mazpie's picture
Initial commit
2d9a728
raw
history blame
11.9 kB
import logging
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from scipy import interpolate
from typing import List
from torch import nn
logger = logging.getLogger(__name__)
def load_temp_embed_with_mismatch(temp_embed_old, temp_embed_new, add_zero=True):
"""
Add/Remove extra temporal_embeddings as needed.
https://arxiv.org/abs/2104.00650 shows adding zero paddings works.
temp_embed_old: (1, num_frames_old, 1, d)
temp_embed_new: (1, num_frames_new, 1, d)
add_zero: bool, if True, add zero, else, interpolate trained embeddings.
"""
# TODO zero pad
num_frms_new = temp_embed_new.shape[1]
num_frms_old = temp_embed_old.shape[1]
logger.info(f"Load temporal_embeddings, lengths: {num_frms_old}-->{num_frms_new}")
if num_frms_new > num_frms_old:
if add_zero:
temp_embed_new[
:, :num_frms_old
] = temp_embed_old # untrained embeddings are zeros.
else:
temp_embed_new = interpolate_temporal_pos_embed(temp_embed_old, num_frms_new)
elif num_frms_new < num_frms_old:
temp_embed_new = temp_embed_old[:, :num_frms_new]
else: # =
temp_embed_new = temp_embed_old
return temp_embed_new
def interpolate_temporal_pos_embed(temp_embed_old, num_frames_new):
"""
temp_embed_old: (1, num_frames_old, 1, d)
Returns:
temp_embed_new: (1, num_frames_new, 1, d)
"""
temp_embed_old = temp_embed_old.squeeze(2).permute(
0, 2, 1
) # (1, d, num_frames_old)
temp_embed_new = F.interpolate(
temp_embed_old, num_frames_new, mode="linear"
) # (1, d, num_frames_new)
temp_embed_new = temp_embed_new.permute(0, 2, 1).unsqueeze(
2
) # (1, num_frames_new, 1, d)
return temp_embed_new
def interpolate_pos_embed(pos_embed_old, pos_embed_new, num_patches_new):
"""
Args:
pos_embed_old: (1, L_old, d), pre-trained
pos_embed_new: (1, L_new, d), newly initialized, to be replaced by interpolated weights
num_patches_new:
"""
# interpolate position embedding
embedding_size = pos_embed_old.shape[-1]
num_extra_tokens = pos_embed_new.shape[-2] - num_patches_new
# height (== width) for the checkpoint position embedding
orig_size = int((pos_embed_old.shape[-2] - num_extra_tokens) ** 0.5)
# height (== width) for the new position embedding
new_size = int(num_patches_new ** 0.5)
if orig_size != new_size:
# class_token and dist_token are kept unchanged
# the extra tokens seems always at the beginning of the position embedding
extra_tokens = pos_embed_old[:, :num_extra_tokens]
# only the position tokens are interpolated
pos_tokens = pos_embed_old[:, num_extra_tokens:]
pos_tokens = pos_tokens.reshape(
-1, orig_size, orig_size, embedding_size
).permute(0, 3, 1, 2)
pos_tokens = torch.nn.functional.interpolate(
pos_tokens, size=(new_size, new_size), mode="bicubic", align_corners=False
)
pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
interpolated_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
logger.info(f"reshape position embedding from {orig_size}**2 to {new_size}**2")
return interpolated_pos_embed
else:
return pos_embed_old
def interpolate_pos_relative_bias_beit(state_dict_old, state_dict_new, patch_shape_new):
"""
Args:
state_dict_old: loaded state dict
state_dict_new: state dict for model with new image size
patch_shape_new: new model patch_shape
ref: https://github.com/microsoft/unilm/blob/master/beit/run_class_finetuning.py
"""
all_keys = list(state_dict_old.keys())
for key in all_keys:
if "relative_position_index" in key:
state_dict_old.pop(key)
if "relative_position_bias_table" in key:
rel_pos_bias = state_dict_old[key]
src_num_pos, num_attn_heads = rel_pos_bias.size()
dst_num_pos, _ = state_dict_new[key].size()
dst_patch_shape = patch_shape_new
if dst_patch_shape[0] != dst_patch_shape[1]:
raise NotImplementedError()
num_extra_tokens = dst_num_pos - (dst_patch_shape[0] * 2 - 1) * (
dst_patch_shape[1] * 2 - 1
)
src_size = int((src_num_pos - num_extra_tokens) ** 0.5)
dst_size = int((dst_num_pos - num_extra_tokens) ** 0.5)
if src_size != dst_size:
# logger.info("Position interpolate for %s from %dx%d to %dx%d" % (
# key, src_size, src_size, dst_size, dst_size))
extra_tokens = rel_pos_bias[-num_extra_tokens:, :]
rel_pos_bias = rel_pos_bias[:-num_extra_tokens, :]
def geometric_progression(a, r, n):
return a * (1.0 - r ** n) / (1.0 - r)
left, right = 1.01, 1.5
while right - left > 1e-6:
q = (left + right) / 2.0
gp = geometric_progression(1, q, src_size // 2)
if gp > dst_size // 2:
right = q
else:
left = q
# if q > 1.090307:
# q = 1.090307
dis = []
cur = 1
for i in range(src_size // 2):
dis.append(cur)
cur += q ** (i + 1)
r_ids = [-_ for _ in reversed(dis)]
x = r_ids + [0] + dis
y = r_ids + [0] + dis
t = dst_size // 2.0
dx = np.arange(-t, t + 0.1, 1.0)
dy = np.arange(-t, t + 0.1, 1.0)
# logger.info("Original positions = %s" % str(x))
# logger.info("Target positions = %s" % str(dx))
all_rel_pos_bias = []
for i in range(num_attn_heads):
z = rel_pos_bias[:, i].view(src_size, src_size).float().numpy()
f = interpolate.interp2d(x, y, z, kind="cubic")
all_rel_pos_bias.append(
torch.Tensor(f(dx, dy))
.contiguous()
.view(-1, 1)
.to(rel_pos_bias.device)
)
rel_pos_bias = torch.cat(all_rel_pos_bias, dim=-1)
new_rel_pos_bias = torch.cat((rel_pos_bias, extra_tokens), dim=0)
state_dict_old[key] = new_rel_pos_bias
return state_dict_old
def tile(x, dim, n_tile):
init_dim = x.size(dim)
repeat_idx = [1] * x.dim()
repeat_idx[dim] = n_tile
x = x.repeat(*repeat_idx)
order_index = torch.LongTensor(
np.concatenate([init_dim * np.arange(n_tile) + i for i in range(init_dim)])
)
return torch.index_select(x, dim, order_index.to(x.device))
def mask_logits(target, mask):
return target * mask + (1 - mask) * (-1e10)
class AllGather(torch.autograd.Function):
"""An autograd function that performs allgather on a tensor."""
@staticmethod
def forward(ctx, tensor, args):
output = [torch.empty_like(tensor) for _ in range(args.world_size)]
torch.distributed.all_gather(output, tensor)
ctx.rank = args.rank
ctx.batch_size = tensor.shape[0]
return torch.cat(output, dim=0)
@staticmethod
def backward(ctx, grad_output):
return (
grad_output[ctx.batch_size * ctx.rank : ctx.batch_size * (ctx.rank + 1)],
None,
)
allgather_wgrad = AllGather.apply
def tie_encoder_decoder_weights(
encoder: nn.Module, decoder: nn.Module, base_model_prefix: str, skip_key: str
):
uninitialized_encoder_weights: List[str] = []
if decoder.__class__ != encoder.__class__:
if issubclass(decoder.__class__, encoder.__class__):
logger.info(
f"decoder ({decoder.__class__}) and encoder ({encoder.__class__}) are not equal, encoder is decoder's father. In this case make sure that all encoder weights are correctly initialized."
)
elif issubclass(encoder.__class__, decoder.__class__):
logger.info(
f"decoder ({decoder.__class__}) and encoder ({encoder.__class__}) are not equal, decoder is encoder's father. In this case make sure that all encoder weights are correctly initialized."
)
else:
raise ValueError(f"decoder ({decoder.__class__}) and encoder ({encoder.__class__}) are not equal!!!")
def tie_encoder_to_decoder_recursively(
decoder_pointer: nn.Module,
encoder_pointer: nn.Module,
module_name: str,
uninitialized_encoder_weights: List[str],
skip_key: str,
depth=0,
):
assert isinstance(decoder_pointer, nn.Module) and isinstance(
encoder_pointer, nn.Module
), f"{decoder_pointer} and {encoder_pointer} have to be of type torch.nn.Module"
if hasattr(decoder_pointer, "weight") and skip_key not in module_name:
assert hasattr(encoder_pointer, "weight")
encoder_pointer.weight = decoder_pointer.weight
if hasattr(decoder_pointer, "bias"):
assert hasattr(encoder_pointer, "bias")
encoder_pointer.bias = decoder_pointer.bias
logger.info(module_name + " is tied")
return
encoder_modules = encoder_pointer._modules
decoder_modules = decoder_pointer._modules
if len(decoder_modules) > 0:
assert (
len(encoder_modules) > 0
), f"Encoder module {encoder_pointer} does not match decoder module {decoder_pointer}"
all_encoder_weights = set(
[module_name + "/" + sub_name for sub_name in encoder_modules.keys()]
)
encoder_layer_pos = 0
for name, module in decoder_modules.items():
if name.isdigit():
encoder_name = str(int(name) + encoder_layer_pos)
decoder_name = name
if not isinstance(
decoder_modules[decoder_name],
type(encoder_modules[encoder_name]),
) and len(encoder_modules) != len(decoder_modules):
# this can happen if the name corresponds to the position in a list module list of layers
# in this case the decoder has added a cross-attention that the encoder does not have
# thus skip this step and subtract one layer pos from encoder
encoder_layer_pos -= 1
continue
elif name not in encoder_modules:
continue
elif depth > 500:
raise ValueError(
"Max depth of recursive function `tie_encoder_to_decoder` reached. It seems that there is a circular dependency between two or more `nn.Modules` of your model."
)
else:
decoder_name = encoder_name = name
tie_encoder_to_decoder_recursively(
decoder_modules[decoder_name],
encoder_modules[encoder_name],
module_name + "/" + name,
uninitialized_encoder_weights,
skip_key,
depth=depth + 1,
)
all_encoder_weights.remove(module_name + "/" + encoder_name)
uninitialized_encoder_weights += list(all_encoder_weights)
# tie weights recursively
tie_encoder_to_decoder_recursively(
decoder, encoder, base_model_prefix, uninitialized_encoder_weights, skip_key
)