Please enable JavaScript.
Coggle requires JavaScript to display documents.
Tensorflow (Eigen (TensorExecutor (NullaryOp, UnaryOp, BinaryOp,…
Tensorflow
Eigen
TensorBase
TensorDevice
TensorConversionOp
TensorAssignOp
TensorExecutor
NullaryOp
UnaryOp
BinaryOp
TernaryOp
TensorEvaluator
Graph
Current Graph
TLS Graph
Global Graph
GraphDef creation
Using nodes registered in Graph in python
GraphDef to Graph
DirectSession->GraphConstructor
Partitioning
Graph partitions based on fetch list
Optimize
Constant Folding
Execute temp graphs to compute
Grappler
Op
OpDefLibrary
OpDef
Attributes
Input
Name
Output
params->OpDev
is_ref for Variable: provided by OpDef
Input names
Data Type
Input parameters
Python Op Creation
Set self as consumer to input tensors
Create output tensor for each output defined in OpDef
Name Scope
with tf.name_scope("scope_name"):
Variable
Constant Op
TensorProto Attribute
VariableV2 Op
Tensor buffer allocated/reused during assign op
set_output_ref+mutex -> mutable ref tensor
Assign Op
Inputs
Constant Output Tensor
mutable input Tensor ref of Variable output
Variable Scope
with tf.variable_scope("scope_name"):
Allocator
Allocator provided by Device on which operation is running
GSP->CPUAllocator
Allocator Registry
Allocator priority
Tensor
Shape
Static Shape
tensor.get_shape().as_list()
Provided by kernel registration SetShapeFn
Dynamic Shape
tf.shape()
Name
Syntax: src_node_name:index
Calculated during operation creation during recompute_node
Significance
Symbolic Handle for computation output
OpKernel
OpKernelContext
Provides context for Operation and allocates output tensors if needed
Quantized
APIs
Public
tf.quantize
tf.nn.quantized_conv2d
tf.nn.quantized_max_pool
tf.dequantize
Private
quantized_matmul
quantized_bias_add
requantize_range
requantize
Device
Device scope
with tf.device("/cpu:0"):
Session
run
Fetch
Fetch list
Tensors and Operations
FetchHandler
Unique elements from fetch list
Recreate output from fetched unique values
Graph->GraphDef