[SCRIPT] add script for converting ElectraForPretrain params to backb…

…one (#1527)

* add script for converting ElectraForPretrain params to backbone

Signed-off-by: Sheng Zha <[email protected]>

* fix pretraining utils for mxnp usage

Signed-off-by: Sheng Zha <[email protected]>

scripts/pretraining/convert_electra_pretrain_backbone.py

@@ -0,0 +1,49 @@
+"""Convert pre-trained model parameters from ElectraForPretrain to ElectraModel"""
+
+import os
+import argparse
+import mxnet as mx
+
+from pretraining_utils import get_electra_pretraining_model
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description=__doc__)
+    group = parser.add_mutually_exclusive_group(required=True)
+    group.add_argument('--model-name', type=str, default='google_electra_small',
+                       help='Name of the pretrained model.')
+    parser.add_argument('--params-file', type=str, required=True,
+                        help='Path to the pretrained parameter file.')
+    parser.add_argument('--out-file', type=str, default=None,
+                        help='Output file path.')
+    parser.add_argument('--generator_units_scale', type=float, default=None,
+                        help='The scale size of the generator units, same as used in pretraining.')
+    parser.add_argument('--generator_layers_scale', type=float, default=None,
+                        help='The scale size of the generator layer, same as used in pretraining.')
+
+    args = parser.parse_args()
+    return args
+
+
+def convert_params(model_name, generator_units_scale, generator_layers_scale,
+                   params_path, out_path):
+    _, _, pretrain_model = get_electra_pretraining_model(model_name, [mx.cpu()],
+                                                         generator_units_scale=generator_units_scale,
+                                                         generator_layers_scale=generator_layers_scale,
+                                                         params_path=params_path)
+    backbone_model = pretrain_model.disc_backbone
+    backbone_model.save_parameters(out_path)
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    out_path = args.out_file
+    if not out_path:
+        params_file = args.params_file
+        file_name_sep = os.path.basename(params_file).split(os.path.extsep)
+        file_name_sep.insert(-1, 'backbone')
+        out_path = os.path.join(
+            os.path.dirname(params_file),
+            os.path.extsep.join(file_name_sep))
+    convert_params(args.model_name, args.generator_units_scale, args.generator_layers_scale,
+                   args.params_file, out_path)

scripts/pretraining/pretraining_utils.py

@@ -7,7 +7,7 @@
 import collections
 
 import numpy as np
-from mxnet import npx
+from mxnet import np as mxnp, npx
 from mxnet.gluon import HybridBlock
 from mxnet.gluon.data import ArrayDataset
 
@@ -16,6 +16,8 @@
 from gluonnlp.data.loading import NumpyDataset, DatasetLoader
 from gluonnlp.data.sampler import SplitSampler, FixedBucketSampler
 from gluonnlp.op import select_vectors_by_position, update_vectors_by_position
+from gluonnlp.initializer import TruncNorm
+from gluonnlp.models.electra import ElectraModel, ElectraForPretrain, get_pretrained_electra
 
 PretrainFeature = collections.namedtuple(
     'PretrainFeature',
@@ -519,39 +521,39 @@ def dynamic_masking(self, input_ids, valid_lengths):
                 np.not_equal(input_ids, ignore_token)
         valid_lengths = valid_lengths.astype(np.float32)
         valid_candidates = valid_candidates.astype(np.float32)
-        num_masked_position = np.maximum(
+        num_masked_position = mxnp.maximum(
             1, np.minimum(N, round(valid_lengths * self._mask_prob)))
 
         # Get the masking probability of each position
         sample_probs = self._proposal_distribution * valid_candidates
-        sample_probs /= np.sum(sample_probs, axis=-1, keepdims=True)
+        sample_probs /= mxnp.sum(sample_probs, axis=-1, keepdims=True)
         sample_probs = npx.stop_gradient(sample_probs)
-        gumbels = np.random.gumbel(np.zeros_like(sample_probs))
+        gumbels = mxnp.random.gumbel(np.zeros_like(sample_probs))
         # Following the instruction of official repo to avoid deduplicate postions
         # with Top_k Sampling as https://github.com/google-research/electra/issues/41
         masked_positions = npx.topk(
-            np.log(sample_probs) + gumbels, k=N,
+            mxnp.log(sample_probs) + gumbels, k=N,
             axis=-1, ret_typ='indices', dtype=np.int32)
 
         masked_weights = npx.sequence_mask(
-            np.ones_like(masked_positions),
+            mxnp.ones_like(masked_positions),
             sequence_length=num_masked_position,
             use_sequence_length=True, axis=1, value=0)
         masked_positions = masked_positions * masked_weights
         length_masks = npx.sequence_mask(
-            np.ones_like(input_ids, dtype=np.float32),
+            mxnp.ones_like(input_ids, dtype=np.float32),
             sequence_length=valid_lengths,
             use_sequence_length=True, axis=1, value=0)
         unmasked_tokens = select_vectors_by_position(
             input_ids, masked_positions) * masked_weights
         masked_weights = masked_weights.astype(np.float32)
         replaced_positions = (
-            np.random.uniform(
-                np.zeros_like(masked_positions),
-                np.ones_like(masked_positions)) < self._replace_prob) * masked_positions
+            mxnp.random.uniform(
+                mxnp.zeros_like(masked_positions),
+                mxnp.ones_like(masked_positions)) < self._replace_prob) * masked_positions
         # dealing with multiple zero values in replaced_positions which causes
         # the [CLS] being replaced
-        filled = np.where(
+        filled = mxnp.where(
             replaced_positions,
             self.vocab.mask_id,
             self.vocab.cls_id).astype(
@@ -568,3 +570,42 @@ def dynamic_masking(self, input_ids, valid_lengths):
                                         masked_positions=masked_positions,
                                         masked_weights=masked_weights)
         return masked_input
+
+
+def get_electra_pretraining_model(model_name, ctx_l,
+                                  max_seq_length=128,
+                                  hidden_dropout_prob=0.1,
+                                  attention_dropout_prob=0.1,
+                                  generator_units_scale=None,
+                                  generator_layers_scale=None,
+                                  params_path=None):
+    """
+    A Electra Pretrain Model is built with a generator and a discriminator, in which
+    the generator has the same embedding as the discriminator but different backbone.
+    """
+    cfg, tokenizer, _, _ = get_pretrained_electra(
+        model_name, load_backbone=False)
+    cfg = ElectraModel.get_cfg().clone_merge(cfg)
+    cfg.defrost()
+    cfg.MODEL.hidden_dropout_prob = hidden_dropout_prob
+    cfg.MODEL.attention_dropout_prob = attention_dropout_prob
+    cfg.MODEL.max_length = max_seq_length
+    # Keep the original generator size if not designated
+    if generator_layers_scale:
+        cfg.MODEL.generator_layers_scale = generator_layers_scale
+    if generator_units_scale:
+        cfg.MODEL.generator_units_scale = generator_units_scale
+    cfg.freeze()
+
+    model = ElectraForPretrain(cfg,
+                               uniform_generator=False,
+                               tied_generator=False,
+                               tied_embeddings=True,
+                               disallow_correct=False,
+                               weight_initializer=TruncNorm(stdev=0.02))
+    if not params_path:
+        model.initialize(ctx=ctx_l)
+    else:
+        model.load_parameters(params_path, ctx=ctx_l)
+    model.hybridize()
+    return cfg, tokenizer, model

scripts/pretraining/run_electra.py

@@ -13,10 +13,10 @@
 from mxnet.lr_scheduler import PolyScheduler
 
 from sklearn import metrics
-from pretraining_utils import ElectraMasker, get_pretrain_data_npz, get_pretrain_data_text
+from pretraining_utils import ElectraMasker, get_pretrain_data_npz, \
+    get_pretrain_data_text, get_electra_pretraining_model
 from gluonnlp.utils.misc import repeat, grouper, set_seed, init_comm, logging_config, naming_convention
-from gluonnlp.initializer import TruncNorm
-from gluonnlp.models.electra import ElectraModel, ElectraForPretrain, get_pretrained_electra
+from gluonnlp.models.electra import ElectraModel, ElectraForPretrain
 from gluonnlp.utils.parameter import clip_grad_global_norm
 try:
     import horovod.mxnet as hvd
@@ -128,41 +128,6 @@ def parse_args():
     return args
 
 
-def get_pretraining_model(model_name, ctx_l,
-                          max_seq_length=128,
-                          hidden_dropout_prob=0.1,
-                          attention_dropout_prob=0.1,
-                          generator_units_scale=None,
-                          generator_layers_scale=None):
-    """
-    A Electra Pretrain Model is built with a generator and a discriminator, in which
-    the generator has the same embedding as the discriminator but different backbone.
-    """
-    cfg, tokenizer, _, _ = get_pretrained_electra(
-        model_name, load_backbone=False)
-    cfg = ElectraModel.get_cfg().clone_merge(cfg)
-    cfg.defrost()
-    cfg.MODEL.hidden_dropout_prob = hidden_dropout_prob
-    cfg.MODEL.attention_dropout_prob = attention_dropout_prob
-    cfg.MODEL.max_length = max_seq_length
-    # Keep the original generator size if not designated
-    if generator_layers_scale:
-        cfg.MODEL.generator_layers_scale = generator_layers_scale
-    if generator_units_scale:
-        cfg.MODEL.generator_units_scale = generator_units_scale
-    cfg.freeze()
-
-    model = ElectraForPretrain(cfg,
-                               uniform_generator=False,
-                               tied_generator=False,
-                               tied_embeddings=True,
-                               disallow_correct=False,
-                               weight_initializer=TruncNorm(stdev=0.02))
-    model.initialize(ctx=ctx_l)
-    model.hybridize()
-    return cfg, tokenizer, model
-
-
 ElectraOutput = collections.namedtuple('ElectraOutput',
                                        ['mlm_scores',
                                         'rtd_scores',
@@ -237,12 +202,12 @@ def train(args):
     logging.info('Training info: num_buckets: {}, '
                  'num_workers: {}, rank: {}'.format(
                      args.num_buckets, num_workers, rank))
-    cfg, tokenizer, model = get_pretraining_model(args.model_name, ctx_l,
-                                                  args.max_seq_length,
-                                                  args.hidden_dropout_prob,
-                                                  args.attention_dropout_prob,
-                                                  args.generator_units_scale,
-                                                  args.generator_layers_scale)
+    cfg, tokenizer, model = get_electra_pretraining_model(args.model_name, ctx_l,
+                                                          args.max_seq_length,
+                                                          args.hidden_dropout_prob,
+                                                          args.attention_dropout_prob,
+                                                          args.generator_units_scale,
+                                                          args.generator_layers_scale)
     data_masker = ElectraMasker(
         tokenizer, args.max_seq_length, mask_prob=args.mask_prob,
         replace_prob=args.replace_prob)