app.py

from fastapi import FastAPI, File, UploadFile
import uvicorn
from typing import List
from io import BytesIO
import numpy as np
import rasterio
from pydantic import BaseModel
import torch
from huggingface_hub import hf_hub_download
from mmcv import Config
from mmseg.apis import init_segmentor
import gradio as gr
from functools import partial
import time
import os

# Initialize the FastAPI app
app = FastAPI()

# Load the model and config
config_path = hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-100M-multi-temporal-crop-classification", 
                              filename="multi_temporal_crop_classification_Prithvi_100M.py", 
                              token=os.environ.get("token"))
ckpt = hf_hub_download(repo_id="ibm-nasa-geospatial/Prithvi-100M-multi-temporal-crop-classification", 
                       filename='multi_temporal_crop_classification_Prithvi_100M.pth', 
                       token=os.environ.get("token"))

config = Config.fromfile(config_path)
config.model.backbone.pretrained = None
model = init_segmentor(config, ckpt, device='cpu')

# Use the test pipeline directly
custom_test_pipeline = model.cfg.data.test.pipeline

# Define the input/output model for FastAPI
class PredictionOutput(BaseModel):
    t1: List[float]
    t2: List[float]
    t3: List[float]
    prediction: List[float]

# Define the inference function
def inference_on_file(file_path, model, custom_test_pipeline):
    with rasterio.open(file_path) as src:
        img = src.read()

    # Apply preprocessing using the custom pipeline
    processed_img = apply_pipeline(custom_test_pipeline, img)
    
    # Run inference
    output = model.inference(processed_img)
    
    # Post-process the output to get the RGB and prediction images
    rgb1 = postprocess_output(output[0])
    rgb2 = postprocess_output(output[1])
    rgb3 = postprocess_output(output[2])
    
    return rgb1, rgb2, rgb3, output

def apply_pipeline(pipeline, img):
    # Implement your custom pipeline processing here
    # This could include normalization, resizing, etc.
    return img

def postprocess_output(output):
    # Convert the model's output into an RGB image or other formats as needed
    return output

@app.post("/predict/", response_model=PredictionOutput)
async def predict(file: UploadFile = File(...)):
    # Read the uploaded file
    target_image = BytesIO(await file.read())
    
    # Save the file temporarily if needed
    with open("temp_image.tif", "wb") as f:
        f.write(target_image.getvalue())

    # Run the prediction
    rgb1, rgb2, rgb3, output = inference_on_file("temp_image.tif", model, custom_test_pipeline)
    
    # Return the results
    return {
        "t1": rgb1.tolist(),
        "t2": rgb2.tolist(),
        "t3": rgb3.tolist(),
        "prediction": output.tolist()
    }

# Optional: Serve the Gradio interface (if you still want to use it with FastAPI)
cdl_color_map = [{'value': 1, 'label': 'Natural vegetation', 'rgb': (233,255,190)},
                 {'value': 2, 'label': 'Forest', 'rgb': (149,206,147)},
                 {'value': 3, 'label': 'Corn', 'rgb': (255,212,0)},
                 {'value': 4, 'label': 'Soybeans', 'rgb': (38,115,0)},
                 {'value': 5, 'label': 'Wetlands', 'rgb': (128,179,179)},
                 {'value': 6, 'label': 'Developed/Barren', 'rgb': (156,156,156)},
                 {'value': 7, 'label': 'Open Water', 'rgb': (77,112,163)},
                 {'value': 8, 'label': 'Winter Wheat', 'rgb': (168,112,0)},
                 {'value': 9, 'label': 'Alfalfa', 'rgb': (255,168,227)},
                 {'value': 10, 'label': 'Fallow/Idle cropland', 'rgb': (191,191,122)},
                 {'value': 11, 'label': 'Cotton', 'rgb':(255,38,38)},
                 {'value': 12, 'label': 'Sorghum', 'rgb':(255,158,15)},
                 {'value': 13, 'label': 'Other', 'rgb':(0,175,77)}]


def apply_color_map(rgb, color_map=cdl_color_map):
    
    
    rgb_mapped = rgb.copy()
    
    for map_tmp in cdl_color_map:
        
        for i in range(3):
            rgb_mapped[i] = np.where((rgb[0] == map_tmp['value']) & (rgb[1] == map_tmp['value']) & (rgb[2] == map_tmp['value']), map_tmp['rgb'][i], rgb_mapped[i])
    
    return rgb_mapped    
    

def stretch_rgb(rgb):
    
    ls_pct=0
    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):
    
    with rasterio.open(fname, "r") as src:
        
        data = src.read()
        
    return data

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 process_rgb(input, mask, indexes):

    
    rgb = stretch_rgb((input[indexes, :, :].transpose((1,2,0))/10000*255).astype(np.uint8))
    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)

    return rgb

def inference_on_file(target_image, model, custom_test_pipeline):

    target_image = target_image.name
    time_taken=-1
    st = time.time()
    print('Running inference...')
    result = inference_segmentor(model, target_image, custom_test_pipeline)
    print("Output has shape: " + str(result[0].shape))

    ##### get metadata mask
    input = open_tiff(target_image)
    meta = get_meta(target_image)
    mask = np.where(input == meta['nodata'], 1, 0)
    mask = np.max(mask, axis=0)[None]
    
    rgb1 = process_rgb(input, mask, [2, 1, 0])
    rgb2 = process_rgb(input, mask, [8, 7, 6])
    rgb3 = process_rgb(input, mask, [14, 13, 12])

    result[0] = np.where(mask == 1, 0, result[0])

    et = time.time()
    time_taken = np.round(et - st, 1)
    print(f'Inference completed in {str(time_taken)} seconds')
    
    output=result[0][0] + 1
    output = np.vstack([output[None], output[None], output[None]]).astype(np.uint8)
    output=apply_color_map(output).transpose((1,2,0))
        
    return rgb1,rgb2,rgb3,output

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 multi temporal crop classification')
    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 classify crop and other land use categories using multi temporal data. More detailes can be found [here](https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M-multi-temporal-crop-classification).\n
    The user needs to provide an HLS geotiff image, including 18 bands for 3 time-step, and each time-step includes the channels described above (Blue, Green, Red, Narrow NIR, SWIR, SWIR 2) in order.
    ''')
    with gr.Row():
        with gr.Column():
            inp = gr.File()
            btn = gr.Button("Submit")
            
    with gr.Row():
        inp1=gr.Image(image_mode='RGB', scale=10, label='T1')
        inp2=gr.Image(image_mode='RGB', scale=10, label='T2')
        inp3=gr.Image(image_mode='RGB', scale=10, label='T3')
        out = gr.Image(image_mode='RGB', scale=10, label='Model prediction')
        # gr.Image(value='Legend.png', image_mode='RGB', scale=2, show_label=False)
    
    btn.click(fn=func, inputs=inp, outputs=[inp1, inp2, inp3, out])
    
    with gr.Row():
        with gr.Column():
            gr.Examples(examples=["chip_102_345_merged.tif",
                             "chip_104_104_merged.tif",
                             "chip_109_421_merged.tif"],
                            inputs=inp,
                            outputs=[inp1, inp2, inp3, out],
                            preprocess=preprocess_example,
                            fn=func,
                            cache_examples=True)
        with gr.Column():
            gr.Markdown(value='### Model prediction legend')
            gr.Image(value='Legend.png', image_mode='RGB', show_label=False)
    

demo.launch() 

if __name__ == "__main__":
    run_gradio_interface()
    uvicorn.run(app, host="0.0.0.0", port=8000)