More small fixes

[hhadian]

This fixes the dim mismatch problem. Please take a look at the changes because I'm not sure this is the right fix.

Currently it runs up to stage 15 of run_tdnn_1a.sh but it can't go further because of this error (related to xconfigs of the top network):

ERROR:root:***Exception caught while parsing the following xconfig line:
***   tdnnf-layer name=tdnnf1 l2-regularize=0.03 dropout-proportion=0.0 bypass-scale=0.66 dim=768 bottleneck-dim=96 time-stride=1
...
File "steps/libs/nnet3/xconfig/utils.py", line 518, in replace_bracket_expressions_in_descriptor
    "invalid or out of range offset.".format(descriptor_string, fields[i]))
RuntimeError: Error tokenizing string '[-1]': expression [-1] has an invalid or out of range offset.

It seems like it's expecting an input layer for the first layer (tdnnf1).

[danpovey]

Thanks. I fixed the xconfig, I think.

[hhadian]

Some notes:

1. What is the 0 in this line: nnet3-init --srand=$srand $dir/configs/default.config 0 -? (L236)
2. Also I can't find this file: steps/chaina/get_model_context.sh
3. It seems that the bottom network (bottom.raw) and transform network (default.ada) are never used in the script after they are created.

[danpovey]

The 0 may be a typo. I believe get_model_context.sh does not exist yet. It would figure out the total context from the bottom and top network context, and the bottom_subsampling_factor (which multiplies the top context). bottom.raw and default.ada should be used in the actual training script, which I still have to write. Just do all you reasonably/easily can-- if there are big gaps I'll fill them in later. Dan

…

On Mon, Jan 7, 2019 at 12:51 PM Hossein Hadian ***@***.***> wrote: Some notes: 1. What is the 0 in this line: nnet3-init --srand=$srand $dir/configs/default.config 0 -? (L236) 2. Also I can't find this file: steps/chaina/get_model_context.sh 3. It seems that the bottom network (bottom.raw) and transform network (default.ada) are never used in the script after they are created. — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#69 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu7U6RAYVdJqmjHDH24tz5UBijLSkks5vA7NggaJpZM4ZzNSo> .