lanking520 opened a new issue #18121:
URL: https://github.com/apache/incubator-mxnet/issues/18121
## Description
I am facing the issue loading the trained Symbolic model in mxnet 1.6.0. It
is reproducible in the most recent nightly build. The error message indicated
that the storage type in my model is not supported. My model is trained in
1.4.1 using Spare NDArray (RowSparse) as weight and Input type is CSR.
I tried to update the model to 1.6.0 and it didn't solve the problem. What
should I do to use Sparse in 1.6.0?
This is partial information of my `symbol.json`
```
{
"nodes": [
{
"op": "null",
"name": "data",
"attrs": {"__storage_type__": "2"},
"inputs": []
},
{
"op": "null",
"name": "weight",
"attrs": {
"__init__": "[\"uniform\", {\"scale\": 0.01}]",
"__shape__": "(1000, 1024)",
"__storage_type__": "1"
},
"inputs": []
}
```
### Error Message
#### Using SymbolBlock to load:
```
imported_net = gluon.nn.SymbolBlock.imports(prefix + '-symbol.json',
['data'],
prefix + '-0000.params')
```
With error
```
line 1304, in __init__
"type is 'row_sparse'." % j.name
AssertionError: SymbolBlock doesn't support Parameter 'weight' because its
storage type is 'row_sparse'.
```
#### Using Module API to load
```
mx.npx.set_np(True, True)
sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch=1)
```
with error
```
line 255, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [22:37:54] src/ndarray/ndarray.cc:1729: Check failed:
Imperative::Get()->is_np_shape() == GlobalOn ||
!Imperative::Get()->is_np_shape(): ndarray was not saved in np shape semantics,
but being loaded in np shape semantics. Please turn off np shape semantics in
Python using `with np_shape(False)` to scope the code of loading the ndarray.
```
----------------------------------------------------------------
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:
[email protected]
