Conv2DKAN giving graph execution error for non-sequential models. #19
Comments
|
I will check this error 😊 Could you please provide more error info (the traceback) about this error and what is the |
|
Please, find below the additional details pertaining to the error: The model fitting function: The model plotting and compilation occurs without raising any error. However, for the model fitting cell raises the following error: Further the execution occurs smoothly when the Conv2DKAN layers are replaced by Conv2D layers. Also using the DenseKAN layer does not result in an error in both Sequential and non-sequential models. The TimeDistributed layer in Keras is a wrapper that applies a specified layer independently to each time step in a sequence of inputs. In my case it allows me to apply the same processing to each frame in the video sequence. Once again thank you for your support 😊 |
|
Let's check the inputs shape together. From my understanding of your description, I believe your input should be a 5-dimensional tensor with import tensorflow as tf
from tfkan.layers import Conv2DKAN
from keras.layers import TimeDistributed
layer = TimeDistributed(Conv2DKAN(64, (3, 3), padding='same'))
x = tf.random.normal((2, 10, 32, 32, 3)) # x is a 5D tensor
y = layer(x)
print(y.shape) # (2, 10, 32, 32, 64)Let me know the inputs tensor you use can help us better find the problem! |
|
The input shape for the example code is as follows: The following is the condensed representation of the model: |
|
I test your code and it works fine in my env as below: import tensorflow as tf
from tfkan.layers import Conv2DKAN, DenseKAN
from keras.layers import TimeDistributed, Input, MaxPooling2D, Dropout, Flatten, LSTM, Dense
from tensorflow.keras.models import Model
def kan_single_branch_non_sequential(do=0.005, rl1=0, rl2=0):
# input for RGB frames
input_rgb = Input(shape=(10, 224, 224, 3),dtype=tf.float32, name='rgb_input')
x_rgb = TimeDistributed(Conv2DKAN(16, (3, 3), padding='same'))(input_rgb)
x_rgb = TimeDistributed(MaxPooling2D((4, 4)))(x_rgb)
x_rgb = TimeDistributed(Dropout(do))(x_rgb)
# ...
x_rgb = TimeDistributed(Flatten())(x_rgb)
x = x_rgb
x = LSTM(32)(x)
x = DenseKAN(32)(x)
# x=ActivityRegularization(l1=rl1,l2=rl2)(x)
output = Dense(6, activation='softmax')(x)
# model definition
model = Model(inputs=[input_rgb], outputs=output)
return model
# build and compile
model = kan_single_branch_non_sequential()
model.build(input_shape=(None, 10, 224, 224, 3))
model.compile(
optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy']
)
# generate mock data
bz = 1
x = tf.random.normal((bz, 10, 224, 224, 3))
y = tf.random.uniform(shape=(bz,), maxval=6, dtype=tf.int32)
# fit the model
model.fit(x, y, batch_size=bz, epochs=1)and the model shows the training progress successfully: 1/1 [==============================] - 1s 1s/step - loss: 1.8257 - accuracy: 0.0000e+00Maybe the graph execution error is caused by OOM? or the version of You might try to reduce the |
|
Thank you very much for your valuable input! The graph execution error was indeed due to OOM. I was able to rectify it by reducing frame size from 224x224 to 64x64. I was struggling to pinpoint the issue for weeks. Thank you very much for your time, effort and interest 😊. |
|
you are welcome, best😄 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Sir/ma'am
When the user fits a non-sequential models (as in the case of multi-branch models), a graph execution error is encountered.
In particular no error is encountered with the model declaration:
model = tf.keras.Sequential([
...
TimeDistributed(Conv2DKAN(16, (3, 3), padding='same')),
...
])
However, using the form
x=TimeDistributed(Conv2DKAN(32, (3, 3), padding='same'))(x)
raises the said error.
Please, look into matter at your earlier convenience.
Thank you for the help in advance.
The text was updated successfully, but these errors were encountered: