尝试将训练数据拟合到 CNN model 时出现 Memory 错误

Question

I am trying to create model on the popular cats and dogs training data using CNN. When I try to fit the training data using fit_generator, I get Memory Error as - self.filepaths is dynamic, is better to call it once outside the loop. Below is my line of code:

model.fit_generator(train_data, steps_per_epoch=10, 
                    validation_data=valid_data, validation_steps=2, epochs=10, verbose=2)

I read the training data using

train_data = ImageDataGenerator().flow_from_directory(train_path, target_size=(224,224), classes=['dogs', 'cats'], batch_size=10)

Kindly suggest a solution. Below is my model:

model = Sequential()
model.add(Conv2D(32, kernel_size=(5, 5),
                 activation='relu',
                  input_shape=(3,224, 224), data_format='channels_first'))
model.add(MaxPooling2D(pool_size=(2, 2), strides=2))
model.add(Conv2D(64, (5, 5), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2), strides=2))
model.add(Dense(32, activation='relu'))
model.add(Dropout(0.4))

Answer 1

Be sure to use python 64-bit. Last time I had a memory error I was using Python 32-bit and I didn't been aware of it.

Answer 2

Try following techniques:

1. Resize image to smaller height and width. 224 size generally works unless you have quite a lot number of images.
2. Use convolution layers and max/min pooling layers. This will help reduce number of neurons.
3. Try reducing the size of hidden layers and number of hidden layers.
4. Use sigmoid classifier since you are trying to build a binary classification (cat / dog).
5. If you are running this on you laptop consider gc.collect() before running your model. Also, try google colab to utilize gpu/tpu.

There can be other possible issues too, if you like to explain a bit more about your model and machine you are using I can provide more specific answer.

Answer 3

I was able to resolve issue using below layers flow: In this flow, I made the data flat and passed the final output to Dense layer. After doing this I got 100% correct prediction. Please let me know if someone has a better solution.

model.add(Conv2D(32, kernel_size=(5, 5), activation='relu', input_shape=(224,224,3)))
model.add(MaxPooling2D(pool_size=(2, 2), strides=2))
model.add(Conv2D(64, (5, 5), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2), strides=2))
model.add(Flatten())
model.add(Dense(32, activation='relu'))
model.add(Dropout(0.4))
model.add(Dense(2, activation='softmax'))

See below execution result:

Epoch 1/10
 - 4s - loss: 222.1019 - accuracy: 0.5800 - val_loss: 23.9488 - val_accuracy: 0.5000
Epoch 2/10
 - 3s - loss: 12.1180 - accuracy: 0.5600 - val_loss: 3.5912 - val_accuracy: 0.5000
Epoch 3/10
 - 3s - loss: 1.3664 - accuracy: 0.7200 - val_loss: 0.5239 - val_accuracy: 0.6000
Epoch 4/10
 - 4s - loss: 1.0074 - accuracy: 0.7200 - val_loss: 0.1986 - val_accuracy: 0.9000
Epoch 5/10
 - 4s - loss: 0.1890 - accuracy: 0.9400 - val_loss: 0.0298 - val_accuracy: 1.0000
Epoch 6/10
 - 3s - loss: 0.1680 - accuracy: 0.9200 - val_loss: 0.1973 - val_accuracy: 0.9000
Epoch 7/10
 - 3s - loss: 0.9097 - accuracy: 0.9600 - val_loss: 0.0020 - val_accuracy: 1.0000
Epoch 8/10
 - 4s - loss: 0.0372 - accuracy: 0.9800 - val_loss: 0.0027 - val_accuracy: 1.0000
Epoch 9/10
 - 3s - loss: 0.0466 - accuracy: 1.0000 - val_loss: 4.7272e-04 - val_accuracy: 1.0000
Epoch 10/10
 - 3s - loss: 0.0172 - accuracy: 1.0000 - val_loss: 2.7418e-07 - val_accuracy: 1.0000