nowcasting.zip

import cv2 import skimage import numpy as np import keras import os import matplotlib.pyplot as plt import keras.backend as K import tensorflow as tf from PIL import Image from keras.models import Sequential,Model,load_model,model_from_json from keras.layers import Input,Add,Conv2D,Conv3D,Concatenate,ConvLSTM2D,BatchNormalization,LeakyReLU,MaxPool2D,UpSampling2D,TimeDistributed from keras.optimizers import Adam,SGD from keras.callbacks import ModelCheckpoint,EarlyStopping from utilsp import RadarData os.environ['CUDA_VISIBLE_DEVICES'] = '1' IMAGE_PATH = 'radar/' #IMAGE_PATH = '/content/drive/MyDrive/nowcasting/train/' INPUT_SEQUENCE = np.array([]) NEXT_SEQUENCE = np.array([]) smooth = 1e-9 class Config(object): WIDTH = 64 HEIGHT = 64 def ssim(x,y,max_val=255.0): return tf.image.ssim(x,y,max_val) def psnr(x,y,max_val=255.0): return tf.image.psnr(x,y,max_val) #recall def POD(x,y): y_pos = K.clip(x,0,1) y_pred_pos = K.clip(y,0,1) y_pred_neg = 1 - y_pred_pos tp = K.sum(y_pos * y_pred_pos) fn = K.sum(y_pos * y_pred_neg) return (tp+smooth)/(tp+fn+smooth) def FAR(x,y): y_pred_pos = K.clip(y,0,1) y_pos = K.clip(x,0,1) y_neg = 1 - y_pos tp = K.sum(y_pos * y_pred_pos) fp = K.sum(y_neg * y_pred_pos) return (fp)/(tp+fp+smooth) def Net(): contentInput = Input(shape=(None,Config.WIDTH,Config.HEIGHT,1),name='content_input') x = ConvLSTM2D(64,(3,3),padding='same',dropout=0.2,return_sequences=True,kernel_initializer='he_normal',name='layer1')(contentInput) x = BatchNormalization()(x) x = ConvLSTM2D(64,(3,3),padding='same',dropout=0.2,return_sequences=True,kernel_initializer='he_normal',name='layer2')(x) x = BatchNormalization()(x) x = ConvLSTM2D(64,(3,3),padding='same',dropout=0.2,return_sequences=True,kernel_initializer='he_normal',name='layer3')(x) x = BatchNormalization()(x) x = ConvLSTM2D(64,(3,3),padding='same',dropout=0.2,return_sequences=True,kernel_initializer='he_normal',name='layer4')(x) x = BatchNormalization()(x) x = ConvLSTM2D(64,(3,3),padding='same',dropout=0.2,return_sequences=True,kernel_initializer='he_normal',name='layer5')(x) x = BatchNormalization()(x) predictions = Conv3D(filters=1,kernel_size=(3,3,3),activation='linear',padding='same',data_format='channels_last')(x) model = Model(inputs=contentInput,outputs=predictions) return model def train(*args): model,epochs,train_gen = args metrics = ['accuracy',ssim,psnr,POD,FAR] optim = Adam(lr=1e-3,beta_1=0.9,beta_2=0.999,amsgrad=False) model.compile(loss='logcosh',optimizer=optim,metrics=metrics) checkpoint = ModelCheckpoint(filepath='weights-2-{epoch:03d}-{loss:.2f}.h5',monitor='loss',save_best_only=False,save_weights_only=True) checkpoint.set_model(model) model.fit_generator(train_gen,epochs=epochs,callbacks=[checkpoint]) if __name__ == '__main__': model = Net() model.summary() tf.keras.utils.plot_model(model,to_file='model.png',show_shapes=True,show_layer_names=True) images = sorted(os.listdir(IMAGE_PATH))[:500] #print(images) """ for image in images: vis(os.path.join('train/',image)) """ #gen = RadarData(images=images) #train(model,80,gen) """ x,y = next(iter(gen)) print(x.shape,y.shape) for i in range(2): for j in range(16): img_x = x[i,j,:,:,:] img_y = y[i,j,:,:,:] cv2.imshow('x',img_x) cv2.imshow('y',img_y) cv2.waitKey(0) """