butterfly-gan

app.py

@@ -1,7 +1,59 @@
 import streamlit as st
 
-st.title("Streamlit First Demo")
-st.markdown("This is a demo of streamlit")
+from utils import carga_modelo, genera
 
-x = st.slider('Selecciona un valor')
-st.write(x, ' al cuadrado es', x * x)
+## Página principal
+st.title("Butterfly GAN (GAN de mariposas)")
+st.write(
+    "Modelo Light-GAN entrenado con 1000 imágenes de mariposas tomadas de la colección del Museo Smithsonian."
+)
+
+## Barra lateral
+st.sidebar.subheader("¡Esta mariposa no existe! Ni en América Latina 🤯.")
+st.sidebar.image("assets/logo.png", width=200)
+st.sidebar.caption(
+    f"[Modelo](https://huggingface.co/ceyda/butterfly_cropped_uniq1K_512) y [Dataset](https://huggingface.co/datasets/huggan/smithsonian_butterflies_subset) usados."
+)
+st.sidebar.caption(f"*Disclaimers:*")
+st.sidebar.caption(
+    "* Este demo es una versión simplificada del creado por [Ceyda Cinarel](https://github.com/cceyda) y [Jonathan Whitaker](https://datasciencecastnet.home.blog/) ([link](https://huggingface.co/spaces/huggan/butterfly-gan)) durante el hackathon [HugGan](https://github.com/huggingface/community-events). Cualquier error se atribuye a [Omar Espejel](https://twitter.com/espejelomar)."
+)
+st.sidebar.caption(
+    "* Modelo basado en el [paper](https://openreview.net/forum?id=1Fqg133qRaI) *Towards Faster and Stabilized GAN Training for High-fidelity Few-shot Image Synthesis*."
+)
+
+## Cargamos modelo
+repo_id = "ceyda/butterfly_cropped_uniq1K_512"
+version_modelo = "57d36a15546909557d9f967f47713236c8288838"
+modelo_gan = carga_modelo(repo_id, version_modelo)
+
+## Generamos 4 mariposas
+n_mariposas = 4
+
+## Función que genera mariposas y lo guarda como un estado de la sesión
+def corre():
+    with st.spinner("Generando, espera un poco..."):
+        ims = genera(modelo_gan, n_mariposas)
+        st.session_state["ims"] = ims
+
+
+## Si no hay una imagen generada entonces generala
+if "ims" not in st.session_state:
+    st.session_state["ims"] = None
+    corre()
+
+## ims contiene las imágenes generadas
+ims = st.session_state["ims"]
+
+## Si la usuaria da click en el botón entonces corremos la función genera()
+corre_boton = st.button(
+    "Genera mariposas, porfa.",
+    on_click=corre,
+    help="Estamos en pleno vuelo, puede tardar.",
+)
+
+if ims is not None:
+    cols = st.columns(n_mariposas)
+    for j, im in enumerate(ims):
+        i = j % n_mariposas
+        cols[i].image(im, use_column_width=True)

requirements.txt

@@ -1 +1,3 @@
+git+https://github.com/huggingface/community-events.git@3fea10c5d5a50c69f509e34cd580fe9139905d04#egg=huggan
+transformers
 streamlit

utils.py

@@ -0,0 +1,78 @@
+import json
+
+import numpy as np
+import torch
+from huggan.pytorch.lightweight_gan.lightweight_gan import LightweightGAN
+from huggingface_hub import hf_hub_download
+
+CONFIG_NAME = "config.json"
+revision = None
+cache_dir = None
+force_download = False
+proxies = None
+resume_download = False
+local_files_only = False
+token = None
+
+
+def carga_modelo(model_name="ceyda/butterfly_cropped_uniq1K_512"):
+    """
+    Loads a pre-trained LightweightGAN model from Hugging Face Model Hub.
+
+    Args:
+        model_name (str): The name of the pre-trained model to load. Defaults to "ceyda/butterfly_cropped_uniq1K_512".
+        model_version (str): The version of the pre-trained model to load. Defaults to None.
+
+    Returns:
+        LightweightGAN: The loaded pre-trained model.
+    """
+    # Load the config
+    config_file = hf_hub_download(
+        repo_id=str(model_name),
+        filename=CONFIG_NAME,
+        revision=revision,
+        cache_dir=cache_dir,
+        force_download=force_download,
+        proxies=proxies,
+        resume_download=resume_download,
+        token=token,
+        local_files_only=local_files_only,
+    )
+    with open(config_file, "r", encoding="utf-8") as f:
+        config = json.load(f)
+
+    # Call the _from_pretrained with all the needed arguments
+    gan = LightweightGAN(latent_dim=256, image_size=512)
+
+    gan = gan._from_pretrained(
+        model_id=str(model_name),
+        revision=revision,
+        cache_dir=cache_dir,
+        force_download=force_download,
+        proxies=proxies,
+        resume_download=resume_download,
+        local_files_only=local_files_only,
+        token=token,
+        use_auth_token=False,
+        config=config,  # usually in **model_kwargs
+    )
+
+    gan.eval()
+    return gan
+
+
+def genera(gan, batch_size=1):
+    """
+    Generates images using the given GAN model.
+
+    Args:
+        gan (nn.Module): The GAN model to use for generating images.
+        batch_size (int, optional): The number of images to generate in each batch. Defaults to 1.
+
+    Returns:
+        numpy.ndarray: A numpy array of generated images.
+    """
+    with torch.no_grad():
+        ims = gan.G(torch.randn(batch_size, gan.latent_dim)).clamp_(0.0, 1.0) * 255
+        ims = ims.permute(0, 2, 3, 1).detach().cpu().numpy().astype(np.uint8)
+    return ims