main_crossval_Nospatial.py

#!/usr/bin/python3

"""
    This script reads kernel_measurements.csv and final llvmir codes and creates pytorch geometric graphs.
    
    @author: mariannatzortzi
"""

from ast import arg
from cProfile import label
from distutils.log import info
import os
from statistics import mode
import sys

#GPU
os.environ['CUDA_VISIBLE_DEVICES'] ="0"

#os.environ['CUDA_VISIBLE_DEVICES'] = ""

#export CUDA_VISIBLE_DEVICES=""


import io
import os  
import glob
import time
import argparse
import datetime
import numpy as np
import pandas as pd
from pandas.io import json
from torch_geometric import data
from tqdm import tqdm
from pathlib import Path
import matplotlib.pyplot as plt
import seaborn as sns

import torch
import torch.nn as nn
import torch_geometric
from torch_geometric.data import Data
from torch_geometric.data import DataLoader


from pyg_crossval import create_pg_graphs, feature_scale
from train_crossval_no_wrapper_l1lossmean import train_one_epoch, evaluate_per_epoch, test_per_epoch
from GTMcrossval_nwrap_nospatial_newparametrized import GTModel


def maincall(trainpaths, devicekernels, args):
    train_pg_graphs_double = []
    for i in trainpaths:
        print('*** train graphs ***')
        print(f'i is {i}')
        # Create a list of pytorch_geometric Data objects containing the kernel measurements    
        train_pg_graphs_sublist, max_number_nodes, max_number_edges = create_pg_graphs(path=i, kernelsdf=devicekernels)
        train_pg_graphs_double.append(train_pg_graphs_sublist)
    train_pg_graphs = [item for sublist in train_pg_graphs_double for item in sublist]
    # print(f'train_pg_graphs is {train_pg_graphs}')
    print(f'length of train pytorch geometric graphs: {len(train_pg_graphs)}')


    # test graphs
    print('**** test graphs ***')
    test_pg_graphs, max_number_nodes, max_number_edges = create_pg_graphs(path=args.test_path, kernelsdf=devicekernels)
    set = "test"
    print(f'length of {set} pytorch geometric graphs: {len(test_pg_graphs)}')

    return train_pg_graphs, test_pg_graphs

if __name__ == '__main__':    

    #--------------------- Parser arguments --------------------------------------
    # Program parameters parsing
    parser = argparse.ArgumentParser(description='Reads two paths the train llvm ir files and measurements of opencl kernels and returns train pg graphs')
    parser.add_argument('-i', '--indent', type=int, default=None, help='If set to a non-negative integer, then json array elements and object members will be pretty-printed with that indent level. An indent level of 0 or negative, will only insert newlines. The default value (None) selects the most compact representation.')
    parser.add_argument('kfolder_path', type=str, default='/home/marianna/Documents/programl/compiled_kernels/final_dataset/kfolder', help='kfolder')
    parser.add_argument('test_path', type=str, default='/home/marianna/Documents/programl/compiled_kernels/final_dataset/ir/test_path', help='The test dataset')

    parser.add_argument('csv_measurements_path', type=str, default='/home/marianna/Documents/programl/compiled_kernels/final_dataset/kernel_measurements.csv', help='The input csv file containing measurements.')
    #parser.add_argument('--save', type=bool, default=False, action='store_true', help='The store_true option automatically creates a default value of False, Likewise, store_false will default to True when the command-line argument is not present.')
    parser.add_argument('--dataset_name', type=str, default='fromCummins', help='name of the dataset that is used for the experiments.')        
    parser.add_argument('experiments_path', type=str, default='/home/marianna/Documents/programl/compiled_kernels/final_dataset/experiments', help='The path containing the experiments.')
    parser.add_argument('--epochs', type=int, default=100, help='The number of epochs to train the model for')
    parser.add_argument('--batch_size', type=int, default=32, help='batch size')
    parser.add_argument('--pct_close', type=float, default=0.05, help='In regression problem, you have to define what a correct prediction is. You define correctness as being within a certain percentage of the true value. pct_close is the percentage.')
    #parser.add_argument('--num_layers', type=int, default=3, help='number of iterations over the neighborhood')
    #parser.add_argument('--norm', type=str, default=None, help='Type of normalization.')
    parser.add_argument('--hidden_dim', type=int, default=64, help='Dimmension of the hidden states.')
    parser.add_argument('--heads', type=int, default=8, help='Number of attention heads')
    parser.add_argument('--num_exps', type=int, default=1, help='Number of experiments')

    parser.add_argument('--lr', type=float, default=0.001, help='learning rate')
    parser.add_argument('--patience', type=int, default=20, help='number of epochs to wait before starting lr decay (early stoppping criterion).')

    parser.add_argument('--feature_scale', type=bool, default=True, help='whether or not do feature scaling')
    parser.add_argument('--scale_type', type=str, default='standardization', help='(or minmax)what type of feature scaling I use')
    parser.add_argument('--devicedataset', type=str, default='GPU', help='(GPU or CPU) the device with which the kernels are running')

    args = parser.parse_args() # args respresents the options we have

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print('available device in this HW is {device}'.format(device=device))
    
    # Load the csv containing the measurements of each opencl kernel for which I also have its corresponding llvm ir code
    kernels_df = pd.read_csv(io.open(args.csv_measurements_path)) #pd.read_csv('/home/marianna/Documents/programl/compiled_kernels/final_dataset/kernel_measurements.csv')
    kernels_filtered = kernels_df.filter(items=['Kernel', 'Device', 'InputBytes', 'OutputBytes', 'End-to-End Time', 'storeOps'])
    print(f'kernels_filtered shape {kernels_filtered.shape}')

    # feature scaling OutputBytes and End-to-End Time 
    y = torch.tensor([kernels_filtered['End-to-End Time'], kernels_filtered['OutputBytes']], dtype=torch.float).permute(1,0)
    ukplot = kernels_filtered[kernels_filtered['Device'] == 'CPU'] #or GPU
    print('unscaled for outputbytes', ukplot['OutputBytes'].describe())
    print('unscaled for exectime', ukplot['End-to-End Time'].describe())
    # sns.displot(data=ukplot, x="OutputBytes", kind="hist", kde=True, aspect=1.5)
    # plt.show()
    # sns.displot(data=ukplot, x="InputBytes", kind="hist", kde=True, aspect=1.5)
    # plt.show()
    # sns.displot(data=ukplot, x="End-to-End Time", kind="hist", kde=True, aspect=1.5)
    # plt.show()
    # sns.displot(data=ukplot, x="storeOps", kind="hist", kde=True, aspect=1.5)
    # plt.show()
    if args.feature_scale:            
        dataoutputs = kernels_filtered['OutputBytes'].values
        print(f'dataoutputs {type(dataoutputs)} {dataoutputs}')
        dataoutputs, meanb, stdb = feature_scale(dataoutputs, args.scale_type)
        kernels_filtered['OutputBytes'] = dataoutputs
        print('check if there is a nan value here:', kernels_filtered['OutputBytes'].isnull().any())
        print('max outputbytes', kernels_filtered['OutputBytes'].max())
        # sns.displot(data=kplot, x="OutputBytes", kind="hist", kde=True, aspect=1.5)
        # plt.show()

        datainputs = kernels_filtered['InputBytes'].values
        print(f'datainputs {type(datainputs)} {datainputs}')
        datainputs, meanb, stdb = feature_scale(datainputs, args.scale_type)
        kernels_filtered['InputBytes'] = datainputs
        print('check if there is a nan value here:', kernels_filtered['InputBytes'].isnull().any())
        print('max inputbytes', kernels_filtered['InputBytes'].max())
        # sns.displot(data=kplot, x="InputBytes", kind="hist", kde=True, aspect=1.5)
        # plt.show()

        dataexectime = kernels_filtered['End-to-End Time'].values
        dataexectime, meant, stdt = feature_scale(dataexectime, args.scale_type)
        kernels_filtered['End-to-End Time'] = dataexectime
        print('check if there is a nan value here:', kernels_filtered['End-to-End Time'].isnull().any())
        print('max execution time', kernels_filtered['End-to-End Time'].max())
        # sns.displot(data=kplot, x="End-to-End Time", kind="hist", kde=True, aspect=1.5)
        # plt.show()

        kplot = kernels_filtered[kernels_filtered['Device'] == 'CPU'] #or GPU
        print('scaled outputbytes', kplot['OutputBytes'].describe())
        print('scaled exectime', kplot['End-to-End Time'].describe())


    print('The shape of the filtered dataframe of the wanted kernels is{mshape}'.format(mshape=kernels_filtered.shape))
    print('The OutputBytes and End-to-End Time are:', kernels_filtered['OutputBytes'], 'and\n', kernels_filtered['End-to-End Time'])

    # Encode the device: GPU -> 1, CPU -> 0
    cleanup = {"Device": {"GPU": 1, "CPU": 0}}
    kernels_filtered = kernels_filtered.replace(cleanup)
    if args.devicedataset == 'GPU':
        devicekernels = kernels_filtered[kernels_filtered['Device'] == 1]
    elif args.devicedataset == 'CPU':
        devicekernels = kernels_filtered[kernels_filtered['Device'] == 0]

    ifold = 0
    all_list = []
    for filename in os.listdir(args.kfolder_path):
        trainpaths = []
        ifold += 1         
        for ff in os.listdir(args.kfolder_path):
            if filename == ff:
                print('****validation graphs*****')
                val_path = os.path.join(args.kfolder_path, ff)
                print(f'the fold used for validation this time is {filename}')
                validation_pg_graphs, max_number_nodes, max_number_edges = create_pg_graphs(path=val_path, kernelsdf=devicekernels)
                set = "validation"
                print(f'length of {set} pytorch geometric graphs: {len(validation_pg_graphs)}')
            else:
                trainpaths.append(os.path.join(args.kfolder_path, ff))
               
        train_pg_graphs, test_pg_graphs = maincall(trainpaths, devicekernels, args)

        for norm in ['None']:#'pairnorm', 'layernorm', 'None', 'neighbornorm']:#, 'batchnorm', 'None', 'layernorm']:#, 'pairnorm', 'lipschitznorm']:
            for num_layers in [1]:#, 2]:#, 3, 5]:
                accuracy_list = []
                for exp in range(args.num_exps):
                    print('-----------------------')
                    print(f'Experiment {exp} with normalization type {norm} and number of layers {num_layers}')
                    print('-----------------------')
                    
                    # model
                    model = GTModel(
                                    num_layers = num_layers,
                                    hidden_dim= args.hidden_dim,
                                    heads=args.heads,
                                    feat_dropout=0.0, 
                                    top_k_pool=5, 
                                    norm=norm,
                                    beta=True
                                    )#.to(device)
                    print(next(model.parameters()).device)
                    print(model)
                    # optimizer
                    optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=0.0005) # args.lr=0.01
                    if args.patience > 0:
                        scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', patience=args.patience, verbose=True, min_lr=0.00000001) #args.patience

                    # Train
                    whole_exp_states = []
                    start_time = time.time()
                    for epoch in tqdm(range(args.epochs)):
                        state, optimizer = train_one_epoch(epoch, model, optimizer, train_pg_graphs, batch_size=args.batch_size, pct_close=args.pct_close)#, device, batch_size=16)
                        whole_exp_states.append(state)

                        val_state = evaluate_per_epoch(epoch, model, validation_pg_graphs, batch_size=args.batch_size, pct_close=args.pct_close)
                        whole_exp_states.append(val_state)

                    
                        # early stopping here comparing with the val loss 

                    test_state = test_per_epoch(meant, stdt, meanb, stdb, epoch, model, test_pg_graphs, batch_size=args.batch_size, pct_close=args.pct_close)
                    whole_exp_states.append(test_state)
        
                    print('End of experiment {exp}.')

        dict_folds = {'ifold': ifold, 'whole_exp_states': whole_exp_states}
        all_list.append(dict_folds)

    # Save the experiments 
    experiments_folder = None
    if True:
        now = time.strftime("%Y%m%d_%H%M%S")
        experiments_folder = os.path.join(args.experiments_path, args.dataset_name, now)
        # Create a new directory at this given path.
        Path(experiments_folder).mkdir(parents=True, exist_ok=True)

    if True:#args.save
        if norm == None:
            norm = "None"
        # A common PyTorch convention is to save models using either a .pt or .pth file extension.
        torch.save(model.state_dict(), os.path.join(experiments_folder, f'model={model.__class__.__name__}_batch={args.batch_size}_nheads={args.heads}_hdim={args.hidden_dim}_kfold={5}_lr={args.lr}.pt'))
        torch.save(all_list, os.path.join(experiments_folder, f'model={model.__class__.__name__}_norm={norm}_nlayers={num_layers}_nheads={args.heads}_hdim={args.hidden_dim}_kfold={5}.pt'))
    end_time = time.time()
    print('Total time elapsed: {:.4f}s'.format(end_time-start_time))