Longformer Torch2Paddle

PaddlePaddle-Longformer-model-base-4096

PyTorch	Shape	Paddle	Shape
embeddings.word_embeddings.weight	[50265, 768]	embeddings.word_embeddings.weight
embeddings.position_embeddings.weight	[4098, 768]	embeddings.position_embeddings.weight
embeddings.token_type_embeddings.weight	[1, 768]	embeddings.token_type_embeddings.weight
embeddings.LayerNorm.weight	[768]	embeddings.layer_norm.weight
embeddings.LayerNorm.bias	[768]	embeddings.layer_norm.bias
encoder.layer.0.attention.self.query.weight	[768, 768]	encoder.layers.0.self_attn.query.weight	T
encoder.layer.0.attention.self.query.bias	[768]	encoder.layers.0.self_attn.query.bias
encoder.layer.0.attention.self.key.weight	[768, 768]	encoder.layers.0.self_attn.key.weight	T
encoder.layer.0.attention.self.key.bias	[768]	encoder.layers.0.self_attn.key.bias
encoder.layer.0.attention.self.value.weight	[768, 768]	encoder.layers.0.self_attn.value.weight	T
encoder.layer.0.attention.self.value.bias	[768]	encoder.layers.0.self_attn.value.bias
encoder.layer.0.attention.self.query_global.weight	[768, 768]	encoder.layers.0.self_attn.query_global.weight	T
encoder.layer.0.attention.self.query_global.bias	[768]	encoder.layers.0.self_attn.query_global.bias
encoder.layer.0.attention.self.key_global.weight	[768, 768]	encoder.layers.0.self_attn.key_global.weight	T
encoder.layer.0.attention.self.key_global.bias	[768]	encoder.layers.0.self_attn.key_global.bias
encoder.layer.0.attention.self.value_global.weight	[768, 768]	encoder.layers.0.self_attn.value_global.weight	T
encoder.layer.0.attention.self.value_global.bias	[768]	encoder.layers.0.self_attn.value_global.bias
encoder.layer.0.attention.output.dense.weight	[768, 768]	encoder.layers.0.self_attn.out.weight	T
encoder.layer.0.attention.output.dense.bias	[768]	encoder.layers.0.self_attn.out.bias
encoder.layer.0.attention.output.LayerNorm.weight	[768]	encoder.layers.0.norm1.weight
encoder.layer.0.attention.output.LayerNorm.bias	[768]	encoder.layers.0.norm1.bias
encoder.layer.0.intermediate.dense.weight	[3072, 768]	encoder.layers.0.linear1.weight	T [768, 3072]
encoder.layer.0.intermediate.dense.bias	[3072]	encoder.layers.0.linear1.bias
encoder.layer.0.output.dense.weight	[768, 3072]	encoder.layers.0.linear2.weight	T [3072, 768]
encoder.layer.0.output.dense.bias	[768]	encoder.layers.0.linear2.bias
encoder.layer.0.output.LayerNorm.weight	[768]	encoder.layers.0.norm2.weight
encoder.layer.0.output.LayerNorm.bias	[768]	encoder.layers.0.norm2.bias
pooler.dense.weight	[768, 768]	pooler.dense.weight	T
pooler.dense.bias	[768]	pooler.dense.bias

more ...

Pytorch tensor.stride & tensor.as_strided

tensor.stride()¶

Stride is the jump necessary to go from one element to the next one in the specified dimension dim.

一个元素到另一个元素，元素粒度

任意维度上的步长，是其低维度乘积。

shape: (12, 512, 768) stride: (512x768x1, 768x1, 1x1)

tensor.as_strided()¶

input (Tensor) – the input tensor.
size (tuple or ints) – the shape of the output tensor
stride (tuple or ints) – the stride of the output tensor
more ...

Pytorch View vs Reshape

torch.view has existed for a long time. It will return a tensor with the new shape. The returned tensor will share the underling data with the original tensor. See the documentation here.

On the other hand, it seems that torch.reshape has been introduced recently in version 0.4. According to the document more ...

pytorch.nn.functional.pad

torch¶

torch.nn.functional.pad

从输入input的最后一个维度向前padding

输入input的$\left\lfloor\frac{\text{len(pad)}}{2}\right\rfloor$个维度进行padding

如果只padding输入张量input的最后1个维度，pad的形式如：(padding_left, padding_right)
如果只padding输入张量input的最后2个维度，pad的形式如：(padding_left, padding_right, padding_top, padding_bottom)
如果只padding输入张量input的最后3个维度，pad的形式如：(padding_left, padding_right, padding_top, padding_bottom, padding_front, padding_back)

more ...

Tokenizer offerset mapping

more ...

paddle implements torch.repeat_interleave/K.repeat_elements using paddle.reshape & paddle.tile

https://pytorch.org/docs/stable/generated/torch.Tensor.repeat.html#torch.Tensor.repeat

https://pytorch.org/docs/stable/generated/torch.repeat_interleave.html

If the repeats is tensor([n1, n2, n3, …]), then the output will be tensor([0, 0, …, 1, 1, …, 2, 2, …, …]) where 0 appears n1 times, 1 appears n2 times, 2 appears n3 times, etc.