[s2s] add create student script (#7290)

Co-authored-by: Suraj Patil <surajp815@gmail.com>
Co-authored-by: Sam Shleifer <sshleifer@gmail.com>

examples/seq2seq/README.md

@@ -369,7 +369,7 @@ runtime: 6H on NVIDIA RTX 24GB GPU
 If you are using `wandb` and comparing the two distillation methods, using this entry point will make your logs consistent,
 because you will have the same hyperparameters logged in every run.
 
-#### With a teacher
+#### With a teacher (Intermediate Supervision)
 *Note* only BART variants are supported
 
 In this method, we use try to enforce that the student and teacher produce similar encoder_outputs, logits, and hidden_states using `BartSummarizationDistiller`.
@@ -378,7 +378,7 @@ This is how `sshleifer/distilbart-xsum*` checkpoints were produced.
 The command that produced `sshleifer/distilbart-xsum-12-6` is:
 
 ```bash
-./train_distilbart_xsum.sh
+./train_distilbart_xsum.sh --logger_name wandb --gpus 1
 ```
 
 runtime: 13H on V-100 16GB GPU.

examples/seq2seq/distillation.py

@@ -4,7 +4,6 @@ import argparse
 import gc
 import os
 import sys
-import warnings
 from pathlib import Path
 from typing import List
 
@@ -15,18 +14,10 @@ from torch.nn import functional as F
 
 from finetune import SummarizationModule, TranslationModule
 from finetune import main as ft_main
-from initialization_utils import copy_layers, init_student
+from make_student import create_student_by_copying_alternating_layers, get_layers_to_supervise
-from transformers import AutoModelForSeq2SeqLM, MBartTokenizer, T5Config, T5ForConditionalGeneration
+from transformers import AutoModelForSeq2SeqLM, MBartTokenizer, T5ForConditionalGeneration
 from transformers.modeling_bart import shift_tokens_right
-from utils import (
+from utils import calculate_bleu, freeze_params, label_smoothed_nll_loss, pickle_load, use_task_specific_params
-    any_requires_grad,
-    assert_all_frozen,
-    calculate_bleu,
-    freeze_params,
-    label_smoothed_nll_loss,
-    pickle_load,
-    use_task_specific_params,
-)
 
 
 # need the parent dir module
@@ -41,87 +32,50 @@ class BartSummarizationDistiller(SummarizationModule):
 
     def __init__(self, hparams):
         assert Path(hparams.data_dir).exists()
-        student, student_cfg, teacher = self.pre_init(hparams)
+        self.output_dir = Path(hparams.output_dir)
+        self.output_dir.mkdir(exist_ok=True)
 
-        super().__init__(hparams, model=student, config=student_cfg)
+        save_dir = self.output_dir.joinpath("student")
+
+        hparams.model_name_or_path = str(save_dir)  # Tell lightning we are training the student
+        teacher = AutoModelForSeq2SeqLM.from_pretrained(hparams.teacher).eval()
+        use_task_specific_params(teacher, hparams.task)  # We copy good generation parameters to student by default
+        student, e_layer_ids, d_layer_ids = create_student_by_copying_alternating_layers(
+            teacher, e=hparams.student_encoder_layers, d=hparams.student_decoder_layers, save_path=save_dir
+        )
+        if hparams.length_penalty != -1:
+            student.config.length_penalty = hparams.length_penalty
+        super().__init__(hparams, model=student, config=student.config)
+        self.e_layer_ids, self.d_layer_ids = e_layer_ids, d_layer_ids  # type: List[int], List[int]
+        self.different_encoder = hparams.student_encoder_layers != teacher.config.encoder_layers
+        self.different_decoder = hparams.student_decoder_layers != teacher.config.decoder_layers
         self.teacher = teacher
-        use_task_specific_params(self.teacher, "summarization")
         freeze_params(self.teacher)
-        self.sanity_check_gradients()
+
+        if not self.different_encoder:  # To save RAM, delete teacher encoder and freeze student encoder.
+            try:
+                del self.teacher.model.encoder
+            except AttributeError:  # T5
+                del self.teacher.encoder
+        # Intermediate supervision: Decide which layers to supervise
+        if hparams.supervise_forward:
+            self.d_matches = get_layers_to_supervise(
+                n_student=len(self.d_layer_ids), n_teacher=self.teacher.config.decoder_layers
+            )
+        else:
+            self.d_matches = self.d_layer_ids
         self.ce_loss_fct = nn.KLDivLoss(reduction="batchmean")
         self.temperature = 2.0
         self.alpha_mlm = hparams.alpha_mlm
         self.alpha_ce = hparams.alpha_ce
         self.alpha_hid = hparams.alpha_hid
-        # self.alpha_cos = hparams.alpha_cos
+        self.alpha_encoder_loss = hparams.alpha_encoder_loss
-        self.alpha_encoder_loss = self.hparams.alpha_encoder_loss
         gc.collect()
         torch.cuda.empty_cache()
 
-    def sanity_check_gradients(self):
-        assert_all_frozen(self.teacher)
-        assert_all_frozen(self.model.model.decoder.embed_tokens)
-        assert_all_frozen(self.model.model.encoder.embed_tokens)
-        if self.different_encoder:
-            assert any_requires_grad(self.model.model.encoder)
-        else:
-            freeze_params(self.model.model.encoder)
-            del self.teacher.model.encoder
-
-    def pre_init(self, hparams):
-        self.output_dir = Path(hparams.output_dir)
-        self.output_dir.mkdir(exist_ok=True)
-        teacher = AutoModelForSeq2SeqLM.from_pretrained(hparams.teacher).eval()
-        student_updates = {
-            "decoder_layers": hparams.student_decoder_layers,
-            "encoder_layers": hparams.student_encoder_layers,
-        }
-        if hparams.length_penalty != -1:
-            student_updates["length_penalty"] = hparams.length_penalty
-        e_layers_to_copy: List = get_layers_to_copy(student_updates["encoder_layers"], teacher.config.encoder_layers)
-        hparams.e_layer_to_copy = e_layers_to_copy
-
-        d_layers_to_copy: List = get_layers_to_copy(student_updates["decoder_layers"], teacher.config.decoder_layers)
-
-        if hparams.supervise_forward:
-            hparams.d_matches = get_layers_to_supervise(
-                student_updates["decoder_layers"], teacher.config.decoder_layers
-            )
-        else:
-            hparams.d_matches = d_layers_to_copy
-        hparams.d_layer_to_copy = d_layers_to_copy
-
-        kw = teacher.config.to_diff_dict()
-        kw.update(student_updates)
-        # Copy weights
-        student_cfg = teacher.config_class(**kw)
-        student = type(teacher)(student_cfg)
-        student, _ = init_student(student, teacher)
-        save_dir = self.output_dir.joinpath("student")
-        self.copy_to_student(d_layers_to_copy, e_layers_to_copy, hparams, student, teacher)
-        student.save_pretrained(save_dir)
-        hparams.model_name_or_path = str(save_dir)
-        return student, student_cfg, teacher
-
-    def copy_to_student(self, d_layers_to_copy, e_layers_to_copy, hparams, student, teacher):
-        if teacher.config.model_type == "t5":
-            return self.copy_t5_to_student(d_layers_to_copy, e_layers_to_copy, hparams, student, teacher)
-        self.different_encoder: bool = hparams.student_encoder_layers != teacher.config.encoder_layers
-        self.different_decoder = hparams.student_decoder_layers != teacher.config.decoder_layers
-        if self.different_decoder:
-            copy_layers(teacher.model.decoder.layers, student.model.decoder.layers, d_layers_to_copy)
-        if self.different_encoder:
-            copy_layers(teacher.model.encoder.layers, student.model.encoder.layers, e_layers_to_copy)
-
-    def copy_t5_to_student(self, d_layers_to_copy, e_layers_to_copy, hparams, student, teacher):
-        self.different_encoder: bool = hparams.student_encoder_layers != teacher.config.num_layers
-        self.different_decoder = hparams.student_decoder_layers != teacher.config.num_layers
-        if self.different_decoder:
-            copy_layers(teacher.decoder.block, student.decoder.block, d_layers_to_copy)
-        if self.different_encoder:
-            copy_layers(teacher.encoder.block, student.encoder.block, e_layers_to_copy)
-
     def calc_mse_loss(self, teacher_outputs: torch.Tensor, student_outputs: torch.Tensor, mask) -> torch.FloatTensor:
+        """Supervise MSE(teacher.encoder_outputs, student.encoder_outputs)."""
+        # raise NotImplementedError()
         if mask is not None:
             # mask has False at padding_idx
             sel_mask = mask[:, :, None].expand_as(student_outputs).bool()
@@ -133,20 +87,15 @@ class BartSummarizationDistiller(SummarizationModule):
         return F.mse_loss(s_logits_slct, t_logits_slct)
 
     def calc_ce_loss(self, mask, s_logits, t_logits):
-        if mask is not None:
+        """Copy pasted from distillbert (transformers/examples/distillation/)"""
-            # mask has False at padding_idx
+
-            sel_mask = mask[:, :, None].expand_as(s_logits)
+        # mask has False at padding_idx
-            s_logits_slct = torch.masked_select(
+        sel_mask = mask[:, :, None].expand_as(s_logits)
-                s_logits, sel_mask
+        vocab_size = s_logits.size(-1)
-            )  # (bs * seq_length * voc_size) modulo the 1s in mask
+        s_logits_slct = torch.masked_select(s_logits, sel_mask)  # (bs * seq_length * voc_size) modulo the 1s in mask
-            t_logits_slct = torch.masked_select(
+        t_logits_slct = torch.masked_select(t_logits, sel_mask)  # (bs * seq_length * voc_size) modulo the 1s in mask
-                t_logits, sel_mask
+        s_logits_slct = s_logits_slct.view(-1, vocab_size)  # (bs * seq_length, voc_size) modulo the 1s in mask
-            )  # (bs * seq_length * voc_size) modulo the 1s in mask
+        t_logits_slct = t_logits_slct.view(-1, vocab_size)  # (bs * seq_length, voc_size) modulo the 1s in mask
-        else:
-            t_logits_slct = t_logits
-            s_logits_slct = s_logits  # (bs * seq_length * voc_size) modulo the 1s in mask
-        s_logits_slct = s_logits_slct.view(-1, s_logits.size(-1))  # (bs * seq_length, voc_size) modulo the 1s in mask
-        t_logits_slct = t_logits_slct.view(-1, s_logits.size(-1))  # (bs * seq_length, voc_size) modulo the 1s in mask
         assert t_logits_slct.size() == s_logits_slct.size()
         loss_ce = (
             self.ce_loss_fct(
@@ -155,7 +104,7 @@ class BartSummarizationDistiller(SummarizationModule):
             )
             * (self.temperature) ** 2
         )
-        return loss_ce, s_logits_slct, t_logits_slct
+        return loss_ce
 
     @staticmethod
     def add_model_specific_args(parser, root_dir):
@@ -164,10 +113,14 @@ class BartSummarizationDistiller(SummarizationModule):
         return parser
 
     def _step(self, batch):
-        # assert is_frozen(self.teacher)
+        # assert is_frozen(self.teacher) copied_decoder_layers
         pad_token_id = self.tokenizer.pad_token_id
-        input_ids, src_mask, tgt_ids = batch["input_ids"], batch["attention_mask"], batch["labels"]
+        input_ids, src_mask, labels = batch["input_ids"], batch["attention_mask"], batch["labels"]
-        decoder_input_ids = shift_tokens_right(tgt_ids, pad_token_id)
+        if isinstance(self.model, T5ForConditionalGeneration):
+            decoder_input_ids = self.model._shift_right(labels)
+        else:
+            decoder_input_ids = shift_tokens_right(labels, pad_token_id)
+
         # noinspection PyCallingNonCallable
         lm_logits, dec_hidden, enc_outputs, enc_hidden_state = self(
             input_ids,
@@ -183,11 +136,11 @@ class BartSummarizationDistiller(SummarizationModule):
         if self.hparams.label_smoothing == 0:
             # Same behavior as modeling_bart.py, besides ignoring pad_token_id
             loss_fct = torch.nn.CrossEntropyLoss(ignore_index=pad_token_id)
-            student_lm_loss = loss_fct(lm_logits.view(-1, lm_logits.shape[-1]), tgt_ids.view(-1))
+            student_lm_loss = loss_fct(lm_logits.view(-1, lm_logits.shape[-1]), labels.view(-1))
         else:
             lprobs = torch.nn.functional.log_softmax(lm_logits, dim=-1)
             student_lm_loss, _ = label_smoothed_nll_loss(
-                lprobs, tgt_ids, self.hparams.label_smoothing, ignore_index=pad_token_id
+                lprobs, labels, self.hparams.label_smoothing, ignore_index=pad_token_id
             )
 
         def zero_tensor():
@@ -196,15 +149,14 @@ class BartSummarizationDistiller(SummarizationModule):
         loss_encoder, hid_loss_enc, hid_loss_dec = zero_tensor(), zero_tensor(), zero_tensor()
         if self.different_encoder:
             with torch.no_grad():
-                teacher_enc_outputs, teacher_enc_hid, _ = self.teacher.model.encoder(
+                teacher_enc_outputs, teacher_enc_hid, _ = self.teacher.get_encoder()(
                     input_ids, attention_mask=src_mask, output_hidden_states=True
                 )
+            # DEPRECATE THIS
             if self.hparams.alpha_encoder_loss > 0:
                 loss_encoder = self.calc_mse_loss(enc_outputs, teacher_enc_outputs, src_mask)
 
-            hid_loss_enc = self.calc_hidden_loss(
+            hid_loss_enc = self.calc_hidden_loss(src_mask, enc_hidden_state, teacher_enc_hid, self.e_layer_ids)
-                src_mask, enc_hidden_state, teacher_enc_hid, self.hparams.e_layer_to_copy
-            )
 
         teacher_enc_outputs = (enc_outputs,)
         assert isinstance(teacher_enc_outputs, tuple), type(teacher_enc_outputs)
@@ -215,13 +167,15 @@ class BartSummarizationDistiller(SummarizationModule):
                 attention_mask=src_mask,
                 encoder_outputs=teacher_enc_outputs,
                 decoder_input_ids=decoder_input_ids,
-                lm_labels=tgt_ids,
+                lm_labels=labels,
                 output_hidden_states=True,
             )
         dec_mask = decoder_input_ids.ne(pad_token_id)
-        loss_ce, s_logits_slct, t_logits_slct = self.calc_ce_loss(dec_mask, lm_logits, tlogits)
+        loss_ce = self.calc_ce_loss(dec_mask, lm_logits, tlogits)
-        if self.alpha_hid > 0:
+        if self.alpha_hid > 0:  # Intermediate supervision of decoder hidden states
-            hid_loss_dec = self.calc_hidden_loss(dec_mask, dec_hidden, tdec_hidden, self.hparams.d_matches)
+            hid_loss_dec = self.calc_hidden_loss(
+                dec_mask, dec_hidden, tdec_hidden, self.d_matches, normalize_hidden=self.hparams.normalize_hidden
+            )
 
         blended_loss = (
             self.alpha_ce * loss_ce
@@ -231,7 +185,9 @@ class BartSummarizationDistiller(SummarizationModule):
         )
         return blended_loss, loss_ce, student_lm_loss, loss_encoder, hid_loss_enc, hid_loss_dec
 
-    def calc_hidden_loss(self, attention_mask, hidden_states, hidden_states_T, matches):
+    @staticmethod
+    def calc_hidden_loss(attention_mask, hidden_states, hidden_states_T, matches, normalize_hidden):
+        """MSE(student_hid, teacher_hid[matches]). Called "Intermediate supervision" in paper. Inspired by TinyBERT."""
         msg = "expected list or tuple for hidden_states, got tensor of shape: "
         assert not isinstance(hidden_states, torch.Tensor), f"{msg}{hidden_states.shape}"
         assert not isinstance(hidden_states_T, torch.Tensor), f"{msg}{hidden_states_T.shape}"
@@ -239,7 +195,7 @@ class BartSummarizationDistiller(SummarizationModule):
         valid_count = mask.sum() * hidden_states[0].size(-1)
         student_states = torch.stack([hidden_states[i] for i in range(len(matches))])
         teacher_states = torch.stack([hidden_states_T[j] for j in matches])
-        if self.hparams.normalize_hidden:
+        if normalize_hidden:
             student_states = F.layer_norm(student_states, student_states.shape[1:])
             teacher_states = F.layer_norm(teacher_states, teacher_states.shape[1:])
         mse = F.mse_loss(student_states, teacher_states, reduction="none")
@@ -287,130 +243,9 @@ class BartTranslationDistiller(BartSummarizationDistiller):
         return parser
 
 
-class T5SummarizationDistiller(BartSummarizationDistiller):
-    def pre_init(self, hparams):
-        raise NotImplementedError("T5 Distillation does not work yet")
-        self.output_dir = Path(hparams.output_dir)
-        self.output_dir.mkdir(exist_ok=True)
-        teacher = T5ForConditionalGeneration.from_pretrained(hparams.teacher)
-        n_layer = hparams.student_decoder_layers
-        assert n_layer == hparams.student_encoder_layers  # TODO(SS): relax this constraint so that we can do 12-6.
-        d_layers_to_copy = get_layers_to_copy(n_layer, len(teacher.decoder.block))
-        e_layers_to_copy: List = get_layers_to_copy(n_layer, len(teacher.encoder.block))
-        student_updates = {"num_layers": n_layer}
-        hparams.d_layer_to_copy = d_layers_to_copy
-        hparams.e_layer_to_copy = e_layers_to_copy
-        kw = teacher.config.to_diff_dict()
-
-        kw.update(student_updates)
-        # Copy weights
-        student_cfg = T5Config(**kw)
-        student = T5ForConditionalGeneration(student_cfg)
-        student, _ = init_student(student, teacher)
-        self.copy_to_student(d_layers_to_copy, e_layers_to_copy, hparams, student, teacher)
-        Path(hparams.output_dir).mkdir(exist_ok=True)
-        task_specific_params = student.config.task_specific_params
-        if task_specific_params is not None:
-            student.config.update(task_specific_params.get("summarization", {}))  # TODO: dont hardcode
-        save_dir = self.output_dir.joinpath("student")
-        save_dir.mkdir(exist_ok=True)
-
-        student.save_pretrained(save_dir)
-        hparams.model_name_or_path = str(save_dir)
-        return student, student_cfg, teacher
-
-    def freeze_embeds(self):
-        freeze_params(self.model.shared)
-        for d in [self.model.encoder, self.model.decoder]:
-            freeze_params(d.embed_tokens)
-
-    def sanity_check_gradients(self):
-        """T5"""
-        assert_all_frozen(self.teacher)
-        assert_all_frozen(self.model.decoder.embed_tokens)
-        assert_all_frozen(self.model.encoder.embed_tokens)
-        if self.different_encoder:
-            assert any_requires_grad(self.model.encoder)
-        else:
-            freeze_params(self.model.encoder)
-            del self.teacher.model.encoder
-        if self.different_decoder:
-            assert any_requires_grad(self.model.decoder)
-        else:
-            freeze_params(self.model.decoder)  # TODO(SS): very suspicious
-
-    def _step(self, batch):
-        pad_token_id = self.tokenizer.pad_token_id
-        source_ids, source_mask, y = batch["input_ids"], batch["attention_mask"], batch["decoder_input_ids"]
-        decoder_input_ids = y[:, :-1].contiguous()
-        labels = y[:, 1:].clone()
-        labels[y[:, 1:] == pad_token_id] = -100
-        # noinspection PyCallingNonCallable
-        dec_mask = decoder_input_ids.ne(pad_token_id)
-
-        sloss, slogits, dec_hidden, enc_outputs, enc_hidden_state = self(
-            source_ids,
-            attention_mask=source_mask,
-            decoder_input_ids=decoder_input_ids,
-            labels=labels,
-            output_hidden_states=True,
-            output_attentions=False,
-            use_cache=False,
-        )
-
-        def zero_tensor():
-            return torch.tensor(0.0).type_as(sloss)
-
-        loss_encoder, hid_loss_enc, hid_loss_dec = zero_tensor(), zero_tensor(), zero_tensor()
-        if self.different_encoder:
-            with torch.no_grad():
-                teacher_enc_outputs, teacher_enc_hid = self.teacher.encoder(
-                    source_ids,
-                    attention_mask=source_mask,
-                    output_hidden_states=True,
-                    use_cache=False,
-                )
-            if self.hparams.alpha_encoder_loss > 0:
-                loss_encoder = self.calc_mse_loss(enc_outputs, teacher_enc_outputs, source_mask)
-
-            hid_loss_enc = self.calc_hidden_loss(
-                source_mask, enc_hidden_state, teacher_enc_hid, self.hparams.e_layer_to_copy
-            )
-
-        teacher_enc_outputs = (enc_outputs,)
-        assert isinstance(teacher_enc_outputs, tuple), type(teacher_enc_outputs)
-
-        with torch.no_grad():
-            tloss, tlogits, tdec_hidden, _ = self.teacher(
-                source_ids,
-                attention_mask=source_mask,
-                encoder_outputs=teacher_enc_outputs,
-                decoder_input_ids=decoder_input_ids,
-                labels=labels,
-                output_hidden_states=True,
-                use_cache=False,
-            )
-
-        loss_ce, s_logits_slct, t_logits_slct = self.calc_ce_loss(dec_mask, slogits, tlogits)
-        if self.alpha_hid > 0:
-            hid_loss_dec = self.calc_hidden_loss(dec_mask, dec_hidden, tdec_hidden, self.hparams.d_matches)
-
-        blended_loss = (
-            self.alpha_ce * loss_ce
-            + self.alpha_mlm * sloss
-            + self.hparams.alpha_encoder_loss * loss_encoder
-            + self.hparams.alpha_hid * (hid_loss_enc + hid_loss_dec)
-        )
-        return blended_loss, loss_ce, sloss, loss_encoder, hid_loss_enc, hid_loss_dec
-
-
 def create_module(args):
-    t5 = "t5" in args.model_name_or_path
     if args.no_teacher:
         module_cls = TranslationModule if "translation" in args.task else SummarizationModule
-    elif t5:  # DISTILL T5 WITH TEACHER FOR SUMMARIZATION
-        assert "translation" not in args.task, "t5 translation distillation not supported"
-        module_cls = T5SummarizationDistiller
     else:  # DISTILL WITH TEACHER
         module_cls = BartTranslationDistiller if "translation" in args.task else BartSummarizationDistiller
     args.setup_cls: str = module_cls.__name__
@@ -443,56 +278,6 @@ def evaluate_checkpoint(ckpt_path: Path, dest_dir=None):
     trainer.test(model)
 
 
-LAYERS_TO_COPY = {
-    # maps  num layers in student -> which teacher layers to copy.
-    # 12: bart, 16: pegasus, 6: marian/Helsinki-NLP
-    12: {
-        1: [0],
-        2: [0, 6],
-        3: [0, 6, 11],
-        4: [0, 4, 8, 11],
-        6: [0, 2, 4, 7, 9, 11],
-        9: [0, 1, 2, 4, 5, 7, 9, 10, 11],
-        12: list(range(12)),
-    },
-    16: {  # maps  num layers in student -> which teacher layers to copy
-        1: [0],
-        2: [0, 8],
-        3: [0, 8, 15],
-        4: [0, 5, 10, 15],
-        6: [0, 3, 6, 9, 12, 15],
-        8: [0, 2, 4, 6, 8, 10, 12, 15],
-        9: [0, 1, 3, 5, 7, 9, 11, 13, 15],
-        12: [0, 1, 2, 3, 4, 5, 6, 7, 9, 11, 13, 15],
-        16: list(range(16)),
-    },
-    6: {1: [0], 2: [0, 5], 3: [0, 2, 5], 4: [0, 1, 3, 5], 6: list(range(6))},
-}
-LAYERS_TO_SUPERVISE = {
-    12: {1: [11], 2: [5, 11], 3: [3, 7, 11], 6: [1, 3, 5, 8, 10, 11]},
-    16: {1: [15], 4: [4, 9, 12, 15], 8: [1, 3, 5, 7, 9, 11, 13, 15]},
-    6: {1: [5], 2: [3, 5], 3: [1, 4, 5], 4: [1, 2, 4, 5]},
-    2: {1: [1], 2: [0, 1]},
-}
-
-
-def get_layers_to_supervise(n_student, n_teacher):
-    return LAYERS_TO_SUPERVISE[n_teacher][n_student]
-
-
-def get_layers_to_copy(n_student, n_teacher):
-    try:
-        val = LAYERS_TO_COPY[n_teacher][n_student]
-        assert len(LAYERS_TO_SUPERVISE[n_teacher][n_student]) == len(val) == n_student
-        return val
-    except KeyError:
-        if n_student != n_teacher:
-            warnings.warn(
-                f"no hardcoded layers to copy for teacher {n_teacher} -> student {n_student}, defaulting to first {n_student}"
-            )
-        return list(range(n_student))
-
-
 def distill_main(args):
     Path(args.output_dir).mkdir(exist_ok=True)
     if len(os.listdir(args.output_dir)) > 3 and args.do_train:

examples/seq2seq/initialization_utils.py

@@ -1,20 +0,0 @@
-from typing import List
-
-from torch import nn
-
-
-def init_student(student, teacher):
-    teacher_state_dict = teacher.state_dict()
-    info = student.load_state_dict(teacher_state_dict, strict=False)
-    assert info.missing_keys == [], info.missing_keys
-    return student, info
-
-
-def copy_decoder_layers(teacher, student, l2copy=[0, 2, 4, 7, 9, 11]):
-    copy_layers(teacher.model.decoder.layers, student.model.decoder.layers, l2copy)
-
-
-def copy_layers(teacher_layers: nn.ModuleList, student_layers: nn.ModuleList, layers_to_copy: List) -> None:
-    layers_to_copy = nn.ModuleList([l for i, l in enumerate(teacher_layers) if i in layers_to_copy])
-    assert len(student_layers) == len(layers_to_copy), f"{len(student_layers)} != {len(layers_to_copy)}"
-    student_layers.load_state_dict(layers_to_copy.state_dict())

examples/seq2seq/make_student.py

@@ -0,0 +1,167 @@
+import warnings
+from pathlib import Path
+from typing import List, Tuple, Union
+
+import fire
+from torch import nn
+
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, PreTrainedModel
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+def copy_layers(src_layers: nn.ModuleList, dest_layers: nn.ModuleList, layers_to_copy: List[int]) -> None:
+    layers_to_copy = nn.ModuleList([l for i, l in enumerate(src_layers) if i in layers_to_copy])
+    assert len(dest_layers) == len(layers_to_copy), f"{len(dest_layers)} != {len(layers_to_copy)}"
+    dest_layers.load_state_dict(layers_to_copy.state_dict())
+
+
+LAYERS_TO_COPY = {
+    # maps  num layers in teacher -> num_layers in student -> which teacher layers to copy.
+    # 12: bart, 16: pegasus, 6: marian/Helsinki-NLP
+    12: {
+        1: [0],  # This says that if the teacher has 12 layers and the student has 1, copy layer 0 of the teacher
+        2: [0, 6],
+        3: [0, 6, 11],
+        4: [0, 4, 8, 11],
+        6: [0, 2, 4, 7, 9, 11],
+        9: [0, 1, 2, 4, 5, 7, 9, 10, 11],
+        12: list(range(12)),
+    },
+    16: {  # maps  num layers in student -> which teacher layers to copy
+        1: [0],
+        2: [0, 8],
+        3: [0, 8, 15],
+        4: [0, 5, 10, 15],
+        6: [0, 3, 6, 9, 12, 15],
+        8: [0, 2, 4, 6, 8, 10, 12, 15],
+        9: [0, 1, 3, 5, 7, 9, 11, 13, 15],
+        12: [0, 1, 2, 3, 4, 5, 6, 7, 9, 11, 13, 15],
+        16: list(range(16)),
+    },
+    6: {1: [0], 2: [0, 5], 3: [0, 2, 5], 4: [0, 1, 3, 5], 6: list(range(6))},
+}
+LAYERS_TO_SUPERVISE = {
+    # maps  num layers in student -> which teacher layers to copy.
+    6: {1: [5], 2: [3, 5], 3: [1, 4, 5], 4: [1, 2, 4, 5]},
+    12: {1: [11], 2: [5, 11], 3: [3, 7, 11], 6: [1, 3, 5, 8, 10, 11]},
+    16: {1: [15], 4: [4, 9, 12, 15], 8: [1, 3, 5, 7, 9, 11, 13, 15]},
+}
+
+
+def pick_layers_to_copy(n_student, n_teacher):
+    try:
+        val = LAYERS_TO_COPY[n_teacher][n_student]
+        return val
+    except KeyError:
+        if n_student != n_teacher:
+            warnings.warn(
+                f"no hardcoded layers to copy for teacher {n_teacher} -> student {n_student}, defaulting to first {n_student}"
+            )
+        return list(range(n_student))
+
+
+def get_layers_to_supervise(n_student, n_teacher) -> List[int]:
+    """Used or the --supervise_forward kwarg"""
+    if n_student > n_teacher:
+        raise ValueError(f"Cannot perform intermediate supervision for student {n_student} > teacher {n_teacher}")
+    elif n_teacher == n_student:
+        return list(range(n_teacher))
+    elif n_student == 1:
+        return [n_teacher - 1]
+    else:
+        return LAYERS_TO_SUPERVISE[n_teacher][n_student]
+
+
+def create_student_by_copying_alternating_layers(
+    teacher: Union[str, PreTrainedModel],
+    save_path: Union[str, Path] = "student",
+    e: Union[int, None] = None,
+    d: Union[int, None] = None,
+    copy_first_teacher_layers=False,
+    **extra_config_kwargs
+) -> Tuple[PreTrainedModel, List[int], List[int]]:
+    """Make a student by copying alternating layers from a teacher, save it to save_path.
+    Args:
+        teacher: str or PreTrainedModel if str, this will call AutoModelForSeq2SeqLM.from_pretrained(teacher) before
+        copying layers
+        save_path: where to save the student, defaults to student directory.
+        e: how many Encoder layers should the student have, default is fully copy of teacher
+        d: how many Decoder layers should the student have, default is fully copy of teacher
+        copy_first_teacher_layers: [bool] dont copy alternating layers, just the first e/d.
+        **extra_config_kwargs: extra kwargs to pass to the student, by default the teacher config is used.
+
+    Returns:
+        student: new, smaller model.  (Also saves it to save_path)
+        e_layers_to_copy: list of which teacher encoder layers were used
+        d_layers_to_copy: list of which teacher decoder layers were used
+    """
+    _msg = "encoder_layers and decoder_layers cannot be both None-- you would just have an identical teacher."
+    assert (e is not None) or (d is not None), _msg
+    if isinstance(teacher, str):
+        AutoTokenizer.from_pretrained(teacher).save_pretrained(save_path)  # purely for convenience
+        teacher = AutoModelForSeq2SeqLM.from_pretrained(teacher).eval()
+    else:
+
+        assert isinstance(teacher, PreTrainedModel), f"teacher must be a model or string got type {type(teacher)}"
+    init_kwargs = teacher.config.to_diff_dict()
+
+    try:
+        teacher_e, teacher_d = teacher.config.encoder_layers, teacher.config.decoder_layers
+        if e is None:
+            e = teacher_e
+        if d is None:
+            d = teacher_d
+        init_kwargs.update({"encoder_layers": e, "decoder_layers": d})
+    except AttributeError:  # T5
+        teacher_e, teacher_d = teacher.config.num_layers, teacher.config.num_hidden_layers
+        assert e == d, "T5 Students must be symmetric"
+        init_kwargs["num_layers"] = e
+
+    # Kwargs to instantiate student = teacher kwargs with updated layer numbers + **extra_config_kwargs
+
+    init_kwargs.update(extra_config_kwargs)
+
+    # Copy weights
+    student_cfg = teacher.config_class(**init_kwargs)
+    student = AutoModelForSeq2SeqLM.from_config(student_cfg)
+    # Start by copying the full teacher state dict this will copy the first N teacher layers to the student.
+    info = student.load_state_dict(teacher.state_dict(), strict=False)
+    assert info.missing_keys == [], info.missing_keys  # every student key should have a teacher keys.
+
+    if copy_first_teacher_layers:  # Our copying is done. We just log and save
+        e_layers_to_copy, d_layers_to_copy = list(range(e)), list(range(d))
+        logger.info(
+            f"Copied encoder layers {e_layers_to_copy} and decoder layers {d_layers_to_copy}. Saving them to {save_path}"
+        )
+        student.save_pretrained(save_path)
+        return student, e_layers_to_copy, d_layers_to_copy
+
+    # Decide which layers of the teacher to copy. Not exactly alternating -- we try to keep first and last layer.
+    e_layers_to_copy: List[int] = pick_layers_to_copy(e, teacher_e)
+    d_layers_to_copy: List[int] = pick_layers_to_copy(d, teacher_d)
+
+    try:
+        copy_layers(teacher.model.encoder.layers, student.model.encoder.layers, e_layers_to_copy)
+        copy_layers(teacher.model.decoder.layers, student.model.decoder.layers, d_layers_to_copy)
+    except AttributeError:  # For t5, student.model.encoder.layers is called student.encoder.block
+        copy_layers(teacher.encoder.block, student.encoder.block, e_layers_to_copy)
+        copy_layers(teacher.decoder.block, student.decoder.block, d_layers_to_copy)
+    logger.info(
+        f"Copied encoder layers {e_layers_to_copy} and decoder layers {d_layers_to_copy}. Saving them to {save_path}"
+    )
+    student.config.init_metadata = dict(
+        teacher_type=teacher.config.model_type,
+        copied_encoder_layers=e_layers_to_copy,
+        copied_decoder_layers=d_layers_to_copy,
+    )
+    student.save_pretrained(save_path)
+    # Save information about copying for easier reproducibility
+
+    return student, e_layers_to_copy, d_layers_to_copy
+
+
+if __name__ == "__main__":
+    fire.Fire(create_student_by_copying_alternating_layers)

examples/seq2seq/save_randomly_initialized_model.py

@@ -0,0 +1,26 @@
+#!/usr/bin/env python
+
+import fire
+
+from transformers import AutoConfig, AutoModelForSeq2SeqLM, AutoTokenizer
+
+
+def save_randomly_initialized_version(config_name: str, save_dir: str, **config_kwargs):
+    """Save a randomly initialized version of a model using a pretrained config.
+    Args:
+        config_name: which config to use
+        save_dir: where to save the resulting model and tokenizer
+        config_kwargs: Passed to AutoConfig
+
+    Usage::
+        save_randomly_initialized_version("facebook/bart-large-cnn", "distilbart_random_cnn_6_3", encoder_layers=6, decoder_layers=3, num_beams=3)
+    """
+    cfg = AutoConfig.from_pretrained(config_name, **config_kwargs)
+    model = AutoModelForSeq2SeqLM.from_config(cfg)
+    model.save_pretrained(save_dir)
+    AutoTokenizer.from_pretrained(config_name).save_pretrained(save_dir)
+    return model
+
+
+if __name__ == "__main__":
+    fire.Fire(save_randomly_initialized_version)

examples/seq2seq/test_make_student.py

@@ -0,0 +1,41 @@
+import tempfile
+import unittest
+
+from make_student import create_student_by_copying_alternating_layers
+from transformers import AutoConfig
+from transformers.file_utils import cached_property
+from transformers.testing_utils import require_torch
+
+
+TINY_BART = "sshleifer/bart-tiny-random"
+TINY_T5 = "patrickvonplaten/t5-tiny-random"
+
+
+@require_torch
+class MakeStudentTester(unittest.TestCase):
+    @cached_property
+    def teacher_config(self):
+        return AutoConfig.from_pretrained(TINY_BART)
+
+    def test_valid_t5(self):
+        student, *_ = create_student_by_copying_alternating_layers(TINY_T5, tempfile.mkdtemp(), e=1, d=1)
+        self.assertEqual(student.config.num_hidden_layers, 1)
+
+    def test_invalid_t5(self):
+        # T5 students must have the same e==d because there is only one config property
+        with self.assertRaises(AssertionError):
+            student, *_ = create_student_by_copying_alternating_layers(TINY_T5, tempfile.mkdtemp(), e=1, d=None)
+
+    def test_same_decoder_small_encoder(self):
+        student, *_ = create_student_by_copying_alternating_layers(TINY_BART, tempfile.mkdtemp(), e=1, d=None)
+        self.assertEqual(student.config.encoder_layers, 1)
+        self.assertEqual(student.config.decoder_layers, self.teacher_config.encoder_layers)
+
+    def test_small_enc_small_dec(self):
+        student, *_ = create_student_by_copying_alternating_layers(TINY_BART, tempfile.mkdtemp(), e=1, d=1)
+        self.assertEqual(student.config.encoder_layers, 1)
+        self.assertEqual(student.config.decoder_layers, 1)
+
+    def test_raises_assert(self):
+        with self.assertRaises(AssertionError):
+            create_student_by_copying_alternating_layers(TINY_BART, tempfile.mkdtemp(), e=None, d=None)

examples/seq2seq/train_distilbart_xsum.sh

@@ -1,21 +1,20 @@
 #!/usr/bin/env bash
 export PYTHONPATH="../":"${PYTHONPATH}"
-export BS=16
-export GAS=2
 python distillation.py \
+  --teacher facebook/bart-large-xsum --data_dir xsum \
+  --student_decoder_layers 6 --student_encoder_layers 12 \
+  --freeze_encoder --freeze_embeds \
   --learning_rate=3e-4 \
   --do_train \
   --do_predict \
-  --fp16 \
+  --fp16 --fp16_opt_level=O1 \
-  --val_check_interval 0.1 --n_val 1000 \
+  --val_check_interval 0.1 --n_val 1000 --eval_beams 2 --length_penalty=0.5 \
-  --teacher facebook/bart-large-xsum --data_dir $XSUM_DIR \
   --max_target_length=60 --val_max_target_length=60 --test_max_target_length=100 \
-  --student_decoder_layers 6 --student_encoder_layers 12 \
-  --freeze_encoder --freeze_embeds \
   --model_name_or_path IGNORED \
-  --alpha_hid=3. --length_penalty=0.5 \
+  --alpha_hid=3. \
-  --train_batch_size=$BS --eval_batch_size=$BS --gradient_accumulation_steps=$GAS --num_train_epochs=6 \
+  --train_batch_size=16 --eval_batch_size=16 --gradient_accumulation_steps=2 \
-  --tokenizer_name facebook/bart-large \
+  --sortish_sampler \
+  --num_train_epochs=6 \
   --warmup_steps 500 \
   --output_dir distilbart_xsum_12_6 \
   "$@"