Update README.md
Browse files
README.md
CHANGED
|
@@ -14,5 +14,76 @@ max_batch_size: 64
|
|
| 14 |
merge_mode: weighted_average
|
| 15 |
---
|
| 16 |
|
| 17 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
|
|
|
|
| 14 |
merge_mode: weighted_average
|
| 15 |
---
|
| 16 |
|
| 17 |
+
First run the following to setup the environment and get the official model code
|
| 18 |
+
|
| 19 |
+
```bash
|
| 20 |
+
# Clone the official repo
|
| 21 |
+
git clone git@github.com:facebookresearch/HighResCanopyHeight.git
|
| 22 |
+
|
| 23 |
+
# Install dependencies
|
| 24 |
+
pip install stac-model[torch]
|
| 25 |
+
|
| 26 |
+
# Download the official pretrained checkpoints
|
| 27 |
+
mkdir checkpoints && aws s3 --no-sign-request sync s3://dataforgood-fb-data/forests/v1/models/saved_checkpoints/ checkpoints/
|
| 28 |
+
```
|
| 29 |
+
|
| 30 |
+
Export the model using the following:
|
| 31 |
+
|
| 32 |
+
```python
|
| 33 |
+
from pathlib import Path
|
| 34 |
+
import sys
|
| 35 |
+
sys.path.append("HighResCanopyHeight")
|
| 36 |
+
|
| 37 |
+
import torch
|
| 38 |
+
import torch.nn as nn
|
| 39 |
+
import torchvision.transforms.v2 as T
|
| 40 |
+
from stac_model.torch.export import export, package
|
| 41 |
+
|
| 42 |
+
import src.transforms
|
| 43 |
+
from inference import SSLAE
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
# Create model and load checkpoint
|
| 47 |
+
class TreeCanopyHeightModel(nn.Module):
|
| 48 |
+
def __init__(self, classify=True, huge=True):
|
| 49 |
+
super().__init__()
|
| 50 |
+
self.model = SSLAE(pretrained=None, classify=classify, huge=huge, n_bins=256)
|
| 51 |
+
|
| 52 |
+
def forward(self, x):
|
| 53 |
+
outputs = self.model(x)
|
| 54 |
+
pred = 10 * outputs + 0.001
|
| 55 |
+
return pred.relu()
|
| 56 |
+
|
| 57 |
+
path = "checkpoints/SSLhuge_satellite.pth"
|
| 58 |
+
ckpt = torch.load(path, map_location="cpu", weights_only=False)
|
| 59 |
+
state_dict = {f"model.{k}": v for k, v in ckpt["state_dict"].items()}
|
| 60 |
+
model = TreeCanopyHeightModel()
|
| 61 |
+
model.load_state_dict(state_dict)
|
| 62 |
+
|
| 63 |
+
# Create exportable transforms
|
| 64 |
+
original_transform = src.transforms.SSLNorm().Trans
|
| 65 |
+
norm = original_transform.transforms[-1]
|
| 66 |
+
|
| 67 |
+
transforms = nn.Sequential(
|
| 68 |
+
T.Normalize(mean=[0], std=[255]), # replace ToTensor() with normalize to 0-1
|
| 69 |
+
T.Normalize(mean=norm.mean, std=norm.std)
|
| 70 |
+
)
|
| 71 |
+
|
| 72 |
+
# Export and save to pt2
|
| 73 |
+
model_program, transforms_program = export(
|
| 74 |
+
input_shape=[-1, 3, 224, 224],
|
| 75 |
+
model=model,
|
| 76 |
+
transforms=transforms,
|
| 77 |
+
device="cpu",
|
| 78 |
+
dtype=torch.float32,
|
| 79 |
+
)
|
| 80 |
+
package(
|
| 81 |
+
output_file=Path("model.pt2"),
|
| 82 |
+
model_program=model_program,
|
| 83 |
+
transforms_program=transforms_program,
|
| 84 |
+
metadata_properties=None,
|
| 85 |
+
aoti_compile_and_package=False
|
| 86 |
+
)
|
| 87 |
+
```
|
| 88 |
+
|
| 89 |
|