mxnet 查看中间层结果

朴灿烈づ我的快乐病毒、 2022-09-05 15:57 145阅读 0赞

import mxnet as mx
    from mxnet import nd
    from mxnet.gluon import nn
     
    mx.cpu(), mx.gpu(), mx.gpu(0)

查看mxnet网络所有节点

import json
        with open('./model-symbol.json', 'r', encoding='utf8') as fp:
            conf = json.load(fp)
    
        # conf = json.loads(symbol.tojson())
        nodes = conf["nodes"]
        heads = set(conf["heads"][0])
    
        symbols = []
        for i, node in enumerate(nodes):
            op = node["op"]
            if op == "null" and i > 0:
                continue
            if op != "null" or i in heads:
                print(node['name'])

cpu模式下，只能返回一层

gpu(0)模式下，能返回多层结果。

查看权重  
在训练过程中，有时候我们为了debug而需要查看中间某一步的权重信息，在mxnet中，我们可以很方便的调用get\_params()方法来得到权重信息。

'''  
查看权重示例代码  
转载时注明地址：http://blog.csdn.net/u010414386?viewmode=contents

'''
    查看权重示例代码
    转载时注明地址：http://blog.csdn.net/u010414386?viewmode=contents
    '''
    import mxnet as mx
    sym, arg_params, aux_params = mx.model.load_checkpoint('resnet-50',0)#载入模型
    mod = mx.mod.Module(symbol=sym,context=mx.gpu()) #创建Module
    mod.bind(for_training=False,data_shapes=[('data',(1,3,224,224))]) #绑定，此代码为预测代码，所以training参数设为False
    mod.set_params(arg_params,aux_params)
    import numpy as np
    import cv2
    def get_image(filename):
        img = cv2.imread(filename)
        img = cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
        img = cv2.resize(img,(224,224))
        img = np.swapaxes(img,0,2)
        img = np.swapaxes(img,1,2)
        img = img[np.newaxis,:]
        return img
    from collections import namedtuple
    Batch = namedtuple('Batch',['data'])
    img = get_image('val_1000/0.jpg') #获取图片
    mod.forward(Batch([mx.nd.array(img)])) #预测结果
    ################################################
    #debug模式下，获取权重信息
    keys = mod.get_params()[0].keys() # 列出所有权重名称
    conv_w = mod.get_params()[0]['conv0_weight'] #获取想要查看的权重信息,如conv_weight
    print conv_w.asnumpy() #查看具体数值
    ################################################
    prob = mod.get_outputs()[0].asnumpy()
    y = np.argsort(np.squeeze(prob))[::-1]
    print('truth label %d; top-1 predict label %d' % (val_label[0], y[0]))

查看中间输出结果  
由于mxnet的网络由symbol组成，而symbol又属于符号式编程，所以我们不能像上面查看权重一样直接查看，我们需要把我们想看的输出结果保存下来。

'''
    方法一
    查看中间结果代码
    转载时注明地址：http://blog.csdn.net/u010414386?viewmode=contents
    '''
    import mxnet as mx
    net = mx.symbol.Variable('data')
    fc1 = mx.symbol.FullyConnected(data=net, name='fc1', num_hidden=128)
    net = mx.symbol.Activation(data=fc1, name='relu1', act_type="relu")
    net = mx.symbol.FullyConnected(data=net, name='fc2', num_hidden=64)
    out = mx.symbol.SoftmaxOutput(data=net, name='softmax')
    # 通过把两个输出组成一个group来得到自己需要查看的中间层输出结果
    group = mx.symbol.Group([fc1, out]) 
    print group.list_outputs()

方法二  
有时候我们使用别人的模型，所以无法像方法一一样在定义模型的时候就确定需要查看的中间层输出结果，  
这时候我们使用get\_internals()方法来查找自己需要查看的中间层  
转载时注明地址：http://blog.csdn.net/u010414386?viewmode=contents

这个出来是list，存放的不同的层的结果。

prob = mod.get\_outputs()

import mxnet as mx
    sym, arg_params, aux_params = mx.model.load_checkpoint('resnet-50',0)#载入模型
    ########################################################################
    args = sym.get_internals().list_outputs() #获得所有中间输出
    internals = model.symbol.get_internals()
    fc1 = internals['fc1_output']
    conv = internals['stage4_unit3_conv1_output']
    group = mx.symbol.Group([fc1, sym, conv])  #把需要输出的结果按group方式组合起来，这样就可以得到中间层的输出
    #########################################################################
    mod = mx.mod.Module(symbol=group,context=mx.gpu()) #创建Module
    mod.bind(for_training=False,data_shapes=[('data',(1,3,224,224))]) #绑定，此代码为预测代码，所以training参数设为False
    mod.set_params(arg_params,aux_params)
    import numpy as np
    import cv2
    def get_image(filename):
        img = cv2.imread(filename)
        img = cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
        img = cv2.resize(img,(224,224))
        img = np.swapaxes(img,0,2)
        img = np.swapaxes(img,1,2)
        img = img[np.newaxis,:]
        return img
    from collections import namedtuple
    Batch = namedtuple('Batch',['data'])
    img = get_image('val_1000/0.jpg') #获取图片
    mod.forward(Batch([mx.nd.array(img)])) #预测结果
    prob = mod.get_outputs()[0].asnumpy()
    y = np.argsort(np.squeeze(prob))[::-1]
    print('truth label %d; top-1 predict label %d' % (val_label[0], y[0]))

原文链接：https://blog.csdn.net/u010414386/article/details/55668880

打印所有层输出：

import mxnet as mx
    
    
    def get_output_symbol(symbol):
      """
      Parameters
      ----------
      symbol: Symbol
          Symbol to be visualized.
    
      """
      import json
      from mxnet.symbol.symbol import Symbol
      if not isinstance(symbol, Symbol):
        raise TypeError("symbol must be Symbol")
    
      conf = json.loads(symbol.tojson())
      nodes = conf["nodes"]
      heads = set(conf["heads"][0])
    
      symbols = []
      for i, node in enumerate(nodes):
        op = node["op"]
        if op == "null" and i > 0:
          continue
        if op != "null" or i in heads:
          symbols.append(node['name'])
      return symbols
    
    def debug_model(model):
      # prepare data 准备输入数据
      input_blob=mx.nd.zeros(shape=(1,3,112,112),ctx=mx.cpu())
      db = mx.io.DataBatch(data=(input_blob,))
      # get output symbol 找到特征层，获取输出节点
      symbols = get_output_symbol(model.symbol)
      symbols = [x for x in symbols if x != 'data']
      arg_params, aux_params = model.get_params()
      internals = model.symbol.get_internals()
      outputs = internals.list_outputs()
      symbols_output_name = [x + '_output' for x in symbols]
      symbols_output = [internals[x] for x in symbols_output_name]
      # 重建符号与模型
      group = mx.symbol.Group(symbols_output)
      mod = mx.mod.Module(symbol=group, context=mx.cpu())
      mod.bind(data_shapes=[('data', (1, 3, 112, 112))])  # 绑定输入shape
      mod.set_params(arg_params, aux_params)
      mod.forward(db, is_train=False)
      output = mod.get_outputs()
      output_dict = {k: v.asnumpy() for k, v in zip(symbols, output)}
      # 保存结果
      import os
      from collections import Iterable
      if not os.path.exists('output'):
        os.mkdir('output')
      for k, v in output_dict.items():
        with open('output/{}.txt'.format(k), 'w') as f:
          print('Shape is {}, data type is {}'.format(v.shape, v.dtype), file=f)
          for i, batch in enumerate(v):
            print('Batch {}:'.format(i), file=f)
            for j, channel in enumerate(batch):
              print('{}Channel {}:'.format(' ' * 4, j), file=f)
              if isinstance(channel, Iterable):
                for k, width in enumerate(channel):
                  print(' ' * 8, file=f, end='')
                  for m, height in enumerate(width):
                    print(height, end='  ', file=f)
                  print(file=f)
              else:
                print(' ' * 8 + str(channel), file=f)
    
    # 加载与训练模型
    def get_model(ctx, image_size, model_str, layer):
      _vec = model_str.split(',')
      assert len(_vec)==2
      prefix = _vec[0]
      epoch = int(_vec[1])
      print('loading',prefix, epoch)
      sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
      all_layers = sym.get_internals()
      sym = all_layers[layer+'_output']
    
      model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
    
      model.bind(data_shapes=[('data', (1, 3, image_size[0], image_size[1]))])
      model.set_params(arg_params, aux_params)
      # 打印输出shape
      arg_shape, out_shape, _ = sym.infer_shape(data=(1, 3, image_size[0], image_size[1]))
      mx.viz.print_summary(sym, {'data': (1, 3, image_size[0], image_size[1])})
      return model
    if __name__=='__main__':
      model = get_model(mx.cpu(), (112, 112), 'model-y1-test2/model,0', 'fc1')	
      debug_model(model)