Elegant way to plot average loss of many trains in tensorflow-CodePudding

I am running many iterations of a train so I can smooth out the loss curves. I would like an elegant way to average all the losses from history.history['loss'] but haven't found an easy way to do it. Here's a minimal example:

import tensorflow as tf
from tensorflow.keras.utils import to_categorical
from matplotlib import pyplot as plt

(x_train, y_train), _ = tf.keras.datasets.mnist.load_data()
x_train = x_train.reshape(60000, 784).astype('float32')/255
y_train = to_categorical(y_train, num_classes=10)

def get_model():
    model = tf.keras.Sequential()
    model.add(tf.keras.layers.Dense(10, activation='sigmoid', 
                input_shape=(784,)))
    model.add(tf.keras.layers.Dense(10, activation='softmax'))
    model.compile(loss="categorical_crossentropy", optimizer="sgd", 
                     metrics = ['accuracy'])
    return model

all_trains = []
for i in range(3):
    model = get_model()
    history = model.fit(x_train, y_train, epochs=2)
    all_trains.append(history)

If I wanted to plot just one example, I would do this:

plt.plot(history.epoch, history.history['loss'])
plt.show()

But instead, I want to average the loss from each train in all_trains and plot them. I can think of many clunky ways to do it but would like to find a clean way.

CodePudding user response：

You could simply do:

import numpy as np
import matplotlib.pyplot as plt

losses = [h.history['loss'] for h in all_trains]
mean_loss = np.mean(losses, axis=0)
std = np.std(losses, axis=0)

plt.errorbar(range(len(mean_loss)), mean_loss, yerr=std, capsize=5, marker='o')
plt.title('Average loss per epoch (± std)')
plt.xlabel('Epoch')
plt.ylabel('Categorical crossentropy')
plt.show()

I also added the standard deviation in this case.