DrSleep / tensorflow-deeplab-resnet

DeepLab-ResNet rebuilt in TensorFlow
MIT License
1.25k stars 429 forks source link

High accuracy 0.843 in train_msc val? #169

Open mokii opened 6 years ago

mokii commented 6 years ago

Train for 20K iterations with original parameters in train_msc.py with about 10K VOC training set and val the model.ckpt-8000 for 1449 val images but got the high accuracy:0.843? The training loss is around 1.5 and it seems very high. Is it rational? step 100 step 200 step 300 step 400 step 500 step 600 step 700 step 800 step 900 step 1000 step 1100 step 1200 step 1300 step 1400 Mean IoU: 0.843

DrSleep commented 6 years ago

Did you initialise from the downloaded model? If so, the released model by DeepLab was pre-trained on trainval, which explains high accuracy on validation.