Skip to content

Commit

Permalink
fix path bug in training script
Browse files Browse the repository at this point in the history
  • Loading branch information
Weizhi Wang committed Aug 5, 2023
1 parent 0fe12e4 commit 93a4bb8
Showing 1 changed file with 4 additions and 2 deletions.
6 changes: 4 additions & 2 deletions train_scripts/train_longmem.sh
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@ fairseq-train ${DATA_DIR} \
--save-interval-updates 10000 --sample-break-mode none \
--tokens-per-sample 1024 \
--batch-size 8 --total-num-update 100000 --seed 42 \
--pretrained-model-path bigscience/bloom-1b7 \
--pretrained-model-path /path/to/gpt2_pretrained_model \
--layer-reduction-factor 2 \
--disable-validation \
--use-external-memory --memory-size 65536 \
Expand All @@ -23,4 +23,6 @@ fairseq-train ${DATA_DIR} \
--no-token-positional-embeddings \
--data-no-shuffle \
--retrieval-layer-index 17 \
--reload-ptm-layer
--reload-ptm-layer

# The --pre-trained-model path refers to the path to reproduced GPT-2-Medium checkpoints. You can find the downloading Google Drive url in README.

0 comments on commit 93a4bb8

Please sign in to comment.