Which version of TF are you using?

[leotam]

I'm encountering
Tensor name "resnet_v1_152/block1/unit_1/bottleneck_v1/conv1/biases" not found in checkpoint files /home/referit/data/resnet_v1_152.ckpt [[node save/RestoreV2 (defined at /home/referit/src/referit/preprocess_data/generic/preprocess_data/extract_img_features.py:32) ]]

which others suggest may be due to TF version naming

[fstrub95]

We were using tf1, with the (deprecated) library slim. The code should work up to tf 1.14. We do not plan to upgrade it to tf v2 as both Harm and I are working on completely different project now. Note that if you want to do so it should now be super easy! Happy to help if necessary.

…

On Tue, Dec 24, 2019, 9:13 PM leotam ***@***.***> wrote: I'm encountering Tensor name "resnet_v1_152/block1/unit_1/bottleneck_v1/conv1/biases" not found in checkpoint files /home/referit/data/resnet_v1_152.ckpt [[node save/RestoreV2 (defined at /home/referit/src/referit/preprocess_data/generic/preprocess_data/extract_img_features.py:32) ]] which others suggest may be due to TF version naming — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=ADF7HBFEX4IQIZCP7MDESC3Q2JUPNA5CNFSM4J7ASF6KYY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4ICRL22Q>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADF7HBFACBWKJWE36RLCFOTQ2JUPNANCNFSM4J7ASF6A> .

[leotam]

Hm I'm using conda 4.8.0 with TF 1.13.1 and python 3.6.8. It seemed similar to this issue: tensorflow/models#2527 which suggested syncing TF versions and barring that setting argscope. Since it seems your argscope is set, I assumed it was TF version.

Am I missing some argscope parameter? Using:
python extract_image_features.py \ -img_dir /home/referitdata/mscoco/images \ -data_dir /home/referit/data/ \ -out_dir /home/referit/data/refer/crop_all_ft \ -ckpt /home/referit/data/resnet_v1_152.ckpt \ -resnet_version 152 \ -feature_name block3 \ -mode crop_all \ -dataset refclef