JVET-AG0156 EE1-Related: On LOP2 training process [D. Rusanovskyy, Y. Li, M. Karczewicz (Qualcomm)]

The LOP2 filter was adopted to NNVC in result of joint training, report provided in JVET-AF0043, with further complexity reduction being studied in new EE1-2.3. During LOP2 joint training, several tests have been conducted on selecting batch size and learning rate combination. Additionally, a change to the LOP2 training strategy were employed with change to the ILF network architecture in EE1-2.3.

In this contribution, results of LOP2 Stage 3 training with modified strategy (batch size 64 and learning rate 0.0008) are reported. Assessment of the validation costs indicates that a new training strategy allows LOP2 network to be trained faster than the anchor strategy, however, it exhibits larger cost value fluctuations.

Inference testing with NNVC7.1 shows that comparing to LOP2 anchor, the float model trained with a new training strategy provides BD-rate change of {−0.20%, 0.18%, -0.77%} in RA and {−0.02%, −0.08%, −0.85%} in AI configurations. A quantized model provides BD-rate change of {−0.09%, 0.12%, −0.73%} and {0.09%, −0.09%, −0.84%}, in RA and AI, respectively.

It is proposed to consider the proposed training strategy for ILF training in LOP category.

It was commented that, as a saturation is observed around epoch 30 (and validation loss in chroma even increasing), it should be considered to switch/drop the learning rate beyond that point (currently, the learnimg rate is dropped around epoch 50, where the validation loss starts becoming lower).

It was agreed to study this in the EE as joint effort, i.e. several parties to run training independently.