_len
     if data_hash is None:
-        data_hash = hash_str(data)+"_"+str(dropout)+"_"+str(learning_rate)+"_"+str(epochs)+"_"+str(emb_size)+"_"+str(inp_len)+"_"+str(kernels_count)+"_"+str(kernel_size)+".keras"
+        data_hash = hash_str(data)+"_"+str(l1)+"_"+str(dropout)+"_"+str(learning_rate)+"_"+str(epochs)+"_"+str(emb_size)+"_"+str(inp_len)+"_"+str(kernels_count)+"_"+str(kernel_size)+".keras"
     elif message == "!getmodelhash":
         return data_hash
     else:
@@ -57,9 +57,9 @@ def train(message: str = "", dropout: float = 0.5, learning_rate: float = 0.001,
         dropout1_layer = Dropout(dropout)(emb_layer)
         attn_layer = MultiHeadAttention(num_heads=4, key_dim=128)(dropout1_layer, dropout1_layer, dropout1_layer)
         noise_layer = GaussianNoise(0.1)(attn_layer)
-        conv1_layer = Conv1D(kernels_count, kernel_size, padding='same', activation='relu', strides=1, input_shape=(64, 128))(noise_layer)
-        conv2_layer = Conv1D(16, 4, padding='same', activation='relu', strides=1)(conv1_layer)
-        conv3_layer = Conv1D(8, 2, padding='same', activation='relu', strides=1)(conv2_layer)
+        conv1_layer = Conv1D(kernels_count, kernel_size, padding='same', activation='relu', strides=1, input_shape=(64, 128), kernel_regularizer=regularizers.L1(regularization))(noise_layer)
+        conv2_layer = Conv1D(16, 4, padding='same', activation='relu', strides=1, kernel_regularizer=regularizers.L1(regularization))(conv1_layer)
+        conv3_layer = Conv1D(8, 2, padding='same', activation='relu', strides=1, kernel_regularizer=regularizers.L1(regularization))(conv2_layer)
         flatten_layer = Flatten()(conv3_layer)
         attn_flatten_layer = Flatten()(attn_layer)
         conv1_flatten_layer = Flatten()(conv1_layer)
@@ -67,16 +67,16 @@ def train(message: str = "", dropout: float = 0.5, learning_rate: float = 0.001,
         conv3_flatten_layer = Flatten()(conv3_layer)
         concat1_layer = Concatenate()([flatten_layer, attn_flatten_layer, conv1_flatten_layer, conv2_flatten_layer, conv3_flatten_layer])
         dropout2_layer = Dropout(dropout)(concat1_layer)
-        dense1_layer = Dense(512, activation="linear")(dropout2_layer)
+        dense1_layer = Dense(512, activation="linear", kernel_regularizer=regularizers.L1(regularization))(dropout2_layer)
         prelu1_layer = PReLU()(dense1_layer)
         dropout3_layer = Dropout(dropout)(prelu1_layer)
-        dense2_layer = Dense(256, activation="tanh")(dropout3_layer)
+        dense2_layer = Dense(256, activation="tanh", kernel_regularizer=regularizers.L1(regularization))(dropout3_layer)
         dropout4_layer = Dropout(dropout)(dense2_layer)
-        dense3_layer = Dense(256, activation="relu")(dropout4_layer)
+        dense3_layer = Dense(256, activation="relu", kernel_regularizer=regularizers.L1(regularization))(dropout4_layer)
         dropout5_layer = Dropout(dropout)(dense3_layer)
-        dense4_layer = Dense(100, activation="tanh")(dropout5_layer)
+        dense4_layer = Dense(100, activation="tanh", kernel_regularizer=regularizers.L1(regularization))(dropout5_layer)
         concat2_layer = Concatenate()([dense4_layer, prelu1_layer, attn_flatten_layer, conv1_flatten_layer])
-        dense4_layer = Dense(resps_len, activation="softmax")(concat2_layer)
+        dense4_layer = Dense(resps_len, activation="softmax", kernel_regularizer=regularizers.L1(regularization))(concat2_layer)
         model = Model(inputs=input_layer, outputs=dense4_layer)
         
         X = []
@@ -101,8 +101,9 @@ def train(message: str = "", dropout: float = 0.5, learning_rate: float = 0.001,
 
 if __name__ == "__main__":
     iface = gr.Interface(fn=train, inputs=["text",
-                                           gr.inputs.Slider(0, 0.5, default=0.5, step=0.00000001, label="Dropout"),
-                                           gr.inputs.Slider(0.00000001, 0.1, default=0.001, step=0.00000001, label="Learning rate"),
+                                           gr.inputs.Slider(0, 0.01, default=0.0001, step=0.00000001, label="Regularization L1"),
+                                           gr.inputs.Slider(0, 0.5, default=0.1, step=0.00000001, label="Dropout"),
+                                           gr.inputs.Slider(0.00000001, 0.01, default=0.001, step=0.00000001, label="Learning rate"),
                                            gr.inputs.Slider(1, 64, default=32, step=1, label="Epochs"),
                                            gr.inputs.Slider(1, 256, default=100, step=1, label="Embedding size"),
                                            gr.inputs.Slider(1, 128, default=16, step=1, label="Input Length"),