continue keras core integration

Author: PatReis

SDP.md SDP.md SDP.md renamed to SDP.md

@@ -48,112 +48,6 @@
     "## Implementaion details"
    ]
   },
-  {
-   "cell_type": "markdown",
-   "id": "fc229b02",
-   "metadata": {},
-   "source": [
-    "Most tensorflow methods already support ragged tensors, which can be looked up here: https://www.tensorflow.org/api_docs/python/tf/ragged. \n",
-    "\n",
-    "For using keras layers, most layers in `kgcnn` inherit from `kgcnn.layers.base.GraphBaseLayer` which adds some utility methods and arguments, such as the `ragged_validate` parameter that is used for ragged tensor creation."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 1,
-   "id": "49ec83af",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "True\n"
-     ]
-    }
-   ],
-   "source": [
-    "import tensorflow as tf\n",
-    "from kgcnn.layers.base import GraphBaseLayer\n",
-    "\n",
-    "class NewLayer(GraphBaseLayer):\n",
-    "    \n",
-    "    def __init__(self, **kwargs):\n",
-    "        super().__init__(**kwargs)\n",
-    "        \n",
-    "    def call(self, inputs, **kwargs):\n",
-    "        # Do something in call.\n",
-    "        return inputs\n",
-    "\n",
-    "new_layer = NewLayer(ragged_validate=False)\n",
-    "print(new_layer._supports_ragged_inputs)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "id": "5c664a24",
-   "metadata": {},
-   "source": [
-    "> **WARNING**: Since ragged tensors can result in quite a performance loss due to shape checks of ragged dimensions on runtime, it is recommended to directly work with the values tensor/information (if possible) and to use `ragged_validate` to `False` in production (for using parallization this might be different). An example is the `tf.ragged.map_flat_values` method. Due to this reason, there are `kgcnn.layers.modules.LazyAdd` and `kgcnn.layers.modules.LazyConcatenate` etc. layers.\n",
-    "\n",
-    "Utility methods of `GraphBaseLayer` to work with values directly are `assert_ragged_input_rank` and `map_values` . Note that this can speed up models and is essentially equal to a disjoint graph tensor representation. \n",
-    "However, with ragged tensors there is also the possibility to try `tf.vectorized_map` or `tf.map_fn` if values can not be accessed.\n",
-    "\n",
-    "Here is an example of how to use `assert_ragged_input_rank` and `map_values` . With `assert_ragged_input_rank` it can be ensured that a ragged tensor is given in `call` by casting padded+mask or normal tensor (for example in case of equal sized graphs) to a ragged version, in order to accesss e.g. `inputs.values` etc.\n",
-    "With `map_values` a function can be directly applied to the values tensor or a list of value tensors. Axis argument can refer to the ragged tensor shape but the `map_values` is restricted to ragged rank of one. Fallback for `map_values` is applying the function directly to its input."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 2,
-   "id": "2a02b24c",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "class NewLayer(GraphBaseLayer):\n",
-    "    \n",
-    "    def __init__(self, **kwargs):\n",
-    "        super().__init__(**kwargs)\n",
-    "        \n",
-    "    def call_v1(self, inputs, **kwargs):\n",
-    "        inputs = self.assert_ragged_input_rank(inputs, ragged_rank=1)\n",
-    "        return tf.RaggedTensor.from_row_splits(\n",
-    "            tf.exp(inputs.values), inputs.row_splits, validate=self.ragged_validate)\n",
-    "\n",
-    "    def call_v2(self, inputs, **kwargs):\n",
-    "        # Possible kwargs for function can be added. Can have axis argument (special case).\n",
-    "        return self.map_values(tf.exp, inputs)\n",
-    "    \n",
-    "new_layer = NewLayer()"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 3,
-   "id": "e4699f10",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "x (v1): (2, 3, 4) (2, None, 4)\n",
-      "x (v2): (2, 3, 4) (2, 3, 4)\n",
-      "x_ragged (v1): (2, None, 4) (2, None, 4)\n",
-      "x_ragged (v2): (2, None, 4) (2, None, 4)\n"
-     ]
-    }
-   ],
-   "source": [
-    "x = tf.ones((2, 3, 4))\n",
-    "x_ragged = tf.ragged.constant([[[1.0, 1.0, 1.0, 1.0]],[[1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0]]], ragged_rank=1)\n",
-    "\n",
-    "print(\"x (v1):\", x.shape, new_layer.call_v1(x).shape)\n",
-    "print(\"x (v2):\", x.shape, new_layer.call_v2(x).shape)\n",
-    "print(\"x_ragged (v1):\", x_ragged.shape, new_layer.call_v1(x_ragged).shape)\n",
-    "print(\"x_ragged (v2):\", x_ragged.shape, new_layer.call_v2(x_ragged).shape)"
-   ]
-  },
   {
    "cell_type": "markdown",
    "id": "212298b7",
@@ -187,7 +81,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.9.7"
+   "version": "3.10.13"
   }
  },
  "nbformat": 4,