1. with tf.name_scope("eval"):
2. correct = tf.nn.in_top_k(logits, y, 1)
3. accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
Sometime, Sandeep says logits are our categories which is 0-9 and sometime he says it is calaculated field. Please explain what is happening in line number 2. I have seen the whole lecture many times where he explains this but I am not able to get it. May be some example can help me to understand it better. I am not able to visualize what is logits inside the argument.
I read the link provided by you, but still i am not clear. Can you explain me the output of below sample codes. I am not getting how these output as True and False.
1) out = tf.nn.in_top_k([[0.8,0.9,0.3], [0.1,0.6,0.4], [0.6,0.9,0.65]], [1,2,3], 1)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
print (sess.run(out))
Output : [ True False False]
2) out = tf.nn.in_top_k([[0.8,0.9,0.3], [0.1,0.6,0.4], [0.6,0.9,0.65]], [1,1,1], 1)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
print (sess.run(out))
Output: [ True True True]
3) out = tf.nn.in_top_k([[0.8,0.9,0.3], [0.1,0.6,0.4], [0.6,0.9,0.65]], [2,2,2], 2)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
print (sess.run(out))
Output: [False True True]
Like we have saver = tf.train.Saver() to save the model and reuse the existing model, do we have such facility in ML libraray scikitlearn for different types of algorithm ? If yes, please let me know.
By steps in DNN, do we mean number of iterations of back propagation and weights adjustment. So, in below line :
dnn_clf.fit(X_train, y_train, batch_size=50, steps=30000), we mean batch size of 50 rows will undergo weights adjustment until 30000 iterations are over ? And same operation is going to repeat until all batch elements are covered.
In an epoch, each training sample gets a chance to update the weights of the classification algorithm. So an epoch is completed after all the batches update the function. Hence, here in the example, if the training set has 100 samples in total, and the batch size is 50, then one epoch is completed after 2 batches of the training data undergo forward propagaation and back propagation. There will be 30000 such epochs in this example.
1. Why don’t we consider negative values for RELU AFunc?
2. how can we say that Relu is non-linear activation function because gradient is a constant (for x>0 like linear function) which does not depend upon the input value x at all. This implies that the weights and biases will be updated during the backpropagation process but the updating factor would be the same. In this scenario, the neural network will not really improve the error since the gradient is the same for every iteration. The network will not be able to train well and capture the complex patterns from the data.
For the first question, I am not sure what you mean by "negative values for ReLU function", I am assuming you meant why cannot ReLU output negative values. Here's a very good explanation of the same:
Linear means to move ahead in a straight line. Please check the entire graph of ReLU and not a portion of the same because the entire graph constitutes the ReLU activation function and not just a part of it, you will notice that the output of ReLU is not linear, or in other words, not a straight line.
Thank you very much for asking me to do activity in prvious query. I can relate your answer with that RELU is adding non-linearity. this is not as same as linear activation func.
I have completed my course 72% and all mandetory projects and also applied for the certificate. Please tell me how much time it will take to get my certificate. Please let me know asap.
Please ignore the 2 questions below as I was finally abe to execute those.
Just one query though!!
Why executing this code " !ls -l model_ckps/ " gives total as 2248 (pls refer to the screenshot)? While the lecture file shows the value as 600? Please comment. Thanks!
It may found some silly question but i really wanted to ask that logistic regression algorithm functionali is quit similar to sigmoid activation function? are they similar? if not then what is difference between them? Functionality wise i am finding them very similar.
What is the difference between making Neural Network using Tensorflow coding and using Tensorflow API like Keras? Which will be more usefull?. As I am very new to coding so thinking of having some easy way to code the neural network? Please suggest.
Hello sir, i wanted to Know what isthe difference between n_inputs and n_neurons, i mean X is the input matrix - so is how are they related numerically with it? and on what basis is the bias b value computed?
Hi, I m running this code in tfv2 but showing this error please help me to resolve
import tensorflow as tf
# from tensorflow.python.feature_column import dense_features
config = tf.estimator.RunConfig(tf_random_seed=42) # not shown in the config
Hello. The codes provided in this course, unfortunately, won't run in Tensorflow version 2.x. I use Google Colab to run my code. You can use Tensorflow 1.x in Colab You just have to run this command on the cell prior to running any underlying code :
%tensorflow_version 1.x
Ps: by default it will run the tf 2.x version so u need to run the above command otherwise it will show error.
Hi When I opened model_ckps in deep learning folder after all implementation of notebook of making neural network,I faced an issue as shown in screenshot. I am not able to download as shown in recording session. Plz help. Thanks
Would request you to pull the latest notebooks from our repository, also, could you please check in which part of the code are you unable to create the checkpoint.
Hi I have done git pull but still showing same error. And code is doing well, there is no issue of creating checkpoint in code.What should i do now? Thanks
Please note that you would not be able to view the saved checkpoints since they are a binary file which cannot be opened in Jupyter notebook. However, the file is saved properly because of which you are able to use the same for prediction.
Hi After making neural network, it has been told in recording that we can use this as API on server side. Can you please tell me if we want to use on mobile on computer how can we use this on server side? Thanks
what is difference between name_scope and variable scope? do we need to initialize and run all the variables inside name_scope and variable_scope in a session or by default they get initialized? what is actually happening when we initialize the variables and why do we need to run them?
Scoping is a lightweight encapsulation mechanism for TensorFlow to share variables. The difference between the tf.name_scope() and tf.variable_scope(0 is very subtle.
- tf.variable_scope() seems to be the preferred mechanism for variable sharing. - name_scope() is more used for name spacing operators, tf.get_variable() ignores name_scopes() unless the with operator is used. Hope that answers your queries.
I am havinf python 2.7 in console. I would like to switch to 3.6 but when I am trying activate python36, it is displaying command not found. I followed steps mentioned in video, but I couldn't make it.
Please login to comment
67 Comments
Hi Team,
I could not understood this part of code:
1. with tf.name_scope("eval"):
2. correct = tf.nn.in_top_k(logits, y, 1)
3. accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
Sometime, Sandeep says logits are our categories which is 0-9 and sometime he says it is calaculated field. Please explain what is happening in line number 2. I have seen the whole lecture many times where he explains this but I am not able to get it. May be some example can help me to understand it better. I am not able to visualize what is logits inside the argument.
Regards,
Birendra Singh
Upvote ShareHi,
In tf.nn.in_top_k, logits are the final predictions, y is the actual class, and 1 means the k value. Feel free to check https://www.tensorflow.org/api_docs/python/tf/math/in_top_k and https://stackoverflow.com/questions/57149583/tf-nn-in-top-klogits-y-1-out-of-range-error-but-equal-actually
Thanks.
Upvote ShareHi Vagdevi,
I read the link provided by you, but still i am not clear. Can you explain me the output of below sample codes. I am not getting how these output as True and False.
1) out = tf.nn.in_top_k([[0.8,0.9,0.3], [0.1,0.6,0.4], [0.6,0.9,0.65]], [1,2,3], 1)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
print (sess.run(out))
Output : [ True False False]
2) out = tf.nn.in_top_k([[0.8,0.9,0.3], [0.1,0.6,0.4], [0.6,0.9,0.65]], [1,1,1], 1)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
print (sess.run(out))
Output: [ True True True]
3) out = tf.nn.in_top_k([[0.8,0.9,0.3], [0.1,0.6,0.4], [0.6,0.9,0.65]], [2,2,2], 2)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
print (sess.run(out))
Output: [False True True]
Regards,
Birendra Singh
Upvote ShareHi,
tf.nn.in_top_k says whether the targets are in the top K predictions.
Thanks.
Upvote ShareHi Team,
Like we have saver = tf.train.Saver() to save the model and reuse the existing model, do we have such facility in ML libraray scikitlearn for different types of algorithm ? If yes, please let me know.
Regards,
Birendra Singh
Upvote ShareHi,
It is very much possible with pickle module. Feel free to go through https://scikit-learn.org/stable/modules/model_persistence.html.
Thanks.
1 Upvote ShareHi Team,
By steps in DNN, do we mean number of iterations of back propagation and weights adjustment. So, in below line :
dnn_clf.fit(X_train, y_train, batch_size=50, steps=30000), we mean batch size of 50 rows will undergo weights adjustment until 30000 iterations are over ? And same operation is going to repeat until all batch elements are covered.
Regards,
Birendra Singh
Upvote ShareHi,
In an epoch, each training sample gets a chance to update the weights of the classification algorithm. So an epoch is completed after all the batches update the function. Hence, here in the example, if the training set has 100 samples in total, and the batch size is 50, then one epoch is completed after 2 batches of the training data undergo forward propagaation and back propagation. There will be 30000 such epochs in this example.
Thanks.
1 Upvote ShareThanks!
Upvote Share1. Why don’t we consider negative values for RELU AFunc?
2. how can we say that Relu is non-linear activation function because gradient is a constant (for x>0 like linear function) which does not depend upon the input value x at all. This implies that the weights and biases will be updated during the backpropagation process but the updating factor would be the same. In this scenario, the neural network will not really improve the error since the gradient is the same for every iteration. The network will not be able to train well and capture the complex patterns from the data.
Upvote ShareHi,
For the first question, I am not sure what you mean by "negative values for ReLU function", I am assuming you meant why cannot ReLU output negative values. Here's a very good explanation of the same:
https://stats.stackexchange.com/questions/362588/how-can-a-network-with-only-relu-nodes-output-negative-values
Linear means to move ahead in a straight line. Please check the entire graph of ReLU and not a portion of the same because the entire graph constitutes the ReLU activation function and not just a part of it, you will notice that the output of ReLU is not linear, or in other words, not a straight line.
Thanks.
Upvote ShareThank you very much for asking me to do activity in prvious query. I can relate your answer with that RELU is adding non-linearity. this is not as same as linear activation func.
1 Upvote ShareHi,
What I have learned over years is that self-learning is the best form of learning.
Thanks.
Upvote Shareat slide 109, if i don't use relu activation then can i define neuron layer to design neural network?
Upvote ShareHi,
Instead of me answering this for you, I would encourage you to try this on our lab and see the results. Don't forget to share your findings.
Thanks.
Upvote ShareIt works without activation function but act as linear regression means non-linearity will not get added without activation function.
even same result in case of MNIST and accuracy get reduced without Act Func.
1 Upvote ShareHow much course I have to complete to get certificate?? Is Optional Projects need to complete to get the certificate?
Upvote ShareHi Manish
You need to complete 2 projects in Big data , 3 projects in machine leanring, 2 in deep learning and one in python.
Also note that optional projects are not counted.
Regards
Upvote ShareHi Malavika,
I have completed my course 72% and all mandetory projects and also applied for the certificate. Please tell me how much time it will take to get my certificate. Please let me know asap.
Thanks,
Manish Kumar
Upvote ShareHi Manish
I have replied to your mail, Do check it
Regards
Upvote ShareAny updates regarding my certificate???
Upvote ShareHi Manish
We have shared the certificates.
Do check your mail.
Upvote SharePlease ignore the 2 questions below as I was finally abe to execute those.
Just one query though!!
Why executing this code " !ls -l model_ckps/ " gives total as 2248 (pls refer to the screenshot)? While the lecture file shows the value as 600? Please comment. Thanks!
Hi,
There is a possiblity that other files have been saved in the same folder.
Thanks.
Upvote ShareIssue executing the 2nd one. Please help with the resolution.
Thanks
Upvote ShareHello,
I am facing error executing the 2 codes. Details of 1st one below and 2nd to follow in next comment.
Please help trouble shoot.
Thanks
Hello sir,
It may found some silly question but i really wanted to ask that logistic regression algorithm functionali is quit similar to sigmoid activation function? are they similar? if not then what is difference between them? Functionality wise i am finding them very similar.
1 Upvote ShareHi,
This is a very good question.
Logistic Regression actually uses the Sigmoid Function as a cost function. You can read more about it here:
https://www.analyticssteps.com/blogs/introduction-logistic-regression-sigmoid-function-code-explanation
https://towardsdatascience.com/introduction-to-logistic-regression-66248243c148
Thanks.
Upvote ShareHello Sir,
What is the difference between making Neural Network using Tensorflow coding and using Tensorflow API like Keras? Which will be more usefull?. As I am very new to coding so thinking of having some easy way to code the neural network? Please suggest.
1 Upvote ShareHi,
Good question!
It does not matter which library you use as long as you know the concept. It's like using a different programming language to solve the same task.
Thanks.
Upvote ShareThank you sir....
Upvote ShareOn what factor the no of hidden units depends?
Upvote ShareHi,
There are no hard and fast rules, however, usually it should 2/3 the size of the input layer, plus the size of the output layer.
Thanks.
Upvote ShareHello sir, i wanted to Know what isthe difference between n_inputs and n_neurons, i mean X is the input matrix - so is how are they related numerically with it? and on what basis is the bias b value computed?
Upvote Sharethis
Hi,
n_input here is the input dataset. n_neurons is the number of neurons for that layer. In this function, b is a vector of zeros.
Thanks.
Upvote ShareThank you for responding sir. However i would like to confirm one more thing, is n_input total no. of inputs provided initially?
like 2 in this case?
Hi,
Instead of me telling you, let me help you find it yourself. Right after the n_inputs statement, add 2 lines of code:
print(n_inputs)
print(n_neurons)
This will help you find out what these 2 variables contitutes.
Thanks.
1 Upvote ShareHi, How is the output of different training batches consolidated to get final weights?
Upvote ShareHi,
That is taken care by the Tensorflow library.
Thanks.
Upvote ShareHi, I m running this code in tfv2 but showing this error please help me to resolve
import tensorflow as tf
# from tensorflow.python.feature_column import dense_features
config = tf.estimator.RunConfig(tf_random_seed=42) # not shown in the config
feature_cols = tf.contrib.learn.infer_real_valued_columns_from_input(X_train)
Upvote Sharednn_clf = tf.contrib.learn.DNNClassifier(hidden_units=[200, 200, 200], n_classes=10,
feature_columns=feature_cols, config=config)
dnn_clf = tf.contrib.learn.SKCompat(dnn_clf) # if TensorFlow >= 1.1
dnn_clf.fit(X_train, y_train, batch_size=50, steps=40000)
other replacement for v2 i have found but please tell be replacement for
.contrib.learn
contrib as it is not present in v2 now
Upvote Shareplease reply sir as i am unable to proceed further in the course
Upvote ShareHi,
As suggested, please create a separate virtual environment for TF1 and TF2 and then proceed with this code. Check your mail for further instructions.
Thanks.
Upvote Sharesir i have uninstalled the version 1.14 then using v2 alone its showing these errors there are few changes in api's for version 2 as
replacement for v2 i have found but please tell be replacement for
.contrib.learn
contrib as it is not present in v2 now
so please help me to find the replacement
Upvote ShareThis comment has been removed.
Hello. The codes provided in this course, unfortunately, won't run in Tensorflow version 2.x. I use Google Colab to run my code. You can use Tensorflow 1.x in Colab You just have to run this command on the cell prior to running any underlying code :
%tensorflow_version 1.x
Ps: by default it will run the tf 2.x version so u need to run the above command otherwise it will show error.
This will solve your problem. All the best.
Upvote ShareHi..
The ppt does not contains all slides explained by you. Please update.
Thanks and Regards..
Upvote ShareHi,
Could you please help me by telling me which slides from which part of the view are you referring to?
Thanks.
Upvote ShareHello
I was talking about the ppt of Artificial Neural Network. It contains only 77 slides in which the explained part is not there.
Please do the needful
Thanks n Regards...
Upvote ShareHi,
Are you referring to the DNN slides? If yes, then you can find them in the DNN topic.
Thanks.
Upvote ShareHi
Upvote Sharethere is also a problem in showing graph at end of notebook.
Hi,
Please note that that code is deprecated, if you check the following URL does not work anymore:
https://tensorboard.appspot...
We would be removing it from our notebooks soon.
Thanks.
-- Rajtilak Bhattacharjee
Upvote ShareHi
Upvote ShareWhen I opened model_ckps in deep learning folder after all implementation of notebook of making neural network,I faced an issue as shown in screenshot. I am not able to download as shown in recording session. Plz help.
Thanks
Hi,
Could you please upload the screenshot that you referred to in your comment.
Thanks.
-- Rajtilak Bhattacharjee
Upvote ShareHi
I m sorry, i just forgot to attach. Plz find
Hi,
Would request you to pull the latest notebooks from our repository, also, could you please check in which part of the code are you unable to create the checkpoint.
Thanks.
-- Rajtilak Bhattacharjee
Upvote ShareHi
Upvote ShareI have done git pull but still showing same error. And code is doing well, there is no issue of creating checkpoint in code.What should i do now?
Thanks
Hi,
Please note that you would not be able to view the saved checkpoints since they are a binary file which cannot be opened in Jupyter notebook. However, the file is saved properly because of which you are able to use the same for prediction.
Thanks.
-- Rajtilak Bhattacharjee
Upvote ShareHi
Upvote ShareAfter making neural network, it has been told in recording that we can use this as API on server side. Can you please tell me if we want to use on mobile on computer how can we use this on server side?
Thanks
Hi,
For mobile and IoT devices you can use TensorFlow Lite. Find more about it from the link below link:
https://www.tensorflow.org/...
Thanks.
-- Rajtilak Bhattacharjee
Upvote ShareHi
Upvote ShareUsing tf.layers.dense, if we can make neuron layers then there is no need of making neuron function?
Hi,
Yes, the tf.layer.dense adds a single layer to your model.
Thanks.
-- Rajtilak Bhattacharjee
Upvote Sharewhat is difference between name_scope and variable scope?
Upvote Sharedo we need to initialize and run all the variables inside name_scope and variable_scope in a session or by default they get initialized?
what is actually happening when we initialize the variables and why do we need to run them?
Hi,
Scoping is a lightweight encapsulation mechanism for TensorFlow to share variables. The difference between the tf.name_scope() and tf.variable_scope(0 is very subtle.
- tf.variable_scope() seems to be the preferred mechanism for variable sharing.
- name_scope() is more used for name spacing operators, tf.get_variable() ignores name_scopes() unless the with operator is used.
Hope that answers your queries.
Thanks.
-- Rajtilak Bhattacharjee
Upvote ShareI am havinf python 2.7 in console.
I would like to switch to 3.6
but when I am trying activate python36, it is displaying command not found.
I followed steps mentioned in video, but I couldn't make it.
Please help me to activate py36
Upvote Share