1127 update to latest

Amadeus/symbolic_yamls/config-accelerate.yaml

@@ -1,7 +1,9 @@
 defaults:
   # - nn_params: nb8_embSum_NMT
   # - nn_params: remi8
-  - nn_params: oct8_embSum_diff_t2m_150M_pretrainingv2
+  # - nn_params: oct8_embSum_diff_t2m_300M_pretrainingv3
+  # - nn_params: oct8_embSum_diff_t2m_150M_pretrainingv2
+  - nn_params: oct8_embSum_har_t2m_600M_pretrainingv3
     # - nn_params: nb8_embSum_diff_t2m_600M_pretrainingv2
   # -  nn_params: nb8_embSum_diff_t2m_600M_finetunningv2
   # - nn_params: nb8_embSum_subPararell
@@ -15,7 +17,7 @@ defaults:
   # - nn_params: remi8_main12_head_16_dim512
     # - nn_params: nb5_embSum_diff_main12head16dim768_sub3
 
-dataset: Melody  # Pop1k7, Pop909, SOD, LakhClean,PretrainingDataset FinetuneDataset
+dataset: msmidi  # Pop1k7, Pop909, SOD, LakhClean,PretrainingDataset FinetuneDataset
 captions_path: dataset/midicaps/train_set.json
 
 # dataset:  SymphonyNet_Dataset # Pop1k7, Pop909, SOD, LakhClean
@@ -23,28 +25,28 @@ captions_path: dataset/midicaps/train_set.json
 
 use_ddp: True # True, False | distributed data parallel
 use_fp16: True # True, False | mixed precision training
-use_diff: True # True,use diffusion in subdecoder
+use_diff: False # True,use diffusion in subdecoder
 diff_steps: 8 # number of diffusion steps
-use_dispLoss: True
+use_dispLoss: False
 lambda_weight: 0.5
 tau: 0.5
 
 train_params:
   device: cuda
-  batch_size: 10
+  batch_size: 9
   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: 3000 # number of iterations for validation process
-  input_length: 3072 # input sequence length3072
+  input_length: 2048 # 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: 0.00001
+  initial_lr: 0.0003
   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