| 55 | |
| 56 | |
| 57 | def shard_features(features, device_list): |
| 58 | num_datashards = len(device_list) |
| 59 | sharded_features = {} |
| 60 | |
| 61 | with tf.device("/cpu:0"): |
| 62 | for k, v in six.iteritems(features): |
| 63 | v = tf.convert_to_tensor(v) |
| 64 | |
| 65 | if not v.shape.as_list(): |
| 66 | v = tf.expand_dims(v, axis=-1) |
| 67 | v = tf.tile(v, [num_datashards]) |
| 68 | |
| 69 | batch_size = tf.shape(v)[0] |
| 70 | size_splits = [] |
| 71 | |
| 72 | for i in range(num_datashards): |
| 73 | size_splits.append( |
| 74 | tf.cond(tf.greater(tf.mod(batch_size, num_datashards), i), |
| 75 | lambda: batch_size // num_datashards + 1, |
| 76 | lambda: batch_size // num_datashards) |
| 77 | ) |
| 78 | |
| 79 | sharded_features[k] = tf.split(v, size_splits, 0) |
| 80 | |
| 81 | datashard_to_features = [] |
| 82 | |
| 83 | for d in range(num_datashards): |
| 84 | feat = { |
| 85 | k: v[d] for k, v in six.iteritems(sharded_features) |
| 86 | } |
| 87 | datashard_to_features.append(feat) |
| 88 | |
| 89 | return datashard_to_features |
| 90 | |
| 91 | |
| 92 | def parallel_model(model_fn, features, devices): |