Tutorial on CNN implementation for own data set in keras(TF & Theano backend)-part-1 

This is "THE" Tutorial !! Seriously among the best.

bro...this is one of the most awesome lecture I've ever heard on machine learning and convolution neutral networks.... did a great work... this is helping me alot in projects...

best video for learning CNNs as well as Keras! Especially loved the simple and detailed explanation of the image preprocessing part, which many others have ignored.

Thanks a looooot Anuj. I wanted to try my own dataset rather than the easily available pre-processed datasets. Your tutorial helped me a lot to do the same. Keep sharing!

A really good tutorial. I struggled to find a step by step process that allows me to understand. Thanks

Great work Anuj Shah! Must say, for begginers, this is THE video tutorial on YT! Keep going!

Excellent work Anuj shah

This is the tutorial that I'm searching for so long....thank you so much sir...:). Please include data augmentation part in it as well. Thanks.

Good & precise explanation. Nice work Anuj.

Thanks Anuj! Great job!

Excellent.... Thanks a lot Anuj

Cool stuff, implemented following your video, thanks man!

Excellent explanation anuj.... A big thanks to you .... Keep posting the videos like this

best explanation on CNN on youtube!!!

Excellent, very easy to understand. Thanks

Thank you very much Anuj.Excellent explanation of code and functions :)

Thank you. This is the best explanation.

Thank you a lot. A really good tutorial and vers good explanation Keep sharing anuj !

really v good, thanks Anuj shah Very helpful

me ayudaste mucho, ¡muchas gracias! saludo de colombia

Great explanation. it's so kind of you to share your knowledge that way, and even more to make the effort of explaining it. Thank you. Subscribing.

Awesome tutorial...............

Love ur all video....thankuuuu so mch for sharing sch a knowledge video.....please make more videos on different accuracy parameters....u r the best .

Amazing job Anuj!

Nice work

Best cnn tutorial 👍👍

Thank you a lot Anuj

Nice job Anuj Shah , it worked perfectly and helped me alot for my project .

Hey Anuj thanks again for another good video. Learnt a lot from you channel :). Specially your initial video on MNIST - Digit recognition tutorial video inspired me to implement my own CNN based Gesture Recognition project. I do have one question regarding input image set. In your video and many other tutorials over net I have noticed that we input only 1 channel image even if image is RGB colored we convert it to GrayScale. This way works fine, but I tried to input 3 channel image and made the required changes in neural net architecture and trained the model. However when I try to visualize the internal layers its not similar to what we get for grayscaled images and image spans across multiple kernels (for 32 kernels) instead of single kernel. I think its due to some stride factor that I have overlooked. But wanted to know if you have tried inputting 3 channel image ever ?

It is awesome Anju thank you

Very nice tutorial. Thank you once again. There a lot of tutorials for image classification however I am looking for pixel classification. I have doubt regarding the size of labels and image data in case of semantic segmentation using CNN. Will the size of labels be (number of samples, img_row, img_cols, number of classes) to train the mode? Also in case of semantic segmentation, can I use Dense layer or I should use upscale layer? Your help will be highly appreciated. If there is any tutorial for this please post a link. Thank you in advance.

Thnkuu so much sir... Very gud lucture

Hi Anuj, nice video, I encountered an error during visualizing the intermediate output... (np.shape(activations)) = (32,126,126) rather than (32,128,128)... do you know why that happen?

what if I have less dataset to train...it would just effect the accuracy or anything else as well

for the Convolution2D layer the documentation says that the sample axis should not be included in the input_shape, why did you include ?

Thank you for good explain.^^.I have an one Question....Is there a reason that you construct the network that way(In lecture) or other note that you referred??? Thanks.

hi, i got an error, how can i solve this?please help me . AttributeError: 'Activation' object has no attribute 'get_output'

how to extract feature of each & individual layers in keras??

Shouldnt expand_dim be at axis 3 instead of 4?

I can just train the CNN with images of faces for facial recognition?

thank you sir......

Hi Anuj. I have a model to classify RGB_D video sequence comprises from two branches (models) one it is inputted color images to extracted features from its, and the other depth images and then the outputs of two branches merged and inputted to the third model. If I want to get intermediate layers of the third model. from where I can get on the input layer of that model to use it in this functio def get_featuresmap (model,layer_idx,x_batch): get_activations=k.function([model.layers[0].input, k.learning_phase()],[model.layers[layer_idx].output]) activations=get_activations([x_batch,0]) return activations

I have uploaded own_dataset folder in google drive, from google drive how to load dataset? please tell how to specify the path

Hey Anuj...your video was quite helpful. Could you tell me how to add an classifier such as an svm to this model by replacing the FC layer...Im workng on ensemble learning and need help

I am getting "None type object has no attribute to append" while appending input_img_resize, i m novice to python please help

Hi Anuj, your are amazing man. I have one question, what if i want to implement it with webcam to recognize in real time and draw label with accuracy score and show in real time also. Thanks

Can you please make video on localization.? is there any built in functions for localization in keras?

python 2.7.6 Keras 1.2.2 and I get this error while building CNN : model.add(Convolution2D(32, 3,3,border_mode='same',input_shape=input_shape)) Traceback (most recent call last): File "", line 1, in File "/usr/local/lib/python2.7/dist-packages/keras/models.py", line 299, in add layer.create_input_layer(batch_input_shape, input_dtype) File "/usr/local/lib/python2.7/dist-packages/keras/engine/topology.py", line 401, in create_input_layer self(x) File "/usr/local/lib/python2.7/dist-packages/keras/engine/topology.py", line 529, in __call__ self.assert_input_compatibility(x) File "/usr/local/lib/python2.7/dist-packages/keras/engine/topology.py", line 469, in assert_input_compatibility str(K.ndim(x))) ValueError: Input 0 is incompatible with layer convolution2d_6: expected ndim=4, found ndim=3

nice work bro

How can this be use to train using GoogleNet and ResNet Architecture?

Hi, this is a great video! But I'm having issue uploading my own dataset. I have a folder with five sub-folders of data, and when I follow your instructions, it only uploads one of the folders. Any advice would be helpful!

Good Explanation, how to handle text file to extract the data using word2vec?

Hi, can you explain how to make the output continuous?

sir , i got Error when checking input: expected conv2d_32_input to have 4 dimensions, but got array with shape (646, 128, 128), someone can help me ? thakns

great work! please do you have a tutorial on how to fine tune inception v3? (that is transfer learning with inception v3)

Error when checking target: expected activation_5 to have shape (1,) but got array with shape (2,) I got this error pls help

Anuj shah can I use this code to build a biometric ear recognition model with a dataset having 125 subjects, atleast 5 samples in each?

thank you!

anuj bro u are great, your tutorials are timee saving, can u plz make a tutorial on how to make custome keras layer

Hai, do you have to resize all the image in your dataset before you go with all that stuff?

Can I use the same approach for recognizing static gesture? I already have my dataset, but could one also test using webcam or video feed?

very good work thank you! but i have this error: np.shape(model.layers[0].get_weights()[0]) i just changed channel from 1 to 3, and changet the cnn model with the vgg16 model. Can you help me?

Hey Anuj.. Thanks a lot for this wonderful video. Helps me a lot.. I do have a question that do you have video like this to apply LSTM model on video dataset. Actually, I am totally new to coding and facing a problem in that LSTM Model. So, can you please help me out ? Thanks again.

Hi Anuj, fist of all thanks to u for making such presentation. as you suggest i followed the same steps train model for 4 classes (cat,dog,horse,rider) & each class having 202 images. i tried so many permutation of convolution,activation,maxpooling, dropout but i could not increase the validation accuracy of model beyond 63%. could you suggest me what all things need to take care to enhance the accuracy of model or suggest useful links. your help would be highly appreciated.

Once again thanks for the grat tutorial. One fast question. What shffles from sklearn.cross_validation does is making train and test data, right? When we say test_data we mean our network ahs never seen that data, right? I mean it's not like k-folding in which you for instance break the dataset into 80% training data and 20% validation and do it for instance 5 times and each time assign data to these two sets randomly which mean at the end the network is also trained by the data in the other 20%. Is shuffle like the action of k-folding or is it really hiding those test data? Another question which I have is, what does epoch generally mean in the context of training a network? Lets say we ahve a 20 epoch of batch_zie 200. Does it mean at each epoch only 200 data is used to train the network, or does it mean we forward and back-propagate the whole training data 20 times? (I assume number of epochs as 20.) I really appreciate if you kindly shed light on these two questions and concepts. Once more really two thumbs up!

Great

bro am getting the below error when running code for getting activations could you please help me with it: c_api.TF_GetCode(self.status.status)) FailedPreconditionError: Error while reading resource variable conv1d_38/bias from Container: localhost. This could mean that the variable was uninitialized. Not found: Container localhost does not exist. (Could not find resource: localhost/conv1d_38/bias) [[{{node conv1d_38/BiasAdd/ReadVariableOp}} = ReadVariableOp[dtype=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:CPU:0"](conv1d_38/bias)]]

one question, are data_dir_list and img_list a vector of string or are they a built in function that is used to read the number of file in the directory? cheers

Hi, thanks for your great tutorial. I can say it was the most practical tutorial I've seen regarding IMPLEMENTATION of Keras and Tensorflow. I am using windows 10 pro, python 3.6 and I have installed Tensorflow and Keras by using pip. So I guess they have to bethe latest version. I have downloaded your code and after unzipping everything, executed it to see how it works, but I face the following warning and error. I am absolutely new to python. But it seems some of the methods are depricated. Can you also add what changes we need to apply to your code to solve this problem? Once again thank you for the great tutorial. The error: Warning (from warnings module): File "C:\Users\User\AppData\Local\Programs\Python\Python36\lib\site-packages\sklearn\cross_validation.py", line 44 "This module will be removed in 0.20.", DeprecationWarning) DeprecationWarning: This module was deprecated in version 0.18 in favor of the model_selection module into which all the refactored classes and functions are moved. Also note that the interface of the new CV iterators are different from that of this module. This module will be removed in 0.20. Using TensorFlow backend. Loaded the images of dataset-data Traceback (most recent call last): File "E:\Apps\own_data_cnn_implementation_keras-master\own_data_cnn_implementation_keras-master\custom_data_cnn.py", line 45, in input_img=cv2.cvtColor(input_img, cv2.COLOR_BGR2GRAY) cv2.error: D:\Build\OpenCV\opencv-3.2.0\modules\imgproc\src\color.cpp:9748: error: (-215) scn == 3 || scn == 4 in function cv::cvtColor

i am getting this error: code: num_channel=1 num_epoch=20 # Define the number of classes num_classes = 2 img_data_list=[] for dataset in data_dir_list: img_list=os.listdir(data_path+'/'+ dataset) print ('Loaded the images of dataset-'+'{} '.format(dataset)) console: img_data_list=[] for dataset in data_dir_list: File "", line 1 for dataset in data_dir_list: ^ SyntaxError: unexpected EOF while parsing can you tell ,whats the problem?

Awesome. thanx for the post. Could you maybe make a tut on Tensorflow serving with Keras sometime?

Can you please tell me if i have a dataset containing labelled data of whether two images are similar or not, how can i train that dataset using cnn. what will be the architecture.

why do we use 32 filters always ( odel.add(Conv2D(32, (3,3)) ???? , why should use 32 only at initial Layer ????

this can work to achive super resolution on images? i mean modifying the code

i use spyder to create the program but i cannot import cv2 and i also cant install opencv on the environment i used on my anaconda navigator. please help

I WILL SAY ONE THING: YOUR CHANNEL IS WORTH SUBSCRIBING

How did you pass the data folder for images? All the images in only one folder? Thank you!!

Good tutorial. How can I connect with you for some advice on my CNN project

Hello sir! While running this code, we are getting a value error: x and y must have same first dimension, but have shapes (20,) and (2,). Kindly help us!

sir, thanks for the video. I am working in NLP. My CNN model is following the structure as ['main_input', 'embedding_1', 'sequential_1', 'sequential_2', 'sequential_3', 'sequential_4', 'merge_1', 'dense_1', 'dropout_5', 'batch_normalization_5', 'dense_2', 'dropout_6', 'batch_normalization_6', 'dense_3', 'dropout_7', 'batch_normalization_7', 'dense_4'] I want to get the output of 'batch_normalization_7' layer i trainned the model. now after training when i followed your procedure. def getfeaturemaps(model,layer_idx,X_batch): get_activations=K.function([model.layers[0].input,K.learning_phase()],model.layers[layer_idx].output,) activation = get_activations([X_batch,0]) return activation Exception is coming ResourceExhaustedError: OOM when allocating tensor with shape[82080,1,100,256] [[Node: sequential_1/conv1d_1/convolution/Conv2D = Conv2D[T=DT_FLOAT, data_format="NHWC", padding="SAME", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true, _device="/job:localhost/replica:0/task:0/device:CPU:0"](sequential_1/conv1d_1/convolution/ExpandDims, sequential_1/conv1d_1/convolution/ExpandDims_1)]]

Hey Anuj...your video was quite helpful... but when i run your code "hist = model.fit(X_train, y_train, batch_size=16, nb_epoch=num_epoch, verbose=1, validation_data=(X_test, y_test))" i get an error as .. Error: ValueError: Error when checking input: expected conv2d_1_input to have 4 dimensions, but got array with shape (646, 128, 128).. Please Help..

why do we take test image path like : test_image = cv2.imread('data/Humans/rider-8.jpg')to Predict . i have taken my own file path to test_image = cv2.imread('data/riders123.jpg')- it is not Predicting right Value , why ? what will be the error....??????

Hi..I want to use only keras with theano in your code. How to disable tensorflow?(I havent installed it).

Very helpful . however i've some queries. How will u design bounding box? the certain object should be detected using a bounding box (if we want make your coding in real time apps.) and the box should slide over the entire image.

i have 25 different classes of images is there any change in convolution layer?

At 2:56, shouldn't it be `from sklearn.model_selection import train_test_split`?

Hello there. You have the best tutorials ever. You explain why every step is as it is. Thank you. I have a problem trough. I ran your code with this dataset with keras 2(backend TF) and i am getting an error on training. error: "Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 1 array(s), but instead got the following list of 658 arrays: [array([[[255, 255, 255]"

helpful

Sir if i keep all the images in a single folder then it shows an error of shape mismatch..how to fix it?..please reply

will the same code work without modification in keras 2?

how do we decide img_row, img_cols values? can we mention any values ? ???

Great video. If there are 808 images i.e., 202 per class why are they split up into labels like this: labels[0:102]=0 labels[102:204]=1 labels[204:606]=2 labels[606:]=3 This says 102 for class 1, 102 for class 2, 202 for class 3 and 202 for class 3??

I want to deploy your model on Android. I tried converting it in Tflite format. Buhh the app crashes when I capture image to identify the image. Can you send me code for conversion of model into Tflite?

00:00 = Intro 01:30 = overview of the entire lesson 02:40 = Loading and Prepocessing the Dataset 13:43 = Building the CNN model 17:03 = Training time 19:38 = Ploting the loss and accuracy curve 24:15 = Visualizing the intermediate layer output of CNN 28:47 = Plotting the confusion matrix for your result 31:01 = Saving and Loading the trained weights and model 33:27 = Wrapping up 34:23 = Thank you

Hello Anuj, what is the purpose of normalisation???

Sir can we do same things using matlab coding??

Sir is this model is working for real time camera to detect and classify the objects .I really like your way of explanation 😎😎 u r really really awesome keep teach us like this ❤️❤️

I have got the following error: File "C:\Users\user\Downloads\own_data_cnn_implementation_keras-master\own_data_cnn_implementation_keras-master\updated_custom_data_cnn.py", line 46, in label = labels_name[dataset] KeyError: 'data'

how you have taken the dataset in this video. i think its a numpy array ...

Sir , could you please tell me how to split a image by n number(user input) without any Inbuild function.