3
I4W^  ã               @   s¾   d dl mZ d dl mZ d dl mZ d dl mZ d dlmZ d dlmZ d dl	m
Z
 d dljjZd dlZd dlZd dlZejeƒZd	d
„ Zddd„Zdd„ Zddd„Zddd„Zddd„ZdS )é    )Úabsolute_import)Údivision)Úprint_function)Úunicode_literals)Ú
caffe2_pb2)Únamedtupledict)Ú	WorkspaceNc                s¬   t ƒ ‰ t|tƒr8xf|jƒ D ]\}}ˆ j||| jƒ qW n@t| jƒt|ƒksNt‚x(t	| j|ƒD ]\}}ˆ j||| jƒ q\W ˆ j
| ƒ | j}‡ fdd„|D ƒ}ˆ td|ƒ|Ž fS )Nc                s   g | ]}ˆ j |ƒ‘qS © )Ú	FetchBlob)Ú.0Úname)Úwsr	   ú:/tmp/pip-build-e66j97b2/torch/caffe2/python/onnx/helper.pyú
<listcomp>#   s    z$c2_native_run_op.<locals>.<listcomp>ÚOutputs)r   Ú
isinstanceÚdictÚitemsÚFeedBlobÚdevice_optionÚlenÚinputÚAssertionErrorÚzipZRunOperatorOnceÚoutputr   )Zop_defÚinputsÚkeyÚvalueÚoutput_namesÚoutput_valuesr	   )r   r   Úc2_native_run_op   s    

r    c       
         sj  t ƒ ‰ | rˆ j| ƒ t|tƒrFx$|jƒ D ]\}}ˆ j|||jƒ q(W nð‡ fdd„|jD ƒ}t|ƒt|ƒkr”xÊt	||ƒD ]\}}ˆ j|||jƒ qvW n¢t|ƒt|jƒkrìt
dt|ƒdt|jƒƒ t
d|ƒ t
dt|ƒd|ƒ t
dt|ƒd|ƒ t|ƒt|jƒk�st‚x0tt|ƒƒD ] }ˆ j|j| || |jƒ �qW ˆ j|ƒ |j}‡ fd	d„|D ƒ}	ˆ td
|ƒ|	Ž fS )Nc                s   g | ]}ˆ j |ƒs|‘qS r	   )ZHasBlob)r   Z
input_name)r   r	   r   r   0   s   z%c2_native_run_net.<locals>.<listcomp>z&c2_native_run_net assert. len(inputs)=z len(predict_net.external_input)=zdebug_arg: zpredict_net ú:zinputs c                s   g | ]}ˆ j |ƒ‘qS r	   )r
   )r   r   )r   r	   r   r   I   s    r   )r   Ú
RunNetOncer   r   r   r   r   Zexternal_inputr   r   ÚprintÚtyper   ÚrangeZexternal_outputr   )
Úinit_netÚpredict_netr   Z	debug_argr   r   ZuninitializedÚir   r   r	   )r   r   Úc2_native_run_net'   s4    






r)   c          	   C   s0   t jƒ }t| dƒ�}|j|jƒ ƒ W d Q R X |S )NÚrb)r   ZNetDefÚopenZParseFromStringÚread)ÚfileÚnetÚfr	   r	   r   Úload_caffe2_netM   s    r0   Fc             C   sT   t |dƒ�}|j| jƒ ƒ W d Q R X |rPt |d dƒ�}|jt| ƒƒ W d Q R X d S )NÚwbÚtxtÚw)r+   ÚwriteZSerializeToStringÚstr)r.   r-   Z
output_txtr/   r	   r	   r   Úsave_caffe2_netT   s
    r6   é   é
   Tc             C   s:   t ƒ }| r|j| ƒ |j|ƒ |j|j|||ƒ}~|d S )zw
        Run the benchmark net on the target model.
        Return the execution time per iteration (millisecond).
    r   )r   r"   Z	CreateNetZBenchmarkNetr   )r&   r'   Úwarmup_itersÚ
main_itersZlayer_detailsr   Úresultsr	   r	   r   Úbenchmark_caffe2_model\   s    

r<   c       
      C   s~   xt |ƒD ]}| |Ž  q
W d}x4t |ƒD ](}tjƒ }| |Ž  tjƒ }	||	| 7 }q(W tjdj|| d || ƒƒ |d | S )z‰
        Run the model several times, and measure the execution time.
        Return the execution time per iteration (millisecond).
    g        zRThe PyTorch model execution time per iter is {} milliseconds, {} iters per second.iè  )r%   ÚtimeÚlogÚinfoÚformat)
Úmodelr   Ztrainingr9   r:   ÚverboseZ_iZtotal_pytorch_timeÚtsÚter	   r	   r   Úbenchmark_pytorch_modelj   s    
rE   )N)F)r7   r8   T)Fr7   r8   F)Ú
__future__r   r   r   r   Zcaffe2.protor   Zonnx.backend.baser   Zcaffe2.python.onnx.workspacer   Z!caffe2.python._import_c_extensionÚpythonZ_import_c_extensionÚCÚioÚloggingr=   Ú	getLoggerÚ__name__r>   r    r)   r0   r6   r<   rE   r	   r	   r	   r   Ú<module>   s$   

&

 