Files
spoterembedding/train.sh
Mathias Claassen 81bbf66aab Initial codebase (#1)
* Add project code

* Logger improvements

* Improvements to web demo code

* added create_wlasl_landmarks_dataset.py and xtract_mediapipe_landmarks.py

* Fix rotation augmentation

* fixed error in docstring, and removed unnecessary replace -1 -> 0

* Readme updates

* Share base notebooks

* Add notebooks and unify for different datasets

* requirements update

* fixes

* Make evaluate more deterministic

* Allow training with clearml

* refactor preprocessing and apply linter

* Minor fixes

* Minor notebook tweaks

* Readme updates

* Fix PR comments

* Remove unneeded code

* Add banner to Readme

---------

Co-authored-by: Gabriel Lema <gabriel.lema@xmartlabs.com>
2023-03-03 10:07:54 -03:00

25 lines
682 B
Bash
Executable File

#!/bin/sh
python -m train \
--save_checkpoints_every -1 \
--experiment_name "augment_rotate_75_x8" \
--epochs 10 \
--optimizer "SGD" \
--lr 0.001 \
--batch_size 32 \
--dataset_name "wlasl" \
--training_set_path "WLASL100_train.csv" \
--validation_set_path "WLASL100_test.csv" \
--vector_length 32 \
--epoch_iters -1 \
--scheduler_factor 0 \
--hard_triplet_mining "in_batch" \
--filter_easy_triplets \
--triplet_loss_margin 1 \
--dropout 0.2 \
--start_mining_hard=200 \
--hard_mining_pre_batch_multipler=16 \
--hard_mining_pre_batch_mining_count=5 \
--augmentations_prob=0.75 \
--hard_mining_scheduler_triplets_threshold=0 \
# --normalize_embeddings \