Commit
•
cf42adb
1
Parent(s):
ceaa103
Upload P5Pretraining
Browse files- config.json +67 -0
- pretrain_model.py +133 -0
- pytorch_model.bin +3 -0
config.json
ADDED
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "t5-small",
|
3 |
+
"activation_dropout": 0.1,
|
4 |
+
"architectures": [
|
5 |
+
"P5Pretraining"
|
6 |
+
],
|
7 |
+
"attention_dropout": 0.1,
|
8 |
+
"auto_map": {
|
9 |
+
"AutoModel": "pretrain_model.P5Pretraining"
|
10 |
+
},
|
11 |
+
"d_ff": 2048,
|
12 |
+
"d_kv": 64,
|
13 |
+
"d_model": 512,
|
14 |
+
"decoder_start_token_id": 0,
|
15 |
+
"dense_act_fn": "relu",
|
16 |
+
"dropout": 0.1,
|
17 |
+
"dropout_rate": 0.1,
|
18 |
+
"eos_token_id": 1,
|
19 |
+
"feed_forward_proj": "relu",
|
20 |
+
"initializer_factor": 1.0,
|
21 |
+
"is_encoder_decoder": true,
|
22 |
+
"is_gated_act": false,
|
23 |
+
"layer_norm_epsilon": 1e-06,
|
24 |
+
"losses": "rating,sequential,explanation,review,traditional",
|
25 |
+
"model_type": "t5",
|
26 |
+
"n_positions": 512,
|
27 |
+
"num_decoder_layers": 6,
|
28 |
+
"num_heads": 8,
|
29 |
+
"num_layers": 6,
|
30 |
+
"output_past": true,
|
31 |
+
"pad_token_id": 0,
|
32 |
+
"relative_attention_max_distance": 128,
|
33 |
+
"relative_attention_num_buckets": 32,
|
34 |
+
"task_specific_params": {
|
35 |
+
"summarization": {
|
36 |
+
"early_stopping": true,
|
37 |
+
"length_penalty": 2.0,
|
38 |
+
"max_length": 200,
|
39 |
+
"min_length": 30,
|
40 |
+
"no_repeat_ngram_size": 3,
|
41 |
+
"num_beams": 4,
|
42 |
+
"prefix": "summarize: "
|
43 |
+
},
|
44 |
+
"translation_en_to_de": {
|
45 |
+
"early_stopping": true,
|
46 |
+
"max_length": 300,
|
47 |
+
"num_beams": 4,
|
48 |
+
"prefix": "translate English to German: "
|
49 |
+
},
|
50 |
+
"translation_en_to_fr": {
|
51 |
+
"early_stopping": true,
|
52 |
+
"max_length": 300,
|
53 |
+
"num_beams": 4,
|
54 |
+
"prefix": "translate English to French: "
|
55 |
+
},
|
56 |
+
"translation_en_to_ro": {
|
57 |
+
"early_stopping": true,
|
58 |
+
"max_length": 300,
|
59 |
+
"num_beams": 4,
|
60 |
+
"prefix": "translate English to Romanian: "
|
61 |
+
}
|
62 |
+
},
|
63 |
+
"torch_dtype": "float32",
|
64 |
+
"transformers_version": "4.25.1",
|
65 |
+
"use_cache": true,
|
66 |
+
"vocab_size": 32100
|
67 |
+
}
|
pretrain_model.py
ADDED
@@ -0,0 +1,133 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import numpy as np
|
5 |
+
|
6 |
+
from modeling_p5 import P5
|
7 |
+
|
8 |
+
class P5Pretraining(P5):
|
9 |
+
def __init__(self, config):
|
10 |
+
super().__init__(config)
|
11 |
+
|
12 |
+
self.losses = self.config.losses.split(',')
|
13 |
+
|
14 |
+
def train_step(self, batch):
|
15 |
+
|
16 |
+
device = next(self.parameters()).device
|
17 |
+
input_ids = batch['input_ids'].to(device)
|
18 |
+
whole_word_ids = batch['whole_word_ids'].to(device)
|
19 |
+
|
20 |
+
lm_labels = batch["target_ids"].to(device)
|
21 |
+
|
22 |
+
loss_weights = batch["loss_weights"].to(device)
|
23 |
+
|
24 |
+
output = self(
|
25 |
+
input_ids=input_ids,
|
26 |
+
whole_word_ids=whole_word_ids,
|
27 |
+
labels=lm_labels,
|
28 |
+
return_dict=True
|
29 |
+
)
|
30 |
+
assert 'loss' in output
|
31 |
+
|
32 |
+
lm_mask = lm_labels != -100
|
33 |
+
lm_mask = lm_mask.float()
|
34 |
+
B, L = lm_labels.size()
|
35 |
+
|
36 |
+
loss = output['loss']
|
37 |
+
|
38 |
+
loss = loss.view(B, L) * lm_mask
|
39 |
+
|
40 |
+
loss = loss.sum(dim=1) / lm_mask.sum(dim=1).clamp(min=1)
|
41 |
+
|
42 |
+
task_counts = {task: 0 for task in self.losses}
|
43 |
+
task_loss = {task: 0 for task in self.losses}
|
44 |
+
|
45 |
+
results = {}
|
46 |
+
|
47 |
+
results['loss'] = (loss * loss_weights).mean()
|
48 |
+
results['total_loss'] = loss.detach().sum()
|
49 |
+
results['total_loss_count'] = len(loss)
|
50 |
+
|
51 |
+
task_counts = {task: 0 for task in self.losses}
|
52 |
+
task_loss = {task: 0 for task in self.losses}
|
53 |
+
|
54 |
+
for _loss, task in zip(loss.detach(), batch['task']):
|
55 |
+
task_loss[task] += _loss
|
56 |
+
task_counts[task] += 1
|
57 |
+
|
58 |
+
for task in self.losses:
|
59 |
+
if task_counts[task] > 0:
|
60 |
+
results[f'{task}_loss'] = task_loss[task]
|
61 |
+
results[f'{task}_loss_count'] = task_counts[task]
|
62 |
+
|
63 |
+
return results
|
64 |
+
|
65 |
+
@torch.no_grad()
|
66 |
+
def valid_step(self, batch):
|
67 |
+
self.eval()
|
68 |
+
device = next(self.parameters()).device
|
69 |
+
input_ids = batch['input_ids'].to(device)
|
70 |
+
|
71 |
+
lm_labels = batch["target_ids"].to(device)
|
72 |
+
|
73 |
+
loss_weights = batch["loss_weights"].to(device)
|
74 |
+
|
75 |
+
output = self(
|
76 |
+
input_ids=input_ids,
|
77 |
+
labels=lm_labels,
|
78 |
+
return_dict=True
|
79 |
+
)
|
80 |
+
assert 'loss' in output
|
81 |
+
|
82 |
+
lm_mask = lm_labels != -100
|
83 |
+
lm_mask = lm_mask.float()
|
84 |
+
B, L = lm_labels.size()
|
85 |
+
|
86 |
+
loss = output['loss']
|
87 |
+
|
88 |
+
loss = loss.view(B, L) * lm_mask
|
89 |
+
|
90 |
+
loss = loss.sum(dim=1) / lm_mask.sum(dim=1).clamp(min=1)
|
91 |
+
|
92 |
+
results = {}
|
93 |
+
|
94 |
+
results['loss'] = (loss * loss_weights).mean()
|
95 |
+
results['total_loss'] = loss.detach().sum()
|
96 |
+
results['total_loss_count'] = len(loss)
|
97 |
+
|
98 |
+
task_counts = {task: 0 for task in self.losses}
|
99 |
+
task_loss = {task: 0 for task in self.losses}
|
100 |
+
|
101 |
+
for _loss, task in zip(loss.detach(), batch['task']):
|
102 |
+
task_loss[task] += _loss
|
103 |
+
task_counts[task] += 1
|
104 |
+
|
105 |
+
for task in self.losses:
|
106 |
+
if task_counts[task] > 0:
|
107 |
+
results[f'{task}_loss'] = task_loss[task]
|
108 |
+
results[f'{task}_loss_count'] = task_counts[task]
|
109 |
+
|
110 |
+
if 'rating' in self.losses:
|
111 |
+
output = self.generate(
|
112 |
+
input_ids=input_ids
|
113 |
+
)
|
114 |
+
|
115 |
+
generated_score = self.tokenizer.batch_decode(output, skip_special_tokens=True)
|
116 |
+
|
117 |
+
results['rating_pred'] = generated_score
|
118 |
+
|
119 |
+
return results
|
120 |
+
|
121 |
+
@torch.no_grad()
|
122 |
+
def generate_step(self, batch):
|
123 |
+
self.eval()
|
124 |
+
device = next(self.parameters()).device
|
125 |
+
input_ids = batch['input_ids'].to(device)
|
126 |
+
|
127 |
+
output = self.generate(
|
128 |
+
input_ids=input_ids,
|
129 |
+
)
|
130 |
+
|
131 |
+
generated_sents = self.tokenizer.batch_decode(output, skip_special_tokens=True)
|
132 |
+
|
133 |
+
return generated_sents
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:1077d7f9d23ba28012cef084386d83491f9fe01da05089780da3530ac74123d8
|
3 |
+
size 243063197
|