Wrong params file is not reported as error in C API

[bputrycz]

Hi,

I would like to report a possible problem in C API.

I played with C API with a toy model (just adding two numbers) like:

net = mx.gluon.nn.Dense(1, in_units=2)
x = mx.sym.var('data')
y = net(x)

mod = mx.mod.Module(symbol=y, context=mx.cpu(), data_names = ['data'], label_names=[])
mod.bind(data_shapes=[("data", (1,2))], for_training=False)
mod.init_params(mx.init.Constant(1))
mod.save_checkpoint('mx_add', 0)

Then using code for prediction like:

PredictorHandle pred_handler = NULL;

const char* input_keys[] = { "data" };
const mx_uint input_shape_indptr[] = { 0, 2 };
const mx_uint input_shape_data[] = { 1, 2 };

const mx_float input[] = {3, 5};
mx_float output[1];

int result;

// In symbol/params_map we have appropriate files mapped into memory
result = MXPredCreate(symbol_map->mem,
                params_map->mem,
                params_map->mapsize,
                1, /* dev_type: cpu */
                0, /* dev_id */
                1, /* num_input_nodes */
                input_keys,
                input_shape_indptr,
                input_shape_data,
                &pred_handler);
printf("%d, '%s'\n", result, MXGetLastError());
  
result = MXPredSetInput(pred_handler, input_keys[0], input, 2);
printf("%d, '%s'\n", result, MXGetLastError());

result = MXPredForward(pred_handler);
printf("%d, '%s'\n", result, MXGetLastError());

result = MXPredGetOutput(pred_handler, 0, output, 1);
printf("%d, '%s'\n", result, MXGetLastError());

printf("sum: %f\n", output[0]);

result = MXPredFree(pred_handler);
printf("%d, '%s'\n", result, MXGetLastError());

I was able to obtain, as expected, sum to be 8:

0, ''
0, ''
0, ''
0, ''
sum: 8.000000
0, ''

BUT: when I put into params file something wrong, like weight names with not matching names,
then the prediction doesn’t report any error, and results with unspecified output:

0, ''
0, ''
0, ''
0, ''
sum: 18980598159864999268117905408.000000
0, ''

Shouldn’t predictor creation fail with some error about loading the params file?

Bartosz

[vdantu]

@bputrycz :.Seems like there is a check to see if data and names sizes are the same. Do you have a script that is easily compilable, we can look into this in depth?

[vdantu]

@mxnet-label-bot update [C API, Question, Pending Requester Info]

[bputrycz]

I wasn't blocked by this issue, just noticed it and reported.
I haven't invested more time into it.
I haven't tested it with the newest MXNet, and I don't want to do it now.
It is your call now to decide if it is worth to check again.

I think, I put above easily replicable code: python for generation of the sample model and C code with prediction from it. It misses only main() and includes.
Then to see the problem, you need to modify params file with changing the names of the keys in the dictionary.