OpenRAM/compiler/characterizer/analytical_util.py

326 lines
10 KiB
Python

# See LICENSE for licensing information.
#
# Copyright (c) 2016-2022 Regents of the University of California and The Board
# of Regents for the Oklahoma Agricultural and Mechanical College
# (acting for and on behalf of Oklahoma State University)
# All rights reserved.
#
import os
import csv
import math
import numpy as np
from openram import debug
process_transform = {'SS':0.0, 'TT': 0.5, 'FF':1.0}
def get_data_names(file_name, exclude_area=True):
"""
Returns just the data names in the first row of the CSV
"""
with open(file_name, newline='') as csvfile:
csv_reader = csv.reader(csvfile, delimiter=' ', quotechar='|')
row_iter = 0
# reader is iterable not a list, probably a better way to do this
for row in csv_reader:
# Return names from first row
names = row[0].split(',')
break
if exclude_area:
try:
area_ind = names.index('area')
except ValueError:
area_ind = -1
if area_ind != -1:
names = names[:area_ind] + names[area_ind+1:]
return names
def get_data(file_name):
"""
Returns data in CSV as lists of features
"""
with open(file_name, newline='') as csvfile:
csv_reader = csv.reader(csvfile, delimiter=' ', quotechar='|')
row_iter = 0
removed_items = 1
for row in csv_reader:
row_iter += 1
if row_iter == 1:
feature_names = row[0].split(',')
input_list = [[] for _ in range(len(feature_names)-removed_items)]
try:
# Save to remove area
area_ind = feature_names.index('area')
except ValueError:
area_ind = -1
try:
process_ind = feature_names.index('process')
except:
debug.error('Process not included as a feature.')
continue
data = []
split_str = row[0].split(',')
for i in range(len(split_str)):
if i == process_ind:
data.append(process_transform[split_str[i]])
elif i == area_ind:
continue
else:
data.append(float(split_str[i]))
data[0] = math.log(data[0], 2)
for i in range(len(data)):
input_list[i].append(data[i])
return input_list
def apply_samples_to_data(all_data, algo_samples):
# Take samples from algorithm and match them to samples in data
data_samples, unused_data = [], []
sample_positions = set()
for sample in algo_samples:
sample_positions.add(find_sample_position_with_min_error(all_data, sample))
for i in range(len(all_data)):
if i in sample_positions:
data_samples.append(all_data[i])
else:
unused_data.append(all_data[i])
return data_samples, unused_data
def find_sample_position_with_min_error(data, sampled_vals):
min_error = 0
sample_pos = 0
count = 0
for data_slice in data:
error = squared_error(data_slice, sampled_vals)
if min_error == 0 or error < min_error:
min_error = error
sample_pos = count
count += 1
return sample_pos
def squared_error(list_a, list_b):
error_sum = 0;
for a,b in zip(list_a, list_b):
error_sum+=(a-b)**2
return error_sum
def get_max_min_from_datasets(dir):
if not os.path.isdir(dir):
debug.warning("Input Directory not found:{}".format(dir))
return [], [], []
# Assuming all files are CSV
data_files = [f for f in os.listdir(dir) if os.path.isfile(os.path.join(dir, f))]
maxs,mins,sums,total_count = [],[],[],0
for file in data_files:
data = get_data(os.path.join(dir, file))
# Get max, min, sum, and count from every file
data_max, data_min, data_sum, count = [],[],[], 0
for feature_list in data:
data_max.append(max(feature_list))
data_min.append(min(feature_list))
data_sum.append(sum(feature_list))
count = len(feature_list)
# Aggregate the data
if not maxs or not mins or not sums:
maxs,mins,sums,total_count = data_max,data_min,data_sum,count
else:
for i in range(len(maxs)):
maxs[i] = max(data_max[i], maxs[i])
mins[i] = min(data_min[i], mins[i])
sums[i] = data_sum[i]+sums[i]
total_count+=count
avgs = [s/total_count for s in sums]
return maxs,mins,avgs
def get_max_min_from_file(path):
if not os.path.isfile(path):
debug.warning("Input file not found: {}".format(path))
return [], [], []
data = get_data(path)
# Get max, min, sum, and count from every file
data_max, data_min, data_sum, count = [],[],[], 0
for feature_list in data:
data_max.append(max(feature_list))
data_min.append(min(feature_list))
data_sum.append(sum(feature_list))
count = len(feature_list)
avgs = [s/count for s in data_sum]
return data_max, data_min, avgs
def get_data_and_scale(file_name, sample_dir):
maxs,mins,avgs = get_max_min_from_datasets(sample_dir)
# Get data
all_data = get_data(file_name)
# Scale data from file
self_scaled_data = [[] for _ in range(len(all_data[0]))]
self_maxs,self_mins = [],[]
for feature_list, cur_max, cur_min in zip(all_data,maxs, mins):
for i in range(len(feature_list)):
self_scaled_data[i].append((feature_list[i]-cur_min)/(cur_max-cur_min))
return np.asarray(self_scaled_data)
def rescale_data(data, old_maxs, old_mins, new_maxs, new_mins):
# unscale from old values, rescale by new values
data_new_scaling = []
for data_row in data:
scaled_row = []
for val, old_max,old_min, cur_max, cur_min in zip(data_row, old_maxs,old_mins, new_maxs, new_mins):
unscaled_data = val*(old_max-old_min) + old_min
scaled_row.append((unscaled_data-cur_min)/(cur_max-cur_min))
data_new_scaling.append(scaled_row)
return data_new_scaling
def sample_from_file(num_samples, file_name, sample_dir=None):
"""
Get a portion of the data from CSV file and scale it based on max/min of dataset.
Duplicate samples are trimmed.
"""
if sample_dir:
maxs,mins,avgs = get_max_min_from_datasets(sample_dir)
else:
maxs,mins,avgs = [], [], []
# Get data
all_data = get_data(file_name)
# Get algorithms sample points, assuming hypercube for now
num_labels = 1
inp_dims = len(all_data) - num_labels
samples = np.random.rand(num_samples, inp_dims)
# Scale data from file
self_scaled_data = [[] for _ in range(len(all_data[0]))]
self_maxs,self_mins = [],[]
for feature_list in all_data:
max_val = max(feature_list)
self_maxs.append(max_val)
min_val = min(feature_list)
self_mins.append(min_val)
for i in range(len(feature_list)):
self_scaled_data[i].append((feature_list[i]-min_val)/(max_val-min_val))
# Apply algorithm sampling points to available data
sampled_data, unused_data = apply_samples_to_data(self_scaled_data,samples)
#unscale values and rescale using all available data (both sampled and unused points rescaled)
if len(maxs)!=0 and len(mins)!=0:
sampled_data = rescale_data(sampled_data, self_maxs,self_mins, maxs, mins)
unused_new_scaling = rescale_data(unused_data, self_maxs,self_mins, maxs, mins)
return np.asarray(sampled_data), np.asarray(unused_new_scaling)
def get_scaled_data(file_name):
"""Get data from CSV file and scale it based on max/min of dataset"""
if file_name:
maxs,mins,avgs = get_max_min_from_file(file_name)
else:
maxs,mins,avgs = [], [], []
# Get data
all_data = get_data(file_name)
# Data is scaled by max/min and data format is changed to points vs feature lists
self_scaled_data = scale_data_and_transform(all_data)
data_np = np.asarray(self_scaled_data)
return data_np
def scale_data_and_transform(data):
"""
Assume data is a list of features, change to a list of points and max/min scale
"""
scaled_data = [[] for _ in range(len(data[0]))]
for feature_list in data:
max_val = max(feature_list)
min_val = min(feature_list)
for i in range(len(feature_list)):
if max_val == min_val:
scaled_data[i].append(0.0)
else:
scaled_data[i].append((feature_list[i]-min_val)/(max_val-min_val))
return scaled_data
def scale_input_datapoint(point, file_path):
"""
Input data has no output and needs to be scaled like the model inputs during
training.
"""
maxs, mins, avgs = get_max_min_from_file(file_path)
debug.info(3, "maxs={}".format(maxs))
debug.info(3, "mins={}".format(mins))
debug.info(3, "point={}".format(point))
scaled_point = []
for feature, mx, mn in zip(point, maxs, mins):
if mx == mn:
scaled_point.append(0.0)
else:
scaled_point.append((feature-mn)/(mx-mn))
return scaled_point
def unscale_data(data, file_path, pos=None):
if file_path:
maxs,mins,avgs = get_max_min_from_file(file_path)
else:
debug.error("Must provide reference data to unscale")
return None
# Hard coded to only convert the last max/min (i.e. the label of the data)
if pos == None:
maxs,mins,avgs = maxs[-1],mins[-1],avgs[-1]
else:
maxs,mins,avgs = maxs[pos],mins[pos],avgs[pos]
unscaled_data = []
for data_row in data:
unscaled_val = data_row*(maxs-mins) + mins
unscaled_data.append(unscaled_val)
return unscaled_data
def abs_error(labels, preds):
total_error = 0
for label_i, pred_i in zip(labels, preds):
cur_error = abs(label_i[0]-pred_i[0])/label_i[0]
total_error += cur_error
return total_error/len(labels)
def max_error(labels, preds):
mx_error = 0
for label_i, pred_i in zip(labels, preds):
cur_error = abs(label_i[0]-pred_i[0])/label_i[0]
mx_error = max(cur_error, mx_error)
return mx_error
def min_error(labels, preds):
mn_error = 1
for label_i, pred_i in zip(labels, preds):
cur_error = abs(label_i[0]-pred_i[0])/label_i[0]
mn_error = min(cur_error, mn_error)
return mn_error