Feat: implement hybrid fine-tuning of encoder and decoder networks

separately

.gitignore

@@ -0,0 +1 @@
+*.zip

end_to_end/.gitignore

@@ -2,3 +2,4 @@ models
 raw_data.csv
 train_all.csv
 __pycache__
+*.pt

end_to_end/deduplication.py

@@ -31,8 +31,9 @@ class BertEmbedder:
         model = AutoModelForSequenceClassification.from_pretrained(model_checkpoint)
         self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         # device = "cpu"
-        model.to(self.device)
+        self.model = model.to(self.device)
-        self.model = model.eval()
+        self.model = self.model.eval()
+        # self.model = torch.compile(self.model)
 
 
     def make_embedding(self, batch_size=128):
@@ -75,10 +76,11 @@ class T5Embedder:
         self.tokenizer.add_special_tokens({"additional_special_tokens": additional_special_tokens})
 
         model = AutoModelForSeq2SeqLM.from_pretrained(model_checkpoint)
-        self.device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         # device = "cpu"
         model.to(self.device)
         self.model = model.eval()
+        self.model = torch.compile(self.model)
 
 
 
@@ -251,10 +253,26 @@ def run_deduplication(
 
     # generate your embeddings
     checkpoint_path = 'models/bert_model'
+
+    # cache embeddings
+    file_path = "train_embeds.pt"
+    if os.path.exists(file_path):
+        # Load the tensor if the file exists
+        tensor = torch.load(file_path, weights_only=True)
+        print("Loaded tensor")
+    else:
+        # Create and save the tensor if the file doesn't exist
+        print('generate train embeddings')
+        train_embedder = Embedder(input_df=train_df, batch_size=batch_size)
+        tensor = train_embedder.make_embedding(checkpoint_path)
+        torch.save(tensor, file_path, weights_only=True)
+        print("Tensor saved to file.")
+    
+    train_embeds = tensor
+
+
+    # if we can, we can cache the train embeddings and load directly
     # we can generate the train embeddings once and re-use for every ship
-    print('generate train embeddings')
-    train_embedder = Embedder(input_df=train_df, batch_size=batch_size)
-    train_embeds = train_embedder.make_embedding(checkpoint_path)
 
     # generate new embeddings for each ship
     print('generate test embeddings')

end_to_end/mapper.py

@@ -3,11 +3,14 @@ from torch.utils.data import DataLoader
 from transformers import (
     T5TokenizerFast,
     AutoModelForSeq2SeqLM,
+    StoppingCriteria,
+    StoppingCriteriaList,
 )
 import os
 from tqdm import tqdm
 from datasets import Dataset
 import numpy as np
+from torch.nn.utils.rnn import pad_sequence
 
 os.environ['TOKENIZERS_PARALLELISM'] = 'false'
 
@@ -35,9 +38,11 @@ class Mapper():
         # load model
         # Define the directory and the pattern
         model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint_path)
-        model = torch.compile(model)
         # set model to eval
         self.model = model.eval()
+        self.model.cuda()
+        # self.model = torch.compile(self.model)
+
 
 
 
@@ -59,8 +64,8 @@ class Mapper():
                 desc = f"<DESC>{row['tag_description']}<DESC>"
                 unit = f"<UNIT>{row['unit']}<UNIT>"
                 element = {
-                    'input' : f"{desc}{unit}",
+                    'input' : f"{desc}{unit}"
-                    'output': f"<THING_START>{row['thing']}<THING_END><PROPERTY_START>{row['property']}<PROPERTY_END>",
+                    # 'output': f"<THING_START>{row['thing']}<THING_END><PROPERTY_START>{row['property']}<PROPERTY_END>",
                 }
                 output_list.append(element)
 
@@ -68,16 +73,16 @@ class Mapper():
 
         def _preprocess_function(example):
             input = example['input']
-            target = example['output']
+            # target = example['output']
             # text_target sets the corresponding label to inputs
             # there is no need to create a separate 'labels'
             model_inputs = self.tokenizer(
                 input,
-                text_target=target, 
+                # text_target=target, 
-                max_length=max_length,
+                # max_length=max_length,
-                return_tensors="pt",
+                padding=True,
-                padding='max_length',
                 truncation=True,
+                return_tensors="pt",
             )
             return model_inputs
 
@@ -93,19 +98,55 @@ class Mapper():
             remove_columns=test_dataset.column_names,
         )
         # datasets = _preprocess_function(test_dataset)
-        datasets.set_format(type='torch', columns=['input_ids', 'attention_mask', 'labels'])
+        datasets.set_format(type='torch', columns=['input_ids', 'attention_mask'])
+
+        def _custom_collate_fn(batch):
+            # Extract data and targets separately if needed
+            inputs = [item['input_ids'] for item in batch]
+            attention_masks = [item['attention_mask'] for item in batch]
+
+            # Pad data to the same length
+            padded_inputs = pad_sequence(inputs, batch_first=True)
+            padded_attention_masks = pad_sequence(attention_masks, batch_first=True)
+
+            return {'input_ids': padded_inputs, 'attention_mask': padded_attention_masks}
+
 
         # create dataloader
-        self.dataloader = DataLoader(datasets, batch_size=batch_size)
+        self.dataloader = DataLoader(
+            datasets, 
+            batch_size=batch_size,
+            collate_fn=_custom_collate_fn
+        )
 
 
     def generate(self):
         device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
-        MAX_GENERATE_LENGTH = 128
+        MAX_GENERATE_LENGTH = 120
 
         pred_generations = []
-        pred_labels = []
+        # pred_labels = []
-        self.model.cuda()
+        # self.model already assigned to device
+        # self.model.cuda()
+
+        # introduce early stopping so that it doesn't have to generate max length
+        class StopOnEndToken(StoppingCriteria):
+            def __init__(self, end_token_id):
+                self.end_token_id = end_token_id
+
+            def __call__(self, input_ids, scores, **kwargs):
+                # Check if the last token in any sequence is the end token
+                batch_stopped = input_ids[:, -1] == self.end_token_id
+                # only stop if all have reached end token
+                if batch_stopped.all():
+                    return True  # Stop generation for the entire batch
+                return False
+
+        # Define the end token ID (e.g., the ID for <eos>)
+        end_token_id = 32103  # property end token
+
+        # Use the stopping criteria
+        stopping_criteria = StoppingCriteriaList([StopOnEndToken(end_token_id)])
 
         print("start generation")
         for batch in tqdm(self.dataloader):
@@ -113,7 +154,7 @@ class Mapper():
             input_ids = batch['input_ids']
             attention_mask = batch['attention_mask']
             # save labels too
-            pred_labels.extend(batch['labels'])
+            # pred_labels.extend(batch['labels'])
             
 
             # Move to GPU if available
@@ -126,7 +167,11 @@ class Mapper():
                 with torch.no_grad():
                     outputs = self.model.generate(input_ids,
                                             attention_mask=attention_mask,
-                                            max_length=MAX_GENERATE_LENGTH)
+                                            max_length=MAX_GENERATE_LENGTH,
+                                            # eos_token_id=32103,
+                                            # early_stopping=True,
+                                            use_cache=True,
+                                            stopping_criteria=stopping_criteria)
                 
                     # Decode the output and print the results
                     pred_generations.extend(outputs.to("cpu"))

end_to_end/run.py

@@ -7,8 +7,8 @@ from preprocess import Abbreviator
 from deduplication import run_deduplication
 
 # global config
-BATCH_SIZE = 1024
+BATCH_SIZE = 512
-SHIPS_LIST = [1000,1001,1003]
+SHIPS_LIST = [1000,1001,1003,1004]
 
 # %%
 # START: we import the raw data csv and extract only a few ships from it to simulate incoming json
@@ -17,7 +17,8 @@ full_df = pd.read_csv(data_path, skipinitialspace=True)
 # subset ships only to that found in SHIPS_LIST
 df = full_df[full_df['ships_idx'].isin(SHIPS_LIST)].reset_index(drop=True)
 
-num_rows = 2000
+# test parameters
+num_rows = len(df) - 1
 df = df[:num_rows]
 print(len(df))
 
@@ -58,7 +59,7 @@ df = run_deduplication(
     test_df=df,
     train_df=train_df,
     batch_size=BATCH_SIZE,
-    threshold=0.9,
+    threshold=0.85,
     diagnostic=True)
 
 # %%

train/hybrid_t5_complete_desc_unit/.gitignore

@@ -0,0 +1,2 @@
+checkpoint*
+tensorboard-log

train/hybrid_t5_complete_desc_unit/custom_t5/.gitignore

@@ -0,0 +1 @@
+__pycache__

train/hybrid_t5_complete_desc_unit/custom_t5/modeling_t5.py

@@ -0,0 +1,125 @@
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers import (
+    T5PreTrainedModel,
+    T5Model
+    
+)
+
+from transformers.modeling_outputs import (
+    SequenceClassifierOutput,
+)
+
+def mean_pooling(encoder_outputs, attention_mask):
+    """
+    Perform mean pooling over encoder outputs, considering the attention mask.
+    """
+    hidden_states = encoder_outputs.last_hidden_state  # Shape: (batch_size, seq_length, hidden_size)
+    mask = attention_mask.unsqueeze(-1)  # Shape: (batch_size, seq_length, 1)
+    masked_hidden_states = hidden_states * mask  # Zero out padding tokens
+    sum_hidden_states = masked_hidden_states.sum(dim=1)  # Sum over sequence length
+    sum_mask = mask.sum(dim=1)  # Sum the mask (number of non-padding tokens)
+    return sum_hidden_states / sum_mask  # Mean pooling
+
+
+class T5EncoderForSequenceClassification(T5PreTrainedModel):
+
+    def __init__(self, checkpoint, tokenizer, config, num_labels):
+        super().__init__(config)
+        self.num_labels = num_labels
+        self.config = config
+
+        # we force the loading of a pre-trained model here
+        self.t5 = T5Model.from_pretrained(checkpoint)
+        self.t5.resize_token_embeddings(len(tokenizer))
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, self.num_labels)
+
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+
+        # encoder_outputs = self.t5.encoder(
+        #     input_ids,
+        #     attention_mask=attention_mask,
+        #     head_mask=head_mask,
+        #     inputs_embeds=inputs_embeds,
+        #     output_attentions=output_attentions,
+        #     output_hidden_states=output_hidden_states,
+        #     return_dict=return_dict,
+        # )
+
+
+        encoder_outputs = self.t5.encoder(input_ids, attention_mask=attention_mask)
+        # last_hidden_state = encoder_outputs.last_hidden_state
+        # use mean of hidden state
+        # pooled_output = mean_pooling(encoder_outputs, attention_mask)
+
+        # Use the hidden state of the first token as the sequence representation
+        pooled_output = encoder_outputs.last_hidden_state[:, 0, :]  # Shape: (batch_size, hidden_size)
+
+        # pooled_output = encoder_outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+        if not return_dict:
+            output = (logits,) + encoder_outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )

train/hybrid_t5_complete_desc_unit/mapping_prediction/.gitignore

@@ -0,0 +1,2 @@
+__pycache__
+exports/

train/hybrid_t5_complete_desc_unit/mapping_prediction/inference.py

@@ -0,0 +1,168 @@
+import torch
+from torch.utils.data import DataLoader
+from transformers import (
+    T5TokenizerFast,
+    AutoModelForSeq2SeqLM,
+)
+import os
+from tqdm import tqdm
+from datasets import Dataset
+import numpy as np
+
+os.environ['TOKENIZERS_PARALLELISM'] = 'false'
+
+
+class Inference():
+    tokenizer: T5TokenizerFast
+    model: torch.nn.Module
+    dataloader: DataLoader
+
+    def __init__(self, checkpoint_path):
+        self._create_tokenizer()
+        self._load_model(checkpoint_path)
+
+
+    def _create_tokenizer(self):
+        # %%
+        # load tokenizer
+        self.tokenizer = T5TokenizerFast.from_pretrained("t5-small", return_tensors="pt", clean_up_tokenization_spaces=True)
+        # Define additional special tokens
+        additional_special_tokens = ["<THING_START>", "<THING_END>", "<PROPERTY_START>", "<PROPERTY_END>", "<NAME>", "<DESC>", "SIG", "UNIT", "DATA_TYPE"]
+        # Add the additional special tokens to the tokenizer
+        self.tokenizer.add_special_tokens({"additional_special_tokens": additional_special_tokens})
+
+    def _load_model(self, checkpoint_path: str):
+        # load model
+        # Define the directory and the pattern
+        model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint_path)
+        model = torch.compile(model)
+        # set model to eval
+        self.model = model.eval()
+
+
+
+
+    def prepare_dataloader(self, input_df, batch_size, max_length):
+        """
+        *arguments*
+        - input_df: input dataframe containing fields 'tag_description', 'thing', 'property'
+        - batch_size: the batch size of dataloader output
+        - max_length: length of tokenizer output
+        """
+        print("preparing dataloader")
+        # convert each dataframe row into a dictionary
+        # outputs a list of dictionaries
+
+        def _process_df(df):
+            output_list = []
+            for _, row in df.iterrows():
+                desc = f"<DESC>{row['tag_description']}<DESC>"
+                unit = f"<UNIT>{row['unit']}<UNIT>"
+                element = {
+                    'input' : f"{desc}{unit}",
+                    'output': f"<THING_START>{row['thing']}<THING_END><PROPERTY_START>{row['property']}<PROPERTY_END>",
+                }
+                output_list.append(element)
+
+            return output_list
+
+        def _preprocess_function(example):
+            input = example['input']
+            target = example['output']
+            # text_target sets the corresponding label to inputs
+            # there is no need to create a separate 'labels'
+            model_inputs = self.tokenizer(
+                input,
+                text_target=target, 
+                max_length=max_length,
+                return_tensors="pt",
+                padding="max_length",
+                truncation=True,
+            )
+            return model_inputs
+
+        test_dataset = Dataset.from_list(_process_df(input_df))
+
+
+        # map maps function to each "row" in the dataset
+        # aka the data in the immediate nesting
+        datasets = test_dataset.map(
+            _preprocess_function,
+            batched=True,
+            num_proc=1,
+            remove_columns=test_dataset.column_names,
+        )
+        # datasets = _preprocess_function(test_dataset)
+        datasets.set_format(type='torch', columns=['input_ids', 'attention_mask', 'labels'])
+
+        # create dataloader
+        self.dataloader = DataLoader(datasets, batch_size=batch_size)
+
+
+    def generate(self):
+        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        MAX_GENERATE_LENGTH = 128
+
+        pred_generations = []
+        pred_labels = []
+
+        print("start generation")
+        for batch in tqdm(self.dataloader):
+            # Inference in batches
+            input_ids = batch['input_ids']
+            attention_mask = batch['attention_mask']
+            # save labels too
+            pred_labels.extend(batch['labels'])
+            
+
+            # Move to GPU if available
+            input_ids = input_ids.to(device)
+            attention_mask = attention_mask.to(device)
+            self.model.to(device)
+
+            # Perform inference
+            with torch.no_grad():
+                outputs = self.model.generate(input_ids,
+                                        attention_mask=attention_mask,
+                                        max_length=MAX_GENERATE_LENGTH)
+                
+                # Decode the output and print the results
+                pred_generations.extend(outputs.to("cpu"))
+
+
+
+        # %%
+        # extract sequence and decode
+        def extract_seq(tokens, start_value, end_value):
+            if start_value not in tokens or end_value not in tokens:
+                return None  # Or handle this case according to your requirements
+            start_id = np.where(tokens == start_value)[0][0]
+            end_id = np.where(tokens == end_value)[0][0]
+
+            return tokens[start_id+1:end_id]
+
+
+        def process_tensor_output(tokens):
+            thing_seq = extract_seq(tokens, 32100, 32101) # 32100 = <THING_START>, 32101 = <THING_END>
+            property_seq = extract_seq(tokens, 32102, 32103) # 32102 = <PROPERTY_START>, 32103 = <PROPERTY_END>
+            p_thing = None
+            p_property = None
+            if (thing_seq is not None):
+                p_thing =  self.tokenizer.decode(thing_seq, skip_special_tokens=False)
+            if (property_seq is not None):
+                p_property =  self.tokenizer.decode(property_seq, skip_special_tokens=False)
+            return p_thing, p_property
+
+        # decode prediction labels
+        def decode_preds(tokens_list):
+            thing_prediction_list = []
+            property_prediction_list = []
+            for tokens in tokens_list:
+                p_thing, p_property = process_tensor_output(tokens)
+                thing_prediction_list.append(p_thing)
+                property_prediction_list.append(p_property)
+            return thing_prediction_list, property_prediction_list 
+
+        thing_prediction_list, property_prediction_list = decode_preds(pred_generations)
+        return thing_prediction_list, property_prediction_list
+

train/hybrid_t5_complete_desc_unit/mapping_prediction/output.txt

@@ -0,0 +1,6 @@
+
+Accuracy for fold 1: 0.9427354472314246
+Accuracy for fold 2: 0.8859813084112149
+Accuracy for fold 3: 0.9683734939759037
+Accuracy for fold 4: 0.9762131303520457
+Accuracy for fold 5: 0.907924874026569

train/hybrid_t5_complete_desc_unit/mapping_prediction/predict.py

@@ -0,0 +1,73 @@
+
+import pandas as pd
+import os
+import glob
+from inference import Inference
+
+checkpoint_directory =  '../'
+
+BATCH_SIZE = 512
+
+def infer_and_select(fold):
+    print(f"Inference for fold {fold}")
+    # import test data
+    data_path = f"../../../data_preprocess/exports/dataset/group_{fold}/test_all.csv"
+    df = pd.read_csv(data_path, skipinitialspace=True)
+
+    # get target data
+    data_path = f"../../../data_preprocess/exports/dataset/group_{fold}/train_all.csv"
+    train_df = pd.read_csv(data_path, skipinitialspace=True)
+    # processing to help with selection later
+    train_df['thing_property'] = train_df['thing'] + " " + train_df['property']
+
+
+    ##########################################
+    # run inference
+    # checkpoint
+    # Use glob to find matching paths
+    directory = os.path.join(checkpoint_directory, f'checkpoint_fold_{fold}b')
+    # Use glob to find matching paths
+    # path is usually checkpoint_fold_1/checkpoint-<step number>
+    # we are guaranteed to save only 1 checkpoint from training
+    pattern = 'checkpoint-*'
+    checkpoint_path = glob.glob(os.path.join(directory, pattern))[0]
+
+
+    infer = Inference(checkpoint_path)
+    infer.prepare_dataloader(df, batch_size=BATCH_SIZE, max_length=128)
+    thing_prediction_list, property_prediction_list = infer.generate()
+
+    # add labels too
+    # thing_actual_list, property_actual_list = decode_preds(pred_labels)
+    # Convert the list to a Pandas DataFrame
+    df_out = pd.DataFrame({
+        'p_thing': thing_prediction_list, 
+        'p_property': property_prediction_list
+    })
+    # df_out['p_thing_correct'] = df_out['p_thing'] == df_out['thing']
+    # df_out['p_property_correct'] = df_out['p_property'] == df_out['property']
+    df = pd.concat([df, df_out], axis=1)
+
+    # we can save the t5 generation output here
+    df.to_csv(f"exports/result_group_{fold}.csv", index=False)
+
+    # here we want to evaluate mapping accuracy within the valid in mdm data only
+    in_mdm = df['MDM']
+    condition_correct_thing = df['p_thing'] == df['thing']
+    condition_correct_property = df['p_property'] == df['property']
+    prediction_mdm_correct = sum(condition_correct_thing & condition_correct_property & in_mdm)
+    pred_correct_proportion = prediction_mdm_correct/sum(in_mdm)
+
+    # write output to file output.txt
+    with open("output.txt", "a") as f:
+        print(f'Accuracy for fold {fold}: {pred_correct_proportion}', file=f)
+
+###########################################  
+# Execute for all folds
+
+# reset file before writing to it
+with open("output.txt", "w") as f:
+    print('', file=f)
+
+for fold in [1,2,3,4,5]:
+    infer_and_select(fold)

train/hybrid_t5_complete_desc_unit/train_decoder.py

@@ -0,0 +1,227 @@
+# %%
+
+# from datasets import load_from_disk
+import os
+import glob
+
+os.environ['NCCL_P2P_DISABLE'] = '1'
+os.environ['NCCL_IB_DISABLE'] = '1'
+os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
+os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"
+
+import torch
+from custom_t5.modeling_t5 import T5EncoderForSequenceClassification
+
+from safetensors.torch import load_file
+from transformers import (
+    T5Config,
+    T5TokenizerFast,
+    AutoModelForSeq2SeqLM,
+    DataCollatorForSeq2Seq,
+    Seq2SeqTrainer,
+    EarlyStoppingCallback,
+    Seq2SeqTrainingArguments,
+    T5ForConditionalGeneration,
+    T5Model
+)
+import evaluate
+import numpy as np
+import pandas as pd
+# import matplotlib.pyplot as plt
+from datasets import Dataset, DatasetDict
+
+
+
+torch.set_float32_matmul_precision('high')
+
+# outputs a list of dictionaries
+def process_df_to_dict(df):
+    output_list = []
+    for _, row in df.iterrows():
+        desc = f"<DESC>{row['tag_description']}<DESC>"
+        unit = f"<UNIT>{row['unit']}<UNIT>"
+        element = {
+            'input' : f"{desc}{unit}",
+            'output': f"<THING_START>{row['thing']}<THING_END><PROPERTY_START>{row['property']}<PROPERTY_END>",
+        }
+        output_list.append(element)
+
+    return output_list
+
+
+def create_split_dataset(fold):
+    # train 
+    data_path = f"../../data_preprocess/exports/dataset/group_{fold}/train_all.csv"
+    train_df = pd.read_csv(data_path, skipinitialspace=True)
+
+    # valid
+    data_path = f"../../data_preprocess/exports/dataset/group_{fold}/valid.csv"
+    validation_df = pd.read_csv(data_path, skipinitialspace=True)
+
+    combined_data = DatasetDict({
+        'train': Dataset.from_list(process_df_to_dict(train_df)),
+        'validation' : Dataset.from_list(process_df_to_dict(validation_df)),
+    })
+    return combined_data
+
+
+# function to perform training for a given fold
+def train(fold):
+    save_path = f'checkpoint_fold_{fold}b'
+    split_datasets = create_split_dataset(fold)
+
+    # prepare tokenizer
+    model_checkpoint = "t5-small"
+    tokenizer = T5TokenizerFast.from_pretrained(model_checkpoint, return_tensors="pt", clean_up_tokenization_spaces=True)
+    # Define additional special tokens
+    additional_special_tokens = ["<THING_START>", "<THING_END>", "<PROPERTY_START>", "<PROPERTY_END>", "<NAME>", "<DESC>", "<SIG>", "<UNIT>", "<DATA_TYPE>"]
+    # Add the additional special tokens to the tokenizer
+    tokenizer.add_special_tokens({"additional_special_tokens": additional_special_tokens})
+
+    max_length = 120
+
+    # given a dataset entry, run it through the tokenizer
+    def preprocess_function(example):
+        input = example['input']
+        target = example['output']
+        # text_target sets the corresponding label to inputs
+        # there is no need to create a separate 'labels'
+        model_inputs = tokenizer(
+            input,
+            text_target=target, 
+            max_length=max_length,
+            truncation=True,
+            padding="max_length"
+        )
+        return model_inputs
+
+    # map maps function to each "row" in the dataset
+    # aka the data in the immediate nesting
+    tokenized_datasets = split_datasets.map(
+        preprocess_function,
+        batched=True,
+        num_proc=8,
+        remove_columns=split_datasets["train"].column_names,
+    )
+
+    # https://github.com/huggingface/transformers/pull/28414
+    # model_checkpoint = "google/t5-efficient-tiny"
+    # device_map set to auto to force it to load contiguous weights 
+    # model = AutoModelForSeq2SeqLM.from_pretrained(model_checkpoint, device_map='auto')
+
+    directory = os.path.join(".", f'checkpoint_fold_{fold}a')
+    # Use glob to find matching paths
+    # path is usually checkpoint_fold_1/checkpoint-<step number>
+    # we are guaranteed to save only 1 checkpoint from training
+    pattern = 'checkpoint-*'
+    prev_checkpoint = glob.glob(os.path.join(directory, pattern))[0]
+    # t5_classify = T5Model.from_pretrained(prev_checkpoint)
+    # Load the checkpoint
+    checkpoint_path = f"{prev_checkpoint}/model.safetensors"
+    checkpoint = load_file(checkpoint_path)
+    # Filter out weights related to the classification head
+    t5_weights = {key: value for key, value in checkpoint.items() if "classifier" not in key}
+
+
+    model = T5ForConditionalGeneration.from_pretrained(model_checkpoint)
+    model.load_state_dict(state_dict=t5_weights, strict=False)
+    # important! after extending tokens vocab
+    model.resize_token_embeddings(len(tokenizer))
+
+    # Freeze the encoder
+    for param in model.encoder.parameters():
+        param.requires_grad = False
+
+    # Freeze the shared embedding layer
+    for param in model.shared.parameters():
+        param.requires_grad = False
+
+
+    data_collator = DataCollatorForSeq2Seq(tokenizer, model=model)
+    metric = evaluate.load("sacrebleu")
+
+
+    def compute_metrics(eval_preds):
+        preds, labels = eval_preds
+        # In case the model returns more than the prediction logits
+        if isinstance(preds, tuple):
+            preds = preds[0]
+
+        decoded_preds = tokenizer.batch_decode(preds, 
+                                            skip_special_tokens=False)
+
+        # Replace -100s in the labels as we can't decode them
+        labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
+        decoded_labels = tokenizer.batch_decode(labels,
+                                                skip_special_tokens=False)
+
+        # Remove <PAD> tokens from decoded predictions and labels
+        decoded_preds = [pred.replace(tokenizer.pad_token, '').strip() for pred in decoded_preds]
+        decoded_labels = [[label.replace(tokenizer.pad_token, '').strip()] for label in decoded_labels]
+
+        # Some simple post-processing
+        # decoded_preds = [pred.strip() for pred in decoded_preds]
+        # decoded_labels = [[label.strip()] for label in decoded_labels]
+        # print(decoded_preds, decoded_labels)
+
+        result = metric.compute(predictions=decoded_preds, references=decoded_labels)
+        return {"bleu": result["score"]}
+
+
+    # Generation Config
+    # from transformers import GenerationConfig
+    gen_config = model.generation_config
+    gen_config.max_length = 128
+
+    # compile
+    # model = torch.compile(model, backend="inductor", dynamic=True)
+
+
+    # Trainer
+
+    args = Seq2SeqTrainingArguments(
+        f"{save_path}",
+        # eval_strategy="epoch",
+        eval_strategy="no",
+        logging_dir="tensorboard-log",
+        logging_strategy="epoch",
+        # save_strategy="epoch",
+        load_best_model_at_end=False,
+        learning_rate=1e-3,
+        per_device_train_batch_size=64,
+        per_device_eval_batch_size=64,
+        auto_find_batch_size=False,
+        ddp_find_unused_parameters=False,
+        weight_decay=0.01,
+        save_total_limit=1,
+        num_train_epochs=80,
+        predict_with_generate=True,
+        bf16=True,
+        push_to_hub=False,
+        generation_config=gen_config,
+        remove_unused_columns=False,
+    )
+
+
+    trainer = Seq2SeqTrainer(
+        model,
+        args,
+        train_dataset=tokenized_datasets["train"],
+        eval_dataset=tokenized_datasets["validation"],
+        data_collator=data_collator,
+        tokenizer=tokenizer,
+        compute_metrics=compute_metrics,
+        # callbacks=[EarlyStoppingCallback(early_stopping_patience=3)],
+    )
+
+    # uncomment to load training from checkpoint
+    # checkpoint_path = 'default_40_1/checkpoint-5600'
+    # trainer.train(resume_from_checkpoint=checkpoint_path)
+
+    trainer.train()
+
+# execute training
+for fold in [1,2,3,4,5]:
+    print(fold)
+    train(fold)
+

train/hybrid_t5_complete_desc_unit/train_encoder.py

@@ -0,0 +1,228 @@
+# %%
+
+# from datasets import load_from_disk
+import os
+
+os.environ['NCCL_P2P_DISABLE'] = '1'
+os.environ['NCCL_IB_DISABLE'] = '1'
+os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
+os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"
+
+import torch
+from custom_t5.modeling_t5 import T5EncoderForSequenceClassification
+from transformers import (
+    AutoTokenizer,
+    AutoModelForSequenceClassification,
+    DataCollatorWithPadding,
+    Trainer,
+    EarlyStoppingCallback,
+    TrainingArguments,
+    T5Config,
+)
+import evaluate
+import numpy as np
+import pandas as pd
+# import matplotlib.pyplot as plt
+from datasets import Dataset, DatasetDict
+
+
+
+torch.set_float32_matmul_precision('high')
+
+# %%
+
+# we need to create the mdm_list
+# import the full mdm-only file
+data_path = '../../data_import/exports/data_mapping_mdm.csv'
+full_df = pd.read_csv(data_path, skipinitialspace=True)
+mdm_list = sorted(list((set(full_df['pattern']))))
+
+# # rather than use pattern, we use the real thing and property
+# thing_property = full_df['thing'] + full_df['property']
+# thing_property = thing_property.to_list()
+# mdm_list = sorted(list(set(thing_property)))
+
+
+# %%
+id2label = {}
+label2id = {}
+for idx, val in enumerate(mdm_list):
+    id2label[idx] = val
+    label2id[val] = idx
+
+# %%
+
+# outputs a list of dictionaries
+# processes dataframe into lists of dictionaries
+# each element maps input to output
+# input: tag_description
+# output: class label
+def process_df_to_dict(df, mdm_list):
+    output_list = []
+    for _, row in df.iterrows():
+        desc = f"<DESC>{row['tag_description']}<DESC>"
+        unit = f"<UNIT>{row['unit']}<UNIT>"
+        # pattern = f"{row['thing'] + row['property']}"
+        pattern = f"{row['thing_pattern'] + ' ' + row['property_pattern']}"
+        try:
+            index = mdm_list.index(pattern)
+        except ValueError:
+            print("Error: value not found in MDM list")
+            index = -1
+        element = {
+            'text' : f"{desc}{unit}",
+            'label': index,
+        }
+        output_list.append(element)
+
+    return output_list
+
+
+def create_split_dataset(fold, mdm_list):
+    # train 
+    data_path = f"../../data_preprocess/exports/dataset/group_{fold}/train_all.csv"
+    train_df = pd.read_csv(data_path, skipinitialspace=True)
+
+    # valid
+    data_path = f"../../data_preprocess/exports/dataset/group_{fold}/valid.csv"
+    validation_df = pd.read_csv(data_path, skipinitialspace=True)
+
+    combined_data = DatasetDict({
+        'train': Dataset.from_list(process_df_to_dict(train_df, mdm_list)),
+        'validation' : Dataset.from_list(process_df_to_dict(validation_df, mdm_list)),
+    })
+    return combined_data
+
+
+# %%
+
+# function to perform training for a given fold
+def train(fold):
+
+    save_path = f'checkpoint_fold_{fold}a'
+    split_datasets = create_split_dataset(fold, mdm_list)
+
+    # prepare tokenizer
+
+    # model_checkpoint = "distilbert/distilbert-base-uncased"
+    # model_checkpoint = 'google-bert/bert-base-cased'
+    model_checkpoint = "t5-small"
+    tokenizer = AutoTokenizer.from_pretrained(model_checkpoint, return_tensors="pt", clean_up_tokenization_spaces=True)
+    # Define additional special tokens
+    additional_special_tokens = ["<THING_START>", "<THING_END>", "<PROPERTY_START>", "<PROPERTY_END>", "<NAME>", "<DESC>", "<SIG>", "<UNIT>", "<DATA_TYPE>"]
+    # Add the additional special tokens to the tokenizer
+    tokenizer.add_special_tokens({"additional_special_tokens": additional_special_tokens})
+
+    max_length = 120
+
+    # given a dataset entry, run it through the tokenizer
+    def preprocess_function(example):
+        input = example['text']
+        # text_target sets the corresponding label to inputs
+        # there is no need to create a separate 'labels'
+        model_inputs = tokenizer(
+            input,
+            max_length=max_length,
+            truncation=True,
+            padding="max_length"
+        )
+        return model_inputs
+
+    # map maps function to each "row" in the dataset
+    # aka the data in the immediate nesting
+    tokenized_datasets = split_datasets.map(
+        preprocess_function,
+        batched=True,
+        num_proc=8,
+        remove_columns="text",
+    )
+
+    # %% temp
+    # tokenized_datasets['train'].rename_columns()
+
+    # %%
+    # create data collator
+
+    data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
+
+    # %%
+    # compute metrics
+    metric = evaluate.load("accuracy")
+
+
+    def compute_metrics(eval_preds):
+        preds, labels = eval_preds
+        preds = np.argmax(preds, axis=1)
+        return metric.compute(predictions=preds, references=labels)
+
+    # %%
+    # create id2label and label2id
+
+
+    # %%
+    # model = AutoModelForSequenceClassification.from_pretrained(
+    #     model_checkpoint,
+    #     num_labels=len(mdm_list),
+    #     id2label=id2label,
+    #     label2id=label2id)
+    model = T5EncoderForSequenceClassification(
+        checkpoint=model_checkpoint,
+        tokenizer=tokenizer,
+        config=T5Config.from_pretrained(model_checkpoint),
+        num_labels=len(mdm_list)
+    )
+    # important! after extending tokens vocab
+    # model.t5.resize_token_embeddings(len(tokenizer))
+
+    # model = torch.compile(model, backend="inductor", dynamic=True)
+
+
+    # %%
+    # Trainer
+
+    training_args = TrainingArguments(
+        output_dir=f"{save_path}",
+        # eval_strategy="epoch",
+        eval_strategy="no",
+        logging_dir="tensorboard-log",
+        logging_strategy="epoch",
+        # save_strategy="epoch",
+        load_best_model_at_end=False,
+        learning_rate=1e-3,
+        per_device_train_batch_size=128,
+        per_device_eval_batch_size=128,
+        auto_find_batch_size=False,
+        ddp_find_unused_parameters=False,
+        weight_decay=0.01,
+        save_total_limit=1,
+        num_train_epochs=80,
+        bf16=True,
+        push_to_hub=False,
+        remove_unused_columns=False,
+    )
+
+
+    trainer = Trainer(
+        model,
+        training_args,
+        train_dataset=tokenized_datasets["train"],
+        eval_dataset=tokenized_datasets["validation"],
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+        compute_metrics=compute_metrics,
+        # callbacks=[EarlyStoppingCallback(early_stopping_patience=3)],
+    )
+
+    # uncomment to load training from checkpoint
+    # checkpoint_path = 'default_40_1/checkpoint-5600'
+    # trainer.train(resume_from_checkpoint=checkpoint_path)
+
+    trainer.train()
+
+# execute training
+for fold in [1,2,3,4,5]:
+    print(fold)
+    train(fold)
+
+
+# %%