Last active
November 29, 2022 03:43
-
-
Save sandeepnmenon/572f8f6efd1c333934dbbc987d52f4b2 to your computer and use it in GitHub Desktop.
Optimus fitting
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| from scipy.optimize import nnls | |
| from scipy.optimize import curve_fit | |
| import math | |
| def optimus_fitting(df): | |
| """ | |
| This function takes in a csv file and returns b0,b1 and b2 fitting the model | |
| l = 1/(b0*k + b1) + b2 | |
| l is the training loss | |
| k is the number of iterations | |
| Solve for b0,b1 and b2 using non-negative least squares | |
| """ | |
| df = df.dropna() | |
| train_loss = df['Train_Loss'] | |
| epoch_num = df['Epoch_no'] | |
| # Solve for b0,b1 and b2 using non-negative least squares | |
| param_bounds = ([0, 0, 0], [np.inf, np.inf, np.inf]) | |
| # sigma = np.ones(len(epoch_num)) | |
| # NUM_SEGMENTS = 3 | |
| # for i in range(len(epoch_num)): | |
| # exp = int(math.floor(i/(math.ceil(1.0*len(epoch_num)/NUM_SEGMENTS)))) | |
| # sigma[i] /= 4 ** exp | |
| # params = curve_fit(optimus_predict, epoch_num, train_loss, sigma=np.array(sigma), absolute_sigma=True, | |
| # bounds=param_bounds) | |
| params = curve_fit(optimus_predict, epoch_num, train_loss, bounds=param_bounds) | |
| return params[0] | |
| def optimus_predict(k, b0, b1, b2): | |
| """ | |
| This function takes in b0, b1, b2 and k and returns the predicted loss | |
| l = 1/(b0*k + b1) + b2 | |
| l is the training loss | |
| k is the number of iterations | |
| """ | |
| return 1/(b0*k + b1) + b2 | |
| # Plot training loss data points and the fitted curve for each model | |
| import matplotlib.pyplot as plt | |
| #import seaborn as sns | |
| import os | |
| import numpy as np | |
| import pandas as pd | |
| data_dir ="./p4_experiment_data" | |
| #sns.set() | |
| epoch_steps = [[1,82], [83,123], [124,350]] | |
| for model in resnet_models: | |
| for gpu in gpu_models: | |
| csv_file = os.path.join(data_dir, model + "_" + gpu + ".csv") | |
| if not os.path.isfile(csv_file): | |
| continue | |
| df = pd.read_csv(csv_file) | |
| df = df.dropna() | |
| # Split the data into 3 segments based on column Epoch_no | |
| df1 = df.loc[df['Epoch_no'] <= epoch_steps[0][1]] | |
| df2 = df.loc[(df['Epoch_no'] >= epoch_steps[1][0]) & (df['Epoch_no'] <= epoch_steps[1][1])] | |
| df3 = df.loc[df['Epoch_no'] >= epoch_steps[2][0]] | |
| # Fit the model for each segment | |
| b0, b1, b2 = optimus_fitting(df1) | |
| b0_2, b1_2, b2_2 = optimus_fitting(df2) | |
| b0_3, b1_3, b2_3 = optimus_fitting(df3) | |
| # Plot the data points and the fitted curve for each segment | |
| plt.figure(figsize=(10, 6)) | |
| plt.plot(df1['Epoch_no'], df1['Train_Loss'], 'o', label='Data') | |
| plt.plot(df1['Epoch_no'], optimus_predict(df1['Epoch_no'], b0, b1, b2), 'r-', label='Fitted Curve') | |
| plt.plot(df2['Epoch_no'], df2['Train_Loss'], 'o', label='Data') | |
| plt.plot(df2['Epoch_no'], optimus_predict(df2['Epoch_no'], b0_2, b1_2, b2_2), 'r-', label='Fitted Curve') | |
| plt.plot(df3['Epoch_no'], df3['Train_Loss'], 'o', label='Data') | |
| plt.plot(df3['Epoch_no'], optimus_predict(df3['Epoch_no'], b0_3, b1_3, b2_3), 'r-', label='Fitted Curve') | |
| plt.xlabel('Epoch') | |
| plt.ylabel('Training Loss') | |
| plt.title(f"Training Loss vs Epoch for {model} on {gpu}") | |
| plt.legend() | |
| plt.show() | |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment