说话人识别LSTM.zip

import torch import numpy as np import math # model = torch.load('model_best.pth.tar') params = torch.load('/home/huke/python_code/object_detection/point_target/v10/model_compression_WbWtAb/Result/07-10_21-11-55/nin_gc.pth') # 4-layers;max channel 64 print("key in params:") for key in params: print(key + ":") print() print("key in params state_dict:") for key in params['state_dict']: print(key + ":") print() mappp = params['state_dict'] # print(model['best_prec1']) print(params['min_avg_loss']) # ********************* W(模型参数)量化(三/二值) *********************** def meancenter_clampConvParams(w): mean = w.data.mean(1, keepdim=True) w.data.sub(mean) # W中心化(C方向) w.data.clamp(-1.0, 1.0) # W截断 return w def Binary_w(input): output = torch.sign(input) return output def binary(input): output = Binary_w(input) return output def weight_tnn_bin(input): output = meancenter_clampConvParams(input) # W中心化+截断 # **************** channel级 - E(|W|) **************** E = torch.mean(torch.abs(output), (3, 2, 1), keepdim=True) # **************** α(缩放因子) **************** alpha = E # ************** W —— +-1 ************** output = binary(output) return output # ''' print("-"*60) print("conv1:") print("-"*60) # print(mappp['tnn_bin_conv1.tnn_bin_conv.weight'].shape) print('tnn_bin_conv1.tnn_bin_conv.bias') print(mappp['tnn_bin_conv1.tnn_bin_conv.bias'].shape) print(mappp['tnn_bin_conv1.tnn_bin_conv.bias']) print() print("-"*60) print("bn1:") print("-"*60) print('tnn_bin_conv1.bn.weight') print(mappp['tnn_bin_conv1.bn.weight'].shape) print(mappp['tnn_bin_conv1.bn.weight']) print('tnn_bin_conv1.bn.bias') print(mappp['tnn_bin_conv1.bn.bias'].shape) print(mappp['tnn_bin_conv1.bn.bias']) print('tnn_bin_conv1.bn.running_mean') print(mappp['tnn_bin_conv1.bn.running_mean'].shape) print(mappp['tnn_bin_conv1.bn.running_mean']) print('tnn_bin_conv1.bn.running_var') print(mappp['tnn_bin_conv1.bn.running_var'].shape) print(mappp['tnn_bin_conv1.bn.running_var']) print('tnn_bin_conv1.bn.num_batches_tracked') print(mappp['tnn_bin_conv1.bn.num_batches_tracked'].shape) print(mappp['tnn_bin_conv1.bn.num_batches_tracked']) print() print("-"*60) print("conv2:") print("-"*60) # print(mappp['tnn_bin_conv2.tnn_bin_conv.weight'].shape) print('tnn_bin_conv2.tnn_bin_conv.bias') print(mappp['tnn_bin_conv2.tnn_bin_conv.bias'].shape) print(mappp['tnn_bin_conv2.tnn_bin_conv.bias']) print() print("-"*60) print("bn2:") print("-"*60) print('tnn_bin_conv2.bn.weight') print(mappp['tnn_bin_conv2.bn.weight'].shape) print(mappp['tnn_bin_conv2.bn.weight']) print('tnn_bin_conv2.bn.bias') print(mappp['tnn_bin_conv2.bn.bias'].shape) print(mappp['tnn_bin_conv2.bn.bias']) print('tnn_bin_conv2.bn.running_mean') print(mappp['tnn_bin_conv2.bn.running_mean'].shape) print(mappp['tnn_bin_conv2.bn.running_mean']) print('tnn_bin_conv2.bn.running_var') print(mappp['tnn_bin_conv2.bn.running_var'].shape) print(mappp['tnn_bin_conv2.bn.running_var']) print('tnn_bin_conv2.bn.num_batches_tracked') print(mappp['tnn_bin_conv2.bn.num_batches_tracked'].shape) print(mappp['tnn_bin_conv2.bn.num_batches_tracked']) print() print("-"*60) print("conv3:") print("-"*60) # print(mappp['tnn_bin_conv3.tnn_bin_conv.weight'].shape) print('tnn_bin_conv3.tnn_bin_conv.bias') print(mappp['tnn_bin_conv3.tnn_bin_conv.bias'].shape) print(mappp['tnn_bin_conv3.tnn_bin_conv.bias']) print() print("-"*60) print("bn3:") print("-"*60) print('tnn_bin_conv3.bn.weight') print(mappp['tnn_bin_conv3.bn.weight'].shape) print(mappp['tnn_bin_conv3.bn.weight']) print('tnn_bin_conv3.bn.bias') print(mappp['tnn_bin_conv3.bn.bias'].shape) print(mappp['tnn_bin_conv3.bn.bias']) print('tnn_bin_conv3.bn.running_mean') print(mappp['tnn_bin_conv3.bn.running_mean'].shape) print(mappp['tnn_bin_conv3.bn.running_mean']) print('tnn_bin_conv3.bn.running_var') print(mappp['tnn_bin_conv3.bn.running_var'].shape) print(mappp['tnn_bin_conv3.bn.running_var']) print('tnn_bin_conv3.bn.num_batches_tracked') print(mappp['tnn_bin_conv3.bn.num_batches_tracked'].shape) print(mappp['tnn_bin_conv3.bn.num_batches_tracked']) print() print("-"*60) print("conv4:") print("-"*60) # print(mappp['tnn_bin_conv4.tnn_bin_conv.weight'].shape) print('tnn_bin_conv4.tnn_bin_conv.bias') print(mappp['tnn_bin_conv4.tnn_bin_conv.bias'].shape) print(mappp['tnn_bin_conv4.tnn_bin_conv.bias']) print() print("-"*60) print("bn4:") print("-"*60) print('tnn_bin_conv4.bn.weight') print(mappp['tnn_bin_conv4.bn.weight'].shape) print(mappp['tnn_bin_conv4.bn.weight']) print('tnn_bin_conv4.bn.bias') print(mappp['tnn_bin_conv4.bn.bias'].shape) print(mappp['tnn_bin_conv4.bn.bias']) print('tnn_bin_conv4.bn.running_mean') print(mappp['tnn_bin_conv4.bn.running_mean'].shape) print(mappp['tnn_bin_conv4.bn.running_mean']) print('tnn_bin_conv4.bn.running_var') print(mappp['tnn_bin_conv4.bn.running_var'].shape) print(mappp['tnn_bin_conv4.bn.running_var']) print('tnn_bin_conv4.bn.num_batches_tracked') print(mappp['tnn_bin_conv4.bn.num_batches_tracked'].shape) print(mappp['tnn_bin_conv4.bn.num_batches_tracked']) print() # for i in range(16): # for j in range(1): # print('------------------------------') # print('i---j:',i,j) # print('------------------------------') # print(mappp['tnn_bin_conv1.tnn_bin_conv.weight'][i][j]) # # conv1_weight = weight_tnn_bin(mappp['tnn_bin_conv1.tnn_bin_conv.weight']) # # print("-"*50) # print("after binary") # for i in range(16): # for j in range(1): # print('------------------------------') # print('i---j:',i,j) # print('------------------------------') # print(conv1_weight[i][j]) # ''' #Layer1 conv1_weight = weight_tnn_bin(mappp['tnn_bin_conv1.tnn_bin_conv.weight']) conv1_weight = conv1_weight.cpu().numpy() conv1_weight = np.reshape(conv1_weight,(16*1*3,3)) np.savetxt('./float_parameters/conv1_weight.txt',conv1_weight,fmt='%2.0f',delimiter=',') conv1_bias = mappp['tnn_bin_conv1.tnn_bin_conv.bias'].cpu().numpy() np.savetxt('./float_parameters/conv1_bias.txt',conv1_bias,fmt='%2.10f',delimiter=',') BN1_weight = mappp['tnn_bin_conv1.bn.weight'].cpu().numpy() np.savetxt('./float_parameters/BN1_weight.txt',BN1_weight,fmt='%2.8f',delimiter=',') BN1_bias = mappp['tnn_bin_conv1.bn.bias'].cpu().numpy() np.savetxt('./float_parameters/BN1_bias.txt',BN1_bias,fmt='%2.8f',delimiter=',') BN1_running_mean = mappp['tnn_bin_conv1.bn.running_mean'].cpu().numpy() np.savetxt('./float_parameters/BN1_running_mean.txt',BN1_running_mean,fmt='%2.8f',delimiter=',') BN1_running_var = mappp['tnn_bin_conv1.bn.running_var'].cpu().numpy() np.savetxt('./float_parameters/BN1_running_var.txt',BN1_running_var,fmt='%4.8f',delimiter=',') #Layer2 conv2_weight = weight_tnn_bin(mappp['tnn_bin_conv2.tnn_bin_conv.weight']) conv2_weight = conv2_weight.cpu().numpy() conv2_weight = np.reshape(conv2_weight,(32*16*3,3)) np.savetxt('./float_parameters/conv2_weight.txt',conv2_weight,fmt='%2.0f',delimiter=',') conv2_bias = mappp['tnn_bin_conv2.tnn_bin_conv.bias'].cpu().numpy() np.savetxt('./float_parameters/conv2_bias.txt',conv2_bias,fmt='%2.10f',delimiter=',') BN2_weight = mappp['tnn_bin_conv2.bn.weight'].cpu().numpy() np.savetxt('./float_parameters/BN2_weight.txt',BN2_weight,fmt='%2.8f',delimiter=',') BN2_bias = mappp['tnn_bin_conv2.bn.bias'].cpu().numpy() np.savetxt('./float_parameters/BN2_bias.txt',BN2_bias,fmt='%2.8f',delimiter=',') BN2_running_mean = mappp['tnn_bin_conv2.bn.running_mean'].cpu().numpy() np.savetxt('./float_parameters/BN2_running_mean.txt',BN2_running_mean,fmt='%2.8f',delimiter=',') BN2_running_var = mappp['tnn_bin_conv2.bn.running_var'].cpu().numpy() np.savetxt('./float_parameters/BN2_running_var.txt',BN2_running_var,fmt='%4.8f',delimiter=',') #Layer3 conv3_weight = weight_tnn_bin(mappp['tnn_bin_conv3.tnn_bin_conv.weight']) conv3_weight = conv3_weight.cpu().numpy() conv3_weight = np.reshape(co