I am tuning my model with these parameters:
param_grid = {
'unit_1': 128,
'unit_2': 64,
'lr': 3e-4,
'activ': 'relu',
'epochs': 400,
'batch_size': 8}
Which is to say, I already know the parameters I'll be using, but I run these through the search function anyway, like so:
kgs = KerasGridSearch(build_classifier_model, param_grid,
monitor='val_loss', greater_is_better=False)
kgs.search({'text': x_train_text, 'vector': x_train_vector}, y_train,
sample_weight=y_train_weight,
batch_size=param_grid['batch_size'],
validation_data = ({'text': x_val_text, 'vector': x_val_vector},
y_val,y_val_weight),
callbacks=[es])
I then evaluate the model that results from this search and get some results:
Loss: 1.5514
Accuracy: 0.6601
Weighted accuracy: 0.6879
However, when I try to train a new model using the same parameters, this time with the fit() function, I get very different results.
history = classifier_model.fit({'text': x_train_text, 'vector': x_train_vector},
y_train,
sample_weight=y_train_weight,
steps_per_epoch=len(x_train_text) // best_params['batch_size'],
validation_data=({'text': x_val_text,
'vector': x_val_vector},
y_val,y_val_weight),
batch_size = best_params['batch_size'],
epochs=best_params['epochs']+10,
callbacks=[es])
Loss: 1.0226
Accuracy: 0.4915
Weighted accuracy: 0.5092
I have run the notebook repeatedly and the same thing happens. It is not making sense to me. I hope it isn't something obvious I'm missing. Please let me know if there is more information I can include to help solve this issue.
Related
I´m facing a strange behaviour which I can´t figure out why it is happening. I´m getting a really high loss(BinaryCrossentropy) on my validation batch around 20 or even higher while training. But after the training I do a prediction on the tet set and I get a loss which is lower than 1. Why is that? I went through my code over and over and can´t find the problem.
I´m doing a binary image classification for brian tumors on a dataset provided via kaggle(Link.
And you can find my notebook here: Google-Colab Notebook
My data is loaded this way:
batch_size = 20
train_ds = tf.keras.utils.image_dataset_from_directory(
train_data_path,
subset='training',
seed=42,
color_mode='grayscale',
batch_size=batch_size,
validation_split=0.30
)
valid_ds = tf.keras.utils.image_dataset_from_directory(
train_data_path,
subset='validation',
seed=42,
batch_size=batch_size,
color_mode='grayscale',
validation_split=0.30
)
test_ds = tf.keras.utils.image_dataset_from_directory(
test_data_path,
color_mode='grayscale',
batch_size=batch_size,
shuffle=False
)
This is my modle strcuture
input_shape = image_batch[0].shape
# set up the model structure
model = tf.keras.Sequential([
layers.Conv2D(32, (3, 3), activation='relu', input_shape=input_shape),
layers.MaxPooling2D((2,2)),
layers.Conv2D(64, (3, 3), activation='relu'),
layers.MaxPooling2D((2, 2)),
layers.Dropout(0.3),
layers.Conv2D(64, (3, 3), activation='relu'),
layers.MaxPooling2D((2, 2)),
layers.Dropout(0.3),
layers.Flatten(),
tf.keras.layers.Dense(32, activation="relu"),
layers.Dropout(0.3),
layers.Dense(1, activation="sigmoid")
])
model.summary()
This is my callback function which returns the plots during training:
class PlotLearning(tf.keras.callbacks.Callback):
"""
Callback to plot the learning curves of the model during training.
"""
def on_train_begin(self, logs={}):
self.metrics = {}
for metric in logs:
self.metrics[metric] = []
def on_epoch_end(self, epoch, logs={}):
# Storing metrics
print(logs)
for metric in logs:
if metric in self.metrics:
self.metrics[metric].append(logs.get(metric))
else:
self.metrics[metric] = [logs.get(metric)]
# Plotting
metrics = [x for x in logs if 'val' not in x]
f, axs = plt.subplots(1, len(metrics), figsize=(15,5))
clear_output(wait=True)
for i, metric in enumerate(metrics):
axs[i].plot(range(1, epoch + 2),
self.metrics[metric],
label=metric)
if logs['val_' + metric]:
axs[i].plot(range(1, epoch + 2),
self.metrics['val_' + metric],
label='val_' + metric)
axs[i].legend()
axs[i].grid()
plt.tight_layout()
plt.show()
callbacks_list = [PlotLearning()]
and this is the part where I start the training
# compile model
optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001)
model.compile(optimizer=optimizer,
loss=tf.keras.losses.BinaryCrossentropy(from_logits=False),
metrics=['accuracy']
)
# fit model
history = model.fit(prep_train_ds,
epochs=30,
validation_data=valid_ds,
callbacks=callbacks_list)
This is the output of the callback function after the last epoch run through:
As you can see the loss is really high and oscillating around 20, so I guess it is overfitting.
But as mentiod above, here is what I get when I make a prediction on the test set and calculate the binary crossentropy. The loss is again less than 1 and at least in the range of the training loss
I tried so many things like, chaning batch size, bcs. not enough samples of one class might be in one batch. Then I wanted to see if it is overfitting and changed the number of filters, applyed droput etc. But I couldn´t get the loss function down on the validation set. I´m quite new in the field of image classification and maybe I´m oversseing a thing.
I am training a VAE (using federated learning, but that is not so important) and wanted to keep the loss and train functions simple to exchange. The initial approach was to have a tf.function as loss function and a tf.function as train function as follows:
#tf.function
def kl_reconstruction_loss(model, model_input, beta):
x, y = model_input
mean, logvar = model.encode(x, y)
z = model.reparameterize(mean, logvar)
x_logit = model.decode(z, y)
cross_ent = tf.nn.sigmoid_cross_entropy_with_logits(logits=x_logit, labels=x)
reconstruction_loss = tf.reduce_mean(tf.reduce_sum(cross_ent, axis=[1, 2, 3]), axis=0)
kl_loss = tf.reduce_mean(0.5 * tf.reduce_sum(tf.exp(logvar) + tf.square(mean) - 1. - logvar, axis=-1), axis=0)
loss = reconstruction_loss + beta * kl_loss
return loss, kl_loss, reconstruction_loss
#tf.function
def train_fn(model: tf.keras.Model, batch, optimizer, kl_beta):
"""Trains the model on a single batch.
Args:
model: The VAE model.
batch: A batch of inputs [images, labels] for the vae.
optimizer: The optimizer to train the model.
beta: Weighting of KL loss
Returns:
The loss.
"""
def vae_loss():
"""Does the forward pass and computes losses for the generator."""
# N.B. The complete pass must be inside loss() for gradient tracing.
return kl_reconstruction_loss(model, batch, kl_beta)
with tf.GradientTape() as tape:
loss, kl_loss, rc_loss = vae_loss()
grads = tape.gradient(loss, model.trainable_variables)
grads_and_vars = zip(grads, model.trainable_variables)
optimizer.apply_gradients(grads_and_vars)
return loss
For my dataset this results in an epoch duration of approx. 25 seconds. However, since I have to call those functions directly in my code, I would have to enter different ones if I would want to try out different loss/train functions.
So, alternatively, I followed https://github.com/google-research/federated/tree/master/gans and wrapped the loss function in a class and the train function in another function. Now I have:
class VaeKlReconstructionLossFns(AbstractVaeLossFns):
#tf.function
def vae_loss(self, model, model_input, labels, global_round):
# KL Reconstruction loss
mean, logvar = model.encode(model_input, labels)
z = model.reparameterize(mean, logvar)
x_logit = model.decode(z, labels)
cross_ent = tf.nn.sigmoid_cross_entropy_with_logits(logits=x_logit, labels=model_input)
reconstruction_loss = tf.reduce_mean(tf.reduce_sum(cross_ent, axis=[1, 2, 3]), axis=0)
kl_loss = tf.reduce_mean(0.5 * tf.reduce_sum(tf.exp(logvar) + tf.square(mean) - 1. - logvar, axis=-1), axis=0)
loss = reconstruction_loss + self._get_beta(global_round) * kl_loss
if model.losses:
loss += tf.add_n(model.losses)
return loss, kl_loss, reconstruction_loss
def create_train_vae_fn(
vae_loss_fns: vae_losses.AbstractVaeLossFns,
vae_optimizer: tf.keras.optimizers.Optimizer):
"""Create a function that trains VAE, binding loss and optimizer.
Args:
vae_loss_fns: Instance of gan_losses.AbstractVAELossFns interface,
specifying the VAE training loss.
vae_optimizer: Optimizer for training the VAE.
Returns:
Function that executes one step of VAE training.
"""
# We check that the optimizer has not been used previously, which ensures
# that when it is bound the train fn isn't holding onto a different copy of
# the optimizer variables then the copy that is being exchanged b/w server and
# clients.
if vae_optimizer.variables():
raise ValueError(
'Expected vae_optimizer to not have been used previously, but '
'variables were already initialized.')
#tf.function
def train_vae_fn(model: tf.keras.Model,
model_inputs,
labels,
global_round,
new_optimizer_state=None):
"""Trains the model on a single batch.
Args:
model: The VAE model.
model_inputs: A batch of inputs (usually images) for the VAE.
labels: A batch of labels corresponding to the inputs.
global_round: The current glob al FL round for beta calculation
new_optimizer_state: A possible optimizer state to overwrite the current one with.
Returns:
The number of examples trained on.
The loss.
The updated optimizer state.
"""
def vae_loss():
"""Does the forward pass and computes losses for the generator."""
# N.B. The complete pass must be inside loss() for gradient tracing.
return vae_loss_fns.vae_loss(model, model_inputs, labels, global_round)
# Set optimizer vars
optimizer_state = get_optimizer_state(vae_optimizer)
if new_optimizer_state is not None:
# if optimizer is uninitialised, initialise vars
try:
tf.nest.assert_same_structure(optimizer_state, new_optimizer_state)
except ValueError:
initialize_optimizer_vars(vae_optimizer, model)
optimizer_state = get_optimizer_state(vae_optimizer)
tf.nest.assert_same_structure(optimizer_state, new_optimizer_state)
tf.nest.map_structure(lambda a, b: a.assign(b), optimizer_state, new_optimizer_state)
with tf.GradientTape() as tape:
loss, kl_loss, rc_loss = vae_loss()
grads = tape.gradient(loss, model.trainable_variables)
grads_and_vars = zip(grads, model.trainable_variables)
vae_optimizer.apply_gradients(grads_and_vars)
return tf.shape(model_inputs)[0], loss, optimizer_state
return train_vae_fn
This new formulation takes about 86 seconds per epoch.
I am struggling to understand why the second version performs so much worse than the first one. Does anyone have a good explanation for this?
Thanks in advance!
EDIT: My Tensorflow version is 2.5.0
After training a sarimax model, I had hoped to be able to preform forecasts in future using it with new observations without having to retrain it. However, I noticed that the number of observations i use in the newly applied forecast change the predictions.
From my understanding, provided that enough observations are given to allow the autoregression and moving average to be calculated correctly, the model would not even use the earlier historic observations to inform itself as the coefficients are not being retrained. In a (3,0,1) example i would have thought it would need atleast 3 observations to apply its trained coefficients. However this does not seem to be the case and i am questioning whether i have understood the model correctly.
as an example and test, i have applied a trained sarimax to the exact same data with the initial few observations removed to test the effect of the number of rows on the prediction with the following code:
import pandas as pd
from statsmodels.tsa.statespace.sarimax import SARIMAX, SARIMAXResults
y = [348, 363, 435, 491, 505, 404, 359, 310, 337, 360, 342, 406, 396, 420, 472, 548, 559, 463, 407, 362, 405, 417, 391, 419, 461, 472, 535, 622, 606, 508, 461, 390, 432]
ynew = y[10:]
print(ynew)
model = SARIMAX(endog=y, order=(3,0,1))
model = model.fit()
print(model.params)
pred1 = model.predict(start=len(y), end = len(y)+7)
model2 = model.apply(ynew)
print(model.params)
pred2 = model2.predict(start=len(ynew), end = len(ynew)+7)
print(pd.DataFrame({'pred1': pred1, 'pred2':pred2}))
The results are as follows:
pred1 pred2
0 472.246996 472.711770
1 494.753955 495.745968
2 498.092585 499.427285
3 489.428531 490.862153
4 477.678527 479.035869
5 469.023243 470.239459
6 465.576002 466.673790
7 466.338141 467.378903
Based on this, it means that if I were to produce a forecast from a trained model with new observations, the change in the number of observations itself would impact the integrity of the forecast.
What is the explanation for this? What is the standard practice for applying a trained model on new observations given the change in the number of them?
If i wanted to update the model but could not control for whether or not i had all of the original observations from the very start of my training set, this test would indicate that my forecast might as well be random numbers.
Main issue
The main problem here is that you are not using your new results object (model2) for your second set of predictions. You have:
pred2 = model.predict(start=len(ynew), end = len(ynew)+7)
but you should have:
pred2 = model2.predict(start=len(ynew), end = len(ynew)+7)
If you fix this, you get very similar predictions:
pred1 pred2
0 472.246996 472.711770
1 494.753955 495.745968
2 498.092585 499.427285
3 489.428531 490.862153
4 477.678527 479.035869
5 469.023243 470.239459
6 465.576002 466.673790
7 466.338141 467.378903
To understand why they're not identical, there is a second issue (which is not a problem in your code, but just a statistical feature of your data/model).
Secondary issue
Your estimated parameters imply an extremely persistent model:
print(params)
gives
ar.L1 2.134401
ar.L2 -1.683946
ar.L3 0.549369
ma.L1 -0.874801
sigma2 1807.187815
with is associated with a near-unit-root process (largest eigenvalue
= 0.99957719).
What this means is that it takes a very long time for the effects of a particular datapoint on the forecast to die out. In your case, this just means that there are still small effects on the forecasts from the first 10 periods.
This isn't a problem, it's just the way this particular estimated model works.
I am trying to implement an autoencoder for prediction of multiple labels using Keras. This is a snippet:
input = Input(shape=(768,))
hidden1 = Dense(512, activation='relu')(input)
compressed = Dense(256, activation='relu', activity_regularizer=l1(10e-6))(hidden1)
hidden2 = Dense(512, activation='relu')(compressed)
output = Dense(768, activation='sigmoid')(hidden2) # sigmoid is used because output of autoencoder is a set of probabilities
model = Model(input, output)
model.compile(optimizer='adam', loss='categorical_crossentropy') # categorical_crossentropy is used because it's prediction of multiple labels
history = model.fit(x_train, x_train, epochs=100, batch_size=50, validation_split=0.2)
I ran this in Jupyter Notebook (CPU) and I am getting loss and validation loss as:
loss: 193.8085 - val_loss: 439.7132
but when I ran it in Google Colab (GPU), I am getting very high loss and validation loss:
loss: 28383285849773932.0000 - val_loss: 26927464965996544.0000.
What could be the reason for this behavior?
After some research, I was able to predict the future value using the LSTM code below. I have also attached the Dmd1ahr.csv file in the github link that I am using.
https://github.com/ukeshchawal/hello-world/blob/master/Dmd1ahr.csv
As you all can see below, 90 data points are training sets and 91st to 100th are future value prediction.
However some of the questions that I still have are:
In order to predict these values I had to originally take more than hundred data sets (here, I have taken 500 data sets) which is not exactly what my primary goal is. Is there a way that given 500 data sets, it will predict the rest 10 or 20 out of sample data points? If yes, will you please write me a sample code where you can just take 500 data points from Dmd1ahr.csv file attached below and it will predict some future values (say 501 to 520) based on those 500 points?
The prediction are way off compared to the one who have in your blogs (definitely indicates for parameter tuning - I tried changing epochs, LSTM layers, Activation, optimizer). What other parameter tuning I can do to make it more robust?
Thank you'll in advance.
import numpy as np
import matplotlib.pyplot as plt
import pandas
# By twaking the architecture it could be made more robust
np.random.seed(7)
numOfSamples = 500
lengthTrain = 90
lengthValidation = 100
look_back = 1 # Can be set higher, in my experiments it made performance worse though
transientTime = 90 # Time to "burn in" time series
series = pandas.read_csv('Dmd1ahr.csv')
def generateTrainData(series, i, look_back):
return series[i:look_back+i+1]
trainX = np.stack([generateTrainData(series, i, look_back) for i in range(lengthTrain)])
testX = np.stack([generateTrainData(series, lengthTrain + i, look_back) for i in range(lengthValidation)])
trainX = trainX.reshape((lengthTrain,look_back+1,1))
testX = testX.reshape((lengthValidation, look_back + 1, 1))
trainY = trainX[:,1:,:]
trainX = trainX[:,:-1,:]
testY = testX[:,1:,:]
testX = testX[:,:-1,:]
############### Build Model ###############
import keras
from keras.models import Model
from keras import layers
from keras import regularizers
inputs = layers.Input(batch_shape=(1,look_back,1), name="main_input")
inputsAux = layers.Input(batch_shape=(1,look_back,1), name="aux_input")
# this layer makes the actual prediction, i.e. decides if and how much it goes up or down
x = layers.recurrent.LSTM(300,return_sequences=True, stateful=True)(inputs)
x = layers.recurrent.LSTM(200,return_sequences=True, stateful=True)(inputs)
x = layers.recurrent.LSTM(100,return_sequences=True, stateful=True)(inputs)
x = layers.recurrent.LSTM(50,return_sequences=True, stateful=True)(inputs)
x = layers.wrappers.TimeDistributed(layers.Dense(1, activation="linear",
kernel_regularizer=regularizers.l2(0.005),
activity_regularizer=regularizers.l1(0.005)))(x)
# auxillary input, the current input will be feed directly to the output
# this way the prediction from the step before will be used as a "base", and the Network just have to
# learn if it goes a little up or down
auxX = layers.wrappers.TimeDistributed(layers.Dense(1,
kernel_initializer=keras.initializers.Constant(value=1),
bias_initializer='zeros',
input_shape=(1,1), activation="linear", trainable=False
))(inputsAux)
outputs = layers.add([x, auxX], name="main_output")
model = Model(inputs=[inputs, inputsAux], outputs=outputs)
model.compile(optimizer='adam',
loss='mean_squared_error',
metrics=['mean_squared_error'])
#model.summary()
#model.fit({"main_input": trainX, "aux_input": trainX[look_back-1,look_back,:]},{"main_output": trainY}, epochs=4, batch_size=1, shuffle=False)
model.fit({"main_input": trainX, "aux_input": trainX[:,look_back-1,:].reshape(lengthTrain,1,1)},{"main_output": trainY}, epochs=100, batch_size=1, shuffle=False)
############### make predictions ###############
burnedInPredictions = np.zeros(transientTime)
testPredictions = np.zeros(len(testX))
# burn series in, here use first transitionTime number of samples from test data
for i in range(transientTime):
prediction = model.predict([np.array(testX[i, :, 0].reshape(1, look_back, 1)), np.array(testX[i, look_back - 1, 0].reshape(1, 1, 1))])
testPredictions[i] = prediction[0,0,0]
burnedInPredictions[:] = testPredictions[:transientTime]
# prediction, now dont use any previous data whatsoever anymore, network just has to run on its own output
for i in range(transientTime, len(testX)):
prediction = model.predict([prediction, prediction])
testPredictions[i] = prediction[0,0,0]
# for plotting reasons
testPredictions[:np.size(burnedInPredictions)-1] = np.nan
############### plot results ###############
#import matplotlib.pyplot as plt
plt.plot(testX[:, 0, 0])
plt.show()
plt.plot(burnedInPredictions, label = "training")
plt.plot(testPredictions, label = "prediction")
plt.legend()
plt.show()