def apply_rotary(x, sinusoidal_pos=None):
    if sinusoidal_pos is None:
        return x
    sin, cos = sinusoidal_pos
    # x.shape [batch, seq_len, 2]
    x1, x2 = x[..., 0::2], x[..., 1::2]
    # [cos_nθ, -sin_nθ] [x1]
    # [sin_nθ,  cos_nθ] [x2]
    # => [x1 * cos_nθ - x2 * sin_nθ, x1 * sin_nθ + x2 * cos_nθ]
    # 苏神的rotary，使用了下面的计算方法。
    # return torch.stack([x1 * cos - x2 * sin, x1 * sin + x2 * cos], dim=-1).flatten(-2, -1)
    # 考虑到矩阵乘法torch.einsum("bhmd,bhnd->bhmn", q, k)，因此可以直接在最后一个维度拼接（无需奇偶交错）
    return torch.cat([x1 * cos - x2 * sin, x1 * sin + x2 * cos], dim=-1)

class RoFormerClassificationHead(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)

        self.config = config

    def forward(self, features, **kwargs):
        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
        x = self.dropout(x)
        x = self.dense(x)
        x = ACT2FN[self.config.hidden_act](x) # 这里是relu
        x = self.dropout(x)
        x = self.out_proj(x)
        return x

import torch
import numpy as np
from roformer import RoFormerForCausalLM, RoFormerConfig
from transformers import BertTokenizer

device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# 可选以下几个。
# junnyu/roformer_chinese_sim_char_small, junnyu/roformer_chinese_sim_char_base
# junnyu/roformer_chinese_sim_char_ft_small, roformer_chinese_sim_char_ft_base
pretrained_model = "junnyu/roformer_chinese_sim_char_base"
tokenizer = BertTokenizer.from_pretrained(pretrained_model)
config = RoFormerConfig.from_pretrained(pretrained_model)
config.is_decoder = True
config.eos_token_id = tokenizer.sep_token_id
config.pooler_activation = "linear"
model = RoFormerForCausalLM.from_pretrained(pretrained_model, config=config)
model.to(device)
model.eval()

def gen_synonyms(text, n=100, k=20):
    ''''含义： 产生sent的n个相似句，然后返回最相似的k个。
    做法：用seq2seq生成，并用encoder算相似度并排序。
    '''
    # 寻找所有相似的句子
    r = []
    inputs1 = tokenizer(text, return_tensors="pt")
    for _ in range(n):
        inputs1.to(device)
        output = tokenizer.batch_decode(model.generate(**inputs1, top_p=0.95, do_sample=True, max_length=128), skip_special_tokens=True)[0].replace(" ","").replace(text, "") # 去除空格，去除原始text文本。
        r.append(output)
    
    # 对相似的句子进行排序
    r = [i for i in set(r) if i != text and len(i) > 0]
    r = [text] + r
    inputs2 = tokenizer(r, padding=True, return_tensors="pt")
    with torch.no_grad():
        inputs2.to(device)
        outputs = model(**inputs2)
        Z = outputs.pooler_output.cpu().numpy()
    Z /= (Z**2).sum(axis=1, keepdims=True)**0.5
    argsort = np.dot(Z[1:], -Z[0]).argsort()
    
    return [r[i + 1] for i in argsort[:k]]

out = gen_synonyms("广州和深圳哪个好？")
print(out)
# ['深圳和广州哪个好？',
#  '广州和深圳哪个好',
#  '深圳和广州哪个好',
#  '深圳和广州哪个比较好。',
#  '深圳和广州哪个最好？',
#  '深圳和广州哪个比较好',
#  '广州和深圳那个比较好',
#  '深圳和广州哪个更好？',
#  '深圳与广州哪个好',
#  '深圳和广州，哪个比较好',
#  '广州与深圳比较哪个好',
#  '深圳和广州哪里比较好',
#  '深圳还是广州比较好？',
#  '广州和深圳哪个地方好一些？',
#  '广州好还是深圳好？',
#  '广州好还是深圳好呢？',
#  '广州与深圳哪个地方好点？',
#  '深圳好还是广州好',
#  '广州好还是深圳好',
#  '广州和深圳哪个城市好？']

import torch
import tensorflow as tf
from transformers import BertTokenizer
from roformer import RoFormerForMaskedLM, TFRoFormerForMaskedLM

text = "今天[MASK]很好，我[MASK]去公园玩。"
tokenizer = BertTokenizer.from_pretrained("junnyu/roformer_v2_chinese_char_base")
pt_model = RoFormerForMaskedLM.from_pretrained("junnyu/roformer_v2_chinese_char_base")
tf_model = TFRoFormerForMaskedLM.from_pretrained(
    "junnyu/roformer_v2_chinese_char_base", from_pt=True
)
pt_inputs = tokenizer(text, return_tensors="pt")
tf_inputs = tokenizer(text, return_tensors="tf")
# pytorch
with torch.no_grad():
    pt_outputs = pt_model(**pt_inputs).logits[0]
pt_outputs_sentence = "pytorch: "
for i, id in enumerate(tokenizer.encode(text)):
    if id == tokenizer.mask_token_id:
        tokens = tokenizer.convert_ids_to_tokens(pt_outputs[i].topk(k=5)[1])
        pt_outputs_sentence += "[" + "||".join(tokens) + "]"
    else:
        pt_outputs_sentence += "".join(
            tokenizer.convert_ids_to_tokens([id], skip_special_tokens=True)
        )
print(pt_outputs_sentence)
# tf
tf_outputs = tf_model(**tf_inputs, training=False).logits[0]
tf_outputs_sentence = "tf: "
for i, id in enumerate(tokenizer.encode(text)):
    if id == tokenizer.mask_token_id:
        tokens = tokenizer.convert_ids_to_tokens(tf.math.top_k(tf_outputs[i], k=5)[1])
        tf_outputs_sentence += "[" + "||".join(tokens) + "]"
    else:
        tf_outputs_sentence += "".join(
            tokenizer.convert_ids_to_tokens([id], skip_special_tokens=True)
        )
print(tf_outputs_sentence)
# small
# pytorch: 今天[的||，||是||很||也]很好，我[要||会||是||想||在]去公园玩。
# tf: 今天[的||，||是||很||也]很好，我[要||会||是||想||在]去公园玩。
# base
# pytorch: 今天[我||天||晴||园||玩]很好，我[想||要||会||就||带]去公园玩。
# tf: 今天[我||天||晴||园||玩]很好，我[想||要||会||就||带]去公园玩。
# large
# pytorch: 今天[天||气||我||空||阳]很好，我[又||想||会||就||爱]去公园玩。
# tf: 今天[天||气||我||空||阳]很好，我[又||想||会||就||爱]去公园玩。

import torch
import tensorflow as tf
from transformers import RoFormerForMaskedLM, RoFormerTokenizer, TFRoFormerForMaskedLM

text = "今天[MASK]很好，我[MASK]去公园玩。"
tokenizer = RoFormerTokenizer.from_pretrained("junnyu/roformer_chinese_base")
pt_model = RoFormerForMaskedLM.from_pretrained("junnyu/roformer_chinese_base")
tf_model = TFRoFormerForMaskedLM.from_pretrained(
    "junnyu/roformer_chinese_base", from_pt=True
)
pt_inputs = tokenizer(text, return_tensors="pt")
tf_inputs = tokenizer(text, return_tensors="tf")
# pytorch
with torch.no_grad():
    pt_outputs = pt_model(**pt_inputs).logits[0]
pt_outputs_sentence = "pytorch: "
for i, id in enumerate(tokenizer.encode(text)):
    if id == tokenizer.mask_token_id:
        tokens = tokenizer.convert_ids_to_tokens(pt_outputs[i].topk(k=5)[1])
        pt_outputs_sentence += "[" + "||".join(tokens) + "]"
    else:
        pt_outputs_sentence += "".join(
            tokenizer.convert_ids_to_tokens([id], skip_special_tokens=True)
        )
print(pt_outputs_sentence)
# tf
tf_outputs = tf_model(**tf_inputs, training=False).logits[0]
tf_outputs_sentence = "tf: "
for i, id in enumerate(tokenizer.encode(text)):
    if id == tokenizer.mask_token_id:
        tokens = tokenizer.convert_ids_to_tokens(tf.math.top_k(tf_outputs[i], k=5)[1])
        tf_outputs_sentence += "[" + "||".join(tokens) + "]"
    else:
        tf_outputs_sentence += "".join(
            tokenizer.convert_ids_to_tokens([id], skip_special_tokens=True)
        )
print(tf_outputs_sentence)
# pytorch: 今天[天气||天||心情||阳光||空气]很好，我[想||要||打算||准备||喜欢]去公园玩。
# tf:      今天[天气||天||心情||阳光||空气]很好，我[想||要||打算||准备||喜欢]去公园玩。

small版本
bert4keras vs pytorch
mean diff : tensor(5.9108e-07)
max diff : tensor(5.7220e-06)
bert4keras vs tf2.0
mean diff : tensor(4.5976e-07)
max diff : tensor(3.5763e-06)

base版本
python compare_model.py
bert4keras vs pytorch
mean diff : tensor(4.3340e-07)
max diff : tensor(5.7220e-06)
bert4keras vs tf2.0
mean diff : tensor(3.4319e-07)
max diff : tensor(5.2452e-06)