That won't work, I'm afraid: asctime() will be evaluated once when you build the graph in Python, and all log entries will include the same timestamp. The tf.Print() op uses the standard logging implementation in TensorFlow, which does not have a particularly rich feature set.

It might be useful to add a timestamp to all log entries, but that would be a fairly invasive change. Alternatively, one could add some ops (perhaps in tf.contrib) that returned/formatted the current time, so that you could build whatever log you wanted.

I'll mark this one contributions welcome.

@eriophora you can use a placeholder, no need to add code

 cti=tf.placeholder(tf.string,shape=(1,1))   
 z=tf.constant(10)                               
 y=tf.Print(z,[cti,z])                    
 with tf.Session() as sess:                                
         for i in range(10):
         time.sleep(1)                          
         ti=time.asctime()    
         xx=np.array([ti],ndmin=2)         
         y.eval(feed_dict={cti:xx}) 

blocked by #5682

Are we ready to proceed on this?

drpngx@ should be merged soon.

One second granularity is a bit too coarse for benchmarking, it would be nice to get microsecond granularity in LOG statements , ie, using NowMicros instead of NowSeconds. Maybe something like this in GenerateLogMessage?

char time_buffer[time_buffer_size];
uint64 micros = env_time->NowMicros()
time_t now = static_cast<time_t>(micros/1000000);
strftime(time_buffer, time_buffer_size, "%Y-%m-%d %H:%M:%S:%d", localtime(&now));
fprintf(stderr, "%s:%ld: %c %s:%d] %s\n", time_buffer, micros%1000000, "IWEF"[severity_], fname_, line_, str().c_str());