Update example configs and README

Author: dfulu

README.md

@@ -25,8 +25,7 @@ pip install ".[dev]"
 In order to run PVNet summation, we assume that you are already set up with
 [PVNet](https://github.com/openclimatefix/pvnet) and have a trained PVNet model already available either locally or pushed to HuggingFace.
 
-Before running any code, copy the example configuration to a
-configs directory:
+Before running any code, copy the example configuration to a configs directory:
 
 ```
 cp -r configs.example configs
@@ -41,31 +40,26 @@ The datasets required are the same as documented in
 data for the national sum i.e. GSP ID 0.
 
 
-### Set up and config example for batch creation
-
-
-The concurrent batches created in the step above will be augmented with a few additional pieces of
-data required for the summation model. Within your copy of `PVNet_summation/configs` make sure you
-have replaced all of the items marked with `PLACEHOLDER`
-
 ### Training PVNet_summation
 
 How PVNet_summation is run is determined by the extensive configuration in the config files. The
-configs stored in `PVNet/configs.example` should work with batches created using the steps and
-batch creation config mentioned above.
+configs stored in `configs.example`.
 
 Make sure to update the following config files before training your model:
 
-1. In `configs/datamodule/default.yaml`:
-    - update `batch_dir` to point to the directory you stored your concurrent batches in during
-      batch creation.
-    - update `gsp_zarr_path` to point to the PVLive data containing the national estimate
-2. In `configs/model/default.yaml`:
-    - update the PVNet model for which you are training a summation model for. A new summation model
-      should be trained for each PVNet model
-    - update the hyperparameters and structure of the summation model
-3. In `configs/trainer/default.yaml`:
-    - set `accelerator: 0` if running on a system without a supported GPU
+
+1. At the very start of training we loop over all of the input samples and make predictions for them using PVNet. These predictions are saved to disk and will be loaded in the training loop for more efficient training. In `configs/config.yaml` update `sample_save_dir` to set where the predictions will be saved to.
+
+2. In `configs/datamodule/default.yaml`:
+  - Update `pvnet_model.model_id` and `pvnet_model.revision` to point to the Huggingface commit or local directory where the exported PVNet model is.
+  - Update `configuration` to point to a data configuration compatible with the PVNet model whose outputs will be fed into the summation model.
+  - Set `train_period` and `val_period` to control the time ranges of the train and val period
+  - Optionally set `max_num_train_samples` and `max_num_val_samples` to limit the number of possible train and validation example which will be used.
+
+3. In `configs/model/default.yaml`:
+    - Update the hyperparameters and structure of the summation model
+4. In `configs/trainer/default.yaml`:
+    - Set `accelerator: 0` if running on a system without a supported GPU
 
 
 Assuming you have updated the configs, you should now be able to run:
@@ -74,7 +68,6 @@ Assuming you have updated the configs, you should now be able to run:
 python run.py
 ```
 
-This will then use the pretrained PVNet model to run inference on the concurrent batches, the outputs from this inference will then be used as the training data for the summation model alongside the national PVLive data (GSP ID 0).
 
 ## Testing
 

configs.example/callbacks/default.yaml

@@ -19,9 +19,3 @@ model_checkpoint:
   dirpath: "PLACEHOLDER/${model_name}"
   auto_insert_metric_name: False
   save_on_train_epoch_end: False
-
-#stochastic_weight_averaging:
-#  _target_: pvnet_summation.callbacks.StochasticWeightAveraging
-#  swa_lrs: 0.0000001
-#  swa_epoch_start: 0.8
-#  annealing_epochs: 5

configs.example/config.yaml

@@ -10,7 +10,7 @@ defaults:
   - hydra: default.yaml
 
 # Before training, the samples are run through PVNet. The samples are saved under this directory
-sample_save_dir: /home/jamesfulton/repos/PVNet_summation/presaved_saved_samples
+sample_save_dir: PLACEHOLDER
 
 work_dir: ${hydra:runtime.cwd}
 

configs.example/datamodule/default.yaml

@@ -2,8 +2,19 @@ configuration: PLACEHOLDER
 batch_size: 32
 num_workers: 20
 prefetch_factor: 2
-
+persistent_workers:  true
 
 # Set these to integers to limit the number of train/val samples precomputed and saved
 max_num_train_samples: null
-max_num_val_samples: null
+max_num_val_samples: null
+
+pvnet_model:
+    model_id: null
+    revision: null
+
+train_period:
+    - null
+    - '2021-12-31'
+val_period:
+    - '2022-01-01'
+    - '2022-12-31'

configs.example/model/default.yaml

@@ -1,31 +1,32 @@
-_target_: pvnet_summation.models.flat_model.DenseModel
+_target_: pvnet_summation.training.lightning_module.PVNetSummationLightningModule
 
-output_quantiles: null
+model:
+  _target_: pvnet_summation.models.dense_model.DenseModel
 
-model_name: "openclimatefix/pvnet_v2"
-model_version: "898630f3f8cd4e8506525d813dd61c6d8de86144"
+  output_quantiles: [0.02, 0.1, 0.25, 0.5, 0.75, 0.9, 0.98]
 
-#--------------------------------------------
-# Tabular network settings
-#--------------------------------------------
 
-output_network:
-  _target_: pvnet.models.late_fusion.linear_networks.networks.ResFCNet
-  _partial_: True
-  fc_hidden_features: 128
-  n_res_blocks: 2
-  res_block_layers: 2
-  dropout_frac: 0.0
+  #--------------------------------------------
+  # Summation network architecture
+  #--------------------------------------------
 
-predict_difference_from_sum: False
+  output_network:
+    _target_: pvnet.models.late_fusion.linear_networks.networks.ResFCNet
+    _partial_: True
+    fc_hidden_features: 128
+    n_res_blocks: 2
+    res_block_layers: 2
+    dropout_frac: 0.1
 
-# ----------------------------------------------
+  predict_difference_from_sum: True
+
+  # ----------------------------------------------
 
 optimizer:
   _target_: pvnet_summation.optimizers.AdamWReduceLROnPlateau
   lr: 0.0001
   weight_decay: 0.25
   amsgrad: True
-  patience: 20
+  patience: 10
   factor: 0.1
   threshold: 0.00