Tips Tricks 21 - Understanding the keras-trained model saved as hdf5 (or h5)

This video explains the process of exploring keras model saved as hdf5 (or .h5).
To download the HDF Viewer: www.hdfgroup.o...

[@SathyanarayananKulasekaran]

19:09 "3*3 kernel, meaning 9 weights"...it will be 27 weights since the input number of channels is 27. Thats how your params in the first layer are 448 (27*16(weights)+16(bias))

[@ashishbisht5273]

Sir, the lecture is wondeful. one doubt I have, is it necessary to save weights separately before saving the entire model ? or weights also get saved , when we do save model command in keras? The reason i asked coz i saved a model using save model function and used it for prediction. The results were not good, so i suspect if weights are saved or not. please help. thank you!

[@Thoughts_of_a_900_ELO]

Sir, I am an undergrad student. As a project I am trying to build a novel automated system to count leaf stomata in a microscopic image, and hopefully measure the dimensions. I have already tried rcnn object detection and VGG16 based semantic segmentation. But result were far from satisfactory, may be due to the busy background of images. Can you suggest me an approach to solve this problem?

[@DigitalSreeni]

Stomata in leaves are usually well separated from each other but often difficult to be separated from the background. Therefore, I recommend a Unet based semantic segmentation approach that can segment stomata from the background. If there are any stomata that are very close, you can try watershed segmentation. In any case, once you have them segmented, you can easily convert the semantic segmented image to labeled image. This gives you something similar to instance segmentation. Now you can count and extract measurements.

[@Tapan_k]

Hello sir, could we know what the code would look like for the pix2pix satellite to maps code. I've tried everything mentioned in the video but I'm unable to get the output

[@hemasowjanyamamidi6123]

Hi @DigitalSreeni, when we are giving input of shape a*b*3 the filter size will be x*y*3 so the last axis of the filter depends on the last axis of the input we are giving. So the filter shape differs according to inputs in the respective layer. That's according to my understanding. Correct me if I am wrong. And coming to bias each filter will have a bias. If we have 32 filters in a conv layer 32 biases will be present. Thanks for the great tutorial...!

[@spaceinfotech5539]

God bless you DigitalSreeni. Your video is a life-saver. Continue sharing and we continue to learn.

[@rukshakkapoor3049]

Sreeni Sir I've trained TCN model and converted that model into h5 file and now I want to do the predictions using that trained model in Real Time. Can you tell me how to that ??😄

[@omarjallow7102]

Thank you. Nice presentation

[@TanTran-gn6hj]

Thanks for the information. You saved me with loading these models. One question: How do you load a model with 2 KerasLayer as custom objects? In the dictionary, if you put 'KerasLayer': layerA, and 'KerasLayer': layerB, how does Tensorflow know which is which?

[@prajnankarmakar4862]

hey i have a doubt regarding saving and loading model in .h5 format. I am unable to save the model. Is there any way i can reach you out for help? Please let me know.

[@chrysatsimperi6437]

Hello! Nice video! Is there a way that I can convert an h5 file to a .weights file?

[@JustMe-zt3fs]

Can you make a chatbot with only the trained model?

[@sorasora3611]

Zakat of knowledge is an instruction🤨

[@debasishjana3125]

Hi, thank you very much for all of your amazing content. I was wondering if I have lots of data, and all the data is making the RAM exhausted, can I use model.fit in for loop. I did one segmentation task but failed to train the model.

[@DigitalSreeni]

You provide batch size when using model. fit method. If you want to define your own batches and train the model in your own for loop then look into model.train_on_batch

[@lalitsingh5150]

Sir please make a video on Capsule networks

[@rameshwarsingh5859]

Thanks Sreeni Sir.

[@DigitalSreeni]

You are welcome.

[@lalitsingh5150]

Thanks again

[@DigitalSreeni]

Always welcome