sagarvegad / video-classification-cnn-and-lstm- Goto Github PK

Hello, can you tell us about the data directory structure in the flow_from_directory() method?
thank you so much!

Hello

I have some questions regarding the pipeline.
I've collected the images from Youtube and extracted features using VGG16, stored them into .npy files.

• When i start training CNN_LSTM the on epoch 5 the acc reaches 95%, but the validation is very low, so after epoch 5 basically no changes.
• How big the dataset should be? For example i've trained the jet class for 800 images. how many did you've used?
• There 2 training networks. So should I first train CNN with bottleneck and then train CNN_RNN?

1. code train_CNN_RNN.py what do numbers 56021 and 3974 indicate ? Is it number of images?

for x,y in train_generator:
if batch == (56021/batch_size):

for x,y in validation_generator:
if batch == (3974/batch_size):

2. When I execute code train_CNN_RNN.py I receive error
   File "train_CNN_RNN.py", line 53, in extract_features_and_store
   if x_generator==None:
   ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

Hi, I am new to LSTM, and I think there are some details in your codes I find it hard to understand.

First, I don't understand why we should set the shuffle=True, after all, we will feed the frame sequence into LSTM later, and I guess the sequence should be arrange by the time order, so it seems unreasonable to shuffle the image before training as it will break the order.

Also, the batch size is set as 128, I am wondering if you mean a batch of images belongs to the same video? After training those batches on vgg, the result batches are sent to LSTM, so I guess a batch of images are seen as frame sequence from the same video, am I right?

Please help me, thank you!

can you please share your data ?

can you please share the pretrained model?