Dilate_conv2D

import numpy as np
import sys
from Convolution1 import Convolution2D
import os
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"   # see issue #152
os.environ["CUDA_VISIBLE_DEVICES"] = ""
def Dilate_conv2D(input_array,kernel,kernel_b=None,dilate_rate=(1,1),padding='same'):
    img_h, img_w, n_channels = input_array.shape
    feature_num, k_h, k_w, k_n_channels = kernel.shape
    d_h,d_w = dilate_rate
    if n_channels != k_n_channels:
        print('n_channels_error')
        sys.exit()
    if len(input_array.shape) != len(kernel.shape) - 1:
        print ('length_error')
        sys.exit()
    pad_h = d_h-1
    pad_w = d_w-1
    if padding == 'same':
        new_kh = (pad_h)*(k_h+1)+k_h
        new_kw = (pad_w)*(k_w+1)+k_w
        y_range = range(pad_h,new_kh,d_h)
        x_range = range(pad_w,new_kw,d_w)
        new_kernel = np.zeros(shape=(feature_num,new_kh,new_kw,k_n_channels),dtype='float32')
        for j in y_range:
            for i in x_range:
                new_kernel[:,j,i,:] = kernel[:,j//d_h,i//d_w,:]
        return Convolution2D(input_array,kenel=new_kernel,kernel_b=None,stride=(1,1),padding=padding)
    else:
        new_kh = pad_h*(k_h-1)+k_h
        new_kw = pad_w*(k_w-1)+k_w
        y_range=range(0,new_kh,d_h)
        x_range=range(0,new_kw,d_w)
        new_kernel = np.zeros(shape=(feature_num,new_kh,new_kw,k_n_channels),dtype='float32')
        for j in y_range:
            for i in x_range:
                new_kernel[:,j,i,:] = kernel[:,j//d_h,i//d_w,:]
        return Convolution2D(input_array,kenel=new_kernel,kernel_b=None,stride=(1,1),padding=padding)
if __name__ == '__main__':
    input_data = np.random.randint(0,10,(12,12,1))
    kernel = np.random.randint(0,10,(2,3,3,1))
    d_h,d_w = 3,3
    result = Dilate_conv2D(input_data,kernel,kernel_b=None,dilate_rate=(d_h,d_w),padding='same')
    print(result)
    import tensorflow as tf
    data = np.asarray(input_data,dtype='float32')
    kernel = np.asarray(kernel,dtype='float32')
    data = np.expand_dims(data,axis=0)
    kernel = np.transpose(kernel,(1,2,3,0))
    const_input = tf.constant(data, tf.float32)
    const_weights = tf.constant(kernel, tf.float32)
    y2 = tf.nn.atrous_conv2d(data, kernel,rate=3, padding="SAME")
    sess = tf.Session()
    print(sess.run(y2).squeeze(axis=0))
    print(result.shape)
    print(sess.run(y2).shape)