I am new to CNN transfer learning. I am trying to use ResNet50 transfer learning by combining MINST dataset & Kaggle A-Z dataset. As I am using Resnet50, I changed channels to 3. Now I am getting the value error. At output avg_pool error is coming. How can I override this or do I need to modify my code? I need to maintain classes as 36. I tried GlobalAvgPooling2D as well, but getting same error.
def load_az_dataset(datasetPath):
data = []
labels = []
for row in open(datasetPath):
row = row.split(",")
label = int(row[0])
image = np.array([int(x) for x in row[1:]], dtype="uint8")
image = image.reshape((28, 28))
data.append(image)
labels.append(label)
data = np.array(data, dtype="float32")
labels = np.array(labels, dtype="int")
return (data, labels)
def load_mnist_dataset():
((trainData, trainLabels), (testData, testLabels)) = mnist.load_data()
data = np.vstack([trainData, testData])
labels = np.hstack([trainLabels, testLabels])
return (data, labels)
EPOCHS = 50
INIT_LR = 1e-1
BS = 128
(azData, azLabels) = load_az_dataset('A_Z Handwritten Data\A_Z Handwritten Data.csv')
(digitsData, digitsLabels) = load_mnist_dataset()
azLabels += 10
data = np.vstack([azData, digitsData])
labels = np.hstack([azLabels, digitsLabels])
data = [cv2.resize(image, (32, 32)) for image in data]
data = np.array(data, dtype="float32")
data = np.expand_dims(data, axis=-1)
data = np.repeat(data, 3, axis=-1)
data /= 255.0
le = LabelBinarizer()
labels = le.fit_transform(labels)
counts = labels.sum(axis=0)
classTotals = labels.sum(axis=0)
classWeight = {}
for i in range(0, len(classTotals)):
classWeight[i] = classTotals.max() / classTotals[i]
(trainX, testX, trainY, testY) = train_test_split(data,labels, test_size=0.20, stratify=labels, random_state=42)
aug = ImageDataGenerator(
rotation_range=10,
zoom_range=0.05,
width_shift_range=0.1,
height_shift_range=0.1,
shear_range=0.15,
horizontal_flip=False,
fill_mode="nearest")
opt = SGD(lr=INIT_LR, decay=INIT_LR / EPOCHS)
model = ResNet50(include_top=False,input_shape=(32,32,3),pooling='avg',classes=len(le.classes_),weights='imagenet')
model.compile(optimizer=opt,loss="categorical_crossentropy",metrics=["accuracy"])
H = model.fit(
aug.flow(trainX, trainY, batch_size=BS),
validation_data=(testX, testY),
steps_per_epoch=len(trainX) // BS,
epochs=EPOCHS,
class_weight=classWeight,
verbose=1)
X_train: (353960, 32, 32, 3) Y_train: (353960, 36) X_test: (88491, 32, 32, 3) Y_test: (88491, 36)
I added only first and last lines of model summary:
Model: "resnet50"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_21 (InputLayer) [(None, 32, 32, 3)] 0
__________________________________________________________________________________________________
conv1_pad (ZeroPadding2D) (None, 38, 38, 3) 0 input_21[0][0]
__________________________________________________________________________________________________
conv1_conv (Conv2D) (None, 16, 16, 64) 9472 conv1_pad[0][0]
__________________________________________________________________________________________________
conv1_bn (BatchNormalization) (None, 16, 16, 64) 256 conv1_conv[0][0]
__________________________________________________________________________________________________
conv1_relu (Activation) (None, 16, 16, 64) 0 conv1_bn[0][0]
__________________________________________________________________________________________________
pool1_pad (ZeroPadding2D) (None, 18, 18, 64) 0 conv1_relu[0][0]
__________________________________________________________________________________________________
######GAP############
conv5_block3_2_relu (Activation (None, 1, 1, 512) 0 conv5_block3_2_bn[0][0]
__________________________________________________________________________________________________
conv5_block3_3_conv (Conv2D) (None, 1, 1, 2048) 1050624 conv5_block3_2_relu[0][0]
__________________________________________________________________________________________________
conv5_block3_3_bn (BatchNormali (None, 1, 1, 2048) 8192 conv5_block3_3_conv[0][0]
__________________________________________________________________________________________________
conv5_block3_add (Add) (None, 1, 1, 2048) 0 conv5_block2_out[0][0]
conv5_block3_3_bn[0][0]
__________________________________________________________________________________________________
conv5_block3_out (Activation) (None, 1, 1, 2048) 0 conv5_block3_add[0][0]
__________________________________________________________________________________________________
avg_pool (GlobalAveragePooling2 (None, 2048) 0 conv5_block3_out[0][0]
==================================================================================================
Total params: 23,587,712
Trainable params: 23,534,592
Non-trainable params: 53,120
______________________________
You correctly loaded the ResNet model without the top layer, that is by setting the parameter include_top=False. But you have to provide a new layer, a new head to your model now.
You just have to concatenate a Dense layer to the ResNet model. The number of nodes provided must be the same as the number of classes you want to classify: in this case 36.
You can fix your error like this:
import tensorflow as tf
from tensorflow.keras.applications.resnet50 import ResNet50
from tensorflow.keras.layers import Flatten, Dense
from tensorflow.keras import Model
...
base = ResNet50(include_top=False, input_shape=(32,32,3), pooling='avg', classes=36, weights='imagenet')
x = Flatten()(base.output)
x = Dense(36, activation='softmax')(x)
model = Model(inputs=base.inputs, outputs=x)
...