Horrible tensorboard graph:

[MadcowD]

Hi,

I'm trying to scope certain components of my tensorflow fold model, and I after rending it in tensorboard, it seems that its ignoring the scope in which I construct different blocks. See this http://i.imgur.com/Vtf9OhP.png

The associated source code which generates this graph is:

with tf.variable_scope("context_free_discriminator"):
    with tf.variable_scope("embedding"):
        pre_process =  (td.InputTransform(lambda tokens: [tokeToVec(s, word2vec, indim) for s in tokens]) 
                        >> td.Map(td.Vector(indim)))

        word_cell = td.ScopedLayer(tf.contrib.rnn.BasicLSTMCell(num_units=300), 'word_cell')
        stacked_cell =  td.ScopedLayer(tf.contrib.rnn.BasicLSTMCell(num_units=100), 'stacked_cell')

        # word lstm convers word sequence into an lstm.
        word_lstm = (pre_process >> td.RNN(word_cell))
        hierarchical_lstm = (word_lstm 
                             >> td.GetItem(0) #hidden states
                             >> td.RNN(stacked_cell))
        embedding =  (hierarchical_lstm  
                                          >> td.GetItem(1) # Final state (out, hidden)
                                          >> td.GetItem(0) # Final output state.
                                          >>  td.FC(outdim))

    with tf.variable_scope("model"):
        logit_output_block = (embedding_block 
                  >> td.FC(428)
                  >> td.FC(328)
                  >> td.FC(num_mocs, activation=None))
        compiler = td.Compiler.create(logit_output_block)
        logit_tensor = compiler.output_tensors

    with tf.variable_scope("training"):
        _desired = tf.placeholder(tf.int32)
        _desired_onehot = tf.one_hot(_desired, num_mocs)
        loss = cross_entropy = tf.reduce_mean(
            tf.nn.softmax_cross_entropy_with_logits(labels=_desired, logits=logit_tensor))

        train_op = tf.train.AdamOptimizer(learning_rate).minimize(cross_entropy)

[fabioasdias]

My understanding is that since td is build over tf, not everything will be accessible as would be the case with "pure" tensorflow, meaning specifically that scopes are probably used by td to deal with the dynamic part of building the graph. I'm not entirely sure that the tf.variable_scope commands are compatible with the whole td.compiler thing. If they were, you wouldn't need a scopedlayer, IMHO.

A good example of that is the results of the 'word_cell' scope, that is "internally" translated to several blocks.

Of course, I'm a random internet dude, not a contributor to the project, so everything I said might be horribly wrong.

[delesley]

Why do you say it's ignoring the scope? The graph is showing the names of various blocks, not the fully qualified names of the variables used by those blocks. Each Layer you create (e.g. FC or ScopedLayer) will create its own variable scope, nested in whatever the current variable scope is, to hold the variables for that layer.

To be honest, what confuses me about the picture is that you have multiple while loops, and multiple instantiations of your LSTM cells; I don't know what's causing that.

If you want to see whether your variables have been created appropriately, call [v.name for v in tf.global_variables()]; that should show you the fully qualified names of all variables.

On Fri, Feb 24, 2017 at 2:12 PM, William Guss notifications@github.com wrote:

Hi,

I'm trying to scope certain components of my tensorflow fold model, and I after rending it in tensorboard, it seems that its ignoring the scope in which I construct different blocks. See this http://i.imgur.com/Vtf9OhP.png

The associated source code which generates this graph is:

with tf.variable_scope("context_free_discriminator"): with tf.variable_scope("embedding"): pre_process = (td.InputTransform(lambda tokens: [tokeToVec(s, word2vec, indim) for s in tokens])

td.Map(td.Vector(indim)))

  word_cell = td.ScopedLayer(tf.contrib.rnn.BasicLSTMCell(num_units=300), 'word_cell')
  stacked_cell =  td.ScopedLayer(tf.contrib.rnn.BasicLSTMCell(num_units=100), 'stacked_cell')

  # word lstm convers word sequence into an lstm.
  word_lstm = (pre_process >> td.RNN(word_cell))
  hierarchical_lstm = (word_lstm
                       >> td.GetItem(0) #hidden states
                       >> td.RNN(stacked_cell))
  embedding =  (hierarchical_lstm
                                    >> td.GetItem(1) # Final state (out, hidden)
                                    >> td.GetItem(0) # Final output state.
                                    >>  td.FC(outdim))

with tf.variable_scope("model"): logit_output_block = (embedding_block

td.FC(428) td.FC(328) td.FC(num_mocs, activation=None)) compiler = td.Compiler.create(logit_output_block) logit_tensor = compiler.output_tensors

with tf.variable_scope("training"): _desired = tf.placeholder(tf.int32) _desired_onehot = tf.one_hot(_desired, num_mocs) loss = cross_entropy = tf.reduce_mean( tf.nn.softmax_cross_entropy_with_logits(labels=_desired, logits=logit_tensor))

  train_op = tf.train.AdamOptimizer(learning_rate).minimize(cross_entropy)

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tensorflow/fold/issues/19, or mute the thread https://github.com/notifications/unsubscribe-auth/AGGbTWb9XH1cEmFARsGoyPrKWN7afbx5ks5rf1XZgaJpZM4MLvGB .

-- DeLesley Hutchins | Software Engineer | delesley@google.com | 505-206-0315

[moshelooks]

Yeah, TensorBoard graph view is not very helpful with Fold, unfortunately. Regarding scoping, if you know what you're doing you can pass in a scope as an argument to td.ScopedLayer and it will get used exactly as provided. If you pass in a string we go through some gymnastics to ensure that every ScopedLayer really does correspond to a unique scope. We do this to avoid some breakage and gotchas in the TF variable scoping mechanism. What Delesley said is a very good suggestion for making sure the right variables are being created. If you are observing undersharing (e.g. a single scoped layer that encapsulates variables that are being created more than once) that is bug, please do report it. But the view of the graph in tensorboard is unfortunately not sufficient to figure out what's going on.

Regarding the while loop, I'm confused; Fold certainly only creates one while loop per compiler; maybe you ran Block.eval() somewhere? (eval() is only for messing around at the repl and unit tests; it shouldn't be used in "real" models)