feat: add demo gallery images and dataset-dir fallback for examples

README.md

@@ -19,5 +19,6 @@ Live demo: [Hugging Face Space](https://huggingface.co/spaces/letitbE/mine2real)
 - `notebooks/lab5.ipynb` — training notebook
 - `checkpoints/cycle_gan_color_fix#0_epoch_10.pt` — trained last checkpoint used by the app
 - `data/` — not in git (Space repo size limit); download the dataset from the link in the Space / course materials, then extract so you have `data/datasets/` (and optionally `data/datasets.tar.gz`) locally
+- `demo_gallery/` — small fixed subset (10 images) shipped in git for the Streamlit example gallery when the full dataset is absent
 - `app.py` — streamlit app
 - `mine2real/` — model definition and inference utilities

app.py

@@ -113,19 +113,19 @@ def get_model():
 @st.cache_data(show_spinner=False)
 def load_example_gallery() -> dict[str, list[Path]]:
     return {
-        "minecraft": list_example_images(MINECRAFT_DIR, limit=8, seed=7),
-        "real": list_example_images(REAL_DIR, limit=8, seed=11),
+        "minecraft": list_example_images(MINECRAFT_DIR, limit=5, seed=7),
+        "real": list_example_images(REAL_DIR, limit=5, seed=11),
     }
 
 
 def render_result(result: TranslationResult, input_label: str, output_label: str) -> None:
     col1, col2, col3 = st.columns(3)
-    col1.image(result.source, caption=input_label, use_container_width=True)
-    col2.image(result.translated, caption=output_label, use_container_width=True)
+    col1.image(result.source, caption=input_label, width="stretch")
+    col2.image(result.translated, caption=output_label, width="stretch")
     col3.image(
         result.reconstructed,
         caption="Cycle reconstruction",
-        use_container_width=True,
+        width="stretch",
     )
 
 
@@ -209,7 +209,7 @@ def app() -> None:
         help="The app accepts exactly one image and runs a full CycleGAN pass with reconstruction.",
     )
 
-    run_button = st.button("Translate image", type="primary", use_container_width=True)
+    run_button = st.button("Translate image", type="primary", width="stretch")
 
     if uploaded is not None:
         image = Image.open(uploaded).convert("RGB")
@@ -243,7 +243,7 @@ def app() -> None:
                 """
             )
         else:
-            st.image(image, caption="Uploaded image", use_container_width=True)
+            st.image(image, caption="Uploaded image", width="stretch")
     else:
         st.info("Upload an image and click `Translate image` to run the model.")
 
@@ -253,7 +253,8 @@ def app() -> None:
         """
         - Notebook: `notebooks/lab5.ipynb`
         - Checkpoint: `checkpoints/cycle_gan_color_fix#0_epoch_10.pt`
-        - Dataset: not stored in this repo (HF Space storage limit); obtain the archive from the published link and extract to `data/datasets/` for local training and example galleries
+        - Dataset link: [Yandex Disk](https://disk.yandex.ru/d/N_G-t-oirnLynw)
+        - Example gallery in repo: `demo_gallery/` (5+5 images); full training data stays under `data/datasets/` when unpacked locally
         """
     )
 
@@ -265,7 +266,7 @@ def app() -> None:
     galleries = load_example_gallery()
     if not galleries["minecraft"] and not galleries["real"]:
         st.info(
-            "Example gallery is empty: add `data/datasets/` (see README) or the dataset from your course link."
+            "Example gallery is empty: ensure `demo_gallery/` is present in the app checkout, or restore `data/datasets/` from the Yandex Disk link above."
         )
     block_a, block_b = st.columns(2)
 
@@ -282,7 +283,7 @@ def app() -> None:
         st.write("")
         cols = st.columns(2)
         for idx, image_path in enumerate(galleries["minecraft"]):
-            cols[idx % 2].image(str(image_path), caption=image_path.name, use_container_width=True)
+            cols[idx % 2].image(str(image_path), caption=image_path.name, width="stretch")
 
     with block_b:
         st.markdown(
@@ -297,7 +298,7 @@ def app() -> None:
         st.write("")
         cols = st.columns(2)
         for idx, image_path in enumerate(galleries["real"]):
-            cols[idx % 2].image(str(image_path), caption=image_path.name, use_container_width=True)
+            cols[idx % 2].image(str(image_path), caption=image_path.name, width="stretch")
 
 
 if __name__ == "__main__":

mine2real/inference.py

@@ -7,15 +7,16 @@ import random
 import numpy as np
 from PIL import Image, ImageOps
 import torch
-from torchvision import transforms as T
 
 from mine2real.model import CycleGAN
 
 IMAGE_SIZE = 256
 CHECKPOINT_PATH = Path("checkpoints/cycle_gan_color_fix#0_epoch_10.pt")
-DATASET_ROOT = Path("data/datasets")
-MINECRAFT_DIR = DATASET_ROOT / "minecraft_forestlike"
-REAL_DIR = DATASET_ROOT / "real_nature_landscape"
+_ROOT = Path(__file__).resolve().parent.parent
+_full_mc = _ROOT / "data" / "datasets" / "minecraft_forestlike"
+_full_real = _ROOT / "data" / "datasets" / "real_nature_landscape"
+MINECRAFT_DIR = _full_mc if _full_mc.is_dir() else _ROOT / "demo_gallery" / "minecraft"
+REAL_DIR = _full_real if _full_real.is_dir() else _ROOT / "demo_gallery" / "real"
 
 
 @dataclass(frozen=True)
@@ -34,14 +35,11 @@ def get_device() -> torch.device:
     return torch.device("cpu")
 
 
-def build_transform(image_size: int = IMAGE_SIZE) -> T.Compose:
-    return T.Compose(
-        [
-            T.Resize((image_size, image_size)),
-            T.ToTensor(),
-            T.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-        ]
-    )
+def pil_to_normalized_tensor(image: Image.Image) -> torch.Tensor:
+    array = np.asarray(image, dtype=np.float32) / 255.0
+    tensor = torch.from_numpy(array).permute(2, 0, 1)
+    tensor = (tensor - 0.5) / 0.5
+    return tensor
 
 
 def tensor_to_pil(image_tensor: torch.Tensor) -> Image.Image:
@@ -55,7 +53,7 @@ def tensor_to_pil(image_tensor: torch.Tensor) -> Image.Image:
 def prepare_image(image: Image.Image, image_size: int = IMAGE_SIZE) -> tuple[Image.Image, torch.Tensor]:
     source = image.convert("RGB")
     source = ImageOps.fit(source, (image_size, image_size), method=Image.Resampling.LANCZOS)
-    tensor = build_transform(image_size)(source).unsqueeze(0)
+    tensor = pil_to_normalized_tensor(source).unsqueeze(0)
     return source, tensor
 
 

requirements.txt

@@ -1,4 +1,3 @@
 streamlit>=1.40,<2
 numpy>=1.26,<3
 pillow>=10,<12
-torchvision>=0.20,<0.23