added deepVAR and Transformer notebook

Author: kashif

ConvTSMixer-hyperparameter_tuning.ipynb

@@ -20,7 +20,7 @@
     "from gluonts.evaluation.backtest import make_evaluation_predictions\n",
     "from gluonts.evaluation import MultivariateEvaluator\n",
     "\n",
-    "#from pts.modules import StudentTOutput\n",
+    "# from pts.modules import StudentTOutput\n",
     "\n",
     "from ConvTSMixer import ConvTSMixerEstimator\n",
     "import random\n",
@@ -38,56 +38,63 @@
    },
    "outputs": [],
    "source": [
-    "class ConvTSMixerObjective:  \n",
-    "    def __init__(self, dataset, train_grouper, test_grouper, metric_type=\"m_sum_mean_wQuantileLoss\"):\n",
+    "class ConvTSMixerObjective:\n",
+    "    def __init__(\n",
+    "        self,\n",
+    "        dataset,\n",
+    "        train_grouper,\n",
+    "        test_grouper,\n",
+    "        metric_type=\"m_sum_mean_wQuantileLoss\",\n",
+    "    ):\n",
     "        self.metric_type = metric_type\n",
     "        self.dataset = dataset\n",
     "        self.dataset_train = train_grouper(self.dataset.train)\n",
     "        self.dataset_test = test_grouper(self.dataset.test)\n",
-    "    \n",
+    "\n",
     "    def get_params(self, trial) -> dict:\n",
     "        return {\n",
-    "        \"context_length\": trial.suggest_int(\"context_length\", dataset.metadata.prediction_length, dataset.metadata.prediction_length*10,4),\n",
-    "        \"batch_size\": trial.suggest_int(\"batch_size\", 128, 256, 64),\n",
-    "        \"depth\": trial.suggest_int(\"depth\", 2, 16,4),\n",
-    "        \"dim\": trial.suggest_int(\"dim\", 16, 64, 16),\n",
-    "        \"patch_size\": trial.suggest_int(\"dim\", 2, 16,4),\n",
-    "         \"kernel_size\": trial.suggest_int(\"dim\", 9, 18, 3),\n",
+    "            \"context_length\": trial.suggest_int(\n",
+    "                \"context_length\",\n",
+    "                dataset.metadata.prediction_length,\n",
+    "                dataset.metadata.prediction_length * 10,\n",
+    "                4,\n",
+    "            ),\n",
+    "            \"batch_size\": trial.suggest_int(\"batch_size\", 128, 256, 64),\n",
+    "            \"depth\": trial.suggest_int(\"depth\", 2, 16, 4),\n",
+    "            \"dim\": trial.suggest_int(\"dim\", 16, 64, 16),\n",
+    "            \"patch_size\": trial.suggest_int(\"dim\", 2, 16, 4),\n",
+    "            \"kernel_size\": trial.suggest_int(\"dim\", 9, 18, 3),\n",
     "        }\n",
-    "     \n",
+    "\n",
     "    def __call__(self, trial):\n",
     "        params = self.get_params(trial)\n",
     "        estimator = estimator = ConvTSMixerEstimator(\n",
-    "            #distr_output=StudentTOutput(dim=int(dataset.metadata.feat_static_cat[0].cardinality)),\n",
+    "            # distr_output=StudentTOutput(dim=int(dataset.metadata.feat_static_cat[0].cardinality)),\n",
     "            input_size=int(self.dataset.metadata.feat_static_cat[0].cardinality),\n",
-    "\n",
     "            prediction_length=self.dataset.metadata.prediction_length,\n",
-    "            context_length=self.dataset.metadata.prediction_length*5,\n",
+    "            context_length=self.dataset.metadata.prediction_length * 5,\n",
     "            freq=self.dataset.metadata.freq,\n",
     "            scaling=\"std\",\n",
-    "\n",
     "            depth=params[\"depth\"],\n",
     "            patch_size=(params[\"patch_size\"], params[\"patch_size\"]),\n",
     "            kernel_size=params[\"kernel_size\"],\n",
     "            dim=params[\"dim\"],\n",
-    "\n",
     "            batch_size=params[\"batch_size\"],\n",
     "            num_batches_per_epoch=100,\n",
-    "            trainer_kwargs=dict(accelerator=\"cuda\", max_epochs=30)\n",
+    "            trainer_kwargs=dict(accelerator=\"cuda\", max_epochs=30),\n",
     "        )\n",
     "        predictor = estimator.train(\n",
-    "        training_data=self.dataset_train,\n",
-    "        num_workers=8,\n",
-    "        shuffle_buffer_length=1024\n",
+    "            training_data=self.dataset_train, num_workers=8, shuffle_buffer_length=1024\n",
+    "        )\n",
+    "\n",
+    "        forecast_it, ts_it = make_evaluation_predictions(\n",
+    "            dataset=self.dataset_test, predictor=predictor, num_samples=100\n",
     "        )\n",
-    "        \n",
-    "        forecast_it, ts_it = make_evaluation_predictions(dataset=self.dataset_test,\n",
-    "                                             predictor=predictor,\n",
-    "                                             num_samples=100)\n",
     "        forecasts = list(forecast_it)\n",
     "        tss = list(ts_it)\n",
-    "        evaluator = MultivariateEvaluator(quantiles=(np.arange(20)/20.0)[1:],\n",
-    "                                  target_agg_funcs={'sum': np.sum})\n",
+    "        evaluator = MultivariateEvaluator(\n",
+    "            quantiles=(np.arange(20) / 20.0)[1:], target_agg_funcs={\"sum\": np.sum}\n",
+    "        )\n",
     "        agg_metrics, _ = evaluator(iter(tss), iter(forecasts))\n",
     "        return agg_metrics[self.metric_type]"
    ]
@@ -102,10 +109,14 @@
    "outputs": [],
    "source": [
     "dataset = get_dataset(\"solar_nips\", regenerate=False)\n",
-    "train_grouper = MultivariateGrouper(max_target_dim=int(dataset.metadata.feat_static_cat[0].cardinality))\n",
+    "train_grouper = MultivariateGrouper(\n",
+    "    max_target_dim=int(dataset.metadata.feat_static_cat[0].cardinality)\n",
+    ")\n",
     "\n",
-    "test_grouper = MultivariateGrouper(num_test_dates=int(len(dataset.test)/len(dataset.train)), \n",
-    "                                   max_target_dim=int(dataset.metadata.feat_static_cat[0].cardinality))\n",
+    "test_grouper = MultivariateGrouper(\n",
+    "    num_test_dates=int(len(dataset.test) / len(dataset.train)),\n",
+    "    max_target_dim=int(dataset.metadata.feat_static_cat[0].cardinality),\n",
+    ")\n",
     "dataset_train = train_grouper(dataset.train)\n",
     "dataset_test = test_grouper(dataset.test)"
    ]