Sampler seems to work

high_order_implicit_representation/rendering.py

@@ -255,16 +255,17 @@ def __init__(self, filenames, batch_size):
     def on_train_epoch_end(
         self, trainer: pl.Trainer, pl_module: pl.LightningModule
     ) -> None:
+        self._dataset.reset()
         pl_module.eval()
         with torch.no_grad():
-            print("We are calling this")
             image_count=0
             for caption_embedding, flattened_position, image in self._dataloader:
+                image_count+=1
+                print('image_count', image_count)
                 flattened_position=flattened_position[0]
                 size = len(flattened_position)
                 y_hat_list = []
                 for i in range(0, size, self._batch_size):
-
 
                     embed_single = caption_embedding.to(pl_module.device)
 
@@ -284,14 +285,15 @@ def on_train_epoch_end(
                     image.shape[1], image.shape[2], 3
                 )
 
-                print('ans.shape', ans.shape)
+                #ans = ans.permute(2,0,1)
+                ans = ans.squeeze(0)
 
                 ans = 0.5 * (ans + 1.0)
 
                 f, axarr = plt.subplots(1, 2)
                 axarr[0].imshow(ans.detach().cpu().numpy())
                 axarr[0].set_title("fit")
-                axarr[1].imshow(image.cpu())
+                axarr[1].imshow(image.squeeze(0).cpu())
                 axarr[1].set_title("original")
 
                 for i in range(2):

high_order_implicit_representation/single_image_dataset.py

@@ -402,25 +402,26 @@ def __init__(self, filenames: List[str]):
         super().__init__()
         self.dataset = PickAPic(files=filenames)
         self.sentence_model = SentenceTransformer("all-MiniLM-L6-v2")
-
+        self.reset()
+
     def reset(self) :
-        self.dataset.reset()
+        self.generator = self.gen_data()
 
     def __len__(self):
         return int(1e12)
 
     def gen_data(self):
 
-        caption, image = next(self.dataset())
-        caption_embedding = self.sentence_model.encode(caption)
+        for caption, image in self.dataset():
+            caption_embedding = self.sentence_model.encode(caption)
 
-        flattened_image, flattened_position, image = simple_image_to_dataset(image)
-        return caption_embedding, flattened_position, image
+            flattened_image, flattened_position, image = simple_image_to_dataset(image)
+            yield caption_embedding, flattened_position, image
 
 
     def __getitem__(self, idx):
 
-        return self.gen_data()
+        return next(self.generator)
 
 
 class Text2ImageDataModule(LightningDataModule):

tests/test_rendering.py

@@ -2,10 +2,11 @@
 from high_order_implicit_representation.rendering import (
     neighborhood_sample_generator,
     NeighborGenerator,
+    Text2ImageSampler
 )
 import torch
 from omegaconf import DictConfig
-from high_order_implicit_representation.networks import Net
+from high_order_implicit_representation.networks import Net, GenNet
 from pytorch_lightning import Trainer
 
 
@@ -133,3 +134,52 @@ def test_neighbor_generator():
         samples=2, frames=2, output_size=[64, 64], width=3, outside=3
     )
     generator.on_train_epoch_end(trainer=trainer, pl_module=model)
+
+
+def test_text2image_sampler() :
+    width = 3
+    outside = 3
+
+    input_features = ((width + 2 * outside) * (width + 2 * outside) - width * width) * 3
+    output_size = width * width * 3
+
+    cfg = DictConfig(
+        content={
+            "max_epochs": 1,
+            "gpus": 0,
+            "lr": 1e-4,
+            "batch_size": 16,
+            "segments": 2,
+            "optimizer": {
+                "name": "adam",
+                "lr": 1.0e-3,
+                "scheduler": "plateau",
+                "patience": 10,
+                "factor": 0.1,
+            },
+            "mlp": {
+                "layers": 2,
+                "segments": 2,
+                "scale": 2.0,
+                "width": 4,
+                "periodicity": None,
+                "rescale_output": False,
+            },
+            "embedding_size": 384,
+            "input_size": 2,
+            "output_size": 3,
+            "input_segments": 20,
+            "layer_type": "continuous",
+            "n": 3,
+        }
+    )
+
+    model = GenNet(cfg)
+    trainer = Trainer(
+        max_epochs=cfg.max_epochs,
+        accelerator='cpu',
+    )
+
+    # Just make sure this runs
+    generator = Text2ImageSampler(filenames=["test_data/test.parquet"], batch_size=2000)
+    generator.on_train_epoch_end(trainer=trainer, pl_module=model)