Image transformations require Scipy but Scipy already installed

General Discussion
, ,
#1

Hello, I’m running Tensorflow 2.4.1 on a Windows 10 computer (Python 3.8.5 with Anaconda). I’m using a NVIDIA GeForce RTX 2060 GPU with TF. I’m having problems trying to launch a CNN model using the ImageDataGenerator feeding from a Pandas dataframe. After defining the data generator and the model, I get the following error when running model.fit: “Image transformations require SciPy. Install SciPy.” However, when I can confirm that Scipy is installed by running “import scipy” without errors.

The definition of the image generator is as follows:

train_datagen = ImageDataGenerator(horizontal_flip=False,
                               vertical_flip=False,
                               rescale=1/255.0).flow_from_dataframe(dataframe=X,
                                                                    x_col='image_name',
                                                                    y_col='response',
                                                                    shuffle=False, 
                                                                    directory=src_path,
                                                                    target_size=(128, 128),
                                                                    class_mode=None
                                                                   )

And error comes when running:
model.fit(train_datagen,
epochs=5)

Please help!

#2

PS. I believe the error results from my GPU configuration. I am running a CPU version of the code and training is running (although extremely slow). I had to change my imagegenerator declaration to:
class_mode='raw',
as I am running a regression model.

#3

It looks like you’ve worked it out. If you need help with getting the code to train faster with your RTX 2060, let us know.

#4

Thanks! Yes, any tips would be appreciated.

#5

If you can share some of the code, users here may be able to see what can be optimized :+1: