Add mlp_encoder_model (MLPs are used to process input data)

mlp_encoder_model.py

@@ -0,0 +1,93 @@
+import keras
+from keras.layers import *
+from keras.models import Sequential, Model
+from keras.utils import Sequence
+from qkeras import *
+
+import tensorflow as tf
+from tensorflow.keras import datasets, layers, models
+
+def var_network(var, hidden=10, output=2):
+    var = Flatten(name="flatten")(var)
+    var = QDense(
+        hidden,
+        kernel_quantizer=quantized_bits(8, 0, alpha=1),
+        bias_quantizer=quantized_bits(8, 0, alpha=1),
+        kernel_regularizer=tf.keras.regularizers.L1L2(0.01),
+        activity_regularizer=tf.keras.regularizers.L2(0.01),
+        name="dense_1"
+    )(var)
+    #var = keras.activations.tanh(var)
+    var = QActivation("quantized_tanh(8, 0, 1)", name="activation_tanh_2")(var)
+    var = QDense(
+        hidden,
+        kernel_quantizer=quantized_bits(8, 0, alpha=1),
+        bias_quantizer=quantized_bits(8, 0, alpha=1),
+        kernel_regularizer=tf.keras.regularizers.L1L2(0.01),
+        activity_regularizer=tf.keras.regularizers.L2(0.01),
+        name="dense_2"
+    )(var)
+    #var = keras.activations.tanh(var)
+    var = QActivation("quantized_tanh(8, 0, 1)", name="activation_tanh_3")(var)
+    return QDense(
+        output,
+        kernel_quantizer=quantized_bits(8, 0, alpha=1),
+        bias_quantizer=quantized_bits(8, 0, alpha=1),
+        kernel_regularizer=tf.keras.regularizers.L1L2(0.01),
+        name="dense_3"
+    )(var)
+
+def mlp_encoder_network(var, hidden=16, hidden_dimx=16, hidden_dimy=16):
+    proj_x = AveragePooling2D(
+        pool_size=(1, 21), 
+        strides=None, 
+        padding="valid", 
+        data_format=None,        
+    )(var)
+    proj_x = Flatten()(proj_x)
+
+    proj_y = AveragePooling2D(
+        pool_size=(13, 1), 
+        strides=None, 
+        padding="valid", 
+        data_format=None,        
+    )(var)
+    proj_y = Flatten()(proj_y)
+
+    proj_x = QDense(
+        hidden_dimx,
+        kernel_quantizer=quantized_bits(8, 0, alpha=1),
+        bias_quantizer=quantized_bits(8, 0, alpha=1),
+        kernel_regularizer=tf.keras.regularizers.L1L2(0.01),
+        activity_regularizer=tf.keras.regularizers.L2(0.01),
+    )(proj_x)
+    proj_x = QActivation("quantized_relu(bits=13, integer=5)(x)")(var)
+
+    proj_y = QDense(
+        hidden_dimy,
+        kernel_quantizer=quantized_bits(8, 0, alpha=1),
+        bias_quantizer=quantized_bits(8, 0, alpha=1),
+        kernel_regularizer=tf.keras.regularizers.L1L2(0.01),
+        activity_regularizer=tf.keras.regularizers.L2(0.01),
+    )(proj_y)
+    proj_y = QActivation("quantized_relu(bits=13, integer=5)(x)")(var)
+
+    var = Concatenate(axis=1)([proj_x, proj_y])
+
+    var = QDense(
+        hidden,
+        kernel_quantizer=quantized_bits(8, 0, alpha=1),
+        bias_quantizer=quantized_bits(8, 0, alpha=1),
+        kernel_regularizer=tf.keras.regularizers.L1L2(0.01),
+        activity_regularizer=tf.keras.regularizers.L2(0.01),
+    )(var)
+
+    var = QActivation("quantized_tanh(8, 0, 1)")(var)
+    return var
+
+def CreateModel(shape):
+    x_base = x_in = Input(shape, name="input_pxls/")
+    stack = mlp_encoder_network(x_base)
+    stack = var_network(stack, hidden=16, output=3) # this network should only be used with 'slim' (3) or 'diagonal' (8) regression targets
+    model = Model(inputs=x_in, outputs=stack, name="smrtpxl_regression")
+    return model

train_model_optimized_giuseppe.ipynb

@@ -132,6 +132,8 @@
    "source": [
     "if model_type == \"xysum\":\n",
     "    from xysum_model import *\n",
+    "elif model_type == \"mlp_encoder\"\n",
+    "    from mlp_encoder_model import *\n",
     "else:\n",
     "    assert False, \"Selected model cannot be found\""
    ]
@@ -593,7 +595,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.10.15"
+   "version": "3.10.18"
   }
  },
  "nbformat": 4,