(xn, batch_mean, batch_var,
moving_mean, moving_var, decay)
| 45 | |
| 46 | |
| 47 | def internal_update_bn_ema(xn, batch_mean, batch_var, |
| 48 | moving_mean, moving_var, decay): |
| 49 | update_op1 = moving_averages.assign_moving_average( |
| 50 | moving_mean, batch_mean, decay, zero_debias=False, |
| 51 | name='mean_ema_op') |
| 52 | update_op2 = moving_averages.assign_moving_average( |
| 53 | moving_var, batch_var, decay, zero_debias=False, |
| 54 | name='var_ema_op') |
| 55 | |
| 56 | # When sync_statistics is True, always enable internal_update. |
| 57 | # Otherwise the update ops (only executed on main tower) |
| 58 | # will hang when some BatchNorm layers are unused (https://github.com/tensorpack/tensorpack/issues/1078) |
| 59 | with tf.control_dependencies([update_op1, update_op2]): |
| 60 | return tf.identity(xn, name='output') |
| 61 | |
| 62 | |
| 63 | def get_sync_bn_mean_var(inputs, red_axis, sync_statistics): |
no outgoing calls
no test coverage detected
searching dependent graphs…