visualizing activations of layers in TensorBoard

user_5759234 0
Viewed 2469

I have a layer, layer3, that is of type:

Tensor("vgg_16/conv3/conv3_3/Relu:0", shape=(1, 500, 700, 120), dtype=float32, device=/device:GPU:0)

I'd like to visualize the activations of this layer. How can I process layer3 to do that? What would I have to add to tf.summary.histogram() to visualize this?

activation-function tensorflow neural-network
1 Answers

First, define your summary over the required layer:

tf.summary.histogram("layer3_hist_summary", layer3)

Next define the summary writer which will be used to write your summaries to disk:

LOGDIR = 'path/to/logsdir' # define your required summary output folder
summary_file_writer = tf.summary.FileWriter(logdir=LOGDIR)

Assuming you may have multiple summaries, merge them to a single op:

summary_op = tf.summary.merge_all()

Now, in your training loop, write the summary result:

for i in range(NUM_ITR):
    _, summary_res = sess.run([train_op, summary_op])
    summary_file_writer.add_summary(summary_res, global_step=i)

To view these summaries, load tensorflow with logsdir=LOGDIR.

user_39322120
Related