Final convolutional layer tensor for activation map with Grad-CAM

[hnguyentt]

Hello,

Thank you for sharing your work!
I am trying to make the activation map to see the important features from your two trained models (COVIDNet-CXR Small and COVIDNet-CXR Large).
To do that I would like to know the name of the final convolutional layer tensor. I have checked your document train_eval_inference.md but have found that tensor name.

I listed all tensor names in your models by this code:

tensors = [t.name for op in tf.get_default_graph().get_operations() for t in op.values()]
for t in tensors:
    print(t)

And this is a subset of the tensor names:

...
conv5_block3_preact_relu/Relu:0
conv5_block3_1_conv/convolution:0
conv5_block3_1_bn/FusedBatchNorm_1:0
conv5_block3_1_bn/FusedBatchNorm_1:1
conv5_block3_1_bn/FusedBatchNorm_1:2
conv5_block3_1_bn/FusedBatchNorm_1:3
conv5_block3_1_bn/FusedBatchNorm_1:4
conv5_block3_1_relu/Relu:0
conv5_block3_2_pad/Pad/paddings:0
conv5_block3_2_pad/Pad:0
conv5_block3_2_conv/convolution:0
conv5_block3_2_bn/FusedBatchNorm_1:0
conv5_block3_2_bn/FusedBatchNorm_1:1
conv5_block3_2_bn/FusedBatchNorm_1:2
conv5_block3_2_bn/FusedBatchNorm_1:3
conv5_block3_2_bn/FusedBatchNorm_1:4
conv5_block3_2_relu/Relu:0
conv5_block3_3_conv/convolution:0
conv5_block3_3_conv/BiasAdd:0
conv5_block3_out/add:0
post_bn/FusedBatchNorm_1:0
post_bn/FusedBatchNorm_1:1
post_bn/FusedBatchNorm_1:2
post_bn/FusedBatchNorm_1:3
post_bn/FusedBatchNorm_1:4
post_relu/Relu:0
flatten_1/Shape:0
flatten_1/strided_slice/stack:0
flatten_1/strided_slice/stack_1:0
flatten_1/strided_slice/stack_2:0
flatten_1/strided_slice:0
flatten_1/Const:0
flatten_1/Prod:0
flatten_1/stack/0:0
flatten_1/stack:0
flatten_1/Reshape:0
dense_1/MatMul:0
dense_1/BiasAdd:0
dense_1/Relu:0
dense_2/MatMul:0
dense_2/BiasAdd:0
dense_2/Relu:0
dense_3/MatMul:0
dense_3/BiasAdd:0
dense_3/Softmax:0
loss/mul/x:0
loss/dense_3_loss/Sum/reduction_indices:0
loss/dense_3_loss/Sum:0
...

Based on that, I guess the final convolutional layer tensor is conv5_block3_out/add:0 and make the activation map based on that.

To confirm what I have done, my question: Is the final convolutional layer tensor is actually conv5_block3_out/add:0?

Thank you for your time.

[lindawangg]

Should be either conv5_block3_3_conv/BiasAdd:0 or conv5_block3_out/add:0. Can print out the shapes to verify which one.

[hnguyentt]

Both of them have shape (<num of samples>, 7, 7, 2048). They match the shape with the architecture in your document: https://github.com/lindawangg/COVID-Net/blob/master/assets/COVIDNet_CXR.pdf

[borjaMinano]

@nguyenhoa93 , did you finally managed to run Grad-Cam?
I have tried to implement it, but it seems GradientTape does not register operations runs with Session.run

with tf.GradientTape() as tape:
	tape.watch(image_tensor)
	pred = sess.run(pred_tensor, feed_dict={image_tensor: np.expand_dims(x, axis=0)})
	convOutputs = sess.run(conv_op, feed_dict={image_tensor: np.expand_dims(x, axis=0)})
	loss = pred[:, np.argmax(pred[0])]

If I print the watched variables, there is no variables watched, so grads = tape.gradient(loss, convOutputs) returns none. However loss and convOutputs have data.
What confuses me is that those variables are numpy arrays instead of tensors.

I am pretty new to tensorflow and a bit lost yet. I'll appreciate any help.

[hnguyentt]

@borjaMinano tf.GradientTape() is applicable for Tensorflow 2. The model was trained with TensorFlow v 1.x`. Did you retrain the model with TF 2.?

Edit: After reading @haydengunraj 's comment, I read again the document of tf.GradientTap() and realized that it's available in Eager Execution - TF 1.15 as well. Like @haydengunraj said, we need to sess.run() to get the values.

First, define last_conv_tensor and grads_tensor:
lass_conv_tensor = graph.get_tensor_by_name(<last_conv_layer_tensor_name>)

with tf.GradientTape() as tape:
	tape.watch(image_tensor)
	one_hot = tf.sparse_to_dense(classIdx, [len(self.classes)], 1.0)
	signal = tf.multiply(graph.get_tensor_by_name(<output_tensor_name>),one_hot)
	loss_tensor = tf.reduce_mean(signal)
	grads_tensor = tape.gradient(loss_tensor, last_conv_tensor)

last_conv, grads = sess.run([last_conv_tensor, grad_tensors], feed_dict={image_tensor: np.expand_dims(x, axis=0)})

I did the activation map already but want to make sure that I select the last convolutional layer correctly. You can see my code here: https://gist.github.com/nguyenhoa93/d49564875234f6722ff89e65db34a00b
To run the code you should create a json file with the keys:

{
  "weightspath": ,
  "metaname": ,
  "ckptname": ,
  "input_tensor": "input_1:0",
  "output_tensor": "dense_3/Softmax:0",
  "final_conv_tensor": "conv5_block3_out/add:0",
  "input_size":
  "top_percent": 
}

I took conv5_block3_out/add:0 as final_conv_tensor.

This is the GradCAM result of COVID-19 when the model predict the image as COVID-19.

[haydengunraj]

@borjaMinano I don't have much experience with Grad-CAM, but the main issue in the code snippet you posted above is more of a general Tensorflow thing. When you use session.run(), you are running the operation in the graph, and Tensorflow returns the result as a numpy array. That numpy array is completely separate from the Tensorflow graph, so calling tape.gradient(loss, convOutputs) does nothing. You need to pass Tensorflow Tensors to tape.gradient(), which should return a gradient Tensor that you can then compute using session.run().

TL;DR, you need to do something like:

with tf.GradientTape() as tape:
    tape.watch(image_tensor)
    loss_tensor = ...define your loss tensor here
    grad_tensor = tape.gradient(loss_tensor, conv_op)

grads = sess.run(grad_tensor, feed_dict={image_tensor: np.expand_dims(x, axis=0)})

[borjaMinano]

Thank you very much for your help.
I adapted @nguyenhoa93 code and get it finally working.

[hnguyentt]

My pleasure, @borjaMinano
Happy to hear that it worked for you.

[ankitdata]

hi @nguyenhoa93 i was referring your code

on line no. 144 what is 'im' ?
im_name = im.split("/")[-1])

I got below error :
im_name = (im.split("/")[-1])
NameError: name 'im' is not defined

thanks

[hnguyentt]

@ankitdata it should be args.impath.split(“/“)[-1]
I editted here: https://gist.github.com/nguyenhoa93/d49564875234f6722ff89e65db34a00b#file-covid-net-gradcam-py-L144-L145

[ankitdata]

Thanks @nguyenhoa93, it worked for me.

Any idea what will be the 'final convolutional layer tensor' for Model: COVIDNet-CXR3-A
with following inputs

{
"input_tensor": "input_1:0",
  "output_tensor": "**norm_dense_1/Softmax:0**",
  "final_conv_tensor": " ? ",
}  

I checked tensors with below code

tensors = [t.name for op in tf.get_default_graph().get_operations() for t in op.values()]
for t in tensors:
    print(t)

which tensor will work from the list ?
@lindawangg @nguyenhoa93

Thanks

[hnguyentt]

@ankitdata
I am not sure because I didn't build the network. For COVIDNet-CXR3-A, I guess: conv_7b/convolution:0

[ankitdata]

@nguyenhoa93

Do you have any plan to build network for COVIDNet-CXR3-A ?
If you build, please share the heatmap here.

thanks .

[hnguyentt]

@nguyenhoa93

Do you have any plan to build network for COVIDNet-CXR3-A ?
If you build, please share the heatmap here.

thanks .

You can use the same code that I shared and change the name of final conv to conv_7b/convolution:0 . Input size (480, 480, 3).

The code will work for all TensorFlow models (TF 1.15).

[ankitdata]

Hi @nguyenhoa93

I tried with 3 models (small ,Large & COVIDNet-CXR3-A) heatmaps i shared below
in ' COVIDNet-CXR3-A' I used conv_7b/convolution:0 and Input size (480, 480, 3) but heatmap in not correct and prediction is Pneumonia.
for model COVIDNet-CXR3-A sensitivity is
Normal: 0.000, Pneumonia: 0.530, COVID-19: 0.470

for small model and large model sensitivity is higher for COVID-19 so prediction is COVID-19

what is your suggestion on this ?
let me know if you get more accuracy in heatmap for COVIDNet-CXR3-A.

Heatmaps
predictions for :
COVIDNet-CXR Small - COVID-19
COVIDNet-CXR Large - COVID-19
COVIDNet-CXR3-A - Pneumonia

input-image - assets/ex-covid.jpeg
thanks

[borjaMinano]

@ankitdata , I tried with the three new models (and also with Large model) and I see the same detection problem with model-A, but not with models B and C, even though the heatmaps are different from Large model. I think the Large model's heatmap is the most accurate, watching at the raw image.

Model B

Model C

[Electro1111]

@nguyenhoa93 Hello! I was wondering if you had done this with model: COVIDNet-CXR4-A

The naming convention for the convolutional layers seems quite different and I was wondering what to use for the final conv layer name. any help would be so much appreciated! Thank you.

[Electro1111]

using @nguyenhoa93 gradcam code these are the results I get using these input specifications:

using model COVIDNet-CXR4-A from:
https://github.com/lindawangg/COVID-Net/blob/master/docs/models.md

{
"weightspath": "path/to/models/COVIDNet-CXR4-A",
"metaname":"model.meta" ,
"ckptname":"model-18540" ,
"input_tensor": "input_1:0",
"output_tensor": "norm_dense_1/Softmax:0",
"final_conv_tensor": "conv2d_203/convolution:0",
"input_size": 480,
"top_percent": .08,
}

using this model

running on image: /data/test/0088be53-27f2-4c30-882b-a73a3a5c8c71.png

I get these results

Pneumonia

Normal

COVID-19

It could be that my final convolutional layer name is incorrect, can anyone verify? @lindawangg @haydengunraj

EDIT:

ok so I also tried it with:

{
"weightspath": "path/to/models/COVIDNet-CXR4-A",
"metaname":"model.meta" ,
"ckptname":"model-18540" ,
"input_tensor": "input_1:0",
"output_tensor": "norm_dense_1/Softmax:0",
"final_conv_tensor": "conv_7b/convolution:0",
"input_size": (480,480,3),
"top_percent": .08,
}

and actually got some result that looks like an actual heatmap, but I just wanted to confirm that "conv_7b/convolution:0" is the final conv layer and the correct layer for computing gradcams in model COVIDNet-CXR4-A? @lindawangg @haydengunraj @nguyenhoa93

[Electro1111]

@ankitdata , I tried with the three new models (and also with Large model) and I see the same detection problem with model-A, but not with models B and C, even though the heatmaps are different from Large model. I think the Large model's heatmap is the most accurate, watching at the raw image.

Model B

Model C

@borjaMinano what tensornames did you use for final_conv_tensor for the models CXR3 B and C because they seem to be named different from CXR3 A

[Electro1111]

Hi Ankit, Is there any way for you to confirm the layers I posted about and whether they are the final conv layers? I also asked a question in another issue about the training tensors for the small and large models. It would be extremely helpful. Best, Robbie Sadre

…

On Sun, Jan 3, 2021 at 10:50 AM Ankit Chilkalwar ***@***.***> wrote: Thanks for sharing this information. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#64 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AP7YI3TDWTJFZ57C7WJH7WDSYC37LANCNFSM4NBAZGCQ> .

[ankitdata]

Hi Ankit, Is there any way for you to confirm the layers I posted about and whether they are the final conv layers? I also asked a question in another issue about the training tensors for the small and large models. It would be extremely helpful. Best, Robbie Sadre
…
On Sun, Jan 3, 2021 at 10:50 AM Ankit Chilkalwar @.***> wrote: Thanks for sharing this information. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#64 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AP7YI3TDWTJFZ57C7WJH7WDSYC37LANCNFSM4NBAZGCQ .

For Large Model:
input tensorname = 'input_1:0', type=str
output tensorname = 'dense_3/Softmax:0'

final conv layer for heatmaps
conv5_block3_out/add:0

if you want list all tensors use below code.
tensors = [t.name for op in tf.get_default_graph().get_operations() for t in op.values()]
for t in tensors:
print(t)

I am not sure on sharing information about small model.

thanks.!