lingyu123-su/MIDIFoundationModel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

1013 update

Browse Source
This commit is contained in:
FelixChan
2025-10-13 17:56:36 +08:00
parent d077e3210e
commit d6b68ef90b
17 changed files with 815 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
Amadeus/symbolic_yamls/config-accelerate.yaml
View File

@ -2,7 +2,8 @@ defaults:
# - nn_params: nb8_embSum_NMT
# - nn_params: remi8
# - nn_params: nb8_embSum_diff_t2m_150M_finetunning
- nn_params: nb8_embSum_diff_t2m_150M_pretrainingv2
# - nn_params: nb8_embSum_diff_t2m_600M_pretrainingv2
- nn_params: nb8_embSum_diff_t2m_600M_finetunningv2
# - nn_params: nb8_embSum_subPararell
# - nn_params: nb8_embSum_diff_t2m_150M
@ -14,7 +15,7 @@ defaults:
# - nn_params: remi8_main12_head_16_dim512
# - nn_params: nb5_embSum_diff_main12head16dim768_sub3
dataset: LakhClean # Pop1k7, Pop909, SOD, LakhClean,PretrainingDataset FinetuneDataset
dataset: FinetuneDataset # Pop1k7, Pop909, SOD, LakhClean,PretrainingDataset FinetuneDataset
captions_path: dataset/midicaps/train_set.json
# dataset: SymphonyNet_Dataset # Pop1k7, Pop909, SOD, LakhClean
@ -30,20 +31,20 @@ tau: 0.5
train_params:
device: cuda
batch_size: 3
batch_size: 5
grad_clip: 1.0
num_iter: 300000 # total number of iterations
num_cycles_for_inference: 10 # number of cycles for inference, iterations_per_validation_cycle * num_cycles_for_inference
num_cycles_for_model_checkpoint: 1 # number of cycles for model checkpoint, iterations_per_validation_cycle * num_cycles_for_model_checkpoint
iterations_per_training_cycle: 10 # number of iterations for logging training loss
iterations_per_validation_cycle: 5000 # number of iterations for validation process
iterations_per_validation_cycle: 3000 # number of iterations for validation process
input_length: 3072 # input sequence length3072
# you can use focal loss, it it's not used, set focal_gamma to 0
focal_alpha: 1
focal_gamma: 0
# learning rate scheduler: 'cosinelr', 'cosineannealingwarmuprestarts', 'not-using', please check train_utils.py for more details
scheduler : cosinelr
initial_lr: 0.00005
initial_lr: 0.0004
decay_step_rate: 0.8 # means it will reach its lowest point at decay_step_rate * total_num_iter
num_steps_per_cycle: 20000 # number of steps per cycle for 'cosineannealingwarmuprestarts'
warmup_steps: 2000 #number of warmup steps