audit_prediction_shap.mpc

from Compiler.script_utils import output_utils

from Compiler.script_utils.data import data


from Compiler import ml
from Compiler import library

from Compiler.script_utils.audit import shap

from Compiler.script_utils import config, timers




class AuditConfig(config.BaseAuditModel):
    test_var: int = 1

    # n_batches: int = 10

program.options_from_args()
cfg = config.from_program_args(program.args, AuditConfig)

if not cfg.emulate:
    program.use_trunc_pr = cfg.trunc_pr
    # program.use_edabits = True
    # program.use_split(4)

ml.set_n_threads(cfg.n_threads)

library.start_timer(timer_id=timers.TIMER_LOAD_DATA)
input_shape_size = cfg.batch_size * cfg.n_batches if cfg.n_batches > 0 else None
input_loader = data.get_input_loader(dataset=cfg.dataset, audit_trigger_idx=cfg.audit_trigger_idx, batch_size=cfg.batch_size,
                                     debug=cfg.debug, emulate=cfg.emulate, consistency_check=cfg.consistency_check,
                                     sha3_approx_factor=cfg.sha3_approx_factor, input_shape_size=input_shape_size)
library.stop_timer(timer_id=timers.TIMER_LOAD_DATA)

# expect 91 * (13136) + 13136 + (91 * 32) + 2*32 + 1

library.start_timer(timer_id=timers.TIMER_AUDIT)

result, debug_output = shap.audit(input_loader, cfg, debug=cfg.debug)

# explanation algorithm, shap
# ...

library.stop_timer(timer_id=timers.TIMER_AUDIT)

print("Done with audit, outputting results")

# REVEALING MIGHT TAKE A LOT OF COMPILATION TIME / OVERHEAD
for k, v in result.items():
    output_utils.output_value(name=k, value=v)

for k, v in debug_output.items():
    output_utils.output_value_debug(name=k, value=v, repeat=False)