def update_training_scores_history(perf_scores_history, model, loss, epoch,
num_best_scores):
""" Update the list of top training scores achieved so far, and log the best scores so far"""
model_sparsity, _, params_nnz_cnt = distiller.model_params_stats(model)
perf_scores_history.append(
distiller.MutableNamedTuple({
'params_nnz_cnt': -params_nnz_cnt,
'sparsity': model_sparsity,
'loss': loss,
'epoch': epoch
}))
# Keep perf_scores_history sorted from best to worst
# Sort by sparsity as main sort key, then sort by top1, top5 and epoch
def key(entry):
return (
entry['params_nnz_cnt'],
-entry['loss'],
entry['epoch'],
)
perf_scores_history.sort(key=key, reverse=True)
for score in perf_scores_history[:num_best_scores]:
msglogger.info(
'==> Best [Loss: %.3f Sparsity:%.2f Params: %d on epoch: %d]',
score.loss, score.sparsity, -score.params_nnz_cnt, score.epoch)
def update_training_scores_history(perf_scores_history, model, top1, top5, epoch, num_best_scores):
""" Update the list of top training scores achieved so far, and log the best scores so far"""
model_sparsity, _, params_nnz_cnt = distiller.model_params_stats(model)
perf_scores_history.append(distiller.MutableNamedTuple({'params_nnz_cnt': -params_nnz_cnt,
'sparsity': model_sparsity,
'top1': top1, 'top5': top5, 'epoch': epoch}))
# Keep perf_scores_history sorted from best to worst
# Sort by sparsity as main sort key, then sort by top1, top5 and epoch
perf_scores_history.sort(key=operator.attrgetter('params_nnz_cnt', 'top1', 'top5', 'epoch'), reverse=True)
for score in perf_scores_history[:num_best_scores]:
msglogger.info('==> Best [Top1: %.3f Top5: %.3f Sparsity:%.2f Params: %d on epoch: %d]',
score.top1, score.top5, score.sparsity, -score.params_nnz_cnt, score.epoch)
def step(self, model, epoch, **kwargs):
assert all(score in kwargs.keys() for score in ('top1', 'top5'))
model_sparsity, _, params_nnz_cnt = distiller.model_params_stats(model)
self.perf_scores_history.append(distiller.MutableNamedTuple({
'params_nnz_cnt': -params_nnz_cnt,
'sparsity': model_sparsity,
'top1': kwargs['top1'],
'top5': kwargs['top5'],
'epoch': epoch}))
# Keep perf_scores_history sorted from best to worst
self.perf_scores_history.sort(
key=operator.attrgetter('top1', 'top5', 'params_nnz_cnt', 'epoch'),
reverse=True)