Commit
Β·
d668197
1
Parent(s):
5089ae8
Update app
Browse files- app.py +117 -316
- HLS.L30.T13REN.2018013T172747.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T13REN.2018013T172747.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif +0 -0
- HLS.L30.T13REN.2018029T172738.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T13REN.2018029T172738.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif +0 -0
- HLS.L30.T13REN.2018061T172724.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T13REN.2018061T172724.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif +0 -0
- HLS.L30.T17RMP.2018004T155509.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T17RMP.2018004T155509.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif +0 -0
- HLS.L30.T17RMP.2018036T155452.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T17RMP.2018036T155452.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif +0 -0
- HLS.L30.T17RMP.2018068T155438.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T17RMP.2018068T155438.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif +0 -0
- HLS.L30.T18TVL.2018029T154533.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T18TVL.2018029T154533.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif +0 -0
- HLS.L30.T18TVL.2018141T154435.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T18TVL.2018141T154435.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif +0 -0
- HLS.L30.T18TVL.2018189T154446.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T18TVL.2018189T154446.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif +0 -0
app.py
CHANGED
|
@@ -1,215 +1,28 @@
|
|
| 1 |
-
#### pull files from hub
|
| 2 |
-
from huggingface_hub import hf_hub_download
|
| 3 |
-
import os
|
| 4 |
-
yaml_file_path=hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-100M", filename="Prithvi_100M_config.yaml", token=os.environ.get("token"))
|
| 5 |
-
checkpoint=hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-100M", filename='Prithvi_100M.pt', token=os.environ.get("token"))
|
| 6 |
-
model_def=hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-100M", filename='Prithvi.py', token=os.environ.get("token"))
|
| 7 |
-
os.system(f'cp {model_def} .')
|
| 8 |
-
#####
|
| 9 |
-
import argparse
|
| 10 |
-
import functools
|
| 11 |
-
import os
|
| 12 |
-
from typing import List
|
| 13 |
|
| 14 |
-
import
|
| 15 |
-
import rasterio
|
| 16 |
import torch
|
| 17 |
import yaml
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
from Prithvi import MaskedAutoencoderViT
|
| 21 |
import gradio as gr
|
|
|
|
| 22 |
from functools import partial
|
|
|
|
| 23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
PERCENTILES = (0.1, 99.9)
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
def process_channel_group(orig_img, new_img, channels, data_mean, data_std):
|
| 31 |
-
""" Process *orig_img* and *new_img* for RGB visualization. Each band is rescaled back to the
|
| 32 |
-
original range using *data_mean* and *data_std* and then lowest and highest percentiles are
|
| 33 |
-
removed to enhance contrast. Data is rescaled to (0, 1) range and stacked channels_first.
|
| 34 |
-
Args:
|
| 35 |
-
orig_img: torch.Tensor representing original image (reference) with shape = (bands, H, W).
|
| 36 |
-
new_img: torch.Tensor representing image with shape = (bands, H, W).
|
| 37 |
-
channels: list of indices representing RGB channels.
|
| 38 |
-
data_mean: list of mean values for each band.
|
| 39 |
-
data_std: list of std values for each band.
|
| 40 |
-
Returns:
|
| 41 |
-
torch.Tensor with shape (num_channels, height, width) for original image
|
| 42 |
-
torch.Tensor with shape (num_channels, height, width) for the other image
|
| 43 |
-
"""
|
| 44 |
-
|
| 45 |
-
stack_c = [], []
|
| 46 |
-
|
| 47 |
-
for c in channels:
|
| 48 |
-
orig_ch = orig_img[c, ...]
|
| 49 |
-
valid_mask = torch.ones_like(orig_ch, dtype=torch.bool)
|
| 50 |
-
valid_mask[orig_ch == NO_DATA_FLOAT] = False
|
| 51 |
-
|
| 52 |
-
# Back to original data range
|
| 53 |
-
orig_ch = (orig_ch * data_std[c]) + data_mean[c]
|
| 54 |
-
new_ch = (new_img[c, ...] * data_std[c]) + data_mean[c]
|
| 55 |
-
|
| 56 |
-
# Rescale (enhancing contrast)
|
| 57 |
-
min_value, max_value = np.percentile(orig_ch[valid_mask], PERCENTILES)
|
| 58 |
-
|
| 59 |
-
orig_ch = torch.clamp((orig_ch - min_value) / (max_value - min_value), 0, 1)
|
| 60 |
-
new_ch = torch.clamp((new_ch - min_value) / (max_value - min_value), 0, 1)
|
| 61 |
-
|
| 62 |
-
# No data as zeros
|
| 63 |
-
orig_ch[~valid_mask] = 0
|
| 64 |
-
new_ch[~valid_mask] = 0
|
| 65 |
-
|
| 66 |
-
stack_c[0].append(orig_ch)
|
| 67 |
-
stack_c[1].append(new_ch)
|
| 68 |
-
|
| 69 |
-
# Channels first
|
| 70 |
-
stack_orig = torch.stack(stack_c[0], dim=0)
|
| 71 |
-
stack_rec = torch.stack(stack_c[1], dim=0)
|
| 72 |
-
|
| 73 |
-
return stack_orig, stack_rec
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
def read_geotiff(file_path: str):
|
| 77 |
-
""" Read all bands from *file_path* and returns image + meta info.
|
| 78 |
-
Args:
|
| 79 |
-
file_path: path to image file.
|
| 80 |
-
Returns:
|
| 81 |
-
np.ndarray with shape (bands, height, width)
|
| 82 |
-
meta info dict
|
| 83 |
-
"""
|
| 84 |
-
|
| 85 |
-
with rasterio.open(file_path) as src:
|
| 86 |
-
img = src.read()
|
| 87 |
-
meta = src.meta
|
| 88 |
-
|
| 89 |
-
return img, meta
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
def save_geotiff(image, output_path: str, meta: dict):
|
| 93 |
-
""" Save multi-band image in Geotiff file.
|
| 94 |
-
Args:
|
| 95 |
-
image: np.ndarray with shape (bands, height, width)
|
| 96 |
-
output_path: path where to save the image
|
| 97 |
-
meta: dict with meta info.
|
| 98 |
-
"""
|
| 99 |
-
|
| 100 |
-
with rasterio.open(output_path, "w", **meta) as dest:
|
| 101 |
-
for i in range(image.shape[0]):
|
| 102 |
-
dest.write(image[i, :, :], i + 1)
|
| 103 |
-
|
| 104 |
-
return
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
def _convert_np_uint8(float_image: torch.Tensor):
|
| 108 |
-
|
| 109 |
-
image = float_image.numpy() * 255.0
|
| 110 |
-
image = image.astype(dtype=np.uint8)
|
| 111 |
-
image = image.transpose((1, 2, 0))
|
| 112 |
-
|
| 113 |
-
return image
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
def load_example(file_paths: List[str], mean: List[float], std: List[float]):
|
| 117 |
-
""" Build an input example by loading images in *file_paths*.
|
| 118 |
-
Args:
|
| 119 |
-
file_paths: list of file paths .
|
| 120 |
-
mean: list containing mean values for each band in the images in *file_paths*.
|
| 121 |
-
std: list containing std values for each band in the images in *file_paths*.
|
| 122 |
-
Returns:
|
| 123 |
-
np.array containing created example
|
| 124 |
-
list of meta info for each image in *file_paths*
|
| 125 |
-
"""
|
| 126 |
-
|
| 127 |
-
imgs = []
|
| 128 |
-
metas = []
|
| 129 |
-
|
| 130 |
-
for file in file_paths:
|
| 131 |
-
img, meta = read_geotiff(file)
|
| 132 |
-
img = img[:6]*10000 if img[:6].mean() <= 2 else img[:6]
|
| 133 |
-
|
| 134 |
-
# Rescaling (don't normalize on nodata)
|
| 135 |
-
img = np.moveaxis(img, 0, -1) # channels last for rescaling
|
| 136 |
-
img = np.where(img == NO_DATA, NO_DATA_FLOAT, (img - mean) / std)
|
| 137 |
-
|
| 138 |
-
imgs.append(img)
|
| 139 |
-
metas.append(meta)
|
| 140 |
-
|
| 141 |
-
imgs = np.stack(imgs, axis=0) # num_frames, H, W, C
|
| 142 |
-
imgs = np.moveaxis(imgs, -1, 0).astype('float32') # C, num_frames, H, W
|
| 143 |
-
imgs = np.expand_dims(imgs, axis=0) # add batch dim
|
| 144 |
-
|
| 145 |
-
return imgs, metas
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
def run_model(model: torch.nn.Module, input_data: torch.Tensor, mask_ratio: float, device: torch.device):
|
| 149 |
-
""" Run *model* with *input_data* and create images from output tokens (mask, reconstructed + visible).
|
| 150 |
-
Args:
|
| 151 |
-
model: MAE model to run.
|
| 152 |
-
input_data: torch.Tensor with shape (B, C, T, H, W).
|
| 153 |
-
mask_ratio: mask ratio to use.
|
| 154 |
-
device: device where model should run.
|
| 155 |
-
Returns:
|
| 156 |
-
3 torch.Tensor with shape (B, C, T, H, W).
|
| 157 |
-
"""
|
| 158 |
-
|
| 159 |
-
with torch.no_grad():
|
| 160 |
-
x = input_data.to(device)
|
| 161 |
-
|
| 162 |
-
_, pred, mask = model(x, mask_ratio)
|
| 163 |
-
|
| 164 |
-
# Create mask and prediction images (un-patchify)
|
| 165 |
-
mask_img = model.unpatchify(mask.unsqueeze(-1).repeat(1, 1, pred.shape[-1])).detach().cpu()
|
| 166 |
-
pred_img = model.unpatchify(pred).detach().cpu()
|
| 167 |
-
|
| 168 |
-
# Mix visible and predicted patches
|
| 169 |
-
rec_img = input_data.clone()
|
| 170 |
-
rec_img[mask_img == 1] = pred_img[mask_img == 1] # binary mask: 0 is keep, 1 is remove
|
| 171 |
-
|
| 172 |
-
# Switch zeros/ones in mask images so masked patches appear darker in plots (better visualization)
|
| 173 |
-
mask_img = (~(mask_img.to(torch.bool))).to(torch.float)
|
| 174 |
-
|
| 175 |
-
return rec_img, mask_img
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
def save_rgb_imgs(input_img, rec_img, mask_img, channels, mean, std, output_dir, meta_data):
|
| 179 |
-
""" Wrapper function to save Geotiff images (original, reconstructed, masked) per timestamp.
|
| 180 |
-
Args:
|
| 181 |
-
input_img: input torch.Tensor with shape (C, T, H, W).
|
| 182 |
-
rec_img: reconstructed torch.Tensor with shape (C, T, H, W).
|
| 183 |
-
mask_img: mask torch.Tensor with shape (C, T, H, W).
|
| 184 |
-
channels: list of indices representing RGB channels.
|
| 185 |
-
mean: list of mean values for each band.
|
| 186 |
-
std: list of std values for each band.
|
| 187 |
-
output_dir: directory where to save outputs.
|
| 188 |
-
meta_data: list of dicts with geotiff meta info.
|
| 189 |
-
"""
|
| 190 |
-
|
| 191 |
-
for t in range(input_img.shape[1]):
|
| 192 |
-
rgb_orig, rgb_pred = process_channel_group(orig_img=input_img[:, t, :, :],
|
| 193 |
-
new_img=rec_img[:, t, :, :],
|
| 194 |
-
channels=channels, data_mean=mean,
|
| 195 |
-
data_std=std)
|
| 196 |
-
|
| 197 |
-
rgb_mask = mask_img[channels, t, :, :] * rgb_orig
|
| 198 |
-
|
| 199 |
-
# Saving images
|
| 200 |
-
|
| 201 |
-
save_geotiff(image=_convert_np_uint8(rgb_orig),
|
| 202 |
-
output_path=os.path.join(output_dir, f"original_rgb_t{t}.tiff"),
|
| 203 |
-
meta=meta_data[t])
|
| 204 |
-
|
| 205 |
-
save_geotiff(image=_convert_np_uint8(rgb_pred),
|
| 206 |
-
output_path=os.path.join(output_dir, f"predicted_rgb_t{t}.tiff"),
|
| 207 |
-
meta=meta_data[t])
|
| 208 |
-
|
| 209 |
-
save_geotiff(image=_convert_np_uint8(rgb_mask),
|
| 210 |
-
output_path=os.path.join(output_dir, f"masked_rgb_t{t}.tiff"),
|
| 211 |
-
meta=meta_data[t])
|
| 212 |
-
|
| 213 |
|
| 214 |
def extract_rgb_imgs(input_img, rec_img, mask_img, channels, mean, std):
|
| 215 |
""" Wrapper function to save Geotiff images (original, reconstructed, masked) per timestamp.
|
|
@@ -230,24 +43,31 @@ def extract_rgb_imgs(input_img, rec_img, mask_img, channels, mean, std):
|
|
| 230 |
for t in range(input_img.shape[1]):
|
| 231 |
rgb_orig, rgb_pred = process_channel_group(orig_img=input_img[:, t, :, :],
|
| 232 |
new_img=rec_img[:, t, :, :],
|
| 233 |
-
channels=channels,
|
| 234 |
-
|
|
|
|
| 235 |
|
| 236 |
rgb_mask = mask_img[channels, t, :, :] * rgb_orig
|
| 237 |
|
| 238 |
# extract images
|
| 239 |
-
rgb_orig_list.append(_convert_np_uint8(rgb_orig))
|
| 240 |
-
rgb_mask_list.append(_convert_np_uint8(rgb_mask))
|
| 241 |
-
rgb_pred_list.append(_convert_np_uint8(rgb_pred))
|
| 242 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 243 |
outputs = rgb_orig_list + rgb_mask_list + rgb_pred_list
|
| 244 |
|
| 245 |
return outputs
|
| 246 |
|
| 247 |
|
| 248 |
-
def predict_on_images(data_files: list,
|
| 249 |
-
|
| 250 |
-
|
| 251 |
try:
|
| 252 |
data_files = [x.name for x in data_files]
|
| 253 |
print('Path extracted from example')
|
|
@@ -257,24 +77,18 @@ def predict_on_images(data_files: list, mask_ratio: float, yaml_file_path: str,
|
|
| 257 |
# Get parameters --------
|
| 258 |
print('This is the printout', data_files)
|
| 259 |
|
| 260 |
-
with open(
|
| 261 |
-
|
| 262 |
-
|
| 263 |
-
model_params = params["model_args"]
|
| 264 |
-
# data related
|
| 265 |
-
train_params = params["train_params"]
|
| 266 |
-
num_frames = model_params['num_frames']
|
| 267 |
-
img_size = model_params['img_size']
|
| 268 |
-
bands = train_params['bands']
|
| 269 |
-
mean = train_params['data_mean']
|
| 270 |
-
std = train_params['data_std']
|
| 271 |
|
| 272 |
batch_size = 8
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 273 |
|
| 274 |
-
|
| 275 |
-
|
| 276 |
-
# We must have *num_frames* files to build one example!
|
| 277 |
-
assert len(data_files) == num_frames, "File list must be equal to expected number of frames."
|
| 278 |
|
| 279 |
if torch.cuda.is_available():
|
| 280 |
device = torch.device('cuda')
|
|
@@ -289,16 +103,23 @@ def predict_on_images(data_files: list, mask_ratio: float, yaml_file_path: str,
|
|
| 289 |
|
| 290 |
# Create model and load checkpoint -------------------------------------------------------------
|
| 291 |
|
| 292 |
-
|
| 293 |
-
|
|
|
|
|
|
|
|
|
|
| 294 |
|
| 295 |
total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
|
| 296 |
print(f"\n--> Model has {total_params:,} parameters.\n")
|
| 297 |
|
| 298 |
model.to(device)
|
| 299 |
|
| 300 |
-
state_dict = torch.load(checkpoint, map_location=device)
|
| 301 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 302 |
print(f"Loaded checkpoint from {checkpoint}")
|
| 303 |
|
| 304 |
# Running model --------------------------------------------------------------------------------
|
|
@@ -348,37 +169,35 @@ def predict_on_images(data_files: list, mask_ratio: float, yaml_file_path: str,
|
|
| 348 |
for d in meta_data:
|
| 349 |
d.update(count=3, dtype='uint8', compress='lzw', nodata=0)
|
| 350 |
|
| 351 |
-
# save_rgb_imgs(batch[0, ...], rec_imgs_full[0, ...], mask_imgs_full[0, ...],
|
| 352 |
-
# channels, mean, std, output_dir, meta_data)
|
| 353 |
-
|
| 354 |
outputs = extract_rgb_imgs(batch_full[0, ...], rec_imgs_full[0, ...], mask_imgs_full[0, ...],
|
| 355 |
channels, mean, std)
|
| 356 |
|
| 357 |
-
|
| 358 |
print("Done!")
|
| 359 |
|
| 360 |
return outputs
|
| 361 |
|
| 362 |
|
|
|
|
| 363 |
|
|
|
|
| 364 |
|
| 365 |
-
|
|
|
|
|
|
|
| 366 |
|
| 367 |
-
|
| 368 |
-
|
| 369 |
-
|
| 370 |
-
|
| 371 |
-
return example_list
|
| 372 |
-
|
| 373 |
|
| 374 |
-
|
| 375 |
-
|
| 376 |
-
|
| 377 |
-
|
| 378 |
-
|
| 379 |
-
The user needs to provide three HLS geotiff images, including the following channels in reflectance units: Blue, Green, Red, Narrow NIR, SWIR, SWIR 2.
|
| 380 |
|
| 381 |
-
|
|
|
|
| 382 |
''')
|
| 383 |
with gr.Row():
|
| 384 |
with gr.Column():
|
|
@@ -386,73 +205,55 @@ Check out our newest model: [Prithvi-EO-2.0-Demo](https://huggingface.co/spaces/
|
|
| 386 |
# inp_slider = gr.Slider(0, 100, value=50, label="Mask ratio", info="Choose ratio of masking between 0 and 100", elem_id='slider'),
|
| 387 |
btn = gr.Button("Submit")
|
| 388 |
with gr.Row():
|
| 389 |
-
gr.Markdown(value='##
|
| 390 |
-
with gr.Row():
|
| 391 |
-
gr.Markdown(value='T1')
|
| 392 |
-
gr.Markdown(value='T2')
|
| 393 |
-
gr.Markdown(value='T3')
|
| 394 |
-
with gr.Row():
|
| 395 |
-
out1_orig_t1=gr.Image(image_mode='RGB')
|
| 396 |
-
out2_orig_t2 = gr.Image(image_mode='RGB')
|
| 397 |
-
out3_orig_t3 = gr.Image(image_mode='RGB')
|
| 398 |
-
with gr.Row():
|
| 399 |
gr.Markdown(value='## Masked images')
|
| 400 |
-
|
| 401 |
-
|
| 402 |
-
|
| 403 |
-
|
| 404 |
-
|
| 405 |
-
|
| 406 |
-
|
| 407 |
-
|
| 408 |
-
|
| 409 |
-
|
| 410 |
-
|
| 411 |
-
|
| 412 |
-
|
| 413 |
-
|
| 414 |
-
|
| 415 |
-
|
| 416 |
-
|
| 417 |
-
|
| 418 |
-
|
| 419 |
-
|
| 420 |
-
btn.click(fn=func,
|
| 421 |
-
# inputs=[inp_files, inp_slider],
|
| 422 |
inputs=inp_files,
|
| 423 |
-
outputs=
|
| 424 |
-
|
| 425 |
-
out3_orig_t3,
|
| 426 |
-
out4_masked_t1,
|
| 427 |
-
out5_masked_t2,
|
| 428 |
-
out6_masked_t3,
|
| 429 |
-
out7_pred_t1,
|
| 430 |
-
out8_pred_t2,
|
| 431 |
-
out9_pred_t3])
|
| 432 |
with gr.Row():
|
| 433 |
-
gr.Examples(examples=[[[
|
| 434 |
-
|
| 435 |
-
|
| 436 |
-
|
| 437 |
-
|
| 438 |
-
|
| 439 |
-
|
| 440 |
-
|
| 441 |
-
|
| 442 |
-
|
| 443 |
-
|
| 444 |
-
|
| 445 |
-
|
| 446 |
-
|
| 447 |
-
|
| 448 |
-
|
| 449 |
-
|
| 450 |
-
out8_pred_t2,
|
| 451 |
-
out9_pred_t3],
|
| 452 |
-
# preprocess=preprocess_example,
|
| 453 |
-
fn=func,
|
| 454 |
-
cache_examples=True
|
| 455 |
)
|
| 456 |
-
|
| 457 |
|
| 458 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
|
| 2 |
+
import os
|
|
|
|
| 3 |
import torch
|
| 4 |
import yaml
|
| 5 |
+
import numpy as np
|
|
|
|
|
|
|
| 6 |
import gradio as gr
|
| 7 |
+
from einops import rearrange
|
| 8 |
from functools import partial
|
| 9 |
+
from huggingface_hub import hf_hub_download
|
| 10 |
|
| 11 |
+
# pull files from hub
|
| 12 |
+
token = os.environ.get("HF_TOKEN", None)
|
| 13 |
+
config_path = hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-EO-1.0-100M",
|
| 14 |
+
filename="config.json", token=token)
|
| 15 |
+
checkpoint = hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-EO-1.0-100M",
|
| 16 |
+
filename='Prithvi_EO_V1_100M.pt', token=token)
|
| 17 |
+
model_def = hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-EO-1.0-100M",
|
| 18 |
+
filename='prithvi_mae.py', token=token)
|
| 19 |
+
model_inference = hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-EO-1.0-100M",
|
| 20 |
+
filename='inference.py', token=token)
|
| 21 |
+
os.system(f'cp {model_def} .')
|
| 22 |
+
os.system(f'cp {model_inference} .')
|
| 23 |
|
| 24 |
+
from prithvi_mae import PrithviMAE
|
| 25 |
+
from inference import process_channel_group, _convert_np_uint8, load_example, run_model
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
def extract_rgb_imgs(input_img, rec_img, mask_img, channels, mean, std):
|
| 28 |
""" Wrapper function to save Geotiff images (original, reconstructed, masked) per timestamp.
|
|
|
|
| 43 |
for t in range(input_img.shape[1]):
|
| 44 |
rgb_orig, rgb_pred = process_channel_group(orig_img=input_img[:, t, :, :],
|
| 45 |
new_img=rec_img[:, t, :, :],
|
| 46 |
+
channels=channels,
|
| 47 |
+
mean=mean,
|
| 48 |
+
std=std)
|
| 49 |
|
| 50 |
rgb_mask = mask_img[channels, t, :, :] * rgb_orig
|
| 51 |
|
| 52 |
# extract images
|
| 53 |
+
rgb_orig_list.append(_convert_np_uint8(rgb_orig).transpose(1, 2, 0))
|
| 54 |
+
rgb_mask_list.append(_convert_np_uint8(rgb_mask).transpose(1, 2, 0))
|
| 55 |
+
rgb_pred_list.append(_convert_np_uint8(rgb_pred).transpose(1, 2, 0))
|
| 56 |
+
|
| 57 |
+
# Add white dummy image values for missing timestamps
|
| 58 |
+
dummy = np.ones((20, 20), dtype=np.uint8) * 255
|
| 59 |
+
num_dummies = 3 - len(rgb_orig_list)
|
| 60 |
+
if num_dummies:
|
| 61 |
+
rgb_orig_list.extend([dummy] * num_dummies)
|
| 62 |
+
rgb_mask_list.extend([dummy] * num_dummies)
|
| 63 |
+
rgb_pred_list.extend([dummy] * num_dummies)
|
| 64 |
+
|
| 65 |
outputs = rgb_orig_list + rgb_mask_list + rgb_pred_list
|
| 66 |
|
| 67 |
return outputs
|
| 68 |
|
| 69 |
|
| 70 |
+
def predict_on_images(data_files: list, config_path: str, checkpoint: str, mask_ratio: float = None):
|
|
|
|
|
|
|
| 71 |
try:
|
| 72 |
data_files = [x.name for x in data_files]
|
| 73 |
print('Path extracted from example')
|
|
|
|
| 77 |
# Get parameters --------
|
| 78 |
print('This is the printout', data_files)
|
| 79 |
|
| 80 |
+
with open(config_path, 'r') as f:
|
| 81 |
+
config = yaml.safe_load(f)['pretrained_cfg']
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
|
| 83 |
batch_size = 8
|
| 84 |
+
bands = config['bands']
|
| 85 |
+
num_frames = len(data_files)
|
| 86 |
+
mean = config['mean']
|
| 87 |
+
std = config['std']
|
| 88 |
+
img_size = config['img_size']
|
| 89 |
+
mask_ratio = mask_ratio or config['mask_ratio']
|
| 90 |
|
| 91 |
+
assert num_frames <= 3, "Demo only supports up to three timestamps"
|
|
|
|
|
|
|
|
|
|
| 92 |
|
| 93 |
if torch.cuda.is_available():
|
| 94 |
device = torch.device('cuda')
|
|
|
|
| 103 |
|
| 104 |
# Create model and load checkpoint -------------------------------------------------------------
|
| 105 |
|
| 106 |
+
config.update(
|
| 107 |
+
num_frames=num_frames,
|
| 108 |
+
)
|
| 109 |
+
|
| 110 |
+
model = PrithviMAE(**config)
|
| 111 |
|
| 112 |
total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
|
| 113 |
print(f"\n--> Model has {total_params:,} parameters.\n")
|
| 114 |
|
| 115 |
model.to(device)
|
| 116 |
|
| 117 |
+
state_dict = torch.load(checkpoint, map_location=device, weights_only=False)
|
| 118 |
+
# discard fixed pos_embedding weight
|
| 119 |
+
for k in list(state_dict.keys()):
|
| 120 |
+
if 'pos_embed' in k:
|
| 121 |
+
del state_dict[k]
|
| 122 |
+
model.load_state_dict(state_dict, strict=False)
|
| 123 |
print(f"Loaded checkpoint from {checkpoint}")
|
| 124 |
|
| 125 |
# Running model --------------------------------------------------------------------------------
|
|
|
|
| 169 |
for d in meta_data:
|
| 170 |
d.update(count=3, dtype='uint8', compress='lzw', nodata=0)
|
| 171 |
|
|
|
|
|
|
|
|
|
|
| 172 |
outputs = extract_rgb_imgs(batch_full[0, ...], rec_imgs_full[0, ...], mask_imgs_full[0, ...],
|
| 173 |
channels, mean, std)
|
| 174 |
|
|
|
|
| 175 |
print("Done!")
|
| 176 |
|
| 177 |
return outputs
|
| 178 |
|
| 179 |
|
| 180 |
+
run_inference = partial(predict_on_images, config_path=config_path,checkpoint=checkpoint)
|
| 181 |
|
| 182 |
+
with gr.Blocks() as demo:
|
| 183 |
|
| 184 |
+
gr.Markdown(value='# Prithvi-EO-1.0 image reconstruction demo')
|
| 185 |
+
gr.Markdown(value='''
|
| 186 |
+
Check out our newest model: [Prithvi-EO-2.0-Demo](https://huggingface.co/spaces/ibm-nasa-geospatial/Prithvi-EO-2.0-Demo).
|
| 187 |
|
| 188 |
+
Prithvi is a first-of-its-kind temporal Vision transformer pretrained by the IBM and NASA team on continental US Harmonised Landsat Sentinel 2 (HLS) data.
|
| 189 |
+
Particularly, the model adopts a self-supervised encoder developed with a ViT architecture and Masked AutoEncoder learning strategy, with a MSE as a loss function.
|
| 190 |
+
The model includes spatial attention across multiple patchies and also temporal attention for each patch.
|
| 191 |
+
More info about the model and its weights are available [here](https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M).\n
|
|
|
|
|
|
|
| 192 |
|
| 193 |
+
This demo showcases the image reconstruction over one to three timestamps.
|
| 194 |
+
The model randomly masks out some proportion of the images and reconstructs them based on the not masked portion of the images.
|
| 195 |
+
The reconstructed images are merged with the visible unmasked patches.
|
| 196 |
+
We recommend submitting images of size 224 to ~1000 pixels for faster processing time.
|
| 197 |
+
Images bigger than 224x224 are processed using a sliding window approach which can lead to artefacts between patches.\n
|
|
|
|
| 198 |
|
| 199 |
+
The user needs to provide the HLS geotiff images, including the following channels in reflectance units: Blue, Green, Red, Narrow NIR, SWIR, SWIR 2.
|
| 200 |
+
Some example images are provided at the end of this page.
|
| 201 |
''')
|
| 202 |
with gr.Row():
|
| 203 |
with gr.Column():
|
|
|
|
| 205 |
# inp_slider = gr.Slider(0, 100, value=50, label="Mask ratio", info="Choose ratio of masking between 0 and 100", elem_id='slider'),
|
| 206 |
btn = gr.Button("Submit")
|
| 207 |
with gr.Row():
|
| 208 |
+
gr.Markdown(value='## Input time series')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 209 |
gr.Markdown(value='## Masked images')
|
| 210 |
+
gr.Markdown(value='## Reconstructed images*')
|
| 211 |
+
|
| 212 |
+
original = []
|
| 213 |
+
masked = []
|
| 214 |
+
predicted = []
|
| 215 |
+
timestamps = []
|
| 216 |
+
for t in range(3):
|
| 217 |
+
timestamps.append(gr.Column(visible=t == 0))
|
| 218 |
+
with timestamps[t]:
|
| 219 |
+
#with gr.Row():
|
| 220 |
+
# gr.Markdown(value=f"Timestamp {t+1}")
|
| 221 |
+
with gr.Row():
|
| 222 |
+
original.append(gr.Image(image_mode='RGB', show_label=False, show_fullscreen_button=False))
|
| 223 |
+
masked.append(gr.Image(image_mode='RGB', show_label=False, show_fullscreen_button=False))
|
| 224 |
+
predicted.append(gr.Image(image_mode='RGB', show_label=False, show_fullscreen_button=False))
|
| 225 |
+
|
| 226 |
+
gr.Markdown(value='\* The reconstructed images include the ground truth unmasked patches.')
|
| 227 |
+
|
| 228 |
+
btn.click(fn=run_inference,
|
|
|
|
|
|
|
|
|
|
| 229 |
inputs=inp_files,
|
| 230 |
+
outputs=original + masked + predicted)
|
| 231 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 232 |
with gr.Row():
|
| 233 |
+
gr.Examples(examples=[[[
|
| 234 |
+
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T13REN.2018013T172747.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif"),
|
| 235 |
+
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T13REN.2018029T172738.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif"),
|
| 236 |
+
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T13REN.2018061T172724.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif")
|
| 237 |
+
]],[[
|
| 238 |
+
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T17RMP.2018004T155509.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif"),
|
| 239 |
+
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T17RMP.2018036T155452.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif"),
|
| 240 |
+
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T17RMP.2018068T155438.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif")
|
| 241 |
+
]],[[
|
| 242 |
+
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T18TVL.2018029T154533.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif"),
|
| 243 |
+
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T18TVL.2018141T154435.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif"),
|
| 244 |
+
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T18TVL.2018189T154446.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif")
|
| 245 |
+
]]],
|
| 246 |
+
inputs=inp_files,
|
| 247 |
+
outputs=original + masked + predicted,
|
| 248 |
+
fn=run_inference,
|
| 249 |
+
cache_examples=True
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 250 |
)
|
|
|
|
| 251 |
|
| 252 |
+
def update_visibility(files):
|
| 253 |
+
timestamps = [gr.Column(visible=t < len(files)) for t in range(3)]
|
| 254 |
+
|
| 255 |
+
return timestamps
|
| 256 |
+
|
| 257 |
+
inp_files.change(update_visibility, inp_files, timestamps)
|
| 258 |
+
|
| 259 |
+
demo.launch() # share=True, ssr_mode=False
|
HLS.L30.T13REN.2018013T172747.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T13REN.2018013T172747.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif
RENAMED
|
File without changes
|
HLS.L30.T13REN.2018029T172738.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T13REN.2018029T172738.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif
RENAMED
|
File without changes
|
HLS.L30.T13REN.2018061T172724.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T13REN.2018061T172724.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif
RENAMED
|
File without changes
|
HLS.L30.T17RMP.2018004T155509.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T17RMP.2018004T155509.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif
RENAMED
|
File without changes
|
HLS.L30.T17RMP.2018036T155452.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T17RMP.2018036T155452.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif
RENAMED
|
File without changes
|
HLS.L30.T17RMP.2018068T155438.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T17RMP.2018068T155438.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif
RENAMED
|
File without changes
|
HLS.L30.T18TVL.2018029T154533.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T18TVL.2018029T154533.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif
RENAMED
|
File without changes
|
HLS.L30.T18TVL.2018141T154435.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T18TVL.2018141T154435.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif
RENAMED
|
File without changes
|
HLS.L30.T18TVL.2018189T154446.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif β examples/HLS.L30.T18TVL.2018189T154446.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif
RENAMED
|
File without changes
|