st_utils.py

from __future__ import absolute_import
import streamlit as st
import torch
import os
import sys
import pickle
import torch
import json
import random
import logging
import argparse
import numpy as np
from io import open
from itertools import cycle
import torch.nn as nn
from model import Seq2Seq
from tqdm import tqdm, trange
import regex as re
from torch.utils.data import (
    DataLoader,
    Dataset,
    SequentialSampler,
    RandomSampler,
    TensorDataset,
)
from torch.utils.data.distributed import DistributedSampler
from transformers import (
    WEIGHTS_NAME,
    AdamW,
    get_linear_schedule_with_warmup,
    RobertaConfig,
    RobertaModel,
    RobertaTokenizer,
)
from huggingface_hub import hf_hub_download
import io

# def list_files(startpath, prev_level=0):
#     # list files recursively
#     for root, dirs, files in os.walk(startpath):
#         level = root.replace(startpath, "").count(os.sep) + prev_level
#         indent = " " * 4 * (level)

#         print("{}{}/".format(indent, os.path.basename(root)))
#         # st.write("{}{}/".format(indent, os.path.basename(root)))

#         subindent = " " * 4 * (level + 1)
#         for f in files:
#             print("{}{}".format(subindent, f))
#             # st.write("{}{}".format(subindent, f))

#         for d in dirs:
#             list_files(d, level + 1)


class CONFIG:
    max_source_length = 256
    max_target_length = 128
    beam_size = 3
    local_rank = -1
    no_cuda = False

    do_train = True
    do_eval = True
    do_test = True
    train_batch_size = 12
    eval_batch_size = 32

    model_type = "roberta"
    model_name_or_path = "microsoft/codebert-base"
    output_dir = "/content/drive/MyDrive/CodeSummarization"
    load_model_path = None
    train_filename = "dataset/python/train.jsonl"
    dev_filename = "dataset/python/valid.jsonl"
    test_filename = "dataset/python/test.jsonl"
    config_name = ""
    tokenizer_name = ""
    cache_dir = "cache"

    save_every = 5000

    gradient_accumulation_steps = 1
    learning_rate = 5e-5
    weight_decay = 1e-4
    adam_epsilon = 1e-8
    max_grad_norm = 1.0
    num_train_epochs = 3.0
    max_steps = -1
    warmup_steps = 0
    train_steps = 100000
    eval_steps = 10000
    n_gpu = torch.cuda.device_count()


# download model with streamlit cache decorator
@st.cache_resource
def download_model():
    if not os.path.exists(r"models/pytorch_model.bin"):
        os.makedirs("./models", exist_ok=True)
        path = hf_hub_download(
            repo_id="tmnam20/codebert-code-summarization",
            filename="pytorch_model.bin",
            cache_dir="cache",
            local_dir=os.path.join(os.getcwd(), "models"),
            local_dir_use_symlinks=False,
            force_download=True,
        )


# load with streamlit cache decorator
# @st.cache(persist=False, show_spinner=True, allow_output_mutation=True)
@st.cache_resource
def load_tokenizer_and_model(pretrained_path):
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    # Config model
    config_class, model_class, tokenizer_class = (
        RobertaConfig,
        RobertaModel,
        RobertaTokenizer,
    )
    model_config = config_class.from_pretrained(
        CONFIG.config_name if CONFIG.config_name else CONFIG.model_name_or_path,
        cache_dir=CONFIG.cache_dir,
    )
    # model_config.save_pretrained("config")

    # load tokenizer
    tokenizer = tokenizer_class.from_pretrained(
        CONFIG.tokenizer_name if CONFIG.tokenizer_name else CONFIG.model_name_or_path,
        cache_dir=CONFIG.cache_dir,
        # do_lower_case=args.do_lower_case
    )

    # load encoder from pretrained RoBERTa
    encoder = model_class.from_pretrained(
        CONFIG.model_name_or_path, config=model_config, cache_dir=CONFIG.cache_dir
    )

    # build decoder
    decoder_layer = nn.TransformerDecoderLayer(
        d_model=model_config.hidden_size, nhead=model_config.num_attention_heads
    )
    decoder = nn.TransformerDecoder(decoder_layer, num_layers=6)

    # build seq2seq model from pretrained encoder and from-scratch decoder
    model = Seq2Seq(
        encoder=encoder,
        decoder=decoder,
        config=model_config,
        beam_size=CONFIG.beam_size,
        max_length=CONFIG.max_target_length,
        sos_id=tokenizer.cls_token_id,
        eos_id=tokenizer.sep_token_id,
    )

    try:
        state_dict = torch.load(
            os.path.join(os.getcwd(), "models", "pytorch_model.bin"),
            map_location=device,
        )
    except RuntimeError as e:
        print(e)
        try:
            state_dict = torch.load(
                os.path.join(os.getcwd(), "models", "pytorch_model.bin"),
                map_location="cpu",
            )
        except RuntimeError as e:
            print(e)
            state_dict = torch.load(
                os.path.join(os.getcwd(), "models", "pytorch_model_cpu.bin"),
                map_location="cpu",
            )

    del state_dict["encoder.embeddings.position_ids"]
    model.load_state_dict(state_dict)

    # model = model.to("cpu")
    # torch.save(model.state_dict(), os.path.join(os.getcwd(), "models", "pytorch_model_cpu.bin"))

    model = model.to(device)

    return tokenizer, model, device


@st.cache_data
def preprocessing(code_segment):
    # remove newlines
    code_segment = re.sub(r"\n", " ", code_segment)

    # remove docstring
    code_segment = re.sub(r'""".*?"""', "", code_segment, flags=re.DOTALL)

    # remove multiple spaces
    code_segment = re.sub(r"\s+", " ", code_segment)

    # remove comments
    code_segment = re.sub(r"#.*", "", code_segment)

    # remove html tags
    code_segment = re.sub(r"<.*?>", "", code_segment)

    # remove urls
    code_segment = re.sub(r"http\S+", "", code_segment)

    # split special chars into different tokens
    code_segment = re.sub(r"([^\w\s])", r" \1 ", code_segment)

    return code_segment.split()


def generate_docstring(model, tokenizer, device, code_segemnt, max_length=None):
    input_tokens = preprocessing(code_segemnt)
    encoded_input = tokenizer.encode_plus(
        input_tokens,
        max_length=CONFIG.max_source_length,
        pad_to_max_length=True,
        truncation=True,
        return_tensors="pt",
    )

    input_ids = encoded_input["input_ids"].to(device)
    input_mask = encoded_input["attention_mask"].to(device)

    if max_length is not None:
        model.max_length = max_length

    summary = model(input_ids, input_mask)

    # decode summary with tokenizer
    summaries = []
    for i in range(summary.shape[1]):
        summaries.append(tokenizer.decode(summary[0][i], skip_special_tokens=True))
    return summaries
    # return tokenizer.decode(summary[0][0], skip_special_tokens=True)