Accuracy not changing across the epochs

See original GitHub issue

`from keras import regularizers from keras.layers import Input, Dense, Activation, Dropout from keras.models import Model, Sequential from keras.utils import np_utils from keras.optimizers import SGD import numpy import pandas from sklearn import preprocessing from sklearn.model_selection import GridSearchCV from keras.models import Sequential from keras.layers import Dense

dataset = numpy.loadtxt(“Train.txt”, delimiter=“\t”)

split into input (X) and output (Y) variables

X = dataset[:,0:106] Y = dataset[:,106]

model = Sequential() model.add(Dense(64, input_dim=106, init=‘uniform’, activation=‘relu’)) model.add(Dense(32, init=‘uniform’, activation=‘tanh’)) model.add(Dense(12, init=‘uniform’, activation=‘relu’)) model.add(Dense(8, init=‘uniform’, activation=‘sigmoid’)) model.add(Dense(1, init=‘uniform’, activation=‘relu’))

Compile model

optimizer = SGD(lr=0.0001, momentum=0.4) model.compile(loss=‘binary_crossentropy’, optimizer=optimizer, metrics=[‘accuracy’])

Fit the model

model.fit(X, Y, nb_epoch=100, batch_size=300, verbose=1)

evaluate the model

scores = model.evaluate(X, Y, verbose=0) print(“%s: %.2f%%” % (model.metrics_names[1], scores[1]*100))

serialize model to JSON

model_json = model.to_json() with open(“model.json”, “w”) as json_file: json_file.write(model_json)

serialize weights to HDF5

model.save_weights(“model.h5”) print(“Saved model to disk”)`

For the above training, the accuracy remains the same across all the epochs. Using Theano backend. DEBUG: nvcc STDOUT mod.cu Creating library D:/tmp/xyz/theano.NOBACKUP/compiledir_Windows-8.1-6.3.9600-Intel64

Using gpu device 0: Quadro K2200 (CNMeM is disabled, cuDNN not available) DEBUG: nvcc STDOUT mod.cu Creating library D:/tmp/xyz/theano.NOBACKUP/compiledir_Windows-8.1-6.3.9600-Intel64

DEBUG: nvcc STDOUT mod.cu Creating library D:/tmp/xyz/theano.NOBACKUP/compiledir_Windows-8.1-6.3.9600-Intel64

Epoch 1/100 10131566/10131566 [==============================] - 338s - loss: 4.4380 - acc: 0.7247 Epoch 2/100 10131566/10131566 [==============================] - 341s - loss: 4.4380 - acc: 0.7247 Epoch 3/100 10131566/10131566 [==============================] - 340s - loss: 4.4380 - acc: 0.7247 Epoch 4/100 10131566/10131566 [==============================] - 350s - loss: 4.4380 - acc: 0.7247 Epoch 5/100 10131566/10131566 [==============================] - 343s - loss: 4.4380 - acc: 0.7247 Epoch 6/100 10131566/10131566 [==============================] - 336s - loss: 4.4380 - acc: 0.7247 Epoch 7/100 10131566/10131566 [==============================] - 346s - loss: 4.4380 - acc: 0.7247 Epoch 8/100 10131566/10131566 [==============================] - 349s - loss: 4.4380 - acc: 0.7247 Epoch 9/100 10131566/10131566 [==============================] - 345s - loss: 4.4380 - acc: 0.7247 Epoch 10/100 10131566/10131566 [==============================] - 345s - loss: 4.4380 - acc: 0.7247 Epoch 11/100 10131566/10131566 [==============================] - 349s - loss: 4.4380 - acc: 0.7247 Epoch 12/100 10131566/10131566 [==============================] - 349s - loss: 4.4380 - acc: 0.7247 Epoch 13/100 10131566/10131566 [==============================] - 344s - loss: 4.4380 - acc: 0.7247 Epoch 14/100 10131566/10131566 [==============================] - 341s - loss: 4.4380 - acc: 0.7247 Epoch 15/100 10131566/10131566 [==============================] - 344s - loss: 4.4380 - acc: 0.7247 Epoch 16/100 10131566/10131566 [==============================] - 344s - loss: 4.4380 - acc: 0.7247 Epoch 17/100 10131566/10131566 [==============================] - 353s - loss: 4.4380 - acc: 0.7247 Epoch 18/100 10131566/10131566 [==============================] - 349s - loss: 4.4380 - acc: 0.7247 Epoch 19/100 10131566/10131566 [==============================] - 343s - loss: 4.4380 - acc: 0.7247 Epoch 20/100 10131566/10131566 [==============================] - 272s - loss: 4.4380 - acc: 0.7247 Epoch 21/100 10131566/10131566 [==============================] - 270s - loss: 4.4380 - acc: 0.7247 Epoch 22/100 10131566/10131566 [==============================] - 268s - loss: 4.4380 - acc: 0.7247 Epoch 23/100 10131566/10131566 [==============================] - 269s - loss: 4.4380 - acc: 0.7247 Epoch 24/100 10131566/10131566 [==============================] - 270s - loss: 4.4380 - acc: 0.7247 Epoch 25/100 10131566/10131566 [==============================] - 269s - loss: 4.4380 - acc: 0.7247 Epoch 26/100 10131566/10131566 [==============================] - 268s - loss: 4.4380 - acc: 0.7247 Epoch 27/100 10131566/10131566 [==============================] - 270s - loss: 4.4380 - acc: 0.7247 Epoch 28/100 10131566/10131566 [==============================] - 268s - loss: 4.4380 - acc: 0.7247 Epoch 29/100 10131566/10131566 [==============================] - 270s - loss: 4.4380 - acc: 0.7247 Epoch 30/100 10131566/10131566 [==============================] - 271s - loss: 4.4380 - acc: 0.7247 Epoch 31/100 10131566/10131566 [==============================] - 268s - loss: 4.4380 - acc: 0.7247 Epoch 32/100 10131566/10131566 [==============================] - 274s - loss: 4.4380 - acc: 0.7247 Epoch 33/100 10131566/10131566 [==============================] - 268s - loss: 4.4380 - acc: 0.7247 Epoch 34/100 10131566/10131566 [==============================] - 272s - loss: 4.4380 - acc: 0.7247 Epoch 35/100 10131566/10131566 [==============================] - 271s - loss: 4.4380 - acc: 0.7247 Epoch 36/100 10131566/10131566 [==============================] - 271s - loss: 4.4380 - acc: 0.7247 Epoch 37/100 10131566/10131566 [==============================] - 342s - loss: 4.4380 - acc: 0.7247 Epoch 38/100 10131566/10131566 [==============================] - 337s - loss: 4.4380 - acc: 0.7247 Epoch 39/100 10131566/10131566 [==============================] - 336s - loss: 4.4380 - acc: 0.7247 Epoch 40/100 10131566/10131566 [==============================] - 351s - loss: 4.4380 - acc: 0.7247 Epoch 41/100 10131566/10131566 [==============================] - 387s - loss: 4.4380 - acc: 0.7247 Epoch 42/100 10131566/10131566 [==============================] - 388s - loss: 4.4380 - acc: 0.7247 Epoch 43/100 10131566/10131566 [==============================] - 364s - loss: 4.4380 - acc: 0.7247

… this remains the same till 100 epochs?

Issue Analytics

State:
Created 7 years ago
Comments:6 (1 by maintainers)

Top GitHub Comments

17reactions

bstrinercommented, Dec 25, 2016

Can you check if your Ys are all 0 or 1? Your model doesn’t match your loss function.

If your targets are [0,1], use a sigmoid output layer and binary_crossentropy loss.

If your targets are [-1,1], use a linear or tanh output layer and hinge or squared_hinge loss.

If your targets are labels for k categories, use to_categorical to convert to one-hot, use a softmax output layer with k outputs, and use categorical_crossentropy loss.

As a side note, try getting a 2 layer model to work before jumping to 5.

Cheers, Ben

0reactions

KushalDavecommented, Dec 26, 2016

@bstriner @Shuailong Yes, my labels are all 0 or 1. Once I set the output layer to sigmoid, I see accurcy improving across epochs with the binary_loss.

Also, I had tried a NN with 2 layers too, the one I posted was one of many instances that I had tried.

Many thanks for all your generous responses.