File size: 20,483 Bytes
f5ac567 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 |
import logging
import os
from huggingface_hub import hf_hub_download
# pull files from hub
yaml_file_path = hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL",
filename="Prithvi_EO_V2_300M_TL_config.yaml", token=os.environ.get("token"))
checkpoint = hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL",
filename='Prithvi_EO_V2_300M_TL.pt', token=os.environ.get("token"))
model_def = hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL",
filename='prithvi_mae.py', token=os.environ.get("token"))
model_inference = hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL",
filename='inference.py', token=os.environ.get("token"))
os.system(f'cp {model_def} .')
os.system(f'cp {model_inference} .')
import os
import torch
import yaml
import numpy as np
import gradio as gr
from einops import rearrange
from functools import partial
from prithvi_mae import PrithviMAE
from inference import process_channel_group, read_geotiff, save_geotiff, _convert_np_uint8, load_example, run_model
NO_DATA = -9999
NO_DATA_FLOAT = 0.0001
PERCENTILES = (0.1, 99.9)
# def process_channel_group(orig_img, new_img, channels, data_mean, data_std):
# """ Process *orig_img* and *new_img* for RGB visualization. Each band is rescaled back to the
# original range using *data_mean* and *data_std* and then lowest and highest percentiles are
# removed to enhance contrast. Data is rescaled to (0, 1) range and stacked channels_first.
# Args:
# orig_img: torch.Tensor representing original image (reference) with shape = (bands, H, W).
# new_img: torch.Tensor representing image with shape = (bands, H, W).
# channels: list of indices representing RGB channels.
# data_mean: list of mean values for each band.
# data_std: list of std values for each band.
# Returns:
# torch.Tensor with shape (num_channels, height, width) for original image
# torch.Tensor with shape (num_channels, height, width) for the other image
# """
#
# stack_c = [], []
#
# for c in channels:
# orig_ch = orig_img[c, ...]
# valid_mask = torch.ones_like(orig_ch, dtype=torch.bool)
# valid_mask[orig_ch == NO_DATA_FLOAT] = False
#
# # Back to original data range
# orig_ch = (orig_ch * data_std[c]) + data_mean[c]
# new_ch = (new_img[c, ...] * data_std[c]) + data_mean[c]
#
# # Rescale (enhancing contrast)
# min_value, max_value = np.percentile(orig_ch[valid_mask], PERCENTILES)
#
# orig_ch = torch.clamp((orig_ch - min_value) / (max_value - min_value), 0, 1)
# new_ch = torch.clamp((new_ch - min_value) / (max_value - min_value), 0, 1)
#
# # No data as zeros
# orig_ch[~valid_mask] = 0
# new_ch[~valid_mask] = 0
#
# stack_c[0].append(orig_ch)
# stack_c[1].append(new_ch)
#
# # Channels first
# stack_orig = torch.stack(stack_c[0], dim=0)
# stack_rec = torch.stack(stack_c[1], dim=0)
#
# return stack_orig, stack_rec
#
#
# def read_geotiff(file_path: str):
# """ Read all bands from *file_path* and returns image + meta info.
# Args:
# file_path: path to image file.
# Returns:
# np.ndarray with shape (bands, height, width)
# meta info dict
# """
#
# with rasterio.open(file_path) as src:
# img = src.read()
# meta = src.meta
# coords = src.lnglat()
#
# return img, meta, coords
#
#
# def save_geotiff(image, output_path: str, meta: dict):
# """ Save multi-band image in Geotiff file.
# Args:
# image: np.ndarray with shape (bands, height, width)
# output_path: path where to save the image
# meta: dict with meta info.
# """
#
# with rasterio.open(output_path, "w", **meta) as dest:
# for i in range(image.shape[0]):
# dest.write(image[i, :, :], i + 1)
#
# return
#
#
# def _convert_np_uint8(float_image: torch.Tensor):
#
# image = float_image.numpy() * 255.0
# image = image.astype(dtype=np.uint8)
# image = image.transpose((1, 2, 0))
#
# return image
#
#
# def load_example(file_paths: List[str], mean: List[float], std: List[float]):
# """ Build an input example by loading images in *file_paths*.
# Args:
# file_paths: list of file paths .
# mean: list containing mean values for each band in the images in *file_paths*.
# std: list containing std values for each band in the images in *file_paths*.
# Returns:
# np.array containing created example
# list of meta info for each image in *file_paths*
# """
#
# imgs = []
# metas = []
#
# for file in file_paths:
# img, meta = read_geotiff(file)
# img = img[:6]*10000 if img[:6].mean() <= 2 else img[:6]
#
# # Rescaling (don't normalize on nodata)
# img = np.moveaxis(img, 0, -1) # channels last for rescaling
# img = np.where(img == NO_DATA, NO_DATA_FLOAT, (img - mean) / std)
#
# imgs.append(img)
# metas.append(meta)
#
# imgs = np.stack(imgs, axis=0) # num_frames, H, W, C
# imgs = np.moveaxis(imgs, -1, 0).astype('float32') # C, num_frames, H, W
# imgs = np.expand_dims(imgs, axis=0) # add batch dim
#
# return imgs, metas
#
#
# def run_model(model: torch.nn.Module, input_data: torch.Tensor, mask_ratio: float, device: torch.device):
# """ Run *model* with *input_data* and create images from output tokens (mask, reconstructed + visible).
# Args:
# model: MAE model to run.
# input_data: torch.Tensor with shape (B, C, T, H, W).
# mask_ratio: mask ratio to use.
# device: device where model should run.
# Returns:
# 3 torch.Tensor with shape (B, C, T, H, W).
# """
#
# with torch.no_grad():
# x = input_data.to(device)
#
# _, pred, mask = model(x, mask_ratio)
#
# # Create mask and prediction images (un-patchify)
# mask_img = model.unpatchify(mask.unsqueeze(-1).repeat(1, 1, pred.shape[-1])).detach().cpu()
# pred_img = model.unpatchify(pred).detach().cpu()
#
# # Mix visible and predicted patches
# rec_img = input_data.clone()
# rec_img[mask_img == 1] = pred_img[mask_img == 1] # binary mask: 0 is keep, 1 is remove
#
# # Switch zeros/ones in mask images so masked patches appear darker in plots (better visualization)
# mask_img = (~(mask_img.to(torch.bool))).to(torch.float)
#
# return rec_img, mask_img
#
#
# def save_rgb_imgs(input_img, rec_img, mask_img, channels, mean, std, output_dir, meta_data):
# """ Wrapper function to save Geotiff images (original, reconstructed, masked) per timestamp.
# Args:
# input_img: input torch.Tensor with shape (C, T, H, W).
# rec_img: reconstructed torch.Tensor with shape (C, T, H, W).
# mask_img: mask torch.Tensor with shape (C, T, H, W).
# channels: list of indices representing RGB channels.
# mean: list of mean values for each band.
# std: list of std values for each band.
# output_dir: directory where to save outputs.
# meta_data: list of dicts with geotiff meta info.
# """
#
# for t in range(input_img.shape[1]):
# rgb_orig, rgb_pred = process_channel_group(orig_img=input_img[:, t, :, :],
# new_img=rec_img[:, t, :, :],
# channels=channels, data_mean=mean,
# data_std=std)
#
# rgb_mask = mask_img[channels, t, :, :] * rgb_orig
#
# # Saving images
#
# save_geotiff(image=_convert_np_uint8(rgb_orig),
# output_path=os.path.join(output_dir, f"original_rgb_t{t}.tiff"),
# meta=meta_data[t])
#
# save_geotiff(image=_convert_np_uint8(rgb_pred),
# output_path=os.path.join(output_dir, f"predicted_rgb_t{t}.tiff"),
# meta=meta_data[t])
#
# save_geotiff(image=_convert_np_uint8(rgb_mask),
# output_path=os.path.join(output_dir, f"masked_rgb_t{t}.tiff"),
# meta=meta_data[t])
def extract_rgb_imgs(input_img, rec_img, mask_img, channels, mean, std):
""" Wrapper function to save Geotiff images (original, reconstructed, masked) per timestamp.
Args:
input_img: input torch.Tensor with shape (C, T, H, W).
rec_img: reconstructed torch.Tensor with shape (C, T, H, W).
mask_img: mask torch.Tensor with shape (C, T, H, W).
channels: list of indices representing RGB channels.
mean: list of mean values for each band.
std: list of std values for each band.
output_dir: directory where to save outputs.
meta_data: list of dicts with geotiff meta info.
"""
rgb_orig_list = []
rgb_mask_list = []
rgb_pred_list = []
for t in range(input_img.shape[1]):
rgb_orig, rgb_pred = process_channel_group(orig_img=input_img[:, t, :, :],
new_img=rec_img[:, t, :, :],
channels=channels,
mean=mean,
std=std)
rgb_mask = mask_img[channels, t, :, :] * rgb_orig
# extract images
rgb_orig_list.append(_convert_np_uint8(rgb_orig).transpose(1, 2, 0))
rgb_mask_list.append(_convert_np_uint8(rgb_mask).transpose(1, 2, 0))
rgb_pred_list.append(_convert_np_uint8(rgb_pred).transpose(1, 2, 0))
outputs = rgb_orig_list + rgb_mask_list + rgb_pred_list
return outputs
def predict_on_images(data_files: list, mask_ratio: float, yaml_file_path: str, checkpoint: str):
try:
data_files = [x.name for x in data_files]
print('Path extracted from example')
except:
print('Files submitted through UI')
# Get parameters --------
print('This is the printout', data_files)
with open(yaml_file_path, 'r') as f:
config = yaml.safe_load(f)
batch_size = 8
bands = config['DATA']['BANDS']
num_frames = len(data_files)
mean = config['DATA']['MEAN']
std = config['DATA']['STD']
coords_encoding = config['MODEL']['COORDS_ENCODING']
img_size = config['DATA']['INPUT_SIZE'][-1]
mask_ratio = mask_ratio or config['DATA']['MASK_RATIO']
if num_frames > 4:
# TODO: Check if we can limit this via UI
logging.warning("Model was only trained with only four timestamps.")
if torch.cuda.is_available():
device = torch.device('cuda')
else:
device = torch.device('cpu')
print(f"Using {device} device.\n")
# Loading data ---------------------------------------------------------------------------------
input_data, temporal_coords, location_coords, meta_data = load_example(file_paths=data_files, mean=mean, std=std)
if len(temporal_coords) != num_frames and 'time' in coords_encoding:
coords_encoding.pop('time')
if not len(location_coords) and 'location' in coords_encoding:
coords_encoding.pop('location')
# Create model and load checkpoint -------------------------------------------------------------
model = PrithviMAE(img_size=config['DATA']['INPUT_SIZE'][-2:],
patch_size=config['MODEL']['PATCH_SIZE'],
num_frames=num_frames,
in_chans=len(bands),
embed_dim=config['MODEL']['EMBED_DIM'],
depth=config['MODEL']['DEPTH'],
num_heads=config['MODEL']['NUM_HEADS'],
decoder_embed_dim=config['MODEL']['DECODER_EMBED_DIM'],
decoder_depth=config['MODEL']['DECODER_DEPTH'],
decoder_num_heads=config['MODEL']['DECODER_NUM_HEADS'],
mlp_ratio=config['MODEL']['MLP_RATIO'],
norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
norm_pix_loss=config['MODEL']['NORM_PIX_LOSS'],
coords_encoding=coords_encoding,
coords_scale_learn=config['MODEL']['COORDS_SCALE_LEARN'])
total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f"\n--> Model has {total_params:,} parameters.\n")
model.to(device)
state_dict = torch.load(checkpoint, map_location=device, weights_only=False)
# discard fixed pos_embedding weight
for k in list(state_dict.keys()):
if 'pos_embed' in k:
del state_dict[k]
model.load_state_dict(state_dict, strict=False)
print(f"Loaded checkpoint from {checkpoint}")
# Running model --------------------------------------------------------------------------------
model.eval()
channels = [bands.index(b) for b in ['B04', 'B03', 'B02']] # BGR -> RGB
# Reflect pad if not divisible by img_size
original_h, original_w = input_data.shape[-2:]
pad_h = img_size - (original_h % img_size)
pad_w = img_size - (original_w % img_size)
input_data = np.pad(input_data, ((0, 0), (0, 0), (0, 0), (0, pad_h), (0, pad_w)), mode='reflect')
# Build sliding window
batch = torch.tensor(input_data, device='cpu')
windows = batch.unfold(3, img_size, img_size).unfold(4, img_size, img_size)
h1, w1 = windows.shape[3:5]
windows = rearrange(windows, 'b c t h1 w1 h w -> (b h1 w1) c t h w', h=img_size, w=img_size)
# Split into batches if number of windows > batch_size
num_batches = windows.shape[0] // batch_size if windows.shape[0] > batch_size else 1
windows = torch.tensor_split(windows, num_batches, dim=0)
# Run model
rec_imgs = []
mask_imgs = []
for x in windows:
temp_coords = torch.Tensor([temporal_coords] * len(x))
loc_coords = torch.Tensor([location_coords[0]] * len(x))
rec_img, mask_img = run_model(model, x, temp_coords, loc_coords, mask_ratio, device)
rec_imgs.append(rec_img)
mask_imgs.append(mask_img)
rec_imgs = torch.concat(rec_imgs, dim=0)
mask_imgs = torch.concat(mask_imgs, dim=0)
# Build images from patches
rec_imgs = rearrange(rec_imgs, '(b h1 w1) c t h w -> b c t (h1 h) (w1 w)',
h=img_size, w=img_size, b=1, c=len(bands), t=num_frames, h1=h1, w1=w1)
mask_imgs = rearrange(mask_imgs, '(b h1 w1) c t h w -> b c t (h1 h) (w1 w)',
h=img_size, w=img_size, b=1, c=len(bands), t=num_frames, h1=h1, w1=w1)
# Cut padded images back to original size
rec_imgs_full = rec_imgs[..., :original_h, :original_w]
mask_imgs_full = mask_imgs[..., :original_h, :original_w]
batch_full = batch[..., :original_h, :original_w]
# Build RGB images
for d in meta_data:
d.update(count=3, dtype='uint8', compress='lzw', nodata=0)
outputs = extract_rgb_imgs(batch_full[0, ...], rec_imgs_full[0, ...], mask_imgs_full[0, ...],
channels, mean, std)
print("Done!")
return outputs
func = partial(predict_on_images, yaml_file_path=yaml_file_path,checkpoint=checkpoint)
with gr.Blocks() as demo:
gr.Markdown(value='# Prithvi-EO-2.0 image reconstruction demo')
gr.Markdown(value='''
Prithvi-EO-2.0 is the second generation EO foundation model developed by the IBM and NASA team.
The temporal ViT is train on 4.2M Harmonised Landsat Sentinel 2 (HLS) samples with four timestamps each, using the Masked AutoEncoder learning strategy.
The model includes spatial and temporal attention across multiple patches and timestamps.
Additionally, temporal and location information is added to the model input via embeddings.
More info about the model are available [here](https://huggingface.co/ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL).\n
This demo showcases the image reconstruction over one to four timestamps.
The model randomly masks out some proportion of the images and then reconstructing them based on the not masked portion of the images.\n
The user needs to provide the HLS geotiff images, including the following channels in reflectance units: Blue, Green, Red, Narrow NIR, SWIR, SWIR 2.
Optionally, the location information is extracted from the tif files while the temporal information can be provided in the filename in the format `<date>T<time>` or `<year><julian day>T<time>` (HLS format).
We recommend submitting images of size 224 to 1000 pixels for faster processing time. Some example images are provided at the end of this page.
''')
with gr.Row():
with gr.Column():
inp_files = gr.Files(elem_id='files')
# inp_slider = gr.Slider(0, 100, value=50, label="Mask ratio", info="Choose ratio of masking between 0 and 100", elem_id='slider'),
btn = gr.Button("Submit")
with gr.Row():
gr.Markdown(value='## Original images')
with gr.Row():
gr.Markdown(value='T1')
gr.Markdown(value='T2')
gr.Markdown(value='T3')
with gr.Row():
out1_orig_t1 = gr.Image(image_mode='RGB')
out2_orig_t2 = gr.Image(image_mode='RGB')
out3_orig_t3 = gr.Image(image_mode='RGB')
with gr.Row():
gr.Markdown(value='## Masked images')
with gr.Row():
gr.Markdown(value='T1')
gr.Markdown(value='T2')
gr.Markdown(value='T3')
with gr.Row():
out4_masked_t1 = gr.Image(image_mode='RGB')
out5_masked_t2 = gr.Image(image_mode='RGB')
out6_masked_t3 = gr.Image(image_mode='RGB')
with gr.Row():
gr.Markdown(value='## Reonstructed images')
with gr.Row():
gr.Markdown(value='T1')
gr.Markdown(value='T2')
gr.Markdown(value='T3')
with gr.Row():
out7_pred_t1 = gr.Image(image_mode='RGB')
out8_pred_t2 = gr.Image(image_mode='RGB')
out9_pred_t3 = gr.Image(image_mode='RGB')
btn.click(fn=func,
# inputs=[inp_files, inp_slider],
inputs=inp_files,
outputs=[out1_orig_t1,
out2_orig_t2,
out3_orig_t3,
out4_masked_t1,
out5_masked_t2,
out6_masked_t3,
out7_pred_t1,
out8_pred_t2,
out9_pred_t3])
with gr.Row():
gr.Examples(examples=[[[os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T13REN.2018013T172747.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif"),
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T13REN.2018029T172738.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif"),
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T13REN.2018061T172724.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif")]],
[[os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T17RMP.2018004T155509.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif"),
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T17RMP.2018036T155452.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif"),
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T17RMP.2018068T155438.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif")]],
[[os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T18TVL.2018029T154533.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif"),
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T18TVL.2018141T154435.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif"),
os.path.join(os.path.dirname(__file__), "examples/HLS.L30.T18TVL.2018189T154446.v2.0.B02.B03.B04.B05.B06.B07_cropped.tif")]]],
inputs=inp_files,
outputs=[out1_orig_t1,
out2_orig_t2,
out3_orig_t3,
out4_masked_t1,
out5_masked_t2,
out6_masked_t3,
out7_pred_t1,
out8_pred_t2,
out9_pred_t3],
fn=func,
cache_examples=True
)
demo.launch()
|