search

tensorflow / models

Models and examples built with TensorFlow
Other
77.16k stars 45.75k forks source link

miou of hnasnet/pnasnet is always 0.4524. It doesn't change #7222

Open vyyxi opened 5 years ago

vyyxi commented 5 years ago

Hello,I use auto-deeplab,such as hnasnet/pnasnet. But the result is always o.452439815 when I train 10000 steps / 30000 steps / 60000 steps, it still doesn't change. I want to know how can I deal with it.

What is the top-level directory of the model you are using: deeplab Have I written custom code: no OS Platform and Distribution:NVIDIA-SMI 410.104 Driver Version: 410.104 TensorFlow installed from：pip online TensorFlow version: tensorflow-gpu==1.10.0 Bazel version: CUDA/cuDNN version:10.0 GPU model and memory: name: Tesla V100-SXM2-16GB major: 7 minor: 0 memoryClockRate(GHz): 1.53

tensorflowbutler commented 5 years ago

Thank you for your post. We noticed you have not filled out the following field in the issue template. Could you update them if they are relevant in your case, or leave them as N/A? Thanks. Exact command to reproduce

vyyxi commented 5 years ago

CUDA_VISIBLE_DEVICES=6 python train.py --logtostderr --training_number_of_steps=10000 --train_split='train' --model_variant="nas_hnasnet" --atrous_rates=6 --atrous_rates=12 --atrous_rates=18 --output_stride=16 --decoder_output_stride=4 --train_crop_size="385,385" --train_batch_size=4 --dataset='cityscapes' --train_logdir="/workspace/models-master/cityscapes/checkpoint" --dataset_dir="/workspace/models-master/cityscapes/tfrecord" --fine_tune_batch_norm=False CUDA_VISIBLE_DEVICES=7 python eval.py --logtostderr --eval_split='val' --model_variant="nas_hnasnet" --atrous_rates=6 --atrous_rates=12 --atrous_rates=18 --output_stride=16 --decoder_output_stride=4 --dataset="cityscapes" --eval_crop_size="385,385" --checkpoint_dir="/workspace/models-master/cityscapes/checkpoint" --eval_logdir="/workspace/models-master/cityscapes/eval" --dataset_dir="/workspace/models-master/cityscapes/tfrecord" @tensorflowbutler

vyyxi commented 5 years ago

And I add this in eval.py to print miou, I don't know if it's right : print_miou = tf.Print(miou,[miou],'miou is:') tf.summary.scalar('print_miou',print_miou)

AmeetR commented 5 years ago

following

AmeetR commented 5 years ago

Additionally, whenever I run hnasnet, it generates the same graph. Is it supposed to do this? I thought the point of architecture search was that the graph would have differences... or are there things in the cells that I'm missing.

DianaLi96 commented 5 years ago

CUDA_VISIBLE_DEVICES=6 python train.py --logtostderr --training_number_of_steps=10000 --train_split='train' --model_variant="nas_hnasnet" --atrous_rates=6 --atrous_rates=12 --atrous_rates=18 --output_stride=16 --decoder_output_stride=4 --train_crop_size="385,385" --train_batch_size=4 --dataset='cityscapes' --train_logdir="/workspace/models-master/cityscapes/checkpoint" --dataset_dir="/workspace/models-master/cityscapes/tfrecord" --fine_tune_batch_norm=False CUDA_VISIBLE_DEVICES=7 python eval.py --logtostderr --eval_split='val' --model_variant="nas_hnasnet" --atrous_rates=6 --atrous_rates=12 --atrous_rates=18 --output_stride=16 --decoder_output_stride=4 --dataset="cityscapes" --eval_crop_size="385,385" --checkpoint_dir="/workspace/models-master/cityscapes/checkpoint" --eval_logdir="/workspace/models-master/cityscapes/eval" --dataset_dir="/workspace/models-master/cityscapes/tfrecord" @tensorflowbutler

Have you solved the problem? Could you share the solution about this problem. I also met the problem.

Riqzz commented 5 years ago

Hello,I use auto-deeplab,such as hnasnet/pnasnet. But the result is always o.452439815 when I train 10000 steps / 30000 steps / 60000 steps, it still doesn't change. I want to know how can I deal with it.

What is the top-level directory of the model you are using: deeplab Have I written custom code: no OS Platform and Distribution:NVIDIA-SMI 410.104 Driver Version: 410.104 TensorFlow installed from：pip online TensorFlow version: tensorflow-gpu==1.10.0 Bazel version: CUDA/cuDNN version:10.0 GPU model and memory: name: Tesla V100-SXM2-16GB major: 7 minor: 0 memoryClockRate(GHz): 1.53

Have you solved the problem?

JINGTING92 commented 5 years ago

Cannot reproduce the results of hnasnet, neither. Following

shokohigol commented 3 years ago

Have you solved the problem?

omarsinnno commented 3 years ago

Has anyone found a solution to the problem of mIOU not changing during training?