add vignette on fine-tuning

Nixtla · Dec 19, 2024 · c2ee9cc · c2ee9cc
1 parent 1386791
commit c2ee9cc
Show file tree

Hide file tree

Showing 13 changed files with 917 additions and 18 deletions.
diff --git a/R/nixtla_client_cross_validation.R b/R/nixtla_client_cross_validation.R
@@ -11,9 +11,9 @@
 #' @param quantiles Quantiles to forecast. Should be between 0 and 1.
 #' @param n_windows Number of windows to evaluate.
 #' @param step_size Step size between each cross validation window. If NULL, it will equal the forecast horizon (h).
-#' @param finetune_steps Number of steps used to finetune 'TimeGPT' in the new data.
-#' @param finetune_depth The depth of the fine-tuning. Accepts an integer value from 1 to 5, where 1 (default) means little fine-tuning and 5 means that the entire model is fine-tuned to your dataset.
-#' @param finetune_loss Loss function to use for finetuning. Options are: "default", "mae", "mse", "rmse", "mape", and "smape".
+#' @param finetune_steps Number of steps used to fine-tune 'TimeGPT' in the new data.
+#' @param finetune_depth The depth of the fine-tuning. Uses a scale from 1 to 5, where 1 means little fine-tuning and 5 means that the entire model is fine-tuned.
+#' @param finetune_loss Loss function to use for fine-tuning. Options are: "default", "mae", "mse", "rmse", "mape", and "smape".
 #' @param clean_ex_first Clean exogenous signal before making the forecasts using 'TimeGPT'.
 #' @param model Model to use, either "timegpt-1" or "timegpt-1-long-horizon". Use "timegpt-1-long-horizon" if you want to forecast more than one seasonal period given the frequency of the data.
 #'

diff --git a/R/nixtla_client_forecast.R b/R/nixtla_client_forecast.R
@@ -10,9 +10,9 @@
 #' @param X_df A tsibble or a data frame with future exogenous variables.
 #' @param level The confidence levels (0-100) for the prediction intervals.
 #' @param quantiles Quantiles to forecast. Should be between 0 and 1.
-#' @param finetune_steps Number of steps used to finetune 'TimeGPT' in the new data.
-#' @param finetune_depth The depth of the fine-tuning. Accepts an integer value from 1 to 5, where 1 (default) means little fine-tuning and 5 means that the entire model is fine-tuned to your dataset.
-#' @param finetune_loss Loss function to use for finetuning. Options are: "default", "mae", "mse", "rmse", "mape", and "smape".
+#' @param finetune_steps Number of steps used to fine-tune 'TimeGPT' in the new data.
+#' @param finetune_depth The depth of the fine-tuning. Uses a scale from 1 to 5, where 1 means little fine-tuning and 5 means that the entire model is fine-tuned.
+#' @param finetune_loss Loss function to use for fine-tuning. Options are: "default", "mae", "mse", "rmse", "mape", and "smape".
 #' @param clean_ex_first Clean exogenous signal before making the forecasts using 'TimeGPT'.
 #' @param add_history Return fitted values of the model.
 #' @param model Model to use, either "timegpt-1" or "timegpt-1-long-horizon". Use "timegpt-1-long-horizon" if you want to forecast more than one seasonal period given the frequency of the data.

diff --git a/R/nixtla_client_historic.R b/R/nixtla_client_historic.R
@@ -8,9 +8,9 @@
 #' @param target_col Column that contains the target variable.
 #' @param level The confidence levels (0-100) for the prediction intervals.
 #' @param quantiles Quantiles to forecast. Should be between 0 and 1.
-#' @param finetune_steps Number of steps used to finetune 'TimeGPT' in the new data.
-#' @param finetune_depth The depth of the fine-tuning. Accepts an integer value from 1 to 5, where 1 (default) means little fine-tuning and 5 means that the entire model is fine-tuned to your dataset.
-#' @param finetune_loss Loss function to use for finetuning. Options are: "default", "mae", "mse", "rmse", "mape", and "smape".
+#' @param finetune_steps Number of steps used to fine-tune 'TimeGPT' in the new data.
+#' @param finetune_depth The depth of the fine-tuning. Uses a scale from 1 to 5, where 1 means little fine-tuning and 5 means that the entire model is fine-tuned.
+#' @param finetune_loss Loss function to use for fine-tuning. Options are: "default", "mae", "mse", "rmse", "mape", and "smape".
 #' @param clean_ex_first Clean exogenous signal before making the forecasts using 'TimeGPT'.
 #' @param model Model to use, either "timegpt-1" or "timegpt-1-long-horizon". Use "timegpt-1-long-horizon" if you want to forecast more than one seasonal period given the frequency of the data.
 #'

diff --git a/man/nixtla_client_cross_validation.Rd b/man/nixtla_client_cross_validation.Rd
diff --git a/man/nixtla_client_forecast.Rd b/man/nixtla_client_forecast.Rd
diff --git a/man/nixtla_client_historic.Rd b/man/nixtla_client_historic.Rd
diff --git a/tests/mocks/api.nixtla.io/v2/forecast-038639-POST.json b/tests/mocks/api.nixtla.io/v2/forecast-038639-POST.json
@@ -0,0 +1,131 @@
+{
+    "input_tokens": 600,
+    "output_tokens": 120,
+    "finetune_tokens": 0,
+    "mean": [
+        38.820404,
+        36.293564,
+        34.978855,
+        32.99633,
+        31.583813,
+        33.274483,
+        36.975334,
+        42.980255,
+        47.18004,
+        49.38613,
+        50.58629,
+        50.16604,
+        49.604767,
+        48.710793,
+        47.50801,
+        49.641068,
+        52.359497,
+        58.20144,
+        58.857323,
+        55.131165,
+        49.567757,
+        45.881218,
+        44.496113,
+        43.748726,
+        10.04649,
+        6.3477135,
+        3.8844414,
+        2.3206596,
+        2.5391273,
+        4.168751,
+        4.5570107,
+        7.487301,
+        9.367214,
+        11.309727,
+        11.275341,
+        11.971142,
+        11.837162,
+        12.7635765,
+        13.369236,
+        14.58382,
+        15.6118355,
+        15.449764,
+        15.033131,
+        14.167515,
+        13.30592,
+        12.907833,
+        10.776638,
+        7.633236,
+        52.749832,
+        49.590576,
+        46.05893,
+        41.110207,
+        39.65557,
+        42.497383,
+        48.12867,
+        54.479225,
+        58.289307,
+        60.86104,
+        61.56459,
+        60.438133,
+        59.821243,
+        56.35112,
+        54.18071,
+        55.2239,
+        56.647575,
+        64.53344,
+        67.05695,
+        65.62052,
+        61.02072,
+        56.786224,
+        56.03811,
+        54.446175,
+        49.7386,
+        49.122772,
+        48.87745,
+        48.59867,
+        49.343666,
+        51.27717,
+        53.082943,
+        56.1737,
+        57.441006,
+        57.917213,
+        58.155327,
+        58.4386,
+        58.29986,
+        58.309822,
+        58.473557,
+        59.172974,
+        59.360233,
+        59.138382,
+        57.66301,
+        55.740982,
+        54.20776,
+        53.051807,
+        51.7741,
+        50.458515,
+        22.853998,
+        22.351372,
+        21.971369,
+        22.012585,
+        22.484016,
+        22.69992,
+        26.459377,
+        31.47617,
+        31.518444,
+        30.856718,
+        30.465942,
+        29.211628,
+        27.852182,
+        27.181736,
+        26.28147,
+        26.776318,
+        28.727142,
+        35.49711,
+        35.459,
+        33.863174,
+        32.520737,
+        29.762478,
+        27.170156,
+        24.734856
+    ],
+    "intervals": null,
+    "weights_x": null,
+    "feature_contributions": null,
+    "request_id": "JEAB9GZ9KA"
+}
diff --git a/tests/mocks/api.nixtla.io/v2/forecast-dd4308-POST.json b/tests/mocks/api.nixtla.io/v2/forecast-dd4308-POST.json
@@ -0,0 +1,131 @@
+{
+    "input_tokens": 8280,
+    "output_tokens": 120,
+    "finetune_tokens": 500,
+    "mean": [
+        39.33689,
+        36.81118,
+        35.862137,
+        34.178474,
+        33.005707,
+        34.602592,
+        38.382442,
+        45.29547,
+        49.116364,
+        51.709076,
+        52.31858,
+        51.4066,
+        50.5456,
+        49.365852,
+        48.270195,
+        50.595474,
+        53.19593,
+        59.678688,
+        60.51818,
+        57.22879,
+        51.752975,
+        47.76425,
+        46.002213,
+        44.784706,
+        10.026592,
+        6.836117,
+        4.517971,
+        3.9740334,
+        3.70512,
+        5.399822,
+        6.174061,
+        9.923893,
+        12.239056,
+        14.2149315,
+        13.17947,
+        14.263432,
+        13.346569,
+        14.199467,
+        14.601391,
+        16.600262,
+        19.36082,
+        20.515034,
+        21.571346,
+        21.480137,
+        21.020634,
+        20.655666,
+        18.897038,
+        16.328136,
+        52.39077,
+        49.435287,
+        46.650024,
+        43.17102,
+        42.517525,
+        45.289276,
+        51.680283,
+        58.49974,
+        62.696686,
+        64.67974,
+        63.97027,
+        62.087547,
+        60.991127,
+        57.563072,
+        55.60578,
+        56.84146,
+        58.360245,
+        66.88493,
+        69.96444,
+        69.03388,
+        64.37969,
+        59.964928,
+        58.9301,
+        56.773148,
+        49.328724,
+        48.39244,
+        47.72139,
+        47.343994,
+        47.672806,
+        48.544514,
+        49.517147,
+        51.288704,
+        52.273098,
+        53.232426,
+        53.966618,
+        54.523796,
+        54.51413,
+        54.70507,
+        55.0904,
+        55.68209,
+        56.167664,
+        56.07667,
+        55.04934,
+        53.626564,
+        52.309814,
+        51.268036,
+        50.12602,
+        48.925766,
+        23.074203,
+        22.41476,
+        21.85689,
+        22.067892,
+        22.626034,
+        22.7155,
+        26.250298,
+        31.046175,
+        31.589384,
+        31.64193,
+        31.067324,
+        29.933369,
+        28.635342,
+        27.827301,
+        26.68113,
+        27.277428,
+        29.669315,
+        35.96617,
+        36.442406,
+        35.202442,
+        33.239815,
+        30.23969,
+        27.701052,
+        25.199667
+    ],
+    "intervals": null,
+    "weights_x": null,
+    "feature_contributions": null,
+    "request_id": "AHD4U6ADNU"
+}