Grad from model param undefined

1

Hello I am trying to play around with the multi process implementation (link and I ran into a problem with the grad being None when I initialize my model. Here is a simple example:

from __future__ import print_function

import argparse
import os
import sys

import torch
import torch.multiprocessing as mp
import torch.nn as nn
import torch.nn.functional as F

class MyModel(torch.nn.Module):
    def __init__(self):
        super(MyModel, self).__init__()
        self.main = nn.Sequential(
            nn.Linear(256,6),nn.Softmax())
    def forward(self,x):
        return self.main(x)

def train(model):
    # This for loop will break sharing of gradient buffers. It's not
    # necessary but it reduces the contention, and has a small memory cost
    # (equal to the total size of parameters).
    for param in model.parameters():
        param.grad.data = param.grad.data.clone()
    # Construct data_loader, optimizer, etc.
    for data, labels in data_loader:
        input1 = torch.ones(256)
        optimizer.zero_grad()
        loss_fn(model(input1), torch.zeros(6)).backward()
        optimizer.step()  # This will update the shared parameters

if __name__ == '__main__':
    num_processes = 4
    model = MyModel()
    # NOTE: this is required for the ``fork`` method to work
    model.share_memory()
    processes = []
    for rank in range(num_processes):
        p = mp.Process(target=train, args=(model,))
        p.start()
        processes.append(p)
    for p in processes:
      p.join()

I get the following error:

Process Process-4:
Traceback (most recent call last):
  File "/home/jtremblay/anaconda2/lib/python2.7/multiprocessing/process.py", line 258, in _bootstrap
    self.run()
  File "/home/jtremblay/anaconda2/lib/python2.7/multiprocessing/process.py", line 114, in run
    self._target(*self._args, **self._kwargs)
  File "/home/jtremblay/code/pytorch-a3c/grad_example.py", line 25, in train
    param.grad.data = param.grad.data.clone()
AttributeError: 'NoneType' object has no attribute 'data'
Process Process-2:
Traceback (most recent call last):
  File "/home/jtremblay/anaconda2/lib/python2.7/multiprocessing/process.py", line 258, in _bootstrap
    self.run()
  File "/home/jtremblay/anaconda2/lib/python2.7/multiprocessing/process.py", line 114, in run
    self._target(*self._args, **self._kwargs)
  File "/home/jtremblay/code/pytorch-a3c/grad_example.py", line 25, in train
    param.grad.data = param.grad.data.clone()
AttributeError: 'NoneType' object has no attribute 'data'
Process Process-3:
Traceback (most recent call last):
  File "/home/jtremblay/anaconda2/lib/python2.7/multiprocessing/process.py", line 258, in _bootstrap
    self.run()
  File "/home/jtremblay/anaconda2/lib/python2.7/multiprocessing/process.py", line 114, in run
    self._target(*self._args, **self._kwargs)
  File "/home/jtremblay/code/pytorch-a3c/grad_example.py", line 25, in train
    param.grad.data = param.grad.data.clone()
AttributeError: 'NoneType' object has no attribute 'data'
Process Process-1:
Traceback (most recent call last):
  File "/home/jtremblay/anaconda2/lib/python2.7/multiprocessing/process.py", line 258, in _bootstrap
    self.run()
  File "/home/jtremblay/anaconda2/lib/python2.7/multiprocessing/process.py", line 114, in run
    self._target(*self._args, **self._kwargs)
  File "/home/jtremblay/code/pytorch-a3c/grad_example.py", line 25, in train
    param.grad.data = param.grad.data.clone()
AttributeError: 'NoneType' object has no attribute 'data'

I know I could create them manually but I am not sure this is a wanted behaviour.

2

  1. param.grad can now be None when it used to be zero. We made this change to better support sparse gradients.

  2. When you share a model, the gradients are not shared. That means you don’t need to do param.grad.data = param.grad.data.clone() – it won’t do anything, and as you saw breaks when .grad is None.

So if you just remove the cloning your code should work fine.

3

You can remove the initial for loop now. It’s no longer needed.

4

That is fantastic, do you think we could update the documentation?

5

Yes, it should definitely be updated! Could you please send a PR?