def write_log2file(self, layer, block_count, layer_name):
write_log(
dir_name=os.path.join(self.settings.save_path, "log"),
file_name="log_block-{:0>2d}_{}.txt".format(
block_count, layer_name),
log_str="{:d}\t{:f}\t{:f}\t{:f}\t{:f}\t{:f}\t{:f}\t{:f}\t{:f}\t\n".
format(int(layer.d.sum()), self.record_selection_loss.avg,
self.record_selection_mse_loss.avg,
self.record_selection_softmax_loss.avg,
self.record_sub_problem_loss.avg,
self.record_sub_problem_mse_loss.avg,
self.record_sub_problem_softmax_loss.avg,
self.record_sub_problem_top1_error.avg,
self.record_sub_problem_top5_error.avg))
log_str = "Block-{:0>2d}-{} #channels: [{:0>4d}|{:0>4d}] ".format(
block_count, layer_name, int(layer.d.sum()), layer.d.size(0))
log_str += "[selection]loss: {:4f} mseloss: {:4f} softmaxloss: {:4f} ".format(
self.record_selection_loss.avg, self.record_selection_mse_loss.avg,
self.record_selection_softmax_loss.avg)
log_str += "[subproblem]loss: {:4f} mseloss: {:4f} softmaxloss: {:4f} ".format(
self.record_sub_problem_loss.avg,
self.record_sub_problem_mse_loss.avg,
self.record_sub_problem_softmax_loss.avg)
log_str += "top1error: {:4f} top5error: {:4f} ".format(
self.record_sub_problem_top1_error.avg,
self.record_sub_problem_top5_error.avg)
self.logger.info(log_str)
def run(self):
"""
Learn the parameters of the additional classifier and
fine tune model with the additional losses and the final loss
"""
best_top1 = 100
best_top5 = 100
start_epoch = 0
# if load resume checkpoint
if self.epoch != 0:
start_epoch = self.epoch + 1
self.epoch = 0
self.trainer.val(0)
for epoch in range(start_epoch, self.settings.n_epochs):
train_error, train_loss, train5_error = self.trainer.train(epoch)
val_error, val_loss, val5_error = self.trainer.val(epoch)
# write log
log_str = "{:d}\t".format(epoch)
for i in range(len(train_error)):
log_str += "{:.4f}\t{:.4f}\t{:.4f}\t{:.4f}\t{:.4f}\t{:.4f}\t".format(
train_error[i], train_loss[i], val_error[i], val_loss[i],
train5_error[i], val5_error[i])
write_log(self.settings.save_path, 'log.txt', log_str)
# save model and checkpoint
best_flag = False
if best_top1 >= val_error[-1]:
best_top1 = val_error[-1]
best_top5 = val5_error[-1]
best_flag = True
if best_flag:
self.checkpoint.save_aux_model(self.trainer.model,
self.trainer.aux_fc)
self.logger.info(
"|===>Best Result is: Top1 Error: {:f}, Top5 Error: {:f}\n".
format(best_top1, best_top5))
self.logger.info(
"|==>Best Result is: Top1 Accuracy: {:f}, Top5 Accuracy: {:f}\n"
.format(100 - best_top1, 100 - best_top5))
if self.settings.dataset in ["imagenet"]:
self.checkpoint.save_aux_checkpoint(self.trainer.model,
self.trainer.seg_optimizer,
self.trainer.fc_optimizer,
self.trainer.aux_fc, epoch,
epoch + 1)
else:
self.checkpoint.save_aux_checkpoint(self.trainer.model,
self.trainer.seg_optimizer,
self.trainer.fc_optimizer,
self.trainer.aux_fc, epoch)
def fine_tuning(self):
"""
Conduct network-wise fine-tuning after channel selection
"""
best_top1 = 100
best_top5 = 100
start_epoch = 0
if self.epoch != 0:
start_epoch = self.epoch + 1
self.epoch = 0
for epoch in range(start_epoch, self.settings.n_epochs):
train_error, train_loss, train5_error = self.network_wise_trainer.train(
epoch)
val_error, val_loss, val5_error = self.network_wise_trainer.val(
epoch)
# write log
log_str = "{:d}\t".format(epoch)
for i in range(len(train_error)):
log_str += "{:.4f}\t{:.4f}\t{:.4f}\t{:.4f}\t{:.4f}\t{:.4f}\t".format(
train_error[i], train_loss[i], val_error[i], val_loss[i],
train5_error[i], val5_error[i])
write_log(self.settings.save_path, 'log.txt', log_str)
# save model and checkpoint
best_flag = False
if best_top1 >= val_error[-1]:
best_top1 = val_error[-1]
best_top5 = val5_error[-1]
best_flag = True
if best_flag:
self.checkpoint.save_aux_model(
self.network_wise_trainer.model,
self.network_wise_trainer.aux_fc)
self.logger.info(
"|===>Best Result is: Top1 Error: {:f}, Top5 Error: {:f}\n".
format(best_top1, best_top5))
self.logger.info(
"|==>Best Result is: Top1 Accuracy: {:f}, Top5 Accuracy: {:f}\n"
.format(100 - best_top1, 100 - best_top5))
if "imagenet" in self.settings.dataset:
self.checkpoint.save_aux_checkpoint(
self.network_wise_trainer.model,
self.network_wise_trainer.seg_optimizer,
self.network_wise_trainer.fc_optimizer,
self.network_wise_trainer.aux_fc, epoch, epoch + 1)
else:
self.checkpoint.save_aux_checkpoint(
self.network_wise_trainer.model,
self.network_wise_trainer.seg_optimizer,
self.network_wise_trainer.fc_optimizer,
self.network_wise_trainer.aux_fc, epoch)
