Got different results from inference.py and eval.py

[sylvialee12]

Hello Antoine,

I've applied your code on my own datasets, but some strange things happened. Having input the same data and same model checkpoint, I got different results from eval.py and inference.py. Here's some examples:
From eval.py, numpy.argmax(predictions_val, axis=1) gives:
[17 18 17 18 17 17 17 17 8 17 17 17 0 18 18 17 17 0 17 17 0 17 17 17
17 17 10 17 19 17 17 17 19 10 17 17 17 9 17 7 17 17 17 17 9 17 17 17
10 10 10 0 17 17 17 8 17 17 0 17 17 17 17 17 17 17 17 17 12 17 17 17
17 17 14 17 17 10 17 0 17 16 17 17 17 17 9 0 17 17 17 17 17 17 17 0
17 0 17 17 17 17 17 17 17 0 1 3 5 17 17 9 17 17 10 17 17 17 17 9
17 17 17 10 17 17 17 17]
But from inference.py, numpy.argmax(predictions_val, axis=1) gives:
[11 8 4 18 6 10 10 6 10 0 10 8 0 18 8 10 10 0 10 8 0 8 10 18
8 8 10 10 19 17 18 10 19 10 10 14 9 10 10 8 0 10 10 11 10 3 11 9
10 10 10 0 18 8 10 8 10 0 0 10 7 10 8 0 9 0 10 10 6 10 10 3
10 0 10 6 11 10 10 0 10 10 1 9 0 11 4 3 0 6 0 17 6 17 10 18
7 0 6 7 0 7 10 0 10 0 1 0 16 6 0 9 10 5 10 7 0 11 10 0
0 6 10 10 1 0 6 3]

They are totally different, have you tried this as I did? Looking forward to your reply. Thanks in advance.

Sylvia

[sylvialee12]

The hyperparameters for training are as the following:
--model=NetVLADModelLF --iterations=300 --frame_features=True --feature_names="rgb" --feature_sizes=1024 --netvlad_clusters=64 --netvlad_hidden_size=128 --moe_l2=1e-6 --learning_rate_decay=0.8 --netvlad_relu=False --gating=True --moe_prob_gating=True --max_steps=120000 --num_classes 20 --start_new_model=True --batch_size 128

[huitang]

hihi may I ask did you freeze the pertained model before inference.py, or you directly use the the downloaded files
I did not find "checkpoint" file in the downloaded Zip, and maybe because of that I am not able to load the pertained model
If you frozen the model, do you mind to share the frozen code?
Many many thanks,