I created a simple neural network for understanding how group convolutions can reduce the number of parameters. But when I use the groups parameter in the second convolution layer, I am getting an unimplemented error. However, when groups parameter is not used, everything works fine. Why when using groups parameter, it throws unimplemented error? Does that mean group convolution is not available in keras api?
import tensorflow
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Conv2D,Reshape,MaxPooling2D
from tensorflow.keras.utils import to_categorical
import numpy as np
num_classes = 10
a = np.random.randint(low=0,high=255,size=(100,28,28,1))
b = np.random.randint(low=0,high=10,size=(100,7,7))
a = a.astype('float32')
a = a/255
X_train, Y_train = a[:80], b[:80]
X_test, Y_test = a[80:], b[80:]
num_classes=10
Y_train = to_categorical(Y_train, num_classes)
Y_test = to_categorical(Y_test, num_classes)
# Create the model
model = Sequential()
model.add(Conv2D(8, kernel_size=(3,3),input_shape=(28,28,1),padding='same'))
model.add(Conv2D(8, kernel_size=(3,3),groups=4,input_shape=(28,28,1),padding='same'))
# model.add(Dense(10, input_shape=input_shape, activation='relu'))
model.add(Dense(50, activation='relu'))
model.add(Dense(num_classes, activation='softmax'))
model.add(MaxPooling2D())
model.add(MaxPooling2D())
# model.add(Reshape(target_shape=(10,)))
model.summary()
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
model.fit(X_train, Y_train, epochs=10, batch_size=250, verbose=1, validation_split=0.2)
# model.save_weights("model.h5")
# # Test the model after training
# test_results = model.evaluate(X_test, Y_test, verbose=1)
# print(f'Test results - Loss: {test_results[0]} - Accuracy: {test_results[1]}%')
Error
UnimplementedError: Fused conv implementation does not support grouped convolutions for now.
[[node sequential_38/conv2d_37/BiasAdd (defined at <ipython-input-42-e7c1c931a421>:50) ]] [Op:__inference_train_function_8596]
Function call stack:
train_function
Here is the colab file for your code. According to the doc
A positive integer specifying the number of groups in which the input is split along the channel axis. Each group is convolved separately with filters / groups filters. The output is the concatenation of all the groups results along the channel axis. Input channels and filters must both be divisible by groups.
In your code, I found no conflict with that. It should work. Otherwise, it may issue with something else.