Runtime Error asking all parameters to have requires_grad=True

[avinath1998]

Describe the bug
I'm trying to finetune a alexnet model and i've set the parameters except for the final layer of the model to requires_grad=False and have created a new classification layer with the desired outputs i want. However the .send() function keeps throwing a runtime error `RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

import syft
import torch
from torchvision import models
import torch.nn as nn

hook = syft.TorchHook(torch)
worker = syft.VirtualWorker(hook, id="worker")

model = models.alexnet(pretrained=True)
for param in model.parameters():
    param.requires_grad=False
model.classifier[6] = nn.Linear(model.classifier[6].in_features, 3)
model.send(worker)
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-15-a250859d9a13> in <module>
----> 1 model.send(worker)

~/implementation/PyGrid/gateway/src/syft/syft/frameworks/torch/hook/hook.py in module_send_(nn_self, force_send, *dest, **kwargs)
    608 
    609             if module_is_missing_grad(nn_self):
--> 610                 create_grad_objects(nn_self)
    611 
    612             for p in nn_self.parameters():

~/implementation/PyGrid/gateway/src/syft/syft/frameworks/torch/hook/hook.py in create_grad_objects(model)
    600             for p in model.parameters():
    601                 o = p.sum()
--> 602                 o.backward()
    603                 if p.grad is not None:
    604                     p.grad -= p.grad

~/implementation/PyGrid/gateway/src/syft/syft/generic/frameworks/hook/trace.py in trace_wrapper(*args, **kwargs)
     81                 syft.hook.trace.logs.append((command, response))
     82             else:
---> 83                 response = func(*args, **kwargs)
     84 
     85             return response

~/implementation/PyGrid/gateway/src/syft/syft/generic/frameworks/hook/hook.py in overloaded_native_method(self, *args, **kwargs)
    436                 except BaseException as e:
    437                     # we can make some errors more descriptive with this method
--> 438                     raise route_method_exception(e, self, args, kwargs)
    439 
    440             else:  # means that there is a wrapper to remove

~/implementation/PyGrid/gateway/src/syft/syft/generic/frameworks/hook/hook.py in overloaded_native_method(self, *args, **kwargs)
    432 
    433                 try:
--> 434                     response = method(*args, **kwargs)
    435 
    436                 except BaseException as e:

~/anaconda3/lib/python3.7/site-packages/torch/tensor.py in backward(self, gradient, retain_graph, create_graph)
    193                 products. Defaults to ``False``.
    194         """
--> 195         torch.autograd.backward(self, gradient, retain_graph, create_graph)
    196 
    197     def register_hook(self, hook):

~/anaconda3/lib/python3.7/site-packages/torch/autograd/__init__.py in backward(tensors, grad_tensors, retain_graph, create_graph, grad_variables)
     97     Variable._execution_engine.run_backward(
     98         tensors, grad_tensors, retain_graph, create_graph,
---> 99         allow_unreachable=True)  # allow_unreachable flag
    100 
    101 

RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn
 

[imraniac]

@avinath1998 If you are trying to change the whole model.classifier block, then your input dimension (which you put 2) is wrong. The input dimension of model.classifier should be 256x6x6.
refer: https://github.com/pytorch/vision/blob/master/torchvision/models/alexnet.py

self.classifier = nn.Sequential(
            nn.Dropout(),
            nn.Linear(256 * 6 * 6, 4096),
            nn.ReLU(inplace=True),
            nn.Dropout(),
            nn.Linear(4096, 4096),
            nn.ReLU(inplace=True),
            nn.Linear(4096, num_classes),
        )

[avinath1998]

@imraniac Oh shoot, i only added this as an example and didnt see that. Regardless, the error still comes up, ive updated the issue accordingly

[imraniac]

Have you still updated the issue with an example? Because it should be 256 * 6 * 6 and not 224

[avinath1998]

@imraniac updated accordingly, still doesn't work, the error occurs

[tudorcebere]

I will check this.

[gmuraru]

@tudorcebere, assigned you