[GitHub] [incubator-mxnet] barry-jin edited a comment on pull request #20262: [2.0] Gluon2.0: switch to use forward interface

[GitBox <gi...@apache.org>]

barry-jin edited a comment on pull request #20262:
URL: https://github.com/apache/incubator-mxnet/pull/20262#issuecomment-849217284


   Looks like GPU memory will not be released after using `asnumpy()` method for numpy ndarray on gpu context. 
   
   Code to reproduce:
   ```python
   import mxnet as mx
   from mxnet import npx, gluon
   from mxnet.gluon import nn
   npx.set_np()
   
   mx.context._current.set(mx.gpu(0))
   
   def check_layer_forward_withinput(net, x):
       x_hybrid = x.copy()
       x.attach_grad()
       x_hybrid.attach_grad()
       net.initialize()
       with mx.autograd.record():
           out1 = net(x)
       out1.backward()
       net.hybridize()
       with mx.autograd.record():
           out2 = net(x_hybrid)
       out2.backward()
       a, b = mx.context.gpu_memory_info(0)
       print("Used memory {} GB, Total memory {} GB.".format((b - a) / (1024 * 1024 * 1024), b / (1024 * 1024 * 1024)))
       mx.test_utils.assert_almost_equal(x.grad, x_hybrid.grad, rtol=1e-5, atol=1e-6)
       mx.test_utils.assert_almost_equal(out1, out2, rtol=1e-5, atol=1e-6)
   
   def test_slice_pooling2d():
       # transpose shape to bring feature dimension 'c' from 2nd position to last
       def transpose(shape):
           return (shape[0],) + shape[2:] + (shape[1],)
   
       for layout in ['NCHW', 'NHWC']:
           max_pooling = nn.MaxPool2D(layout=layout)
           avg_pooling = nn.AvgPool2D(layout=layout)
           global_maxpooling = nn.GlobalMaxPool2D(layout=layout)
           global_avgpooling = nn.GlobalAvgPool2D(layout=layout)
           pooling_layers = [max_pooling, avg_pooling, global_maxpooling, global_avgpooling]
           class Net(gluon.HybridBlock):
               def __init__(self,
                            slice,
                            pooling_layer,
                            **kwargs):
                   super(Net, self).__init__(**kwargs)
                   self.slice = slice
                   self.pool0 = pooling_layer
   
               def forward(self, x):
                   x_slice = mx.npx.slice(x, begin=self.slice[0], end=self.slice[1])
                   out = self.pool0(x_slice)
                   return out
   
           xshape = (16, 128, 256, 256)
           slice_shape = (4, 16, 32, 64)
           if layout == 'NHWC':
               xshape = transpose(xshape)
               slice_shape = transpose(slice_shape)
           x = mx.np.random.uniform(size=xshape)
           slice = [(0, 0, 0, 0), slice_shape]
           for i in range(len(pooling_layers)):
               net = Net(slice, pooling_layers[i])
               check_layer_forward_withinput(net, x)
   
   if __name__ == '__main__':
       test_slice_pooling2d() 
   ``` 
   
   Before comment out `mx.test_utils.assert_almost_equal()` these two lines: 
   ``` 
   Used memory 2.142578125 GB, Total memory 14.755615234375 GB.
   Used memory 4.119140625 GB, Total memory 14.755615234375 GB.
   Used memory 4.619140625 GB, Total memory 14.755615234375 GB.
   Used memory 5.119140625 GB, Total memory 14.755615234375 GB.
   Used memory 5.119140625 GB, Total memory 14.755615234375 GB.
   Used memory 6.119140625 GB, Total memory 14.755615234375 GB.
   Used memory 6.619140625 GB, Total memory 14.755615234375 GB.
   Used memory 7.119140625 GB, Total memory 14.755615234375 GB.
   ``` 
   
   After comment out these two lines:
   ``` 
   Used memory 2.142578125 GB, Total memory 14.755615234375 GB.
   Used memory 2.142578125 GB, Total memory 14.755615234375 GB.
   Used memory 2.1171875 GB, Total memory 14.755615234375 GB.
   Used memory 2.1171875 GB, Total memory 14.755615234375 GB.
   Used memory 2.1171875 GB, Total memory 14.755615234375 GB.
   Used memory 2.1171875 GB, Total memory 14.755615234375 GB.
   Used memory 2.1171875 GB, Total memory 14.755615234375 GB.
   Used memory 2.6171875 GB, Total memory 14.755615234375 GB.
   ``` 


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org