prepared for run by prof Filip Gralinski (done)

code/README.md

@@ -20,9 +20,11 @@ pdm install
 ```
 ### 4. Run training code
 ```bash
-pdm run_training
+pdm train
 ```
-or
+
+## Required secrets
 ```
-pdm run src/train_codebert_mlm.py
+export HF_TOKEN=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
+export WANDB_API_KEY=313671f10f2a389b3171b32da8d4abdad91aaa7c
 ```

code/src/config.json

@@ -1,8 +1,9 @@
 {
     "extra_embeddings": true,
-    "run_name": "no-sinusoidal",
+    "run_name": "tree-continued",
     "data_dir": "./data/codeparrot-clean-parsed-starencoder-no-comments/",
-    "output_dir": "./outputs/long-no-comments-starencoder-no-sinusoidal",
+    "output_dir": "./outputs/no-comments-starencoder-tree-2",
+    "checkpoint": null,
     "seed": 420,
     "mlm_probability": 0.15,
     "batch_size": 32,

code/src/training.py

@@ -1,23 +1,22 @@
 import wandb
+
 import json
+import torch
+import random
 import logging
+import numpy as np
 from pathlib import Path
-from datasets import load_from_disk, DatasetDict
+from safetensors.torch import load_file
+from datasets import load_from_disk, DatasetDict, load_dataset
 from transformers import (
-    RobertaConfig,
     AutoConfig,
-    RobertaForMaskedLM,
     AutoTokenizer,
     TrainingArguments,
     Trainer,
     DataCollatorForLanguageModeling,
     AutoModelForMaskedLM
 )
-import random
-import numpy as np
-import torch
 
-from tree_codebert import TreeCodeBERTForPreTraining
 from tree_starencoder import TreeStarEncoderForPreTraining
 
 logging.basicConfig(
@@ -51,22 +50,15 @@ def main():
     set_seed(config['seed'])
 
     # Initialize W&B
-    wandb.init(project='codeparrot-starencoder-no-comments', config=config, name=config['run_name'])
+    wandb.init(project='gralinski', config=config, name=config['run_name'])
     
     # Upload the training files to W&B
     wandb.save(__file__)
-    wandb.save(Path(__file__).parent / 'config.json')
+    wandb.save(current_dir / 'config.json')
     if config['extra_embeddings']:
         wandb.save(current_dir / 'tree_starencoder.py')
 
-    if 'CodeSearchNet' in config['data_dir']:
-        dataset = DatasetDict({
-            'train': load_from_disk(data_dir / 'train'),
-            'valid': load_from_disk(data_dir / 'valid'),
-            'test': load_from_disk(data_dir / 'test')
-        })
-    else:
-        dataset = load_from_disk(data_dir)
+    dataset = load_dataset("patrykbart/codeparrot-clean-no-comments-starencoder-small", split='train', num_proc=16, cache_dir=data_dir.parent)
     if config['num_samples'] > 0:
         dataset = dataset.select(range(config['num_samples']))
     train_testvalid = dataset.train_test_split(test_size=config['test_size'] + config['valid_size'])
@@ -83,11 +75,9 @@ def main():
     
     # Continue with the rest of processing
     columns_to_remove = dataset['train'].column_names
-    columns_to_remove.remove('input_ids')
-    columns_to_remove.remove('attention_mask')
+    columns_to_remove = [col for col in columns_to_remove if col not in ['input_ids', 'attention_mask']]
     if config['extra_embeddings']:
-        columns_to_remove.remove('depths')
-        columns_to_remove.remove('sibling_idxs')
+        columns_to_remove = [col for col in columns_to_remove if col not in ['depths', 'sibling_idxs']]
     dataset = dataset.remove_columns(columns_to_remove)
     logger.info(f'Loaded dataset:\n{dataset}')
     
@@ -102,12 +92,20 @@ def main():
         logger.info("Set padding token to be the same as the EOS token.")
 
     model_config = AutoConfig.from_pretrained('bigcode/starencoder')
-    if config['extra_embeddings']:
-        model = TreeStarEncoderForPreTraining(model_config)
-    else:
-        model = AutoModelForMaskedLM.from_config(model_config)
+    model = TreeStarEncoderForPreTraining(model_config) if config['extra_embeddings'] else AutoModelForMaskedLM.from_config(model_config)
     logger.info(f'Loaded model: {model.__class__.__name__}')
 
+    # Load checkpoint if provided
+    if config['checkpoint'] is not None:
+        checkpoint_path = Path(config['checkpoint']) / 'model.safetensors'
+        logger.info(f'Loading checkpoint from {checkpoint_path}')
+        state_dict = load_file(checkpoint_path)
+        model.load_state_dict(state_dict, strict=False)
+        model.tie_weights()
+        config['warmup_steps'] = 0
+        config['learning_rate'] = 4.8701e-7
+        logger.info('Checkpoint loaded successfully.')
+    
     # Setup training arguments
     training_args = TrainingArguments(
         output_dir=str(output_dir),

code/src/tree_starencoder.py

@@ -13,12 +13,12 @@ class TreeStarEncoderForPreTraining(BertForMaskedLM):
         super().__init__(config)
         self.config = config
 
-        # self.fusion_layer = nn.Sequential(
-        #     nn.Linear(config.hidden_size * 4, config.hidden_size),
-        #     nn.GELU(),
-        #     nn.Dropout(config.hidden_dropout_prob),
-        #     nn.LayerNorm(config.hidden_size)
-        # )
+        self.fusion_layer = nn.Sequential(
+            nn.Linear(config.hidden_size * 3, config.hidden_size),
+            nn.GELU(),
+            nn.Dropout(config.hidden_dropout_prob),
+            nn.LayerNorm(config.hidden_size)
+        )
 
         # Override config to set max_seq_length
         config.max_position_embeddings = max_seq_length
@@ -31,13 +31,13 @@ class TreeStarEncoderForPreTraining(BertForMaskedLM):
         
         self.seq_pos_embeddings = nn.Embedding(max_seq_length, config.hidden_size)
         
-        # # Initialize sequential position embeddings with sinusoidal pattern
-        # position = torch.arange(max_seq_length).unsqueeze(1)
-        # div_term = torch.exp(torch.arange(0, config.hidden_size, 2) * (-math.log(10000.0) / config.hidden_size))
-        # pe = torch.zeros(max_seq_length, config.hidden_size)
-        # pe[:, 0::2] = torch.sin(position * div_term)
-        # pe[:, 1::2] = torch.cos(position * div_term)
-        # self.seq_pos_embeddings.weight.data.copy_(pe)
+        # Initialize sequential position embeddings with sinusoidal pattern
+        position = torch.arange(max_seq_length).unsqueeze(1)
+        div_term = torch.exp(torch.arange(0, config.hidden_size, 2) * (-math.log(10000.0) / config.hidden_size))
+        pe = torch.zeros(max_seq_length, config.hidden_size)
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        self.seq_pos_embeddings.weight.data.copy_(pe)
         
         # New node type embeddings
         self.node_type_embeddings = nn.Embedding(217, config.hidden_size)
@@ -72,10 +72,11 @@ class TreeStarEncoderForPreTraining(BertForMaskedLM):
         # node_type_embeddings = self.node_type_embeddings(node_types)
         
         # combined = torch.cat([token_embeddings, tree_embeddings, seq_embeddings, node_type_embeddings], dim=-1)
-        # combined_embeddings = self.fusion_layer(combined)
+        combined = torch.cat([token_embeddings, tree_embeddings, seq_embeddings], dim=-1)
+        combined_embeddings = self.fusion_layer(combined)
 
         # Add the embeddings instead of concatenating
-        combined_embeddings = token_embeddings + tree_embeddings + seq_embeddings
+        # combined_embeddings = token_embeddings + tree_embeddings + seq_embeddings
         
         combined_embeddings = self.norm(combined_embeddings)