我无法使用 tensorboardX 可视化模型的权重和偏差。这是我的模型(无论如何它很简单):
self.pipe = nn.Sequential(nn.Linear(9, 128),
nn.ReLU(),
nn.Linear(128, 256),
nn.ReLU(),
nn.Linear(256,2),
nn.Softmax()
)
def forward(self, x):
return self.pipe(x)
这是我在 pytorch 中遇到错误的地方
for name, param in net.named_parameters():
writer.add_histogram(name, param, epoch_size, bins='auto')
错误是
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
<ipython-input-70-d060d2df4423> in <module>()
1 for name, param in net.named_parameters():
----> 2 writer.add_histogram(name, param, epoch_size, bins='auto')
~\Anaconda3\lib\site-packages\tensorboardX\writer.py in add_histogram(self, tag, values, global_step, bins, walltime)
403 if isinstance(bins, six.string_types) and bins == 'tensorflow':
404 bins = self.default_bins
--> 405 self.file_writer.add_summary(
406 histogram(tag, values, bins), global_step, walltime)
407
AttributeError: 'NoneType' object has no attribute 'add_summary'
但我真的必须看到权重处于次优状态的直方图。所以我稍微更改了代码以逐步进行
param = np.array(list(net.parameters()))
print(param[0].data)
writer.add_histogram('weight', param[0].data)
繁荣!仍然是同样的错误,也许这根本没有改变。