Fix YAML referencing to allow better CLI overrides (#150)

examples/bert/tests/smoketest_config_classification.yaml

@@ -1,5 +1,5 @@
-tokenizer_name: &tokenizer_name prajjwal1/bert-tiny
-max_seq_len: &max_seq_len 32
+tokenizer_name: prajjwal1/bert-tiny
+max_seq_len: 32
 
 # Run Name
 run_name: test
@@ -8,22 +8,22 @@ run_name: test
 model:
   name: mosaic_bert
   num_labels: 2
-  pretrained_model_name: *tokenizer_name
-  tokenizer_name: *tokenizer_name
+  pretrained_model_name: ${tokenizer_name}
+  tokenizer_name: ${tokenizer_name}
 
 # Dataloaders
 train_loader:
   split: train
-  tokenizer_name: *tokenizer_name
-  max_seq_len: *max_seq_len
+  tokenizer_name: ${tokenizer_name}
+  max_seq_len: ${max_seq_len}
   shuffle: true
   drop_last: true
   num_workers: 4
 
 eval_loader:
   split: validation
-  tokenizer_name: *tokenizer_name
-  max_seq_len: *max_seq_len
+  tokenizer_name: ${tokenizer_name}
+  max_seq_len: ${max_seq_len}
   shuffle: false
   drop_last: false
   num_workers: 4

examples/bert/tests/smoketest_config_glue.yaml

@@ -7,13 +7,13 @@ default_seed: 1111
 precision: fp32
 
 # Tokenizer for dataset creation
-tokenizer_name: &tokenizer_name bert-base-uncased
+tokenizer_name: bert-base-uncased
 
 # Base model config
 model:
   name:
   pretrained_model_name: prajjwal1/bert-tiny
-  tokenizer_name: *tokenizer_name
+  tokenizer_name: ${tokenizer_name}
 
 
 # Loading

examples/bert/tests/smoketest_config_main.yaml

@@ -1,6 +1,6 @@
-tokenizer_name: &tokenizer_name prajjwal1/bert-tiny
-max_seq_len: &max_seq_len 32
-mlm_probability: &mlm_probability 0.15
+tokenizer_name: prajjwal1/bert-tiny
+max_seq_len: 32
+mlm_probability: 0.15
 
 # Run Name
 run_name: test
@@ -9,8 +9,8 @@ run_name: test
 model:
   name:
   use_pretrained: false
-  pretrained_model_name: *tokenizer_name
-  tokenizer_name: *tokenizer_name
+  pretrained_model_name: ${tokenizer_name}
+  tokenizer_name: ${tokenizer_name}
 
 # Dataloaders
 train_loader:
@@ -19,12 +19,12 @@ train_loader:
     remote:
     local:
     split: train
-    tokenizer_name: *tokenizer_name
-    max_seq_len: *max_seq_len
+    tokenizer_name: ${tokenizer_name}
+    max_seq_len: ${max_seq_len}
     group_method: truncate
     predownload: 1000
     shuffle: true
-    mlm_probability: *mlm_probability
+    mlm_probability: ${mlm_probability}
     num_canonical_nodes: 8
   drop_last: true
   num_workers: 4
@@ -35,12 +35,12 @@ eval_loader:
     remote:
     local:
     split: val
-    tokenizer_name: *tokenizer_name
-    max_seq_len: *max_seq_len
+    tokenizer_name: ${tokenizer_name}
+    max_seq_len: ${max_seq_len}
     group_method: truncate
     predownload: 1000
     shuffle: false
-    mlm_probability: *mlm_probability
+    mlm_probability: ${mlm_probability}
     num_canonical_nodes: 8
   drop_last: false
   num_workers: 4

examples/bert/yamls/finetuning/glue/hf-bert-base-uncased.yaml

@@ -10,13 +10,13 @@ default_seed: 19
 precision: bf16
 
 # Tokenizer for dataset creation
-tokenizer_name: &tokenizer_name bert-base-uncased
+tokenizer_name: bert-base-uncased
 
 # Base model config
 model:
   name: hf_bert
-  pretrained_model_name: *tokenizer_name
-  tokenizer_name: *tokenizer_name
+  pretrained_model_name: ${tokenizer_name}
+  tokenizer_name: ${tokenizer_name}
 
 # Loading
 starting_checkpoint_load_path:      # Fill this in with the composer checkpoint from the end of pre-training a HF BERT

examples/bert/yamls/finetuning/glue/mcloud_run.yaml

@@ -39,13 +39,13 @@ parameters:
   precision: bf16
 
   # Tokenizer for dataset creation
-  tokenizer_name: &tokenizer_name bert-base-uncased
+  tokenizer_name: bert-base-uncased
 
   # Base model config
   model:
     name: mosaic_bert
-    pretrained_model_name: *tokenizer_name
-    tokenizer_name: *tokenizer_name
+    pretrained_model_name: ${tokenizer_name}
+    tokenizer_name: ${tokenizer_name}
 
   # Loading
   # (fill this in with the composer checkpoint from the end of pre-training a Mosaic BERT)

examples/bert/yamls/finetuning/glue/mosaic-bert-base-uncased.yaml

@@ -10,13 +10,13 @@ default_seed: 19
 precision: bf16
 
 # Tokenizer for dataset creation
-tokenizer_name: &tokenizer_name bert-base-uncased
+tokenizer_name: bert-base-uncased
 
 # Base model config
 model:
   name: mosaic_bert
-  pretrained_model_name: *tokenizer_name
-  tokenizer_name: *tokenizer_name
+  pretrained_model_name: ${tokenizer_name}
+  tokenizer_name: ${tokenizer_name}
 
 # Loading
 # (fill this in with the composer checkpoint from the end of pre-training a Mosaic BERT)

examples/bert/yamls/finetuning/hf-bert-base-uncased.yaml

@@ -10,8 +10,8 @@
 # Run Name
 run_name: finetune-hf-bert
 
-tokenizer_name: &tokenizer_name bert-base-uncased
-max_seq_len: &max_seq_len 128
+tokenizer_name: bert-base-uncased
+max_seq_len: 128
 
 load_path: # (Optionally) provide a composer checkpoint to use for the starting weights
 
@@ -20,22 +20,22 @@ model:
   name: hf_bert
   num_labels: 2 # <-- Make sure to update these after you modify the starter script!
   use_pretrained: true
-  pretrained_model_name: *tokenizer_name
-  tokenizer_name: *tokenizer_name
+  pretrained_model_name: ${tokenizer_name}
+  tokenizer_name: ${tokenizer_name}
 
 # Dataloaders (make sure to update these after you modify the starter script!)
 train_loader:
   split: train
-  tokenizer_name: *tokenizer_name
-  max_seq_len: *max_seq_len
+  tokenizer_name: ${tokenizer_name}
+  max_seq_len: ${max_seq_len}
   shuffle: true
   drop_last: true
   num_workers: 8
 
 eval_loader:
   split: validation
-  tokenizer_name: *tokenizer_name
-  max_seq_len: *max_seq_len
+  tokenizer_name: ${tokenizer_name}
+  max_seq_len: ${max_seq_len}
   shuffle: true
   drop_last: true
   num_workers: 8

examples/bert/yamls/finetuning/mcloud_run.yaml

@@ -27,31 +27,31 @@ parameters:
   # Run Name
   run_name: *run_name
 
-  tokenizer_name: &tokenizer_name bert-base-uncased
-  max_seq_len: &max_seq_len 128
+  tokenizer_name: bert-base-uncased
+  max_seq_len: 128
 
   load_path: # (Optionally, but recommended) provide a composer checkpoint to use for the starting weights
 
   # Model
   model:
     name: mosaic_bert
     num_labels: 2 # <-- Make sure to update these after you modify the starter script!
-    pretrained_model_name: *tokenizer_name
-    tokenizer_name: *tokenizer_name
+    pretrained_model_name: ${tokenizer_name}
+    tokenizer_name: ${tokenizer_name}
 
   # Dataloaders (make sure to update these after you modify the starter script!)
   train_loader:
     split: train
-    tokenizer_name: *tokenizer_name
-    max_seq_len: *max_seq_len
+    tokenizer_name: ${tokenizer_name}
+    max_seq_len: ${max_seq_len}
     shuffle: true
     drop_last: true
     num_workers: 8
 
   eval_loader:
     split: validation
-    tokenizer_name: *tokenizer_name
-    max_seq_len: *max_seq_len
+    tokenizer_name: ${tokenizer_name}
+    max_seq_len: ${max_seq_len}
     shuffle: true
     drop_last: true
     num_workers: 8

examples/bert/yamls/finetuning/mosaic-bert-base-uncased.yaml

@@ -10,31 +10,31 @@
 # Run Name
 run_name: finetune-mosaic-bert
 
-tokenizer_name: &tokenizer_name bert-base-uncased
-max_seq_len: &max_seq_len 128
+tokenizer_name: bert-base-uncased
+max_seq_len: 128
 
 load_path: # (Optionally) provide a composer checkpoint to use for the starting weights
 
 # Model
 model:
   name: mosaic_bert
   num_labels: 2 # <-- Make sure to update these after you modify the starter script!
-  pretrained_model_name: *tokenizer_name
-  tokenizer_name: *tokenizer_name
+  pretrained_model_name: ${tokenizer_name}
+  tokenizer_name: ${tokenizer_name}
 
 # Dataloaders (make sure to update these after you modify the starter script!)
 train_loader:
   split: train
-  tokenizer_name: *tokenizer_name
-  max_seq_len: *max_seq_len
+  tokenizer_name: ${tokenizer_name}
+  max_seq_len: ${max_seq_len}
   shuffle: true
   drop_last: true
   num_workers: 8
 
 eval_loader:
   split: validation
-  tokenizer_name: *tokenizer_name
-  max_seq_len: *max_seq_len
+  tokenizer_name: ${tokenizer_name}
+  max_seq_len: ${max_seq_len}
   shuffle: true
   drop_last: true
   num_workers: 8

examples/bert/yamls/main/hf-bert-base-uncased.yaml

@@ -3,12 +3,12 @@
 
 # Follow the instructions in the README to set up ./my-copy-c4
 # Or point data paths to your remote C4 dataset
-data_local: &data_local ./my-copy-c4
-data_remote: &data_remote null
+data_local: ./my-copy-c4
+data_remote: null
 
-max_seq_len: &max_seq_len 128
-tokenizer_name: &tokenizer_name bert-base-uncased
-mlm_probability: &mlm_probability 0.15
+max_seq_len: 128
+tokenizer_name: bert-base-uncased
+mlm_probability: 0.15
 
 # Run Name
 run_name: hf-bert-base-uncased
@@ -17,8 +17,8 @@ run_name: hf-bert-base-uncased
 model:
   name: hf_bert
   use_pretrained: false # Train the model from scratch. Set to true to start from the HF off-the-shelf weights.
-  pretrained_model_name: *tokenizer_name
-  tokenizer_name: *tokenizer_name
+  pretrained_model_name: ${tokenizer_name}
+  tokenizer_name: ${tokenizer_name}
   # This implementation generally uses the default architecture values for from the Hugging Face BertConfig object
   # These values can be changed here when pretraining from scratch. Note that these should only be used
   # if used_pretained: false, otherwise the model will not be loaded properly
@@ -32,25 +32,25 @@ model:
 train_loader:
   name: text
   dataset:
-    local: *data_local
-    remote: *data_remote
+    local: ${data_local}
+    remote: ${data_remote}
     split: train
-    tokenizer_name: *tokenizer_name
-    max_seq_len: *max_seq_len
+    tokenizer_name: ${tokenizer_name}
+    max_seq_len: ${max_seq_len}
     group_method: truncate
     shuffle: true
-    mlm_probability: *mlm_probability
+    mlm_probability: ${mlm_probability}
   drop_last: true
   num_workers: 8
 
 eval_loader:
   name: text
   dataset:
-    local: *data_local
-    remote: *data_remote
+    local: ${data_local}
+    remote: ${data_remote}
     split: val
-    tokenizer_name: *tokenizer_name
-    max_seq_len: *max_seq_len
+    tokenizer_name: ${tokenizer_name}
+    max_seq_len: ${max_seq_len}
     group_method: truncate
     shuffle: false
     mlm_probability: 0.15 # We always evaluate at 15% masking for consistent comparison