Code that I used to extract the weights from .npy file

get_weights_mtcnn.py

import numpy as np
#ht width channels filter_nos : provided
#filter_nos channels h w : output
# file_open = open('mtcnn_layer_wise_wts.txt', 'w')
file_rnet_wts = open('mtcnn_rnet_wts.txt','w')
# file_onet_wts = open('mtcnn_onet_wts.txt','w')

def print4d(mat, file_open):

  file_open.write("["+str(mat.shape[0])+"]"+"["+str(mat.shape[1])+"]"+"["+str(mat.shape[2])+"]"+"["+str(mat.shape[3])+"]"+" = "+"{")
  for i in range(mat.shape[0]):

 
    file_open.write("{")

    for j in range(mat.shape[1]):

    
      file_open.write("{")

      for k in range(mat.shape[2]):

        
        file_open.write("{")

        for l in range(mat.shape[3]):
          

          file_open.write(str(mat[i][j][k][l]))

          if l!=mat.shape[3]-1:
            file_open.write(",")
          else:
            file_open.write("}")

        if k!=mat.shape[2]-1:
          file_open.write(",")
        else:
          file_open.write("}")

      if j!=mat.shape[1]-1:
        file_open.write(",")
      else:
        file_open.write("}")

    if i!=mat.shape[0]-1:
      file_open.write(",")

    else:
      file_open.write("};\n")

def print1d(mat,file_open):
    file_open.write("["+str(mat.shape[0])+"]"+" = {")
    for i in range(len(mat)):
        file_open.write(str(mat[i]))
        if i!=len(mat) -1:
            file_open.write(", ")
        else:
            file_open.write("};")

    file_open.write("\n")

def print2d(mat,file_open):

    file_open.write("["+str(mat.shape[0])+"]"+"["+str(mat.shape[1])+"]"+" = {")
    for i in range(mat.shape[0]):
      file_open.write('{')
      for j in range(mat.shape[1]):
          file_open.write(str(mat[i][j]))
          if j!=mat.shape[1] -1:
              file_open.write(", ")
          else:
              file_open.write("}")

      if i!=mat.shape[0]-1:
        file_open.write(", ")
      else:
        file_open.write("}")
    file_open.write(";\n")

a = np.load('mtcnn_weights.npy', allow_pickle=True).tolist()

# # ###################################### o-net #################################

# ################################# conv 1  ####################################
# onet_conv_3d_1=a['onet'][0]
# onet_conv_3d_1 = np.transpose(onet_conv_3d_1,(3,2,1,0))
# file_onet_wts.write('conv1')
# print4d(onet_conv_3d_1,file_onet_wts)

# ############################## bias 1 ######################################
# onet_bias_1 = a['onet'][1]
# print(onet_bias_1.shape)
# file_onet_wts.write('bias_1')
# print1d(onet_bias_1,file_onet_wts)

# ############################### prelu 1 ####################################
# onet_prelu1 = a['onet'][2]
# print(onet_prelu1.shape)
# file_onet_wts.write('prelu_1')
# onet_prelu1 = onet_prelu1.flatten()
# print1d(onet_prelu1,file_onet_wts)

# ################################# conv 2  ####################################
# onet_conv_3d_2=a['onet'][3]
# onet_conv_3d_2 = np.transpose(onet_conv_3d_2,(3,2,1,0))
# file_onet_wts.write('conv2')
# print4d(onet_conv_3d_2,file_onet_wts)

# ############################## bias 2 ######################################
# onet_bias_2 = a['onet'][4]
# print(onet_bias_2.shape)
# file_onet_wts.write('bias_2')
# print1d(onet_bias_2,file_onet_wts)

# ############################### prelu 2 ####################################
# onet_prelu2 = a['onet'][5]
# print(onet_prelu2.shape)
# file_onet_wts.write('prelu_2')
# onet_prelu2 = onet_prelu2.flatten()
# print1d(onet_prelu2,file_onet_wts)

# ################################# conv 3  ####################################
# onet_conv_3d_3=a['onet'][6]
# onet_conv_3d_3 = np.transpose(onet_conv_3d_3,(3,2,1,0))
# file_onet_wts.write('conv3')
# print4d(onet_conv_3d_3,file_onet_wts)

# ############################## bias 3 ######################################
# onet_bias_3 = a['onet'][7]
# print(onet_bias_3.shape)
# file_onet_wts.write('bias_3')
# print1d(onet_bias_3,file_onet_wts)

# ############################### prelu 3 ####################################
# onet_prelu3 = a['onet'][8]
# print(onet_prelu3.shape)
# file_onet_wts.write('prelu_1')
# onet_prelu3 = onet_prelu3.flatten()
# print1d(onet_prelu3,file_onet_wts)

# ################################# conv 4  ####################################
# onet_conv_3d_4=a['onet'][9]
# onet_conv_3d_4 = np.transpose(onet_conv_3d_4,(3,2,1,0))
# file_onet_wts.write('conv4')
# print4d(onet_conv_3d_4,file_onet_wts)

# ############################## bias 4 ######################################
# onet_bias_4 = a['onet'][10]
# print(onet_bias_4.shape)
# file_onet_wts.write('bias_4')
# print1d(onet_bias_4,file_onet_wts)

# ############################### prelu 4 ####################################
# onet_prelu4 = a['onet'][11]
# print(onet_prelu4.shape)
# file_onet_wts.write('prelu_4')
# onet_prelu4 = onet_prelu4.flatten()
# print1d(onet_prelu4,file_onet_wts)

# ############################## Dense-1-mat ###################################
# onet_dense_mat=a['onet'][12]
# onet_dense_mat = np.transpose(onet_dense_mat,(1,0))
# print(onet_dense_mat.shape)
# file_onet_wts.write('dense_1_mat')
# print2d(onet_dense_mat,file_onet_wts)

# ########################### Dense-1-bias ####################################
# onet_dense_bias_1 = a['onet'][13]

# print(onet_dense_bias_1.shape)
# file_onet_wts.write('dense_1_bias')
# onet_dense_bias_1 = onet_dense_bias_1.flatten()
# print1d(onet_dense_bias_1,file_onet_wts)

# ############################### prelu 5 ####################################
# onet_prelu5 = a['onet'][14]
# print(onet_prelu5.shape)
# file_onet_wts.write('prelu_5')
# onet_prelu5 = onet_prelu5.flatten()
# print1d(onet_prelu5,file_onet_wts)

# ############################## Dense-2-mat ###################################
# onet_dense_mat2=a['onet'][15]
# onet_dense_mat2 = np.transpose(onet_dense_mat2,(1,0))
# print(onet_dense_mat2.shape)
# file_onet_wts.write('dense_2_mat')
# print2d(onet_dense_mat2,file_onet_wts)

# ########################### Dense-2-bias ####################################
# onet_dense_bias_2 = a['onet'][16]
# print(onet_dense_bias_2.shape)
# file_onet_wts.write('dense_2_bias')
# onet_dense_bias_2 = onet_dense_bias_2.flatten()
# print1d(onet_dense_bias_2,file_onet_wts)

# ############################## Dense-3-mat ###################################
# onet_dense_mat3=a['onet'][17]
# onet_dense_mat3 = np.transpose(onet_dense_mat3,(1,0))
# print(onet_dense_mat3.shape)
# file_onet_wts.write('dense_3_mat')
# print2d(onet_dense_mat3,file_onet_wts)

# ########################### Dense-3-bias ####################################
# onet_dense_bias_3 = a['onet'][18]
# print(onet_dense_bias_3.shape)
# file_onet_wts.write('dense_3_bias')
# onet_dense_bias_3 = onet_dense_bias_3.flatten()
# print1d(onet_dense_bias_3,file_onet_wts)

# ############################## Dense-4-mat ###################################
# onet_dense_mat4=a['onet'][19]
# onet_dense_mat4 = np.transpose(onet_dense_mat4,(1,0))
# print(onet_dense_mat4.shape)
# file_onet_wts.write('dense_4_mat')
# print2d(onet_dense_mat4,file_onet_wts)

# ########################### Dense-4-bias ####################################
# onet_dense_bias_4 = a['onet'][20]
# print(onet_dense_bias_4.shape)
# file_onet_wts.write('dense_4_bias')
# onet_dense_bias_4 = onet_dense_bias_4.flatten()
# print1d(onet_dense_bias_4,file_onet_wts)

# ################################## weights for o-net ends ####################

# ###################################### r-net ###################################

# ################################# conv 1  ####################################
# rnet_conv_3d_1=a['rnet'][0]
# rnet_conv_3d_1 = np.transpose(rnet_conv_3d_1,(3,2,1,0))
# file_rnet_wts.write('conv1')
# print4d(rnet_conv_3d_1,file_rnet_wts)
# print(rnet_conv_3d_1.shape)

# ############################## bias 1 ######################################
# rnet_bias_1 = a['rnet'][1]
# print(rnet_bias_1.shape)
# file_rnet_wts.write('bias_1')
# print1d(rnet_bias_1,file_rnet_wts)

# ############################### prelu 1 ####################################
# rnet_prelu1 = a['rnet'][2]
# print(rnet_prelu1.shape)
# file_rnet_wts.write('prelu_1')
# rnet_prelu1 = rnet_prelu1.flatten()
# print1d(rnet_prelu1,file_rnet_wts)

# ############################# conv2 ######################################
# rnet_conv_3d_2=a['rnet'][3]
# rnet_conv_3d_2 = np.transpose(rnet_conv_3d_2,(3,2,1,0))
# file_rnet_wts.write('conv2')
# print4d(rnet_conv_3d_2,file_rnet_wts)
# print(rnet_conv_3d_2.shape)

# ############################## bias 2 ######################################
# rnet_bias_2 = a['rnet'][4]
# print(rnet_bias_2.shape)
# file_rnet_wts.write('bias_2')
# print1d(rnet_bias_2,file_rnet_wts)

# ############################ prelu 2 ####################################
# rnet_prelu2 = a['rnet'][5]
# print(rnet_prelu2.shape)
# file_rnet_wts.write('prelu_2')
# rnet_prelu2 = rnet_prelu2.flatten()
# print1d(rnet_prelu2,file_rnet_wts)

# ############################# conv3 ######################################
# rnet_conv_3d_3=a['rnet'][6]
# rnet_conv_3d_3 = np.transpose(rnet_conv_3d_3,(3,2,1,0))
# file_rnet_wts.write('conv3')
# print4d(rnet_conv_3d_3,file_rnet_wts)
# print(rnet_conv_3d_3.shape)

# ############################## bias 3 ######################################
# rnet_bias_3 = a['rnet'][7]
# print(rnet_bias_3.shape)
# file_rnet_wts.write('bias_3')
# print1d(rnet_bias_3,file_rnet_wts)

# ############################ prelu 3 ####################################
# rnet_prelu3 = a['rnet'][8]
# print(rnet_prelu3.shape)
# file_rnet_wts.write('prelu_3')
# rnet_prelu3 = rnet_prelu3.flatten()
# print1d(rnet_prelu3,file_rnet_wts)

# ############################## Dense-1-mat ###################################
rnet_dense_mat=a['rnet'][9]
print(rnet_dense_mat.shape)
file_rnet_wts.write('dense_1_mat')
print2d(rnet_dense_mat,file_rnet_wts)

# ########################### Dense-1-bias ####################################
# rnet_dense_bias = a['rnet'][10]
# print(rnet_dense_bias.shape)
# file_rnet_wts.write('dense_1_bias')
# rnet_dense_bias = rnet_dense_bias.flatten()
# print1d(rnet_dense_bias,file_rnet_wts)

# ############################ prelu 4 ####################################
# rnet_prelu4 = a['rnet'][11]
# print(rnet_prelu4.shape)
# file_rnet_wts.write('prelu_4')
# rnet_prelu4 = rnet_prelu4.flatten()
# print1d(rnet_prelu4,file_rnet_wts)

# ############################## Dense 2a-mat ###################################
# rnet_dense_2a_mat=a['rnet'][12]
# print('rnet dense 2a mat shape: ',rnet_dense_2a_mat.shape)
# file_rnet_wts.write('rnet_dense_2a_mat')
# print2d(rnet_dense_2a_mat,file_rnet_wts)

# ########################### Dense-2a-bias ####################################
# rnet_dense_2a_bias = a['rnet'][13]
# print(rnet_dense_2a_bias.shape)
# file_rnet_wts.write('rnet_dense_2a_bias')
# rnet_dense_2a_bias = rnet_dense_2a_bias.flatten()
# print1d(rnet_dense_2a_bias,file_rnet_wts)

# ############################## Dense 2b-mat ###################################
# rnet_dense_2b_mat=a['rnet'][14]
# print('rnet dense 2b mat shape: ',rnet_dense_2b_mat.shape)
# file_rnet_wts.write('rnet_dense_2b_mat')
# print2d(rnet_dense_2b_mat,file_rnet_wts)

# ########################### Dense-2b-bias ####################################
# rnet_dense_2b_bias = a['rnet'][15]
# print(rnet_dense_2b_bias.shape)
# file_rnet_wts.write('rnet_dense_2b_bias')
# rnet_dense_2b_bias = rnet_dense_2b_bias.flatten()
# print1d(rnet_dense_2b_bias,file_rnet_wts)

# ##################################### 1st conv layer #######################
# conv_2d_1 = a['pnet'][0]
# conv_2d_1 = np.transpose(conv_2d_1,(3,2,1,0))
# file_open.write('conv1')
# print4d(conv_2d_1,file_open)
# print(conv_2d_1.shape)

# ############################################################################

# ##################################### bias conv1 layer #######################
# bias_1 = a['pnet'][1]
# print(bias_1.shape)
# file_open.write('bias_1')
# print1d(bias_1,file_open)

# ############################################################################

# ##################################### prelu 1 layer #######################
# prelu1 = a['pnet'][2]
# print(prelu1.shape)
# file_open.write('prelu_1')
# prelu1 = prelu1.flatten()
# print1d(prelu1,file_open)

# ############################################################################

# ##################################### 2nd conv layer #######################
# conv_2d_2 = a['pnet'][3]
# conv_2d_2 = np.transpose(conv_2d_2,(3,2,1,0))
# file_open.write('conv2')
# print4d(conv_2d_2,file_open)
# print(conv_2d_2.shape)

# ############################################################################
# ##################################### bias conv2 layer #######################
# bias_2 = a['pnet'][4]
# print(bias_2.shape)
# file_open.write('bias_2')
# print1d(bias_2,file_open)

# ############################################################################

# ##################################### prelu 2 layer #######################
# prelu2 = a['pnet'][5]
# print(prelu2.shape)
# file_open.write('prelu_2')
# prelu2 = prelu2.flatten()
# print1d(prelu2,file_open)

# ############################################################################

# ##################################### 3rd conv layer #######################
# conv_2d_3 = a['pnet'][6]
# conv_2d_3 = np.transpose(conv_2d_3,(3,2,1,0))
# file_open.write('conv3')
# print4d(conv_2d_3,file_open)
# print(conv_2d_3.shape)

# ############################################################################

# ##################################### bias conv3 layer #######################
# bias_3 = a['pnet'][7]
# print(bias_3.shape)
# file_open.write('bias_3')
# print1d(bias_3,file_open)

# ############################################################################

# ##################################### prelu 3 layer #######################
# prelu3 = a['pnet'][8]
# print(prelu3.shape)
# file_open.write('prelu_3')
# prelu3 = prelu3.flatten()
# print1d(prelu3,file_open)

# ############################################################################

# ##################################### 4a conv layer #######################
# conv_2d_4a = a['pnet'][9]
# conv_2d_4a = np.transpose(conv_2d_4a,(3,2,1,0))
# file_open.write('conv4a')
# print4d(conv_2d_4a,file_open)
# print(conv_2d_4a.shape)

# ############################################################################

# ##################################### bias conv4a layer #######################
# bias_4a = a['pnet'][10]
# print(bias_4a.shape)
# file_open.write('bias_4a')
# print1d(bias_4a,file_open)

# ############################################################################

# ##################################### 4b conv layer #######################
# conv_2d_4b = a['pnet'][11]
# conv_2d_4b = np.transpose(conv_2d_4b,(3,2,1,0))
# file_open.write('conv4b')
# print4d(conv_2d_4b,file_open)
# print(conv_2d_4b.shape)

# ############################################################################

# ##################################### bias conv4b layer #######################
# bias_4b = a['pnet'][12]
# print(bias_4b.shape)
# file_open.write('bias_4b')
# print1d(bias_4b,file_open)

# ############################################################################