app.py

######### pull files
import os
from huggingface_hub import hf_hub_download
config_path=hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-100M-sen1floods11", filename="sen1floods11_Prithvi_100M.py", token=os.environ.get("token"))
ckpt=hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-100M-sen1floods11", filename='sen1floods11_Prithvi_100M.pth', token=os.environ.get("token"))
##########


import argparse
from mmcv import Config

from mmseg.models import build_segmentor

from mmseg.datasets.pipelines import Compose, LoadImageFromFile

import rasterio
import tifffile
import torch

from mmseg.apis import init_segmentor

from mmcv.parallel import collate, scatter

import numpy as np
import glob
import os

import time

import numpy as np
import gradio as gr
from functools import partial

import pdb

import matplotlib.pyplot as plt

from skimage import exposure 

def stretch_rgb(rgb):
    
    ls_pct=1
    pLow, pHigh = np.percentile(rgb[~np.isnan(rgb)], (ls_pct,100-ls_pct))
    img_rescale = exposure.rescale_intensity(rgb, in_range=(pLow,pHigh))
    
    return img_rescale


def open_tiff(fname):  # it supports images written by rasterio and tifffile
    with rasterio.open(fname, "r") as src:
        img_rasterio = src.read()

    with tifffile.TiffFile(fname) as tif:
        img_tifffile = tif.asarray()
    
    if img_rasterio.shape == img_tifffile.shape:
        return img_rasterio
    if img_rasterio.shape[0] == img_tifffile.shape[-1]:
        return img_rasterio
    return img_tifffile


def write_tiff(img_wrt, filename, metadata):

    """
    It writes a raster image to file.

    :param img_wrt: numpy array containing the data (can be 2D for single band or 3D for multiple bands)
    :param filename: file path to the output file
    :param metadata: metadata to use to write the raster to disk
    :return:
    """

    with rasterio.open(filename, "w", **metadata) as dest:

        if len(img_wrt.shape) == 2:
            
            img_wrt = img_wrt[None]

        for i in range(img_wrt.shape[0]):
            dest.write(img_wrt[i, :, :], i + 1)
    
    return filename
            

def get_meta(fname):
    
    with rasterio.open(fname, "r") as src:
        
        meta = src.meta
        
    return meta

def preprocess_example(example_list):
    
    example_list = [os.path.join(os.path.abspath(''), x) for x in example_list]
    
    return example_list


def inference_segmentor(model, imgs, custom_test_pipeline=None):
    """Inference image(s) with the segmentor.

    Args:
        model (nn.Module): The loaded segmentor.
        imgs (str/ndarray or list[str/ndarray]): Either image files or loaded
            images.

    Returns:
        (list[Tensor]): The segmentation result.
    """
    cfg = model.cfg
    device = next(model.parameters()).device  # model device
    # build the data pipeline
    test_pipeline = [LoadImageFromFile()] + cfg.data.test.pipeline[1:] if custom_test_pipeline == None else custom_test_pipeline
    test_pipeline = Compose(test_pipeline)
    # prepare data
    data = []
    imgs = imgs if isinstance(imgs, list) else [imgs]
    for img in imgs:
        img_data = {'img_info': {'filename': img}}
        img_data = test_pipeline(img_data)
        data.append(img_data)
    # print(data.shape)
    
    data = collate(data, samples_per_gpu=len(imgs))
    if next(model.parameters()).is_cuda:
        # data = collate(data, samples_per_gpu=len(imgs))
        # scatter to specified GPU
        data = scatter(data, [device])[0]
    else:
        # img_metas = scatter(data['img_metas'],'cpu')
        # data['img_metas'] = [i.data[0] for i in data['img_metas']]
        
        img_metas = data['img_metas'].data[0]
        img = data['img']
        data = {'img': img, 'img_metas':img_metas}
    
    with torch.no_grad():
        result = model(return_loss=False, rescale=True, **data)
    return result


def inference_on_file(target_image, model, custom_test_pipeline):

    target_image = target_image.name

    time_taken=-1
    
    st = time.time()
    print('Running inference...')
    try:
        result = inference_segmentor(model, target_image, custom_test_pipeline)
    except:
        print('Error: Try different channels order.')
        model.cfg.data.test.pipeline[0]['channels_last'] = True
        result = inference_segmentor(model, target_image, custom_test_pipeline)
        
    print("Output has shape: " + str(result[0].shape))

    ##### prep outputs
    mask = open_tiff(target_image)
    rgb = stretch_rgb((mask[[3, 2, 1], :, :].transpose((1,2,0))/10000*255).astype(np.uint8))
    meta = get_meta(target_image)
    mask = np.where(mask == meta['nodata'], 1, 0)
    mask = np.max(mask, axis=0)[None]
    rgb = np.where(mask.transpose((1,2,0)) == 1, 0, rgb)
    rgb = np.where(rgb < 0, 0, rgb)
    rgb = np.where(rgb > 255, 255, rgb)
    
    prediction = np.where(mask == 1, 0, result[0]*255)
    et = time.time()
    time_taken = np.round(et - st, 1)
    print(f'Inference completed in {str(time_taken)} seconds')
        
    return rgb, prediction[0]

def process_test_pipeline(custom_test_pipeline, bands=None):
    
    # change extracted bands if necessary
    if bands is not None:
        
        extract_index = [i for i, x in enumerate(custom_test_pipeline) if x['type'] == 'BandsExtract' ]
        
        if len(extract_index) > 0:
            
            custom_test_pipeline[extract_index[0]]['bands'] = eval(bands)
            
    collect_index = [i for i, x in enumerate(custom_test_pipeline) if x['type'].find('Collect') > -1]
    
    # adapt collected keys if necessary
    if len(collect_index) > 0:
        
        keys = ['img_info', 'filename', 'ori_filename', 'img', 'img_shape', 'ori_shape', 'pad_shape', 'scale_factor', 'img_norm_cfg']
        custom_test_pipeline[collect_index[0]]['meta_keys'] = keys
    
    return custom_test_pipeline

config = Config.fromfile(config_path)
config.model.backbone.pretrained=None
model = init_segmentor(config, ckpt, device='cpu')
custom_test_pipeline=process_test_pipeline(model.cfg.data.test.pipeline, None)

func = partial(inference_on_file, model=model, custom_test_pipeline=custom_test_pipeline)

with gr.Blocks() as demo:
   
    gr.Markdown(value='# Prithvi sen1floods11')
    gr.Markdown(value='''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. This demo showcases how the model was finetuned to detect water at a higher resolution than it was trained on (i.e. 10m versus 30m) using Sentinel 2 imagery from on the [sen1floods11 dataset](https://github.com/cloudtostreet/Sen1Floods11). More detailes can be found [here](https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M-sen1floods11).\n
    The user needs to provide a Sentinel 2 image with all the 12 bands (in the usual Sentinel 2) order in reflectance units multiplied by 10,000 (e.g. to save on space), with the code that is going to pull up Blue, Green, Red, Narrow NIR, SWIR, SWIR 2.
    ''')
    with gr.Row():
        with gr.Column():
            inp = gr.File()
            btn = gr.Button("Submit")
    
    with gr.Row():
        gr.Markdown(value='### Input RGB')
        gr.Markdown(value='### Model prediction (Black: Land; White: Water)')
        
    with gr.Row():
        out1=gr.Image(image_mode='RGB')
        out2 = gr.Image(image_mode='L')
    
    btn.click(fn=func, inputs=inp, outputs=[out1, out2])
    
    with gr.Row():
        gr.Examples(examples=["India_900498_S2Hand.tif",
                              "Spain_7370579_S2Hand.tif",
                              "USA_430764_S2Hand.tif"],
                    inputs=inp,
                    outputs=[out1, out2],
                    preprocess=preprocess_example,
                    fn=func,
                    cache_examples=True,
    )

demo.launch()