Very Low Validation Accuracy for Resnet50 and Resnet101 models using TF-TRT

[Selventhiranraj]

I'm getting 0.1% as Top 1 accuracy for the validation of Resnet50 and Resnet101 models using the TF-TRT. Validation script: https://github.com/tensorflow/tensorrt/blob/master/tftrt/examples/image-classification/image_classification.py Dataset : Imagenet Model : https://tfhub.dev/tensorflow/resnet_50/classification/1 Preprocess method : vgg

Why is the accuracy very low?? Any help/suggestions appreciated. Thanks

[DEKHTIARJonathan]

I'm unfortunately unable to reproduce your issue. We are routinely running these benchmarks and we never had such issue.

Here are a few ideas to investigate:

• try not to convert the model with TF-TRT and just use normal to TF to conduct inference:

  ./scripts/resnet_v1_50.sh \
  --use_xla \
  --data_dir=/data/imagenet/train-val-tfrecord --model_dir=/models

  A) If that works perfectly, there might be an issue in TF-TRT. B) If that doesn't, the issue comes from either: your checkpoint/SavedModel is bad or your dataloading.

• if A) above: try using our TF-TRT automated scripts:

  ./scripts/resnet_v1_50.sh \
  --use_xla \
  --data_dir=/data/imagenet/train-val-tfrecord --model_dir=/models \
  --use_tftrt --tftrt_precision="FP16"

If that doesn't solve the issue, I will need a complete reproducer case with step-by-step instructions in a docker container to reproduce the issue (you will need to provide the checkpoint).

[mingxiaoh]

meet same issue