from farmlanddataset import FarmDataset

import torch as tc青岛西海岸今晚比赛预测

from osgeo import gdal

from torchvision import transforms

import png

import numpy as np

use_cuda=True

model=tc.load('./tmp/model30') #torch.save(model,'./tmp/model{}'.format(epoch))

device = tc.device("cuda" if use_cuda else "cpu")

model=model.to(device)

model.eval()

ds=FarmDataset(istrain=False)

def createres(d,outputname):

#创建一个和ds大小相同的灰度图像BMP

driver = gdal.GetDriverByName("BMP")

#driver=ds.GetDriver()

od=driver.Create('./tmp/'+outputname,d.RasterXSize,d.RasterYSize,1)

return od

def createpng(height,width,data,outputname):

w=png.Writer(width,height,bitdepth=2,greyscale=True)

of=open('./tmp/'+outputname,'wb')

w.write_array(of,data.flat)

of.close()

return

def predict(d,outputname='tmp.bmp'):

wx=d.RasterXSize #width

wy=d.RasterYSize #height

print(wx,wy)

od=data=np.zeros((wy,wx),np.uint8)

#od=createres(d,outputname=outputname)

#ob=od.GetRasterBand(1) #得到第一个channnel

blocksize=1024

step=512

for cy in range(step,wy-blocksize,step):

for cx in range(step,wx-blocksize,step):

img=d.ReadAsArray(cx-step,cy-step,blocksize,blocksize)[0:3,:,:] #channel*h*w

if (img.sum()==0): continue

x=tc.from_numpy(img/255.0).float()

#print(x.shape)

x=x.unsqueeze(0).to(device)

r=model.forward(x)

r=tc.argmax(r.cpu()[0],0).byte().numpy() #512*512

#ob.WriteArray(r,cx,cy)

od[cy-step//2:cy+step//2,cx-step//2:cx+step//2]=r[256:step+256,256:step+256]

print(cy,cx)

#del od

createpng(wy,wx,od,outputname)

return

print("start predict.....")

predict(ds[0],'image_3_predict.png')

print("start predict 2 .....")

predict(ds[1],'image_4_predict.png')

相关阅读：