handler.py

from typing import Dict, List, Any
from scipy.special import softmax
import numpy as np
import weakref

from utils import clean_str, clean_str_nopunct
import torch
from utils import MultiHeadModel, BertInputBuilder, get_num_words

import transformers
from transformers import BertTokenizer, BertForSequenceClassification
import psutil
from transformers.utils import logging

transformers.logging.set_verbosity_debug()

UPTAKE_MODEL = 'ddemszky/uptake-model'
REASONING_MODEL = 'ddemszky/student-reasoning'
QUESTION_MODEL = 'ddemszky/question-detection'


class Utterance:
    def __init__(self, speaker, text, uid=None,
                 transcript=None, starttime=None, endtime=None, **kwargs):
        self.speaker = speaker
        self.text = text
        self.uid = uid
        self.starttime = starttime
        self.endtime = endtime
        self.transcript = weakref.ref(transcript) if transcript else None
        self.props = kwargs

        self.uptake = None
        self.reasoning = None
        self.question = None

    def get_clean_text(self, remove_punct=False):
        if remove_punct:
            return clean_str_nopunct(self.text)
        return clean_str(self.text)

    def get_num_words(self):
        return get_num_words(self.text)

    def to_dict(self):
        return {
            'speaker': self.speaker,
            'text': self.text,
            'uid': self.uid,
            'starttime': self.starttime,
            'endtime': self.endtime,
            'uptake': self.uptake,
            'reasoning': self.reasoning,
            'question':  self.question,
            **self.props
        }

    def __repr__(self):
        return f"Utterance(speaker='{self.speaker}'," \
               f"text='{self.text}', uid={self.uid}," \
               f"starttime={self.starttime}, endtime={self.endtime}, props={self.props})"


class Transcript:
    def __init__(self, **kwargs):
        self.utterances = []
        self.params = kwargs

    def add_utterance(self, utterance):
        utterance.transcript = weakref.ref(self)
        self.utterances.append(utterance)

    def get_idx(self, idx):
        if idx >= len(self.utterances):
            return None
        return self.utterances[idx]

    def get_uid(self, uid):
        for utt in self.utterances:
            if utt.uid == uid:
                return utt
        return None

    def length(self):
        return len(self.utterances)

    def to_dict(self):
        return {
            'utterances': [utterance.to_dict() for utterance in self.utterances],
            **self.params
        }

    def __repr__(self):
        return f"Transcript(utterances={self.utterances}, custom_params={self.params})"


class QuestionModel:
    def __init__(self, device, tokenizer, input_builder, max_length=300, path=QUESTION_MODEL):
        print("Loading models...")
        self.device = device
        self.tokenizer = tokenizer
        self.input_builder = input_builder
        self.max_length = max_length
        self.model = MultiHeadModel.from_pretrained(
            path, head2size={"is_question": 2})
        self.model.to(self.device)

    def run_inference(self, transcript):
        self.model.eval()
        with torch.no_grad():
            for i, utt in enumerate(transcript.utterances):
                if "?" in utt.text:
                    utt.question = 1
                else:
                    text = utt.get_clean_text(remove_punct=True)
                    instance = self.input_builder.build_inputs([], text,
                                                               max_length=self.max_length,
                                                               input_str=True)
                    output = self.get_prediction(instance)
                    print(output)
                    utt.question = np.argmax(
                        output["is_question_logits"][0].tolist())

    def get_prediction(self, instance):
        instance["attention_mask"] = [[1] * len(instance["input_ids"])]
        for key in ["input_ids", "token_type_ids", "attention_mask"]:
            instance[key] = torch.tensor(
                instance[key]).unsqueeze(0)  # Batch size = 1
            instance[key].to(self.device)

        output = self.model(input_ids=instance["input_ids"],
                            attention_mask=instance["attention_mask"],
                            token_type_ids=instance["token_type_ids"],
                            return_pooler_output=False)
        return output


class ReasoningModel:
    def __init__(self, device, tokenizer, input_builder, max_length=128, path=REASONING_MODEL):
        print("Loading models...")
        self.device = device
        self.tokenizer = tokenizer
        self.input_builder = input_builder
        self.max_length = max_length
        self.model = BertForSequenceClassification.from_pretrained(path)
        self.model.to(self.device)

    def run_inference(self, transcript, min_num_words=8):
        self.model.eval()
        with torch.no_grad():
            for i, utt in enumerate(transcript.utterances):
                if utt.get_num_words() >= min_num_words:
                    instance = self.input_builder.build_inputs([], utt.text,
                                                               max_length=self.max_length,
                                                               input_str=True)
                    output = self.get_prediction(instance)
                    utt.reasoning = np.argmax(output["logits"][0].tolist())

    def get_prediction(self, instance):
        instance["attention_mask"] = [[1] * len(instance["input_ids"])]
        for key in ["input_ids", "token_type_ids", "attention_mask"]:
            instance[key] = torch.tensor(
                instance[key]).unsqueeze(0)  # Batch size = 1
            instance[key].to(self.device)

        output = self.model(input_ids=instance["input_ids"],
                            attention_mask=instance["attention_mask"],
                            token_type_ids=instance["token_type_ids"])
        return output


class UptakeModel:
    def __init__(self, device, tokenizer, input_builder, max_length=120, path=UPTAKE_MODEL):
        print("Loading models...")
        self.device = device
        self.tokenizer = tokenizer
        self.input_builder = input_builder
        self.max_length = max_length
        self.model = MultiHeadModel.from_pretrained(path, head2size={"nsp": 2})
        self.model.to(self.device)

    def run_inference(self, transcript, min_prev_words, uptake_speaker=None):
        self.model.eval()
        prev_num_words = 0
        prev_utt = None
        with torch.no_grad():
            for i, utt in enumerate(transcript.utterances):
                if ((uptake_speaker is None) or (utt.speaker == uptake_speaker)) and (prev_num_words >= min_prev_words):
                    textA = prev_utt.get_clean_text(remove_punct=False)
                    textB = utt.get_clean_text(remove_punct=False)
                    instance = self.input_builder.build_inputs([textA], textB,
                                                               max_length=self.max_length,
                                                               input_str=True)
                    output = self.get_prediction(instance)

                    utt.uptake = int(
                        softmax(output["nsp_logits"][0].tolist())[1] > .8)
                prev_num_words = utt.get_num_words()
                prev_utt = utt

    def get_prediction(self, instance):
        instance["attention_mask"] = [[1] * len(instance["input_ids"])]
        for key in ["input_ids", "token_type_ids", "attention_mask"]:
            instance[key] = torch.tensor(
                instance[key]).unsqueeze(0)  # Batch size = 1
            instance[key].to(self.device)

        output = self.model(input_ids=instance["input_ids"],
                            attention_mask=instance["attention_mask"],
                            token_type_ids=instance["token_type_ids"],
                            return_pooler_output=False)
        return output


class EndpointHandler():
    def __init__(self, path="."):
        print("Loading models...")
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
        self.input_builder = BertInputBuilder(tokenizer=self.tokenizer)

    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        """
       data args:
            inputs (:obj: `list`):
            List of dicts, where each dict represents an utterance; each utterance object must have a `speaker`,
            `text` and `uid`and can include list of custom properties
            parameters (:obj: `dict`)
       Return:
            A :obj:`list` | `dict`: will be serialized and returned
        """
        # get inputs
        utterances = data.pop("inputs", data)
        params = data.pop("parameters", None)

        print("EXAMPLES")
        for utt in utterances[:3]:
            print("speaker %s: %s" % (utt["speaker"], utt["text"]))

        transcript = Transcript(filename=params.pop("filename", None))
        for utt in utterances:
            transcript.add_utterance(Utterance(**utt))

        print("Running inference on %d examples..." % transcript.length())
        cpu_percent = psutil.cpu_percent()
        logging.set_verbosity_info()
        logger = logging.get_logger("transformers")
        logger.info(f"CPU Usage before models loaded: {cpu_percent}%")
        mem_info = psutil.virtual_memory()
        used_mem = mem_info.used / (1024 ** 3)  # Convert to gigabytes
        total_mem = mem_info.total / (1024 ** 3)  # Convert to gigabytes
        logger.info(f"Used Memory before models loaded: {used_mem:.2f} GB, Total RAM: {total_mem:.2f} GB")
        # Uptake
        uptake_model = UptakeModel(
            self.device, self.tokenizer, self.input_builder)
        uptake_model.run_inference(transcript, min_prev_words=params['uptake_min_num_words'],
                                   uptake_speaker=params.pop("uptake_speaker", None))
        cpu_percent = psutil.cpu_percent()
        mem_info = psutil.virtual_memory()
        used_mem = mem_info.used / (1024 ** 3)  # Convert to gigabytes
        total_mem = mem_info.total / (1024 ** 3)  # Convert to gigabytes
        logger.info(f"Used Memory after model 1 loaded: {used_mem:.2f} GB, Total Mem: {total_mem:.2f} GB")
        logger.info(f"CPU Usage after model 1 loaded: {cpu_percent}%")
        del uptake_model
        cpu_percent = psutil.cpu_percent()
        mem_info = psutil.virtual_memory()
        used_mem = mem_info.used / (1024 ** 3)  # Convert to gigabytes
        total_mem = mem_info.total / (1024 ** 3)  # Convert to gigabytes
        logger.info(f"Used Memory after model 1 deleted: {used_mem:.2f} GB, Total Mem: {total_mem:.2f} GB")
        logger.info(f"CPU Usage after model 1 deleted: {cpu_percent}%")
        # Reasoning
        reasoning_model = ReasoningModel(
            self.device, self.tokenizer, self.input_builder)
        reasoning_model.run_inference(transcript)
        cpu_percent = psutil.cpu_percent()
        mem_info = psutil.virtual_memory()
        used_mem = mem_info.used / (1024 ** 3)  # Convert to gigabytes
        total_mem = mem_info.total / (1024 ** 3)  # Convert to gigabytes
        logger.info(f"Used Memory after model 2 loaded: {used_mem:.2f} GB, Total Mem: {total_mem:.2f} GB")
        logger.info(f"CPU Usage after model 2 loaded: {cpu_percent}%")
        # print(f"CPU Usage after model 2 loaded: {cpu_percent}%")
        del reasoning_model
        cpu_percent = psutil.cpu_percent()
        mem_info = psutil.virtual_memory()
        used_mem = mem_info.used / (1024 ** 3)  # Convert to gigabytes
        total_mem = mem_info.total / (1024 ** 3)  # Convert to gigabytes
        logger.info(f"Used Memory after model 2 deleted: {used_mem:.2f} GB, Total Mem: {total_mem:.2f} GB")
        logger.info(f"CPU Usage after model 2 deleted: {cpu_percent}%")
        # print(f"CPU Usage after model 2 deleted: {cpu_percent}%")
        # Question
        question_model = QuestionModel(
            self.device, self.tokenizer, self.input_builder)
        question_model.run_inference(transcript)
        cpu_percent = psutil.cpu_percent()
        logger.info(f"CPU Usage after model 3 loaded: {cpu_percent}%")
        mem_info = psutil.virtual_memory()
        used_mem = mem_info.used / (1024 ** 3)  # Convert to gigabytes
        total_mem = mem_info.total / (1024 ** 3)  # Convert to gigabytes
        logger.info(f"Used Memory after model 3 loaded: {used_mem:.2f} GB, Total Mem: {total_mem:.2f} GB")
        # print(f"CPU Usage after model 3 loaded: {cpu_percent}%")
        del question_model
        cpu_percent = psutil.cpu_percent()
        logger.info(f"CPU Usage after model 3 deleted: {cpu_percent}%")
        mem_info = psutil.virtual_memory()
        used_mem = mem_info.used / (1024 ** 3)  # Convert to gigabytes
        total_mem = mem_info.total / (1024 ** 3)  # Convert to gigabytes
        logger.info(f"Used Memory after model 3 deleted: {used_mem:.2f} GB, Total Mem: {total_mem:.2f} GB")
        # print(f"CPU Usage after model 3 deleted: {cpu_percent}%")
        return transcript.to_dict()