Age | Commit message (Collapse) | Author |
|
parse.c test code was deleted too
|
|
To make onehot available I had to refactor how the data is stored, the
current implementation supports json files.
|
|
Functions Parameters were replaced by Configs struct to allow
scalability, in the future I might merge parse.c source code into util.c
|
|
|
|
|
|
The CLI option to disable it was also added.
|
|
|
|
Things done:
* config path should search config file in the following order: cli
option, environment, xdg_path
* Implement a retrain command.
* when you require more keys than the ones available in the input,
stop the program.
|
|
Things implemented:
* json_read() must die if the key does not exist or the value type is wrong.
* on predict command input should be shown exactly the same
* float precision CLI option should be added.
|
|
things implemented:
- CLI option --only-out added
- CLI option --format works more intuitively
- csv tsv and json output support
|
|
things implemented:
- read output in false bug was solved.
- Make generic rule added to build test executables
- format option added to the CLI
|
|
things implemented:
- Automatic file format infering
- json_read() small refactor done
- error checking in csv_read() improved
|
|
Error messages are more specific and now columns are read as arrays of
strings.
|
|
There is still much to do
|
|
|
|
|
|
weights update for loop was changed for a cblas routine
|
|
|
|
|
|
Documentation:
Example sections was improved
Files sections was added
Fixes:
default config filepath was fixed
|
|
The man page must have a section specifying where to save configurations
files for loading the network architecture and weights.
|
|
Installation and uninstallation rules must be added on Makefile
|
|
Now is possible to read data from standard input
|
|
|
|
|
|
|
|
train subcommand can read and train the network
TODO:
- Refactor json_read() to parse multiple labels
- Implement a function to save network weights once the network have
trained
- Implement a function to load trained weights to use with predict
subcommand
|
|
|
|
|
|
parameters implemented:
- input_keys
- label_keys
- epochs
- labels
- weights_path
Network weights and architecture loading is not implemented yet.
|
|
The program config parameters where encapsulated in a struct to keep
track of them without implementing a lot of functions
|
|
|
|
|
|
Gdb files will be ignored from now
|
|
|
|
|
|
The point is have this file for new activation function without bloating
nn.c file
|
|
Training function were fixed and new activation and cost function were
implemented
|
|
The function needs to be tested
|
|
|
|
|
|
|
|
It needs to be tested and some backward layer functions were redefined
to improve readability
|
|
|
|
|
|
The function nn_layer_forward was changed from f(X*W) to X*W and
nn_layer_map_activation will map the activation function through the
output of nn_layer_forward.
|
|
|
|
|
|
|
|
|