Training procedure unclear

[lennart-rth]

Hi,
can you give some additional information about the training process?
I would like to train NAG2G from scratch, but lack a training script.
Can I use the same arguments as in validate.sh without the --infer_step flag and with lr and optimizer as in the paper?
Best regards,
Lennart

[lennart-rth]

In case someone is in the same situation as me. This is what I ended up with:

[ -z "${MASTER_PORT}" ] && MASTER_PORT=10086
[ -z "${n_gpu}" ] && n_gpu=1 #$(nvidia-smi -L | wc -l)
[ -z "${OMPI_COMM_WORLD_SIZE}" ] && OMPI_COMM_WORLD_SIZE=1
[ -z "${OMPI_COMM_WORLD_RANK}" ] && OMPI_COMM_WORLD_RANK=0
[ -z "${len_penalty}" ] && len_penalty=0.0
[ -z "${beam_size}" ] && beam_size=10
[ -z "${beam_size_second}" ] && beam_size_second=5
[ -z "${beam_head_second}" ] && beam_head_second=2
# SimpleGenerator SequenceGeneratorBeamSearch SequenceGeneratorBeamSearch_test
[ -z "${search_strategies}" ] && search_strategies=SequenceGeneratorBeamSearch
[ -z "${temperature}" ] && temperature=1
[ -z "${num_workers}" ] && num_workers=12
export TORCH_NCCL_ASYNC_ERROR_HANDLING=1
export OMP_NUM_THREADS=12
export TORCH_USE_CUDA_DSA=1
export CUDA_LAUNCH_BLOCKING=1
TORCH_DISTRIBUTED_DEBUG=DETAIL
run_name=nag2g
save_dir="./save/${run_name}"
mkdir -p ${save_dir}

python -m torch.distributed.run \
       --nnodes=1 --nproc_per_node=$n_gpu $(which unicore-train) ./USPTO50K_brief_20230227  --user-dir . --valid-subset valid,test \
       --num-workers 12 --ddp-backend=no_c10d --user-dir ./NAG2G  \
       --task G2G_unimolv2 --loss G2G --arch NAG2G_G2G  --encoder-type unimolv2 \
       --optimizer adam --adam-betas '(0.9, 0.999)' --adam-eps 1e-8 --clip-norm 1.0 \
       --lr-scheduler polynomial_decay --lr 2.5e-4 --warmup-updates 12000 --total-num-update 10000 --batch-size 16 \
       --update-freq 1 --seed 1 --power 1\
       --tensorboard-logdir $save_dir/tsb \
       --log-interval 100 --log-format simple \
       --save-interval-updates 5000 --validate-interval-updates 5000 --keep-interval-updates 1 --no-epoch-checkpoints --max-epoch 0 \
       --save-dir $save_dir --use-decoder \
       --encoder-embed-dim 768 --encoder-ffn-embed-dim 768 --decoder-embed-dim 768 --decoder-ffn-embed-dim 768 \
       --config_file config.ini \
    --max-update 120000 --decoder_type new --want_decoder_attn --end-learning-rate 1e-09 --total-num-update 120000 --activation-fn gelu  --laplacian_pe_dim 0 --auto-regressive --position-type relative --use_reorder --want_charge_h --shufflegraph randomsmiles --want_h_degree --decoder-layers 6 --decoder-attention-heads 24 --pair-embed-dim 256 --encoder-layers 12  --encoder-attention-heads 48 \
     --fp16 --fp16-init-scale 4 --fp16-scale-window 256 \

I haven't managed to replicate it exactly, as I'm probably still missing something.

[thegodone]

I also try to install it validation works but honestly this is for advanced expert linux. Did you manage to train it ?

[synsis]

I apologize for the delayed response. I have updated the train.sh file.

[mujeebarshad]

@lennart-rth Were you able to train the model with the new script? I am having issue training it as well, could you guide me towards solution? #6