FEAT: Add RMoK (#1148)

* WIP - Add reversible mixture of kan

* WIP - Allows import of RMoK

* AutoRMoK, add it to doc, add parameters

* Fix tests

* Get default config of AutoRMoK

* FIX: timemixer shapes mismatch and doc update (#1138)

* Use math.ceil to prevent shape mismatch

* Show exog support for KAN in doc

* FEAT: TimeLLM is faster and supports more LLMs (#1139)

* Fix issue #950: Reduce TimeLLM setup time for training

* Restore changes on the examples

* Revert changes to nbs/models.ipynb, nbs/models.softs.ipynb and neuralforecast/_modidx.py

* Revert changes to nbs/models.ipynb, nbs/models.softs.ipynb and neuralforecast/_modidx.py

* Refactor code to dynamically load models with AutoModel, AutoTokenizer, and AutoConfig

- Updated load_model_and_tokenizer function to use AutoModel, AutoTokenizer, and AutoConfig for flexible model loading.
- Included default model(gpt2) for cases where the specified model fails to load.
- Kept llm, llm_config, and llm_tokenizer arguments to minimize changes.
- Changed llm from storing pretrained weights to accepting pretrained model path to reduce necessary modifications.

This update enhances the flexibility and reliability of model loading based on received feedback while minimizing necessary changes.

* Refactor code to dynamically load models with AutoModel, AutoTokenizer, and AutoConfig

- Updated load_model_and_tokenizer function to use AutoModel, AutoTokenizer, and AutoConfig for flexible model loading.
- Included default model(gpt2) for cases where the specified model fails to load.
- Kept llm, llm_config, and llm_tokenizer arguments to minimize changes.
- Changed llm from storing pretrained weights to accepting pretrained model path to reduce necessary modifications.

This update enhances the flexibility and reliability of model loading based on received feedback while minimizing necessary changes.

* clear output

* modify test code

* Optimize model loading and add deprecation warning

- Simplify model loading logic
- Add constant for default model name
- Improve error handling for model loading
- Add success messages for model loading
- Implement deprecation warning for 'llm_config' and 'llm_tokenizer' parameters
- Update print messages for clarity
- Remove redundant code

This commit improves code readability, maintainability, and user experience
by providing clearer feedback and warnings about deprecated parameters.

* Resolved conflict in nbs/models.timellm.ipynb

---------

Co-authored-by: ive2go <[email protected]>
Co-authored-by: Olivier Sprangers <[email protected]>

* Consistency with math.ceil

---------

Co-authored-by: Olivier Sprangers <[email protected]>
Co-authored-by: ive2go <[email protected]>

* Add image, docstring, fix typo in comment

---------

Co-authored-by: Olivier Sprangers <[email protected]>
Co-authored-by: ive2go <[email protected]>

nbs/docs/capabilities/01_overview.ipynb

@@ -34,6 +34,7 @@
     "|`NHITS` | `AutoNHITS` | MLP | Univariate | Direct | F/H/S | \n",
     "|`NLinear` | `AutoNLinear` | MLP | Univariate | Direct  | - | \n",
     "|`PatchTST` | `AutoPatchTST` | Transformer | Univariate | Direct  | - | \n",
+    "|`RMoK` | `AutoRMoK` | KAN | Multivariate | Direct | - |\n",
     "|`RNN` | `AutoRNN` | RNN | Univariate | Recursive  | F/H/S | \n",
     "|`SOFTS` | `AutoSOFTS` | MLP | Multivariate | Direct  | - | \n",
     "|`StemGNN` | `AutoStemGNN` | GNN | Multivariate | Direct  | - | \n",

nbs/models.ipynb

@@ -66,6 +66,7 @@
     "from neuralforecast.models.itransformer import iTransformer\n",
     "\n",
     "from neuralforecast.models.kan import KAN\n",
+    "from neuralforecast.models.rmok import RMoK\n",
     "\n",
     "from neuralforecast.models.stemgnn import StemGNN\n",
     "from neuralforecast.models.hint import HINT\n",
@@ -2300,6 +2301,14 @@
     "model.fit(dataset=dataset)"
    ]
   },
+  {
+   "cell_type": "markdown",
+   "id": "8efc1692",
+   "metadata": {},
+   "source": [
+    "## C. KAN-Based"
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -2444,7 +2453,7 @@
    "id": "fd705a56",
    "metadata": {},
    "source": [
-    "## C. Transformer-Based"
+    "## D. Transformer-Based"
    ]
   },
   {
@@ -3427,7 +3436,7 @@
    "id": "57d6cb1f",
    "metadata": {},
    "source": [
-    "## D. CNN Based"
+    "## E. CNN Based"
    ]
   },
   {
@@ -3573,7 +3582,7 @@
    "id": "e6fd22c7",
    "metadata": {},
    "source": [
-    "## E. Multivariate"
+    "## F. Multivariate"
    ]
   },
   {
@@ -4680,6 +4689,158 @@
     "model.fit(dataset=dataset)"
    ]
   },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ab15c4b6",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#| export\n",
+    "class AutoRMoK(BaseAuto):\n",
+    "\n",
+    "    default_config = {\n",
+    "        \"input_size_multiplier\": [1, 2, 3, 4, 5],\n",
+    "        \"h\": None,\n",
+    "        \"n_series\": None,\n",
+    "        \"taylor_order\": tune.choice([3, 4, 5]),\n",
+    "        \"jacobi_degree\": tune.choice([4, 5, 6]),\n",
+    "        \"wavelet_function\": tune.choice(['mexican_hat', 'morlet', 'dog', 'meyer', 'shannon']),\n",
+    "        \"learning_rate\": tune.loguniform(1e-4, 1e-1),\n",
+    "        \"scaler_type\": tune.choice([None, 'robust', 'standard', 'identity']),\n",
+    "        \"max_steps\": tune.choice([500, 1000, 2000]),\n",
+    "        \"batch_size\": tune.choice([32, 64, 128, 256]),\n",
+    "        \"loss\": None,\n",
+    "        \"random_seed\": tune.randint(1, 20),\n",
+    "    }\n",
+    "\n",
+    "    def __init__(self,\n",
+    "                 h,\n",
+    "                 n_series,\n",
+    "                 loss=MAE(),\n",
+    "                 valid_loss=None,\n",
+    "                 config=None, \n",
+    "                 search_alg=BasicVariantGenerator(random_state=1),\n",
+    "                 num_samples=10,\n",
+    "                 refit_with_val=False,\n",
+    "                 cpus=cpu_count(),\n",
+    "                 gpus=torch.cuda.device_count(),\n",
+    "                 verbose=False,\n",
+    "                 alias=None,\n",
+    "                 backend='ray',\n",
+    "                 callbacks=None):\n",
+    "        \n",
+    "        # Define search space, input/output sizes\n",
+    "        if config is None:\n",
+    "            config = self.get_default_config(h=h, backend=backend, n_series=n_series)                 \n",
+    "\n",
+    "        # Always use n_series from parameters, raise exception with Optuna because we can't enforce it\n",
+    "        if backend == 'ray':\n",
+    "            config['n_series'] = n_series\n",
+    "        elif backend == 'optuna':\n",
+    "            mock_trial = MockTrial()\n",
+    "            if ('n_series' in config(mock_trial) and config(mock_trial)['n_series'] != n_series) or ('n_series' not in config(mock_trial)):\n",
+    "                raise Exception(f\"config needs 'n_series': {n_series}\")           \n",
+    "\n",
+    "        super(AutoRMoK, self).__init__(\n",
+    "              cls_model=RMoK, \n",
+    "              h=h,\n",
+    "              loss=loss,\n",
+    "              valid_loss=valid_loss,\n",
+    "              config=config,\n",
+    "              search_alg=search_alg,\n",
+    "              num_samples=num_samples, \n",
+    "              refit_with_val=refit_with_val,\n",
+    "              cpus=cpus,\n",
+    "              gpus=gpus,\n",
+    "              verbose=verbose,\n",
+    "              alias=alias,\n",
+    "              backend=backend,\n",
+    "              callbacks=callbacks,            \n",
+    "        )\n",
+    "\n",
+    "    @classmethod\n",
+    "    def get_default_config(cls, h, backend, n_series):\n",
+    "        config = cls.default_config.copy()        \n",
+    "        config['input_size'] = tune.choice([h * x \\\n",
+    "                        for x in config[\"input_size_multiplier\"]])\n",
+    "\n",
+    "        # Rolling windows with step_size=1 or step_size=h\n",
+    "        # See `BaseWindows` and `BaseRNN`'s create_windows\n",
+    "        config['step_size'] = tune.choice([1, h])\n",
+    "        del config[\"input_size_multiplier\"]\n",
+    "        if backend == 'optuna':\n",
+    "            # Always use n_series from parameters\n",
+    "            config['n_series'] = n_series\n",
+    "            config = cls._ray_config_to_optuna(config)           \n",
+    "\n",
+    "        return config         "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "111d8d3b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "show_doc(AutoRMoK, title_level=3)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "2073d4aa",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%capture\n",
+    "# Use your own config or AutoRMoK.default_config\n",
+    "config = dict(max_steps=1, val_check_steps=1, input_size=12, learning_rate=1e-2)\n",
+    "model = AutoRMoK(h=12, n_series=1, config=config, num_samples=1, cpus=1)\n",
+    "\n",
+    "# Fit and predict\n",
+    "model.fit(dataset=dataset)\n",
+    "y_hat = model.predict(dataset=dataset)\n",
+    "\n",
+    "# Optuna\n",
+    "model = AutoRMoK(h=12, n_series=1, config=None, backend='optuna')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ebe2c500",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#| hide\n",
+    "# Check Optuna\n",
+    "assert model.config(MockTrial())['h'] == 12\n",
+    "\n",
+    "# Unit test to test that Auto* model contains all required arguments from BaseAuto\n",
+    "test_args(AutoRMoK, exclude_args=['cls_model']) \n",
+    "\n",
+    "# Unit test for situation: Optuna with updated default config\n",
+    "my_config = AutoRMoK.get_default_config(h=12, n_series=1, backend='optuna')\n",
+    "def my_config_new(trial):\n",
+    "    config = {**my_config(trial)}\n",
+    "    config.update({'max_steps': 1, 'val_check_steps': 1, 'input_size': 12, 'learning_rate': 1e-1})\n",
+    "    return config\n",
+    "\n",
+    "model = AutoRMoK(h=12, n_series=1, config=my_config_new, backend='optuna', num_samples=1, cpus=1)\n",
+    "model.fit(dataset=dataset)\n",
+    "\n",
+    "# Unit test for situation: Ray with updated default config\n",
+    "my_config = AutoRMoK.get_default_config(h=12, n_series=1, backend='ray')\n",
+    "my_config['max_steps'] = 1\n",
+    "my_config['val_check_steps'] = 1\n",
+    "my_config['input_size'] = 12\n",
+    "my_config['learning_rate'] = 1e-1\n",
+    "model = AutoRMoK(h=12, n_series=1, config=my_config, backend='ray', num_samples=1, cpus=1)\n",
+    "model.fit(dataset=dataset)"
+   ]
+  },
   {
    "attachments": {},
    "cell_type": "markdown",