diff --git a/ddsp/training/gin/papers/icml2020/README.md b/ddsp/training/gin/papers/icml2020/README.md index 50fac2e5..2ec7878c 100644 --- a/ddsp/training/gin/papers/icml2020/README.md +++ b/ddsp/training/gin/papers/icml2020/README.md @@ -48,7 +48,7 @@ ddsp_run \ This command points to datasets on GCP, and the gin_params for file_pattern are redudant with the default values in the gin files, but provided here to show how you would modify them for local dataset paths. -In the paper we train for ~1.2M steps with a batch size of 64. The command above is tuned for a single v100 (max batch size = 32), you would need to use multiple gpus to exactly reproduce the experiement. Given the large amount of pretraining, a pretrained checkpoint [is available here](https://storage.googleapis.com/ddsp-inv/ckpts/synthetic_pretrained_ckpt.zip) +In the paper we train for ~1.2M steps with a batch size of 64. The command above is tuned for a single v100 (max batch size of 32), you would need to use multiple gpus to exactly reproduce the experiement. Given the large amount of pretraining, a pretrained checkpoint [is available here](https://storage.googleapis.com/ddsp-inv/ckpts/synthetic_pretrained_ckpt.zip) or on GCP at `gs://ddsp-inv/ckpts/synthetic_pretrained_ckpt`. ### Eval and Sample @@ -82,8 +82,13 @@ ddsp_run \ ### Train Now we finetune the model from above on a specific dataset. Use the `--restore_dir` flag to point to your pretrained checkpoint. + A pretrained model on 1.2M steps (batch size=64) of synthetic data [is available here](https://storage.googleapis.com/ddsp-inv/ckpts/synthetic_pretrained_ckpt.zip) -or on GCP at `gs://ddsp-inv/ckpts/synthetic_pretrained_ckpt`. +or on GCP. + +```bash +gsutil cp -r gs://ddsp-inv/ckpts/synthetic_pretrained_ckpt /path/to/synthetic_pretrained_ckpt +``` ```bash ddsp_run \ @@ -94,7 +99,7 @@ ddsp_run \ --gin_file=papers/icml2020/finetune_dataset.gin \ --gin_param="SyntheticNotes.file_pattern='gs://ddsp-inv/datasets/notes_t125_h100_m65_v2.tfrecord*'" \ --gin_param="train_data/TFRecordProvider.file_pattern='gs://ddsp-inv/datasets/all_instruments_train.tfrecord*'" \ ---gin_param="test_data/TFRecordProvider.file_pattern = 'gs://ddsp-inv/datasets/all_instruments_test.tfrecord*'" \ +--gin_param="test_data/TFRecordProvider.file_pattern='gs://ddsp-inv/datasets/all_instruments_test.tfrecord*'" \ --gin_param="batch_size=12" \ --alsologtostderr ``` @@ -104,7 +109,7 @@ We have provided sharded TFRecord files for the [URMP dataset](http://www2.ece.r If training on GCP it is fast to directly read from these buckets, but if training locally you will probably want to download the files locally (~16 GB) using the `gsutil` command line utility from the [gcloud sdk](https://cloud.google.com/sdk/docs/downloads-interactive). -In the paper, this model was trained with a batch size of 64 on 8 accelerators (8 per an accelerator), and typically converges after 50-100k iterations. The command above is tuned for a single v100 (max batch size = 12), you would need to use multiple gpus to exactly reproduce the experiement. +In the paper, this model was trained with a batch size of 64 on 8 accelerators (8 per an accelerator), and typically converges after 200-400k iterations. The command above is tuned for a single v100 (max batch size of 12), you would need to use multiple GPUs or TPUs to exactly reproduce the experiement. To use a TPU, start up an instance from the web interface and pass the internal ip address to the tpu flag `--tpu=grpc://`. ### Eval and Sample