overfitting very fast?

[asdfqwer2015]

Hi, sorry to bother you once again. But I'm very interested in your model.

I found the model may overfitting in training very fast.

describe:
I tested a trained model, and it seems has a low accuracy in predict video. And I tried to feed the trained model with random cached pickle numpy(10, 8, 112, 112, 3) samples from both trainset and testset. It predict well when process samples from trainset, but predict incorrect result for almost all sampled samples from testsset. BTW, the sampled samples covers all 25 classes.
And I evaluated the testset, the loss is >> 1 while the loss for trainset is almost 0.

So I modified the script, and monitor val loss while training.
for epoch in range(int(200000 // steps_per_epoch) + 1): gesture_classifier.train(input_fn = train_input_fn,steps = steps_per_epoch,hooks = [logging_hook]); eval_results = gesture_classifier.evaluate(input_fn = eval_input_fn, steps=100);
eval samples are random samples from all the testset.
And I get curves like:

yellow curve - train loss
blue curve - val loss

And the edit distance for samples from testset likes:

val error is still very high.

Is the model overfitting and did I miss something?

[breadbread1984]

one difference between my implement and the description of original paper is how the hidden layers are normalized. I used layer normalization to use small batch which can fit in my computing hardware. you can replace layer normalization with batch normalization and increase the batch size and train c3d and r3dcnn again to test whether the result improves.

besides, the original paper use multiple channel together to detect hand gesture. I couldn't read some channel from the dataset, so I haven't tried it.

furthermore, the nv gesture dataset is a relative small dataset. you may need early stopping during training.

[asdfqwer2015]

ok, I'll try BN istead of LN now.
But I only have a single gpu to train the model, so the batchsize can be max at 4 for my scenario, which may leads BN unstable. Hope to find a mechanism for tf.estimator API similar to subdivision in yolo to calculate mean grad for several batch for large batch size.