Please enable JavaScript.
Coggle requires JavaScript to display documents.
TENSORFLOW_200 (Dataflow graph (STRUCTURE (edges represents the inputs,…
TENSORFLOW_200
Dataflow graph
-
-
-
STRUCTURE
edges represents the inputs, output of the computation
-
-
BUILD GRAPH
-
declare variables, constants... will create an tf.Operation and a tf.Tensor and add them to a graph
-
-
NAMING OPERATION
by default, TF provides unique names for each TF.Operation
-
2 ways
-
remember "name" is the name of the TF.OPERATION while TF.TENSOR is name with "name:index" with index could be 0, 1, 2, ... to make it unique
2nd way is using tf,name_scope
if "name" is already exists in the graph, _1, _2, .. is appended to the name to make it unique
DEVICE PLACEMENT
-
-
ps server, worker server, etc...
tf.Graph
-
Graph Collections
-
such as trainable_collections, global_collections
-
-
by default, tf.Variables is in global_collections and trainable collection
XECUTE GRAPHS
tf.Session
-
-
several other options for placement, device, configuration
by default, run only operations of the default graphs
-
-
-