File size: 14,840 Bytes
8ebda9e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
# coding=utf-8
# Copyright 2022 IDEA-CCNL The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch Della model. """

import os
import torch
import numpy as np
from fengshen.models.deepVAE.deep_vae import DeepVAE
from pytorch_lightning.core.lightning import LightningModule
from transformers.models.gpt2.configuration_gpt2 import GPT2Config
from transformers.models.bert.tokenization_bert import BertTokenizer
from fengshen.models.deepVAE.latent_connector import GPT2ForDecoderLatentConnector, GPT2ForEncoderLatentConnector
from transformers.optimization import AdamW, get_linear_schedule_with_warmup


class DeepVAEModule(LightningModule):
    @classmethod
    def add_module_specific_args(cls, parser):
        group = parser.add_argument_group('vae', 'configurations')
        group.add_argument("--checkpoint_path", type=str, default=None)
        group.add_argument("--gpt2_model_path", type=str)
        group.add_argument("--beta_kl_constraints_start", default=1, type=float,
                           help="min beta for all the latent z posterior vs prior kl loss")
        group.add_argument("--beta_kl_constraints_stop", default=1, type=float,
                           help="max beta for all the latent z posterior vs prior kl loss")
        group.add_argument("--beta_n_cycles", default=30, type=int,
                           help="number of cycles for kl loss ratio within an epoch")
        group.add_argument("--freebit_kl_constraints", default=.1, type=float,
                           help="free bit for all the latent z kl loss")
        group.add_argument("--latent_dim", default=256, type=int,
                           help="latent dimension of deepVAE Z")
        group.add_argument("--learning_rate", default=5e-5, type=float,
                           help="The initial learning rate for Adam.")
        group.add_argument("--weight_decay", default=0.0, type=float,
                           help="Weight deay if we apply some.")
        group.add_argument("--adam_epsilon", default=1e-8, type=float,
                           help="Epsilon for Adam optimizer.")
        group.add_argument("--max_grad_norm", default=1.0, type=float,
                           help="Max gradient norm.")
        group.add_argument("--warmup_steps", default=0, type=int,
                           help="Linear warmup over warmup_steps.")
        group.add_argument("--CVAE", action='store_true',
                           help="specify this argument if finetuning CVAE, otherwise ignore this argument")

        return parser

    @classmethod
    def load_model(cls, args, labels_dict=None):
        checkpoint = torch.load(os.path.join(args.checkpoint_path, 'mp_rank_00_model_states.pt'))

        latent_dim = checkpoint['latent_dim'] if ('latent_dim' in checkpoint.keys()) else args.latent_dim
        labels_dict = checkpoint['label_dict'] if ('label_dict' in checkpoint.keys()) else labels_dict

        enc_config = GPT2Config.from_pretrained(args.gpt2_model_path)
        tokenizer = BertTokenizer.from_pretrained(args.gpt2_model_path)
        special_tokens_dict = {'bos_token': '<BOS>', 'eos_token': '<EOS>'}
        # special_tokens_dict = {'bos_token': '<BOS>', 'eos_token': '<EOS>', 'additional_special_tokens': ['<ENT>', '<ENS>']}
        tokenizer.add_special_tokens(special_tokens_dict)
        encoder_model = GPT2ForEncoderLatentConnector(config=enc_config)
        encoder_model.resize_token_embeddings(len(tokenizer))

        dec_config = GPT2Config.from_pretrained(args.gpt2_model_path)
        decoder_model = GPT2ForDecoderLatentConnector(config=dec_config, latent_dim=latent_dim)
        decoder_model.resize_token_embeddings(len(tokenizer))

        vae_model = DeepVAE(encoder_model, decoder_model, latent_dim=latent_dim,
                            hidden_dim=enc_config.hidden_size, layer_num=enc_config.num_hidden_layers,
                            pad_token_id=tokenizer.pad_token_id, unk_token_id=tokenizer.unk_token_id,
                            bos_token_id=tokenizer.bos_token_id, eos_token_id=tokenizer.eos_token_id,
                            CVAE=args.CVAE)

        # TODO: all the related params should be loaded here! Including latent_nets, posterior_nets, prior_nets, pooling, decoder.transformer.Wv, decoder.transformer.Wz
        anchor = 'module.model.'
        start = len(anchor)
        vae_dict = {key[start:]: val for key, val in checkpoint['module'].items() if anchor in key}
        # comment out if not initialized from VAE
        # if args.CVAE:
        #     # manually load prior and posterior if initialize CVAE model for the first time because of dim mismatch
        #     prior_post_dict = {key: vae_dict.pop(key) for key in list(vae_dict) if ('posterior_nets' in key or 'prior_nets' in key)}
        #     for idx in range(enc_config.num_hidden_layers):
        #         vae_model.posterior_nets[idx].weight.data[:, enc_config.hidden_size:] = prior_post_dict[f"posterior_nets.{idx}.weight"]
        #         vae_model.prior_nets[idx].weight.data[:, enc_config.hidden_size:] = prior_post_dict[f"prior_nets.{idx}.weight"]
        #     enc_wte_shape, dec_wte_shape  = vae_dict['encoder.transformer.wte.weight'].shape[0], vae_dict['decoder.transformer.wte.weight'].shape[0]
        #     vae_model.encoder.transformer.wte.weight.data[:enc_wte_shape, :] = vae_dict.pop('encoder.transformer.wte.weight')
        #     vae_model.decoder.transformer.wte.weight.data[:dec_wte_shape, :] = vae_dict.pop('decoder.transformer.wte.weight')
        #     vae_model.decoder.lm_head.weight.data[:dec_wte_shape, :] = vae_dict.pop('decoder.lm_head.weight')
        missing_keys, unexpected_keys = vae_model.load_state_dict(vae_dict, strict=False)
        print(f"Vae model loading process: missing keys {missing_keys}, unexpected keys {unexpected_keys}")

        return vae_model, tokenizer

    def __init__(
        self,
        args,
        train_steps=0,
        labels_dict=None
    ):
        super().__init__()
        # self.save_hyperparameters()
        self.args = args

        if args.checkpoint_path is not None:
            self.model, self.encoder_tokenizer, self.decoder_tokenizer, self.latent_dim, \
                self.labels_dict, self.args = DeepVAEModule.load_model(self.args, labels_dict=labels_dict)
        else:
            self.encoder_tokenizer = BertTokenizer.from_pretrained(self.args.encoder_model_path)
            encoder_config = GPT2Config.from_pretrained(self.args.encoder_model_path)
            special_tokens_dict = {'bos_token': '<BOS>', 'eos_token': '<EOS>', 'additional_special_tokens': ['<ENT>', '<ENS>']}
            self.encoder_tokenizer.add_special_tokens(special_tokens_dict)
            self.latent_dim = self.args.latent_dim
            encoder = GPT2ForEncoderLatentConnector.from_pretrained(self.args.encoder_model_path, config=encoder_config)
            # Notice: resize_token_embeddings expect to receive the full size of the new vocabulary, i.e. the length of the tokenizer.
            encoder.resize_token_embeddings(len(self.encoder_tokenizer))

            self.decoder_tokenizer = BertTokenizer.from_pretrained(self.args.decoder_model_path)
            self.decoder_tokenizer.add_special_tokens(special_tokens_dict)
            decoder_config = GPT2Config.from_pretrained(self.args.decoder_model_path)
            self.labels_dict = labels_dict
            decoder = GPT2ForDecoderLatentConnector.from_pretrained(self.args.decoder_model_path, config=decoder_config,
                                                                    latent_dim=self.latent_dim)

            # Notice: resize_token_embeddings expect to receive the full size of the new vocabulary, i.e. the length of the tokenizer.
            decoder.resize_token_embeddings(len(self.decoder_tokenizer))
            self.model = DeepVAE(encoder, decoder, latent_dim=self.args.latent_dim,
                                 hidden_dim=encoder_config.hidden_size, layer_num=encoder_config.num_hidden_layers,
                                 pad_token_id=self.decoder_tokenizer.pad_token_id, unk_token_id=self.decoder_tokenizer.unk_token_id,
                                 bos_token_id=self.decoder_tokenizer.bos_token_id, eos_token_id=self.decoder_tokenizer.eos_token_id,
                                 CVAE=args.CVAE)

        self.train_steps = train_steps
        # TODO: adjust the cyclic schedule
        self.beta_kl_constraints_list = self.get_cyclic_linear_beta_list(self.train_steps,
                                                                         start=args.beta_kl_constraints_start, stop=args.beta_kl_constraints_stop,  n_cycle=args.beta_n_cycles)
        # self.mlm_probability_list = self.get_decoder_beta_list(self.train_steps,
        #     start=0., stop=1.,  n_cycle=args.beta_n_cycles)
        # self.beta_kl_constraints_list = self.get_constant_ratio(self.train_steps, args.beta_kl_constraints)
        self.mlm_probability_list = self.get_constant_ratio(self.train_steps, 0.)
        # self.freebit_kl_constraints = args.freebit_kl_constraints

    def get_constant_ratio(self, n_steps, ratio):
        L = np.ones(n_steps)
        L *= ratio
        return L

    def get_decoder_beta_list(self, n_steps, start=0., stop=1.0, n_cycle=4):
        L = np.ones(n_steps)
        t_range = int(n_steps / n_cycle)
        for t_cur in range(n_steps):
            if t_cur > t_range:
                L[t_cur] = 0.
            else:
                ratio = t_cur / t_range
                value = stop - ratio * (stop-start)
                L[t_cur] = value
        return L

    def get_cyclic_linear_beta_list(self, n_steps, start=0.5, stop=1.0, n_cycle=4):
        L = np.ones(n_steps)
        t_range = int(n_steps / n_cycle)
        for t_cur in range(n_steps):
            loc = t_cur % t_range
            split_range = int(t_range * 0.25)
            if loc <= 2*split_range:
                value = start
            elif loc <= 3*split_range:
                ratio = (loc % split_range) / split_range
                value = ratio * (stop-start)
            else:
                value = stop
            L[t_cur] = value
        return L

    #####
    # Torch lightning
    #####

    def on_save_checkpoint(self, checkpoint) -> None:
        checkpoint['label_dict'] = self.labels_dict
        checkpoint['latent_dim'] = self.latent_dim

    def training_step(self, batch, batch_idx):
        if batch is None:
            loss = torch.Tensor([0.]).to(next(self.model.parameters()).device)
            loss.requires_grad = True
            return loss
        inputs, cond_inputs = batch, None
        if self.args.CVAE:
            inputs, cond_inputs = batch

        total_loss, rec_loss, total_kl_loss, layer_kl_loss = \
            self.model(inputs, self.beta_kl_constraints_list[batch_idx], cond_inputs)
        # the logging interval are set by the trainer_args log_every_n_steps
        for idx, pg in enumerate(self.optimizers().param_groups):
            self.log(f"learning_rate_{idx}", pg['lr'])
        unscaled_kl_constraint_loss = 0. if self.beta_kl_constraints_list[batch_idx] == 0. else total_kl_loss/self.beta_kl_constraints_list[batch_idx]
        self.log("total_loss", total_loss)
        self.log("total_kl_constraint_loss", total_kl_loss)
        self.log("unscaled_kl_constraint_loss", unscaled_kl_constraint_loss)
        self.log("beta_kl_constraints", self.beta_kl_constraints_list[batch_idx])
        self.log("beta_mlm_probability", self.mlm_probability_list[batch_idx])
        self.log("rec_loss", rec_loss)
        for idx, kl_loss in enumerate(layer_kl_loss):
            self.log(f"layer_{idx}_kl_loss", kl_loss.mean())

        return total_loss

    def training_step_end(self, batch_parts):
        pass

    def training_epoch_end(self, outputs):
        pass

    def validation_step(self, batch, batch_idx):
        if batch is None:
            loss = torch.Tensor([0.]).to(next(self.model.parameters()).device)
            loss.requires_grad = True
            return loss
        inputs, cond_inputs = batch, None
        if self.args.CVAE:
            inputs, cond_inputs = batch

        total_loss, rec_loss, total_kl_loss, layer_kl_loss = self.model(inputs, 1., cond_inputs)
        # the logging interval are set by the trainer_args log_every_n_steps
        self.log("val_total_loss", total_loss)
        self.log("val_kl_constraint_loss", total_kl_loss)
        self.log("val_recon_loss", rec_loss)
        for idx, kl_loss in enumerate(layer_kl_loss):
            self.log(f"layer_{idx}_kl_loss", kl_loss.mean())
        return total_loss

    def validation_epoch_end(self, outputs):
        pass

    def test_step(self, batch, batch_idx):
        if batch is None:
            loss = torch.Tensor([0.]).to(next(self.model.parameters()).device)
            loss.requires_grad = True
            return loss
        inputs, cond_inputs = batch, None
        if self.args.CVAE:
            inputs, cond_inputs = batch
        total_loss, rec_loss, total_kl_loss, layer_kl_loss = self.model(inputs, 1., cond_inputs)
        self.log("test_total_loss", total_loss)
        self.log("test_recon_loss", rec_loss)
        self.log("test_kl_constraint_loss", total_kl_loss)
        for idx, kl_loss in enumerate(layer_kl_loss):
            self.log(f"layer_{idx}_kl_loss", kl_loss.mean())
        return total_loss

    def configure_optimizers(self):
        no_decay = ['bias', 'LayerNorm.weight']
        optimizer_grouped_parameters = [
            {'params': [p for n, p in self.model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': self.args.weight_decay},
            {'params': [p for n, p in self.model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]

        optimizer = AdamW(optimizer_grouped_parameters, lr=self.args.learning_rate, eps=self.args.adam_epsilon)
        scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=self.args.warmup_steps, num_training_steps=self.train_steps)

        return {'optimizer': optimizer,
                'lr_scheduler': {
                    'scheduler': scheduler,
                    'interval': 'step',
                    'frequency': 1
                }
                }