@@ -7,9 +7,9 @@ During training or serving, we provide the evaluation function to measure the mo
### Evaluator Design
Currently, every operation is expressed in the graph. we divide the evaluator process into three steps.
1. Initialize the metric state necessary and add it into the block.
1. Initialize the metric state and add it into the block.
2. Calculate the statistic of the metric state in every mini-batch. The single operator is only responsible for calculating necessary statistics for one mini-batch. For example, accuracy operator only calculate a minibatch data if run once.\
2. Calculate the statistic of the metric state in every mini-batch. The single operator is only responsible for calculating necessary statistics for one mini-batch. For example, accuracy operator only calculate a minibatch data if run once.
3. Merge the mini-batch statistics to form the evaluation result for multiple mini-batches. When it comes to distributed training/Multi-GPU training, aggregate the value from different devices.
...
...
@@ -20,38 +20,30 @@ This design is shown in python API. There would be an abstract python interface
```python
classEvaluator(object):
"""
Evalutor Base class.
Evaluator Base class.
"""
def__init__(self):
"""
create metric states and append to block
"""
pass
def_clear_state(self):
"""
clear metric states at the begin of each pass
"""
pass
Different evaluator may has different metric states. E.g, Accuracy need two variables, total and right sample counts.
Auc need four variables, `true_positives`,
`true_negatives`, `false_positives` and `false_negatives`. So every evaluator should create its needed variables and append the related mini-batch operator to main_program
def_append_evalutor_op(self):
"""
add mini-batch caculate operators to block
add increment operator to accumulate the metric state
The initialization of Evaluator should be responsible for:
create metric states and append to the main_program
add mini-batch evaluator caculate operators to the main_program
add increment operator to accumulate the metric states
"""
pass
def_merge(self):
defclear(self):
"""
Merge the mini-batch statistics to form the evaluation result for multiple mini-batches.
clear metric states at the begin of each pass/user specified batch
"""
pass
returninit_program
defevaluate(self):
"""
only one exported interface
user calculate the result
Merge the mini-batch statistics to form the evaluation result for multiple mini-batches.