PHP Search Engine

Hw4 Test CNN .

Professor Pollett,

The homework description say "'.....When you do testing with your trained model, use a different data set then what you used to train and validate with...'' #Do you mean we have to create a new python file, let's say conv_test.py to experiment the test? #Or you want to use the same file conv_train.py to run the test? If this is the case, I think we need one more parameter to indicate for the test case.

(Edited: 2017-11-15)

Professor Pollett, The homework description say "'.....When you do testing with your trained model, use a different data set then what you used to train and validate with...'' #Do you mean we have to create a new python file, let's say conv_test.py to experiment the test? #Or you want to use the same file conv_train.py to run the test? If this is the case, I think we need one more parameter to indicate for the test case.

-- Hw4 Test CNN

Hi,

We are currently doing this:

test network_description.txt 0.1 10000 Q model.ckpt test_data

where, 'test' (instead of cross/cross-l1/cross-l2) is given as a parameter and the folder test_data is used and the model file that was saved using model.ckpt is restored to perform the accuracy calculation. Instead of a separate file, we created a function that tested the network.

Please let us know if this is fine.

Hi, We are currently doing this: test network_description.txt 0.1 10000 Q model.ckpt test_data where, 'test' (instead of cross/cross-l1/cross-l2) is given as a parameter and the folder test_data is used and the model file that was saved using model.ckpt is restored to perform the accuracy calculation. Instead of a separate file, we created a function that tested the network. Please let us know if this is fine.

-- Hw4 Test CNN

Can we use any other library apart from tensorflow,numpy and matplotlib ? such as sklearn

-- Hw4 Test CNN

I've updated the description a little to handle thinh's question. I think what I wrote is consistent with what Vasudha's group is doing. I am not sure how sklearn will help with this homework. Let's just stick with tensorflow,numpy and matplotlib. If you want to use sklearn on HW5, you can.

Best,

Chris

I've updated the description a little to handle thinh's question. I think what I wrote is consistent with what Vasudha's group is doing. I am not sure how sklearn will help with this homework. Let's just stick with tensorflow,numpy and matplotlib. If you want to use sklearn on HW5, you can. Best, Chris

-- Hw4 Test CNN

Hi,

I saw this in the description:

"or no training just testing (epsilon max_updates class_letter are then ignored)."

Are we not supposed to consider the class_letter while testing ? We do not store information about the class being trained for anywhere. So, with this being a binary classifier project, how will we know we are supposed to be testing for a particular class ? In other words, how will we form the y vector for the test data if we do not know what to test for ?

Please let us know.

Hi, I saw this in the description: "or no training just testing (epsilon max_updates class_letter are then ignored)." Are we not supposed to consider the class_letter while testing ? We do not store information about the class being trained for anywhere. So, with this being a binary classifier project, how will we know we are supposed to be testing for a particular class ? In other words, how will we form the y vector for the test data if we do not know what to test for ? Please let us know.

-- Hw4 Test CNN

Sorry, I was a little too zealous in highlighting what was to be ignored. You shouldn't ignore the class letter.

Best, Chris

Sorry, I was a little too zealous in highlighting what was to be ignored. You shouldn't ignore the class letter. Best, Chris

-- Hw4 Test CNN

Because the hw4 says "...the last two layers consist of a dense layer of 64 units all of whose ouputs connect to a single sigmoid perceptron...", and hw also says "''Fix a network approximately like LeNet-5...''", but LeNet-5 use softmax for output layer

#So we still uses sigmoid activation for output layer? #Down sampling layers-layer 2 and 5 in LeNet-use softmax, but we will use relu because the hw says "For the purposes of this homework all other units use relu activations". Is it right?

(Edited: 2017-11-19)

Because the hw4 says "...the last two layers consist of a dense layer of 64 units all of whose ouputs connect to a single sigmoid perceptron...", and hw also says "''Fix a network approximately like LeNet-5...''", but LeNet-5 use softmax for output layer #So we still uses sigmoid activation for output layer? #Down sampling layers-layer 2 and 5 in LeNet-use softmax, but we will use relu because the hw says "For the purposes of this homework all other units use relu activations". Is it right?

-- Hw4 Test CNN

(1) we didn't use softmax in HW2, so I wanted to mimic the output from HW2, that's why I wrote it like the above. You can use a sigmoid if you like.

(2) Downsampling should still be done using a max pool layer.

(1) we didn't use softmax in HW2, so I wanted to mimic the output from HW2, that's why I wrote it like the above. You can use a sigmoid if you like. (2) Downsampling should still be done using a max pool layer.