feat: end-to-end code needed for deployment that includes preprocess,

mapping, post-process (de-duplication)

end_to_end/.README.md

@@ -0,0 +1,5 @@
+# End to End deployment code
+
+This code takes in an input dataframe (assuming that the server code already
+processed json to python dataframe), and applies our mapping + threshold +
+de_duplicate code.

end_to_end/.gitignore

@@ -0,0 +1,4 @@
+models
+raw_data.csv
+train_all.csv
+__pycache__

end_to_end/deduplication.py

@@ -0,0 +1,415 @@
+
+# %%
+import pandas as pd
+import os
+import glob
+from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score, confusion_matrix
+import numpy as np
+from tqdm import tqdm
+
+import torch
+from transformers import (
+    AutoTokenizer,
+    AutoModelForSequenceClassification,
+    AutoModelForSeq2SeqLM,
+)
+
+##################
+# global parameters
+##################
+
+
+
+class BertEmbedder:
+    def __init__(self, input_texts, model_checkpoint):
+        # we need to generate the embedding from list of input strings
+        self.embeddings = []
+        self.inputs = input_texts
+        model_checkpoint = model_checkpoint 
+        self.tokenizer = AutoTokenizer.from_pretrained(model_checkpoint, return_tensors="pt", clean_up_tokenization_spaces=True)
+
+        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.eval()
+
+
+    def make_embedding(self, batch_size=128):
+        all_embeddings = self.embeddings
+        input_texts = self.inputs
+
+        for i in range(0, len(input_texts), batch_size):
+            batch_texts = input_texts[i:i+batch_size]
+            # Tokenize the input text
+            inputs = self.tokenizer(batch_texts, return_tensors="pt", padding=True, truncation=True, max_length=120)
+            input_ids = inputs.input_ids.to(self.device)
+            attention_mask = inputs.attention_mask.to(self.device)
+
+
+            # Pass the input through the encoder and retrieve the embeddings
+            with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
+                with torch.no_grad():
+                    encoder_outputs = self.model(input_ids, attention_mask=attention_mask, output_hidden_states=True)
+                    # get last layer
+                    embeddings = encoder_outputs.hidden_states[-1]
+                    # get cls token embedding
+                    cls_embeddings = embeddings[:, 0, :]  # Shape: (batch_size, hidden_size)
+                    all_embeddings.append(cls_embeddings)
+        
+        # remove the batch list and makes a single large tensor, dim=0 increases row-wise
+        all_embeddings = torch.cat(all_embeddings, dim=0)
+
+        self.embeddings = all_embeddings
+
+class T5Embedder:
+    def __init__(self, input_texts, model_checkpoint):
+        # we need to generate the embedding from list of input strings
+        self.embeddings = []
+        self.inputs = input_texts
+        model_checkpoint = model_checkpoint 
+        self.tokenizer = AutoTokenizer.from_pretrained("t5-base", 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})
+
+        model = AutoModelForSeq2SeqLM.from_pretrained(model_checkpoint)
+        self.device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
+        # device = "cpu"
+        model.to(self.device)
+        self.model = model.eval()
+
+
+
+
+    def make_embedding(self, batch_size=128):
+        all_embeddings = self.embeddings
+        input_texts = self.inputs
+
+        for i in range(0, len(input_texts), batch_size):
+            batch_texts = input_texts[i:i+batch_size]
+            # Tokenize the input text
+            inputs = self.tokenizer(batch_texts, return_tensors="pt", padding=True, truncation=True, max_length=128)
+            input_ids = inputs.input_ids.to(self.device)
+            attention_mask = inputs.attention_mask.to(self.device)
+
+
+            # Pass the input through the encoder and retrieve the embeddings
+
+            with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
+                with torch.no_grad():
+                    encoder_outputs = self.model.encoder(input_ids, attention_mask=attention_mask)
+                    embeddings = encoder_outputs.last_hidden_state
+
+            # Compute the mean pooling of the token embeddings
+            # mean_embedding = embeddings.mean(dim=1)
+            mean_embedding = (embeddings * attention_mask.unsqueeze(-1)).sum(dim=1) / attention_mask.sum(dim=1, keepdim=True)
+            all_embeddings.append(mean_embedding)
+        
+        # remove the batch list and makes a single large tensor, dim=0 increases row-wise
+        all_embeddings = torch.cat(all_embeddings, dim=0)
+
+        self.embeddings = all_embeddings
+
+
+def cosine_similarity_chunked(batch1, batch2, chunk_size=1024):
+    device = 'cuda'
+    batch1_size = batch1.size(0)
+    batch2_size = batch2.size(0)
+    batch2.to(device)
+    
+    # Prepare an empty tensor to store results
+    cos_sim = torch.empty(batch1_size, batch2_size, device=device)
+
+    # Process batch1 in chunks
+    for i in range(0, batch1_size, chunk_size):
+        batch1_chunk = batch1[i:i + chunk_size]  # Get chunk of batch1
+        
+        batch1_chunk.to(device)
+        # Expand batch1 chunk and entire batch2 for comparison
+        # batch1_chunk_exp = batch1_chunk.unsqueeze(1)  # Shape: (chunk_size, 1, seq_len)
+        # batch2_exp = batch2.unsqueeze(0)  # Shape: (1, batch2_size, seq_len)
+        batch2_norms = batch2.norm(dim=1, keepdim=True)
+
+        
+        # Compute cosine similarity for the chunk and store it in the final tensor
+        # cos_sim[i:i + chunk_size] = F.cosine_similarity(batch1_chunk_exp, batch2_exp, dim=-1)
+
+        # Compute cosine similarity by matrix multiplication and normalizing
+        sim_chunk = torch.mm(batch1_chunk, batch2.T) / (batch1_chunk.norm(dim=1, keepdim=True) * batch2_norms.T + 1e-8)
+        
+        # Store the results in the appropriate part of the final tensor
+        cos_sim[i:i + chunk_size] = sim_chunk
+    
+    return cos_sim
+
+
+
+
+###################
+# helper functions
+class Embedder():
+    input_df: pd.DataFrame
+    fold: int
+    batch_size: int
+
+    def __init__(self, input_df, batch_size):
+        self.input_df = input_df
+        self.batch_size = batch_size
+
+
+    def make_embedding(self, checkpoint_path):
+
+        def generate_input_list(df):
+            input_list = []
+            for _, row in df.iterrows():
+                desc = f"<DESC>{row['tag_description']}<DESC>"
+                unit = f"<UNIT>{row['unit']}<UNIT>"
+                # name = f"<NAME>{row['tag_name']}<NAME>"
+                element = f"{desc}{unit}"
+                input_list.append(element)
+            return input_list
+
+        # prepare reference embed
+        train_data = list(generate_input_list(self.input_df))
+        # Define the directory and the pattern
+        # embedder = T5Embedder(train_data, checkpoint_path)
+        embedder = BertEmbedder(train_data, checkpoint_path)
+        embedder.make_embedding(batch_size=self.batch_size)
+        return embedder.embeddings
+
+
+
+
+# the selection function takes in the full cos_sim_matrix then subsets the
+# matrix according to the test_candidates_mask and train_candidates_mask that we
+# give it
+# it returns the most likely source candidate index and score among the source
+# candidate list
+# we then map the local idx to the ship-level idx
+def selection(cos_sim_matrix, source_mask, target_mask):
+    # subset_matrix = cos_sim_matrix[condition_source]
+    # except we are subsetting 2D matrix (row, column)
+    subset_matrix = cos_sim_matrix[np.ix_(source_mask, target_mask)]
+    # we select top-k here
+    # Get the indices of the top-k maximum values along axis 1
+    top_k = 1
+    # returns a potential 2d matrix of which columns have the highest values
+    # top_k_indices = np.argsort(subset_matrix, axis=1)[:, -top_k:]  # Get indices of top k values
+    # this partial sorts and ensures we care only top_k are correctly sorted
+    top_k_indices = np.argpartition(subset_matrix, -top_k, axis=1)[:, -top_k:]
+    
+    # Get the values of the top 5 maximum scores
+    top_k_values = np.take_along_axis(subset_matrix, top_k_indices, axis=1)
+    
+    # Calculate the average of the top-k scores along axis 1
+    y_scores = np.mean(top_k_values, axis=1)
+    max_idx = np.argmax(y_scores)
+    max_score = y_scores[max_idx]
+
+    # convert boolean to indices
+    condition_indices = np.where(source_mask)[0]
+    max_idx = condition_indices[max_idx]
+    
+
+    return max_idx, max_score
+
+
+
+####################
+# global level
+# obtain the full mdm_list
+
+#####################
+# fold level
+
+def run_deduplication(
+    test_df,
+    train_df,
+    batch_size=1024,
+    threshold=0.9,
+    diagnostic=False
+    ):
+
+    # TODO: replace this with a list of values to import
+    # too wasteful to just import everything
+    data_path = '../data_import/exports/data_mapping_mdm.csv'
+    full_df = pd.read_csv(data_path, skipinitialspace=True)
+    full_df['mapping'] = full_df['thing'] + ' ' + full_df['property']
+    full_mdm_mapping_list = sorted(list((set(full_df['mapping']))))
+
+    # set the fold
+    # import test data
+    df = test_df
+    df['p_mapping'] = df['p_thing'] + " " + df['p_property']
+
+    # get target data
+    data_path = "train_all.csv"
+    train_df = pd.read_csv(data_path, skipinitialspace=True)
+    train_df['mapping'] = train_df['thing'] + " " + train_df['property']
+
+    # generate your embeddings
+    checkpoint_path = 'models/bert_model'
+    # 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')
+    test_embedder = Embedder(input_df=df, batch_size=batch_size)
+    global_test_embeds = test_embedder.make_embedding(checkpoint_path)
+
+
+    # create global_answer array
+    # the purpose of this array is to track the classification state at the global
+    # level
+    global_answer = np.zeros(len(df), dtype=bool)
+
+    #############################
+    # ship level
+    # we have to split into per-ship analysis
+    ships_list = sorted(list(set(df['ships_idx'])))
+
+    for ship_idx in tqdm(ships_list):
+        # ship_df = df[df['ships_idx'] == ship_idx]
+        # required to map local ship_answer array to global_answer array
+        # map_local_index_to_global_index = ship_df.index.to_numpy()
+
+        # we want to subset the ship and only p_mdm values
+        ship_mask = df['ships_idx'] == ship_idx
+        map_local_index_to_global_index = np.where(ship_mask)[0]
+        ship_df = df[ship_mask].reset_index(drop=True)
+
+        # subset the test embeds
+        test_embeds = global_test_embeds[map_local_index_to_global_index]
+
+        # generate the cosine sim matrix for the ship level
+        cos_sim_matrix = cosine_similarity_chunked(test_embeds, train_embeds, chunk_size=1024).cpu().numpy()
+
+        ##############################
+        # selection level
+        # The general idea:
+        # step 1: keep only pattern generations that belong to mdm list
+        # -> this removes totally wrong datasets that mapped to totally wrong things
+        # step 2: loop through the mdm list and isolate data in both train and test that
+        # belong to the same pattern class
+        # -> this is more tricky, because we have non-mdm mapping to correct classes
+        # -> so we have to find which candidate is most similar to the training data
+
+        # it is very tricky to keep track of classification across multiple stages so we
+        # will use a boolean answer list to map answers back to the global answer list
+
+        # initialize the local answer list
+        ship_answer_list = np.ones(len(ship_df), dtype=bool)
+
+        ###########
+        # STEP 1A: ensure that the predicted mapping labels are valid
+        pattern_match_mask = ship_df['p_mapping'].apply(lambda x: x in full_mdm_mapping_list).to_numpy()
+        pattern_match_mask = pattern_match_mask.astype(bool)
+        # anything not in the pattern_match_mask are hallucinations
+        # this has the same effect as setting any wrong generations as non-mdm
+        ship_answer_list[~pattern_match_mask] = False
+
+        # # STEP 1B: subset our de-duplication to use only predicted_mdm labels
+        # p_mdm_mask = ship_df['p_mdm']
+        # # assign false to any non p_mdm entries
+        # ship_answer_list[~p_mdm_mask] = False
+        # # modify pattern_match_mask to remove any non p_mdm values
+        # pattern_match_mask = pattern_match_mask & p_mdm_mask
+
+        ###########
+        # STEP 2
+        # we now go through each class found in our generated set
+
+        # we want to identify per-ship mdm classes
+        ship_predicted_classes = sorted(set(ship_df['p_mapping'][pattern_match_mask].to_list()))
+
+        # this function performs the selection given a class
+        # it takes in the cos_sim_matrix
+        # it returns the selection by mutating the answer_list
+        # it sets all relevant idxs to False initially, then sets the selected values to True
+        def selection_for_class(select_class, cos_sim_matrix, answer_list):
+
+            # create local copy of answer_list
+            ship_answer_list = answer_list.copy()
+            # sample_df = ship_df[ship_df['p_mapping'] == select_class]
+
+
+            # we need to set all idx of chosen entries as False in answer_list -> assume wrong by default
+            # selected_idx_list = sample_df.index.to_numpy()
+            selected_idx_list = np.where(ship_df['p_mapping'] == select_class)[0]
+
+            # basic assumption check
+
+            # generate the masking arrays for both test and train embeddings
+            # we select a tuple from each group, and use that as a candidate for selection
+            test_candidates_mask = ship_df['p_mapping'] == select_class
+            # we make candidates to compare against in the data sharing the same class
+            train_candidates_mask = train_df['mapping'] == select_class
+
+            if sum(train_candidates_mask) == 0:
+                # it can be the case that the mdm-valid mapping class is not found in training data
+                # print("not found in training data", select_class)
+                ship_answer_list[selected_idx_list] = False
+                return ship_answer_list
+
+            # perform selection
+            # max_idx is the id
+            max_idx, max_score = selection(cos_sim_matrix, test_candidates_mask, train_candidates_mask)
+
+
+            # set the duplicate entries to False
+            ship_answer_list[selected_idx_list] = False
+            # then only set the one unique chosen value as True
+            if max_score > threshold:
+                ship_answer_list[max_idx] = True
+
+            return ship_answer_list 
+
+        # we choose one mdm class
+        for select_class in ship_predicted_classes:
+            # this resulted in big improvement
+            if (sum(ship_df['p_mapping'] == select_class)) > 0:
+                ship_answer_list = selection_for_class(select_class, cos_sim_matrix, ship_answer_list)
+
+        # we want to write back to global_answer
+        # first we convert local indices to global indices
+        ship_local_indices = np.where(ship_answer_list)[0]
+        ship_global_indices = map_local_index_to_global_index[ship_local_indices]
+        global_answer[ship_global_indices] = True
+
+    # we set all unselected values to None
+    df.loc[~global_answer, 'p_thing'] = None
+    df.loc[~global_answer, 'p_property'] = None
+
+
+    if diagnostic:
+        print(80 * '*')
+
+        y_true = df['MDM'].to_list()
+        y_pred = global_answer
+
+        tn, fp, fn, tp = confusion_matrix(y_true, y_pred).ravel()
+        print(f"tp: {tp}")
+        print(f"tn: {tn}")
+        print(f"fp: {fp}")
+        print(f"fn: {fn}")
+
+        # compute metrics
+        accuracy = accuracy_score(y_true, y_pred)
+        f1 = f1_score(y_true, y_pred)
+        precision = precision_score(y_true, y_pred)
+        recall = recall_score(y_true, y_pred)
+
+        # print the results
+        print(f'accuracy: {accuracy:.5f}')
+        print(f'f1 score: {f1:.5f}')
+        print(f'Precision: {precision:.5f}')
+        print(f'Recall: {recall:.5f}')
+
+
+    return df
+
+

end_to_end/mapper.py

@@ -0,0 +1,169 @@
+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 Mapper():
+    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 = []
+        self.model.cuda()
+
+        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)
+
+            # Perform inference
+            # disable if running on gpu's without tensor cores
+            with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
+                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

end_to_end/preprocess.py

@@ -0,0 +1,77 @@
+# %%
+import re
+from replacement_dict import desc_replacement_dict, unit_replacement_dict
+
+class Abbreviator:
+
+    def __init__(self, df):
+        self.df = df
+
+    def _count_abbreviation_occurrences(self, tag_descriptions, abbreviation):
+        """Count the number of occurrences of the abbreviation in the list of machine descriptions."""
+        pattern = re.compile(abbreviation)
+        count = sum(len(pattern.findall(description)) for description in tag_descriptions)
+        return count
+
+    def _replace_abbreviations(self, tag_descriptions, abbreviations):
+        """Replace the abbreviations according to the key-pair value provided."""
+        replaced_descriptions = []
+        for description in tag_descriptions:
+            for abbreviation, replacement in abbreviations.items():
+                description = re.sub(abbreviation, replacement, description)
+
+            replaced_descriptions.append(description)
+        return replaced_descriptions
+
+    def _cleanup_spaces(self, tag_descriptions):
+        # Replace all whitespace with a single space
+        replaced_descriptions = []
+        for description in tag_descriptions:
+            description_clean = re.sub(r'\s+', ' ', description)
+            replaced_descriptions.append(description_clean)
+        return replaced_descriptions
+
+    # remove all dots
+    def _cleanup_dots(self, tag_descriptions):
+        replaced_descriptions = []
+        for description in tag_descriptions:
+            description_clean = re.sub(r'\.', '', description)
+            replaced_descriptions.append(description_clean)
+        return replaced_descriptions
+
+
+    def run(self):
+        df = self.df
+
+        # %%
+        # Replace abbreviations
+        print("running substitution for descriptions")
+        # normalize to uppercase
+        # strip leading and trailing whitespace
+        df['tag_description'] = df['tag_description'].str.strip()
+        df['tag_description'] = df['tag_description'].str.upper()
+        # Replace whitespace-only entries with "NOVALUE"
+        # note that "N/A" can be read as nan
+        # replace whitespace only values as NOVALUE
+        df['tag_description']= df['tag_description'].fillna("NOVALUE")
+        df['tag_description'] = df['tag_description'].replace(r'^\s*$', 'NOVALUE', regex=True)
+
+        # perform actual substitution
+        tag_descriptions = df['tag_description']
+        replaced_descriptions = self._replace_abbreviations(tag_descriptions, desc_replacement_dict)
+        replaced_descriptions = self._cleanup_spaces(replaced_descriptions)
+        replaced_descriptions = self._cleanup_dots(replaced_descriptions)
+        df["tag_description"] = replaced_descriptions
+        # print("Descriptions after replacement:", replaced_descriptions)
+
+        # %%
+        print("running substitutions for units")
+        df['unit'] = df['unit'].fillna("NOVALUE")
+        df['unit'] = df['unit'].replace(r'^\s*$', 'NOVALUE', regex=True)
+        unit_list = df['unit']
+        new_unit = self._replace_abbreviations(unit_list, unit_replacement_dict)
+        new_unit = self._cleanup_spaces(new_unit)
+        df['unit'] = new_unit
+
+        return df
+

end_to_end/replacement_dict.py

@@ -0,0 +1,291 @@
+# substitution mapping for descriptions
+# Abbreviations and their replacements
+desc_replacement_dict = {
+    r'\bLIST\b': 'LIST',
+    # exhaust gas
+    r'\bE\. GAS\b': 'EXHAUST GAS',
+    r'\bEXH\.\b': 'EXHAUST',
+    r'\bEXH\b': 'EXHAUST',
+    r'\bEXHAUST\.\b': 'EXHAUST',
+    r'\bEXHAUST\b': 'EXHAUST',
+    r'\bBLR\.EXH\.\b': 'BOILER EXHAUST',
+    # temperature
+    r'\bTEMP\.\b': 'TEMPERATURE',
+    r'\bTEMP\b': 'TEMPERATURE',
+    r'\bTEMPERATURE\.\b': 'TEMPERATURE',
+    r'\bTEMPERATURE\b': 'TEMPERATURE',
+    # cylinder
+    r'\bCYL(\d+)\b': r'CYLINDER\1',
+    r'\bCYL\.(\d+)\b': r'CYLINDER\1',
+    r'\bCYL(?=\d|\W|$)\b': 'CYLINDER',
+    r'\bCYL\.\b': 'CYLINDER',
+    r'\bCYL\b': 'CYLINDER',
+    # cooling
+    r'\bCOOL\.\b': 'COOLING',
+    r'\bCOOLING\b': 'COOLING',
+    r'\bCOOLER\b': 'COOLER',
+    r'\bCW\b': 'COOLING WATER',
+    r'\bC\.W\b': 'COOLING WATER',
+    r'\bJ\.C\.F\.W\b': 'JACKET COOLING FEED WATER',
+    r'\bJ\.C F\.W\b': 'JACKET COOLING FEED WATER',
+    r'\bJACKET C\.F\.W\b': 'JACKET COOLING FEED WATER',
+    r'\bCOOL\. F\.W\b': 'COOLING FEED WATER',
+    r'\bC\.F\.W\b': 'COOLING FEED WATER',
+    # sea water
+    r'\bC\.S\.W\b': 'COOLING SEA WATER',
+    r'\bCSW\b': 'COOLING SEA WATER',
+    r'\bC.S.W\b': 'COOLING SEA WATER',
+    # water
+    r'\bFEED W\.\b': 'FEED WATER',
+    r'\bFEED W\b': 'FEED WATER',
+    r'\bF\.W\b': 'FEED WATER',
+    r'\bF\.W\.\b': 'FEED WATER',
+    r'\bFW\b': 'FEED WATER',
+    # r'\bWATER\b': 'WATER',
+    r'\bSCAV\.\b': 'SCAVENGE',
+    r'\bSCAV\b': 'SCAVENGE',
+    r'\bINL\.\b': 'INLET',
+    r'\bINLET\b': 'INLET',
+    r'\bOUT\.\b': 'OUTLET',
+    r'\bOUTL\.\b': 'OUTLET',
+    r'\bOUTLET\b': 'OUTLET',
+    # tank
+    r'\bSTOR\.TK\b': 'STORAGE TANK',
+    r'\bSTOR\. TK\b': 'STORAGE TANK',
+    r'\bSERV\. TK\b': 'SERVICE TANK',
+    r'\bSETT\. TK\b': 'SETTLING TANK',
+    r'\bBK\b': 'BUNKER',
+    r'\bTK\b': 'TANK',
+    # PRESSURE
+    r'\bPRESS\b': 'PRESSURE',
+    r'\bPRESS\.\b': 'PRESSURE',
+    r'\bPRESSURE\b': 'PRESSURE',
+    r'PRS\b': 'PRESSURE',  # this is a special replacement - it is safe to replace PRS w/o checks
+    # ENGINE
+    r'\bENG\.\b': 'ENGINE',
+    r'\bENG\b': 'ENGINE',
+    r'\bENGINE\b': 'ENGINE',
+    r'\bENGINE SPEED\b': 'ENGINE SPEED',
+    r'\bENGINE RUNNING\b': 'ENGINE RUNNING',
+    r'\bENGINE RPM PICKUP\b': 'ENGINE RPM PICKUP',
+    r'\bENGINE ROOM\b': 'ENGINE ROOM',
+    r'\bE/R\b': 'ENGINE ROOM',
+    # MAIN ENGINE
+    r'\bM/E NO.(\d+)\b': r'NO\1 MAIN_ENGINE',
+    r'\bM/E NO(\d+)\b': r'NO\1 MAIN_ENGINE',
+    r'\bM/E  NO.(\d+)\b': r'NO\1 MAIN_ENGINE',
+    r'\bME NO.(\d+)\b': r'NO\1 MAIN_ENGINE',
+    r'\bM/E\b': 'MAIN_ENGINE',
+    r'\bM/E(.)\b': r'MAIN_ENGINE \1', # M/E(S/P)
+    r'\bME(.)\b': r'MAIN_ENGINE \1', # ME(S/P)
+    r'\bM_E\b': 'MAIN_ENGINE',
+    r'\bME(?=\d|\W|$)\b': 'MAIN_ENGINE',
+    r'\bMAIN ENGINE\b': 'MAIN_ENGINE',
+    # ENGINE variants
+    r'\bM_E_RPM\b': 'MAIN ENGINE RPM',
+    r'\bM/E_M\.G\.O\.\b': 'MAIN ENGINE MARINE GAS OIL',
+    r'\bM/E_H\.F\.O\.\b': 'MAIN ENGINE HEAVY FUEL OIL',
+    # GENERATOR ENGINE
+    r'\bGEN(\d+)\b': r'NO\1 GENERATOR_ENGINE',
+    r'\bGE(\d+)\b': r'NO\1 GENERATOR_ENGINE',
+    # ensure that we substitute only for terms where following GE is num or special
+    r'\bGE(?=\d|\W|$)\b': 'GENERATOR_ENGINE',
+    r'\bG/E(\d+)\b': r'NO\1 GENERATOR_ENGINE',
+    r'\bG/E\b': r'GENERATOR_ENGINE',
+    r'\bG_E(\d+)\b': r'NO\1 GENERATOR_ENGINE',
+    r'\bG_E\b': 'GENERATOR_ENGINE',
+    r'\bGENERATOR ENGINE\b': 'GENERATOR_ENGINE',
+    r'\bG/E_M\.G\.O\b': 'GENERATOR_ENGINE MARINE GAS OIL',
+    # DG
+    r'\bDG(\d+)\b': r'NO\1 GENERATOR_ENGINE',
+    r'\bDG\b': 'GENERATOR_ENGINE',
+    r'\bD/G\b': 'GENERATOR_ENGINE',
+    r'\bDG(\d+)\((.)\)\b': r'NO\1\2 GENERATOR_ENGINE', # handle DG2(A)
+    r'\bDG(\d+[A-Za-z])\b': r'NO\1 GENERATOR_ENGINE', # handle DG2A
+    # DG variants
+    r'\bDG_CURRENT\b': 'GENERATOR_ENGINE CURRENT',
+    r'\bDG_LOAD\b': 'GENERATOR_ENGINE LOAD',
+    r'\bDG_FREQUENCY\b': 'GENERATOR_ENGINE FREQUENCY',
+    r'\bDG_VOLTAGE\b': 'GENERATOR_ENGINE VOLTAGE',
+    r'\bDG_CLOSED\b': 'GENERATOR_ENGINE CLOSED',
+    r'\bD/G_CURRENT\b': 'GENERATOR_ENGINE CURRENT',
+    r'\bD/G_LOAD\b': 'GENERATOR_ENGINE LOAD',
+    r'\bD/G_FREQUENCY\b': 'GENERATOR_ENGINE FREQUENCY',
+    r'\bD/G_VOLTAGE\b': 'GENERATOR_ENGINE VOLTAGE',
+    r'\bD/G_CLOSED\b': 'GENERATOR_ENGINE CLOSED',
+    # MGE
+    r'\b(\d+)MGE\b': r'NO\1 MAIN_GENERATOR_ENGINE',
+    # generator engine and mgo
+    r'\bG/E_M\.G\.O\.\b': r'GENERATOR_ENGINE MARINE GAS OIL',
+    r'\bG/E_H\.F\.O\.\b': r'GENERATOR_ENGINE HEAVY FUEL OIL',
+    # ultra low sulfur fuel oil
+    r'\bU\.L\.S\.F\.O\b': 'ULTRA LOW SULFUR FUEL OIL',
+    r'\bULSFO\b': 'ULTRA LOW SULFUR FUEL OIL',
+    # marine gas oil
+    r'\bM\.G\.O\b': 'MARINE GAS OIL',
+    r'\bMGO\b': 'MARINE GAS OIL',
+    r'\bMDO\b': 'MARINE DIESEL OIL',
+    # light fuel oil
+    r'\bL\.F\.O\b': 'LIGHT FUEL OIL',
+    r'\bLFO\b': 'LIGHT FUEL OIL',
+    # heavy fuel oil
+    r'\bHFO\b': 'HEAVY FUEL OIL',
+    r'\bH\.F\.O\b': 'HEAVY FUEL OIL',
+    # piston cooling oil
+    r'\bPCO\b': 'PISTON COOLING OIL',
+    r'\bP\.C\.O\.\b': 'PISTON COOLING OIL',
+    r'\bP\.C\.O\b': 'PISTON COOLING OIL',
+    r'PISTION C.O': 'PISTON COOLING OIL',
+    # diesel oil
+    r'\bD.O\b': 'DIESEL OIL',
+    # for remaining fuel oil that couldn't be substituted
+    r'\bF\.O\b': 'FUEL OIL',
+    r'\bFO\b': 'FUEL OIL',
+    # lubricant
+    r'\bLUB\.\b': 'LUBRICANT',
+    r'\bLUBE\b': 'LUBRICANT',
+    r'\bLUBR\.\b': 'LUBRICANT',
+    r'\bLUBRICATING\.\b': 'LUBRICANT',
+    r'\bLUBRICATION\.\b': 'LUBRICANT',
+    # lubricating oil
+    r'\bL\.O\b': 'LUBRICATING OIL',
+    r'\bLO\b': 'LUBRICATING OIL',
+    # lubricating oil pressure
+    r'\bLO_PRESS\b': 'LUBRICATING OIL PRESSURE',
+    r'\bLO_PRESSURE\b': 'LUBRICATING OIL PRESSURE',
+    # temperature
+    r'\bL\.T\b': 'LOW TEMPERATURE',
+    r'\bLT\b': 'LOW TEMPERATURE',
+    r'\bH\.T\b': 'HIGH TEMPERATURE',
+    r'\bHT\b': 'HIGH TEMPERATURE',
+    # BOILER
+    # auxiliary boiler
+    # replace these first before replacing AUXILIARY only
+    r'\bAUX\.BOILER\b': 'AUXILIARY BOILER',
+    r'\bAUX\. BOILER\b': 'AUXILIARY BOILER',
+    r'\bAUX BLR\b': 'AUXILIARY BOILER',
+    r'\bAUX\.\b': 'AUXILIARY',
+    r'\bAUX\b': 'AUXILIARY',
+    # composite boiler
+    r'\bCOMP\. BOILER\b': 'COMPOSITE BOILER',
+    r'\bCOMP\.BOILER\b': 'COMPOSITE BOILER',
+    r'\bCOMP BOILER\b': 'COMPOSITE BOILER',
+    r'\bCOMP\b': 'COMPOSITE',
+    r'\bCMPS\b': 'COMPOSITE',
+    # any other boiler
+    r'\bBLR\.\b': 'BOILER',
+    r'\bBLR\b': 'BOILER',
+    r'\bBOILER W.CIRC.P/P\b': 'BOILER WATER CIRC P/P',
+    # windind
+    r'\bWIND\.\b': 'WINDING',
+    r'\bWINDING\b': 'WINDING',
+    # VOLTAGE/FREQ/CURRENT
+    r'\bVLOT\.': 'VOLTAGE', # correct spelling
+    r'\bVOLT\.': 'VOLTAGE',
+    r'\bVOLTAGE\b': 'VOLTAGE',
+    r'\bFREQ\.': 'FREQUENCY',
+    r'\bFREQUENCY\b': 'FREQUENCY',
+    r'\bCURR\.': 'CURRENT',
+    r'\bCURRENT\b': 'CURRENT',
+    # TURBOCHARGER
+    r'\bTCA\b': 'TURBOCHARGER',
+    r'\bTCB\b': 'TURBOCHARGER',
+    r'\bT/C\b': 'TURBOCHARGER',
+    r'\bT_C\b': 'TURBOCHARGER',
+    r'\bT/C_RPM\b': 'TURBOCHARGER RPM',
+    r'\bTC(\d+)\b': r'TURBOCHARGER\1',
+    r'\bT/C(\d+)\b': r'TURBOCHARGER\1',
+    r'\bTC(?=\d|\W|$)\b': 'TURBOCHARGER',
+    r'\bTURBOCHAGER\b': 'TURBOCHARGER',
+    r'\bTURBOCHARGER\b': 'TURBOCHARGER',
+    r'\bTURBOCHG\b': 'TURBOCHARGER',
+    # misc spelling errors
+    r'\bOPERATOIN\b': 'OPERATION',
+    # wrongly attached terms
+    r'\bBOILERMGO\b': 'BOILER MGO',
+    # additional standardizing replacement
+    # replace # followed by a number with NO
+    r'#(?=\d)\b': 'NO',
+    r'\bNO\.(?=\d)\b': 'NO',
+    r'\bNO\.\.(?=\d)\b': 'NO',
+    # others:
+    # generator
+    r'\bGEN\.\b': 'GENERATOR',
+    # others
+    r'\bGEN\.WIND\.TEMP\b': 'GENERATOR WINDING TEMPERATURE',
+    r'\bFLTR\b': 'FILTER',
+    r'\bCLR\b': 'CLEAR',
+}
+
+# substitution mapping for units
+# Abbreviations and their replacements
+unit_replacement_dict = {
+    r'\b%\b': 'PERCENT',
+    r'\b-\b': '',
+    r'\b-  \b': '',
+    # ensure no character after A
+    r'\bA(?!\w|/)': 'CURRENT',
+    r'\bAmp(?!\w|/)': 'CURRENT',
+    r'\bHz\b': 'HERTZ',
+    r'\bKG/CM2\b': 'PRESSURE',
+    r'\bKG/H\b': 'KILOGRAM PER HOUR',
+    r'\bKNm\b': 'RPM',
+    r'\bKW\b': 'POWER',
+    r'\bKg(?!\w|/)': 'MASS',
+    r'\bKw\b': 'POWER',
+    r'\bL(?!\w|/)': 'VOLUME',
+    r'\bMT/h\b': 'METRIC TONNES PER HOUR',
+    r'\bMpa\b': 'PRESSURE',
+    r'\bPF\b': 'POWER FACTOR',
+    r'\bRPM\b': 'RPM',
+    r'\bV(?!\w|/)': 'VOLTAGE',
+    r'\bbar(?!\w|/)': 'PRESSURE',
+    r'\bbarA\b': 'SCAVENGE PRESSURE',
+    r'\bcST\b': 'VISCOSITY',
+    r'\bcSt\b': 'VISCOSITY',
+    r'\bcst\b': 'VISCOSITY',
+    r'\bdeg(?!\w|/|\.)': 'DEGREE',
+    r'\bdeg.C\b': 'TEMPERATURE',
+    r'\bdegC\b': 'TEMPERATURE',
+    r'\bdegree\b': 'DEGREE',
+    r'\bdegreeC\b': 'TEMPERATURE',
+    r'\bhPa\b': 'PRESSURE',
+    r'\bhours\b': 'HOURS',
+    r'\bkN\b': 'THRUST',
+    r'\bkNm\b': 'TORQUE',
+    r'\bkW\b': 'POWER',
+    # ensure that kg is not followed by anything
+    r'\bkg(?!\w|/)': 'FLOW', # somehow in the data its flow
+    r'\bkg/P\b': 'MASS FLOW',
+    r'\bkg/cm2\b': 'PRESSURE',
+    r'\bkg/cm²\b': 'PRESSURE',
+    r'\bkg/h\b': 'MASS FLOW',
+    r'\bkg/hr\b': 'MASS FLOW',
+    r'\bkg/pulse\b': '',
+    r'\bkgf/cm2\b': 'PRESSURE',
+    r'\bkgf/cm²\b': 'PRESSURE',
+    r'\bkgf/㎠\b': 'PRESSURE',
+    r'\bknots\b': 'SPEED',
+    r'\bkw\b': 'POWER',
+    r'\bl/Hr\b': 'VOLUME FLOW',
+    r'\bl/h\b': 'VOLUME FLOW',
+    r'\bl_Hr\b': 'VOLUME FLOW',
+    r'\bl_hr\b': 'VOLUME FLOW',
+    r'\bM\b': 'DRAFT', # for wind draft
+    r'm': 'm', # wind draft and trim - not useful
+    r'\bm/s\b': 'SPEED',
+    r'\bm3\b': 'VOLUME',
+    r'\bmH2O\b': 'DRAFT',
+    r'\bmWC\b': 'DRAFT',
+    r'\bmbar\b': 'PRESSURE',
+    r'\bmg\b': 'ACCELERATION',
+    r'\bmin-¹\b': '', # data too varied
+    r'\bmm\b': '', # data too varied
+    r'\bmmH2O\b': 'WATER DRUM LEVEL',
+    r'\brev\b': 'RPM',
+    r'\brpm\b': 'RPM',
+    r'\bx1000min-¹\b': '',
+    r'\b°C\b': 'TEMPERATURE',
+    r'\bºC\b': 'TEMPERATURE',
+    r'\b℃\b': 'TEMPERATURE'
+}

end_to_end/run.py

@@ -0,0 +1,64 @@
+# %%
+import pandas as pd
+import os
+import glob
+from mapper import Mapper
+from preprocess import Abbreviator
+from deduplication import run_deduplication
+
+# global config
+BATCH_SIZE = 1024
+SHIPS_LIST = [1000,1001,1003]
+
+# %%
+# START: we import the raw data csv and extract only a few ships from it to simulate incoming json
+data_path = 'raw_data.csv'
+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
+df = df[:num_rows]
+print(len(df))
+
+# pre-process data
+abbreviator = Abbreviator(df)
+df = abbreviator.run()
+
+# %%
+##########################################
+# run mapping
+# checkpoint
+# Use glob to find matching paths
+checkpoint_path = 'models/mapping_model'
+mapper = Mapper(checkpoint_path)
+mapper.prepare_dataloader(df, batch_size=BATCH_SIZE, max_length=128)
+thing_prediction_list, property_prediction_list = mapper.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)
+
+
+# %%
+####################################
+# run de_duplication with thresholding
+data_path = "train_all.csv"
+train_df = pd.read_csv(data_path, skipinitialspace=True)
+train_df['mapping'] = train_df['thing'] + " " + train_df['property']
+
+df = run_deduplication(
+    test_df=df,
+    train_df=train_df,
+    batch_size=BATCH_SIZE,
+    threshold=0.9,
+    diagnostic=True)
+
+# %%

post_process/de_duplication/utils.py

@@ -3,9 +3,7 @@ from transformers import (
     AutoTokenizer,
     AutoModelForSequenceClassification,
     AutoModelForSeq2SeqLM,
-    DataCollatorWithPadding,
 )
-import torch.nn.functional as F
 
 
 
@@ -24,7 +22,7 @@ class BertEmbedder:
         self.model = model.eval()
 
 
-    def make_embedding(self, batch_size=64):
+    def make_embedding(self, batch_size=128):
         all_embeddings = self.embeddings
         input_texts = self.inputs