PVNet 4.1.8__tar.gz → 4.1.10__tar.gz

{pvnet-4.1.8 → pvnet-4.1.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: PVNet
-Version: 4.1.8
+Version: 4.1.10
 Summary: PVNet
 Author-email: Peter Dudfield <info@openclimatefix.org>
 Description-Content-Type: text/markdown

{pvnet-4.1.8 → pvnet-4.1.10}/PVNet.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: PVNet
-Version: 4.1.8
+Version: 4.1.10
 Summary: PVNet
 Author-email: Peter Dudfield <info@openclimatefix.org>
 Description-Content-Type: text/markdown

{pvnet-4.1.8 → pvnet-4.1.10}/PVNet.egg-info/SOURCES.txt

@@ -30,7 +30,6 @@ pvnet/models/model_cards/wind_india_model_card_template.md
 pvnet/models/multimodal/__init__.py
 pvnet/models/multimodal/basic_blocks.py
 pvnet/models/multimodal/multimodal.py
-pvnet/models/multimodal/multimodal_base.py
 pvnet/models/multimodal/unimodal_teacher.py
 pvnet/models/multimodal/encoders/__init__.py
 pvnet/models/multimodal/encoders/basic_blocks.py

pvnet-4.1.10/pvnet/__init__.py

@@ -0,0 +1,2 @@
+"""PVNet"""
+__version__ = "4.1.10"

{pvnet-4.1.8 → pvnet-4.1.10}/pvnet/models/base_model.py

@@ -1,5 +1,5 @@
 """Base model for all PVNet submodels"""
-
+import copy
 import json
 import logging
 import os
@@ -22,6 +22,7 @@ from huggingface_hub.constants import CONFIG_NAME, PYTORCH_WEIGHTS_NAME
 from huggingface_hub.file_download import hf_hub_download
 from huggingface_hub.hf_api import HfApi
 from ocf_data_sampler.torch_datasets.sample.base import copy_batch_to_device
+from torchvision.transforms.functional import center_crop
 
 from pvnet.models.utils import (
     BatchAccumulator,
@@ -535,6 +536,51 @@ class BaseModel(pl.LightningModule, PVNetModelHubMixin):
         # save all validation results to array, so we can save these to weights n biases
         self.validation_epoch_results = []
 
+    def _adapt_batch(self, batch):
+        """Slice batches into appropriate shapes for model.
+
+        Returns a new batch dictionary with adapted data, leaving the original batch unchanged.
+        We make some specific assumptions about the original batch and the derived sliced batch:
+        - We are only limiting the future projections. I.e. we are never shrinking the batch from
+          the left hand side of the time axis, only slicing it from the right
+        - We are only shrinking the spatial crop of the satellite and NWP data
+
+        """
+        # Create a copy of the batch to avoid modifying the original
+        new_batch = {key: copy.deepcopy(value) for key, value in batch.items()}
+
+        if "gsp" in new_batch.keys():
+            # Slice off the end of the GSP data
+            gsp_len = self.forecast_len + self.history_len + 1
+            new_batch["gsp"] = new_batch["gsp"][:, :gsp_len]
+            new_batch["gsp_time_utc"] = new_batch["gsp_time_utc"][:, :gsp_len]
+
+        if self.include_sat:
+            # Slice off the end of the satellite data and spatially crop
+            # Shape: batch_size, seq_length, channel, height, width
+            new_batch["satellite_actual"] = center_crop(
+                new_batch["satellite_actual"][:, : self.sat_sequence_len],
+                output_size=self.sat_encoder.image_size_pixels,
+            )
+
+        if self.include_nwp:
+            # Slice off the end of the NWP data and spatially crop
+            for nwp_source in self.nwp_encoders_dict:
+                # shape: batch_size, seq_len, n_chans, height, width
+                new_batch["nwp"][nwp_source]["nwp"] = center_crop(
+                    new_batch["nwp"][nwp_source]["nwp"],
+                    output_size=self.nwp_encoders_dict[nwp_source].image_size_pixels,
+                )[:, : self.nwp_encoders_dict[nwp_source].sequence_length]
+
+        if self.include_sun:
+            sun_len = self.forecast_len + self.history_len + 1
+            # Slice off end of solar coords
+            for s in ["solar_azimuth", "solar_elevation"]:
+                if s in new_batch.keys():
+                    new_batch[s] = new_batch[s][:, :sun_len]
+
+        return new_batch
+
     def transfer_batch_to_device(self, batch, device, dataloader_idx):
         """Method to move custom batches to a given device"""
         return copy_batch_to_device(batch, device)
@@ -693,6 +739,10 @@ class BaseModel(pl.LightningModule, PVNetModelHubMixin):
     def training_step(self, batch, batch_idx):
         """Run training step"""
         y_hat = self(batch)
+
+        # Batch is adapted in the model forward method, but needs to be adapted here too
+        batch = self._adapt_batch(batch)
+
         y = batch[self._target_key][:, -self.forecast_len :]
 
         losses = self._calculate_common_losses(y, y_hat)
@@ -774,6 +824,8 @@ class BaseModel(pl.LightningModule, PVNetModelHubMixin):
         accum_batch_num = batch_idx // self.trainer.accumulate_grad_batches
 
         y_hat = self(batch)
+        # Batch is adapted in the model forward method, but needs to be adapted here too
+        batch = self._adapt_batch(batch)
 
         y = batch[self._target_key][:, -self.forecast_len :]
 

{pvnet-4.1.8 → pvnet-4.1.10}/pvnet/models/multimodal/multimodal.py

@@ -8,15 +8,15 @@ from omegaconf import DictConfig
 from torch import nn
 
 import pvnet
+from pvnet.models.base_model import BaseModel
 from pvnet.models.multimodal.basic_blocks import ImageEmbedding
 from pvnet.models.multimodal.encoders.basic_blocks import AbstractNWPSatelliteEncoder
 from pvnet.models.multimodal.linear_networks.basic_blocks import AbstractLinearNetwork
-from pvnet.models.multimodal.multimodal_base import MultimodalBaseModel
 from pvnet.models.multimodal.site_encoders.basic_blocks import AbstractSitesEncoder
 from pvnet.optimizers import AbstractOptimizer
 
 
-class Model(MultimodalBaseModel):
+class Model(BaseModel):
     """Neural network which combines information from different sources
 
     Architecture is roughly as follows:
@@ -152,7 +152,7 @@ class Model(MultimodalBaseModel):
         self.use_id_embedding = location_id_mapping is not None
 
         if self.use_id_embedding:
-            num_embeddings = max(location_id_mapping.values())
+            num_embeddings = max(location_id_mapping.values()) + 1
 
         super().__init__(
             history_minutes=history_minutes,

{pvnet-4.1.8 → pvnet-4.1.10}/pvnet/models/multimodal/unimodal_teacher.py

@@ -11,12 +11,12 @@ from pyaml_env import parse_config
 from torch import nn
 
 import pvnet
+from pvnet.models.base_model import BaseModel
 from pvnet.models.multimodal.linear_networks.basic_blocks import AbstractLinearNetwork
-from pvnet.models.multimodal.multimodal_base import MultimodalBaseModel
 from pvnet.optimizers import AbstractOptimizer
 
 
-class Model(MultimodalBaseModel):
+class Model(BaseModel):
     """Neural network which combines information from different sources
 
     The network is trained via unimodal teachers [1].
@@ -106,7 +106,7 @@ class Model(MultimodalBaseModel):
         self.use_id_embedding = location_id_mapping is not None
 
         if self.use_id_embedding:
-            num_embeddings = max(location_id_mapping.values())
+            num_embeddings = max(location_id_mapping.values()) + 1
 
         # This is set but modified later based on the teachers
         self.add_image_embedding_channel = False

pvnet-4.1.8/pvnet/__init__.py

@@ -1,2 +0,0 @@
-"""PVNet"""
-__version__ = "4.1.8"

pvnet-4.1.8/pvnet/models/multimodal/multimodal_base.py

@@ -1,55 +0,0 @@
-"""Base model class for multimodal model and unimodal teacher"""
-import copy
-
-from torchvision.transforms.functional import center_crop
-
-from pvnet.models.base_model import BaseModel
-
-
-class MultimodalBaseModel(BaseModel):
-    """Base model class for multimodal model and unimodal teacher"""
-
-    def _adapt_batch(self, batch):
-        """Slice batches into appropriate shapes for model.
-
-        Returns a new batch dictionary with adapted data, leaving the original batch unchanged.
-        We make some specific assumptions about the original batch and the derived sliced batch:
-        - We are only limiting the future projections. I.e. we are never shrinking the batch from
-          the left hand side of the time axis, only slicing it from the right
-        - We are only shrinking the spatial crop of the satellite and NWP data
-
-        """
-        # Create a copy of the batch to avoid modifying the original
-        new_batch = {key: copy.deepcopy(value) for key, value in batch.items()}
-
-        if "gsp" in new_batch.keys():
-            # Slice off the end of the GSP data
-            gsp_len = self.forecast_len + self.history_len + 1
-            new_batch["gsp"] = new_batch["gsp"][:, :gsp_len]
-            new_batch["gsp_time_utc"] = new_batch["gsp_time_utc"][:, :gsp_len]
-
-        if self.include_sat:
-            # Slice off the end of the satellite data and spatially crop
-            # Shape: batch_size, seq_length, channel, height, width
-            new_batch["satellite_actual"] = center_crop(
-                new_batch["satellite_actual"][:, : self.sat_sequence_len],
-                output_size=self.sat_encoder.image_size_pixels,
-            )
-
-        if self.include_nwp:
-            # Slice off the end of the NWP data and spatially crop
-            for nwp_source in self.nwp_encoders_dict:
-                # shape: batch_size, seq_len, n_chans, height, width
-                new_batch["nwp"][nwp_source]["nwp"] = center_crop(
-                    new_batch["nwp"][nwp_source]["nwp"],
-                    output_size=self.nwp_encoders_dict[nwp_source].image_size_pixels,
-                )[:, : self.nwp_encoders_dict[nwp_source].sequence_length]
-
-        if self.include_sun:
-            sun_len = self.forecast_len + self.history_len + 1
-            # Slice off end of solar coords
-            for s in ["solar_azimuth", "solar_elevation"]:
-                if s in new_batch.keys():
-                    new_batch[s] = new_batch[s][:, :sun_len]
-
-        return new_batch