Upload model.py with huggingface_hub
Browse files
model.py
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
model.py — helpers to load BDAPPV baseline models from HuggingFace Hub.
|
| 3 |
+
|
| 4 |
+
Usage:
|
| 5 |
+
from huggingface_hub import hf_hub_download
|
| 6 |
+
import importlib.util, sys
|
| 7 |
+
|
| 8 |
+
path = hf_hub_download("gabrielkasmi/bdappv-models", "model.py")
|
| 9 |
+
spec = importlib.util.spec_from_file_location("bdappv_model", path)
|
| 10 |
+
mod = importlib.util.module_from_spec(spec); spec.loader.exec_module(mod)
|
| 11 |
+
|
| 12 |
+
seg = mod.load_segmentation_model("google") # or "ign"
|
| 13 |
+
clf = mod.load_classification_model("google") # or "ign"
|
| 14 |
+
"""
|
| 15 |
+
|
| 16 |
+
import torch
|
| 17 |
+
import torch.nn as nn
|
| 18 |
+
from huggingface_hub import hf_hub_download
|
| 19 |
+
from torchvision.models.segmentation import deeplabv3_resnet101
|
| 20 |
+
from torchvision.models import inception_v3
|
| 21 |
+
|
| 22 |
+
REPO = "gabrielkasmi/bdappv-models"
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
def load_segmentation_model(provider: str = "google", device: str = "cpu"):
|
| 26 |
+
"""
|
| 27 |
+
Load the DeepLabV3-ResNet101 segmentation model.
|
| 28 |
+
|
| 29 |
+
Args:
|
| 30 |
+
provider : "google" or "ign"
|
| 31 |
+
device : "cpu", "cuda", or "mps"
|
| 32 |
+
|
| 33 |
+
Returns:
|
| 34 |
+
model in eval mode
|
| 35 |
+
"""
|
| 36 |
+
assert provider in ("google", "ign"), "provider must be 'google' or 'ign'"
|
| 37 |
+
|
| 38 |
+
path = hf_hub_download(REPO, f"deeplab_{provider}_best.pth")
|
| 39 |
+
model = deeplabv3_resnet101(weights=None, aux_loss=False)
|
| 40 |
+
model.classifier[-1] = nn.Conv2d(256, 1, kernel_size=1)
|
| 41 |
+
|
| 42 |
+
state = torch.load(path, map_location=device, weights_only=False)
|
| 43 |
+
model_dict = model.state_dict()
|
| 44 |
+
compatible = {k: v for k, v in state.items()
|
| 45 |
+
if k in model_dict and v.shape == model_dict[k].shape}
|
| 46 |
+
model_dict.update(compatible)
|
| 47 |
+
model.load_state_dict(model_dict)
|
| 48 |
+
|
| 49 |
+
return model.eval().to(device)
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
def load_classification_model(provider: str = "google", device: str = "cpu"):
|
| 53 |
+
"""
|
| 54 |
+
Load the InceptionV3 classification model (panel / no panel).
|
| 55 |
+
|
| 56 |
+
Args:
|
| 57 |
+
provider : "google" or "ign"
|
| 58 |
+
device : "cpu", "cuda", or "mps"
|
| 59 |
+
|
| 60 |
+
Returns:
|
| 61 |
+
model in eval mode
|
| 62 |
+
"""
|
| 63 |
+
assert provider in ("google", "ign"), "provider must be 'google' or 'ign'"
|
| 64 |
+
|
| 65 |
+
path = hf_hub_download(REPO, f"inception_{provider}_best.pth")
|
| 66 |
+
model = inception_v3(weights=None, aux_logits=False)
|
| 67 |
+
model.fc = nn.Linear(model.fc.in_features, 1)
|
| 68 |
+
|
| 69 |
+
state = torch.load(path, map_location=device, weights_only=False)
|
| 70 |
+
model.load_state_dict(state)
|
| 71 |
+
|
| 72 |
+
return model.eval().to(device)
|