Skip to content

Gradient Descent, Signs of Healthy Training: Need to detach() tensor. #28

Description

@ethman

We need to detach a tensor in the "Signs of Healthy Training" section on the Gradient Descent page.

Here's the cell:

fig, ax = plt.subplots(1, 2, figsize=(15, 5))

for j, lr in enumerate(LEARNING_RATES):
    ax[0].plot(grad_norms[j], label=f'Learning rate: {lr}')
    ax[0].legend()
    ax[0].set_xlabel('Iteration')
    ax[0].set_ylabel('grad norm')
    ax[0].set_title('Gradient norm for each learning rate')
    
for j, lr in enumerate(LEARNING_RATES):
    ax[1].plot(np.log10(losses[j]), label=f'Learning rate: {lr}')
    ax[1].legend()
    ax[1].set_xlabel('Iteration')
    ax[1].set_ylabel('log(loss)')
    ax[1].set_title('Loss for each learning rate')
plt.show()

And the cell's current output:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-18-91a4bf5517c3> in <module>
      9 
     10 for j, lr in enumerate(LEARNING_RATES):
---> 11     ax[1].plot(np.log10(losses[j]), label=f'Learning rate: {lr}')
     12     ax[1].legend()
     13     ax[1].set_xlabel('Iteration')

/opt/hostedtoolcache/Python/3.7.10/x64/lib/python3.7/site-packages/torch/tensor.py in __array__(self, dtype)
    619             return handle_torch_function(Tensor.__array__, (self,), self, dtype=dtype)
    620         if dtype is None:
--> 621             return self.numpy()
    622         else:
    623             return self.numpy().astype(dtype, copy=False)

RuntimeError: Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions