CODEWITHSUNDEEP
tensorflowkerasconv-neural-networkresnet
SomeOneInteresting
SomeOneInteresting

Reputation: 87

How to use ResNet152 in tensorflow.keras to train it on my own data?

I am trying to build my own network using ResNet152 in Colab. Here is the part of my code :

(x_train, y_train), (x_test, y_test) = cifar10.load_data()

x_train = x_train.astype('float32') 
x_test = x_test.astype('float32') 

# Normalization 
mean = np.mean(x_train, axis=0)
std = np.std(x_train, axis=0)

x_train = (x_train - mean) / (std+1e-7)
x_test = (x_test - mean) / (std+1e-7)

y_train = to_categorical(y_train, num_labels)  
y_test = to_categorical(y_test, num_labels)

Then data augmentation

datagen = ImageDataGenerator(
          rotation_range=9,
          width_shift_range=.2, height_shift_range=.2,
          horizontal_flip=True, vertical_flip=True,
          rescale=True,
          validation_split=.2
          )

datagen.fit(x_train)

datagen_train = datagen.flow(x_train, y_train, shuffle=True, subset='training')
datagen_val = datagen.flow(x_train, y_train, shuffle=True, subset='validation')

And then model building:

model = Sequential()
model.add(ResNet152(include_top=False, pooling='avg'))
model.add(Dense(num_labels, activation='softmax'))

for layer in model.layers[:]:
    layer.trainable = True

But Colab is giving me very poor results after fitting the generator. Are there any problems in my code?

Upvotes: 0

Views: 1343

Answers (1)

Timbus Calin
Timbus Calin

Reputation: 14993

The only problem in your code is that you set all your layers to trainable: practically you do not benefit from the concept of transfer learning; depending on:

  1. The size of your own dataset.
  2. The degree of similarity between your own dataset and the pretrained dataset.

you must freeze a different number of layers in order to have good results.

enter image description here

The image above summarizes what I have stated before. I would start by training only the last layer (Dense) for 2-3 epochs, and then train, say, the last quarter of layers (starting from 3/4 index up to the last).

Upvotes: 3

Related Questions