navsim/agents/backbones/eva.py

import os
from collections import OrderedDict
from mmcv.runner import _load_checkpoint

import torch
import torch.nn as nn
import torch.nn.functional as F
import math
import numpy as np
import logging
from functools import partial
from scipy import interpolate
from math import pi
from einops import rearrange, repeat
import warnings
import torch.utils.checkpoint as cp
from mmdet.models.builder import BACKBONES

logger = logging.getLogger(__name__)
BatchNorm2d = torch.nn.BatchNorm2d
XFORMERS_ENABLED = os.environ.get("XFORMERS_DISABLED") is None
try:
    if XFORMERS_ENABLED:
        from xformers.ops import memory_efficient_attention, unbind

        XFORMERS_AVAILABLE = True
        warnings.warn("xFormers is available (Attention)")
    else:
        warnings.warn("xFormers is disabled (Attention)")
        raise ImportError
except ImportError:
    XFORMERS_AVAILABLE = False
    warnings.warn("xFormers is not available (Attention)")

class Conv2d(torch.nn.Conv2d):
    """
    A wrapper around :class:`torch.nn.Conv2d` to support empty inputs and more features.
    """

    def __init__(self, *args, **kwargs):
        """
        Extra keyword arguments supported in addition to those in `torch.nn.Conv2d`:
        Args:
            norm (nn.Module, optional): a normalization layer
            activation (callable(Tensor) -> Tensor): a callable activation function
        It assumes that norm layer is used before activation.
        """
        norm = kwargs.pop("norm", None)
        activation = kwargs.pop("activation", None)
        super().__init__(*args, **kwargs)

        self.norm = norm
        self.activation = activation

    def forward(self, x):
        # torchscript does not support SyncBatchNorm yet
        # https://github.com/pytorch/pytorch/issues/40507
        # and we skip these codes in torchscript since:
        # 1. currently we only support torchscript in evaluation mode
        # 2. features needed by exporting module to torchscript are added in PyTorch 1.6 or
        # later version, `Conv2d` in these PyTorch versions has already supported empty inputs.
        if not torch.jit.is_scripting():
            with warnings.catch_warnings(record=True):
                if x.numel() == 0 and self.training:
                    # https://github.com/pytorch/pytorch/issues/12013
                    assert not isinstance(
                        self.norm, torch.nn.SyncBatchNorm
                    ), "SyncBatchNorm does not support empty inputs!"

        x = F.conv2d(
            x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups
        )
        if self.norm is not None:
            x = self.norm(x)
        if self.activation is not None:
            x = self.activation(x)
        return x


def window_partition(x, window_size):
    """
    Partition into non-overlapping windows with padding if needed.
    Args:
        x (tensor): input tokens with [B, H, W, C].
        window_size (int): window size.
    Returns:
        windows: windows after partition with [B * num_windows, window_size, window_size, C].
        (Hp, Wp): padded height and width before partition
    """
    B, H, W, C = x.shape

    pad_h = (window_size - H % window_size) % window_size
    pad_w = (window_size - W % window_size) % window_size
    if pad_h > 0 or pad_w > 0:
        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
    Hp, Wp = H + pad_h, W + pad_w

    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    return windows, (Hp, Wp)


def window_unpartition(windows, window_size, pad_hw, hw):
    """
    Window unpartition into original sequences and removing padding.
    Args:
        x (tensor): input tokens with [B * num_windows, window_size, window_size, C].
        window_size (int): window size.
        pad_hw (Tuple): padded height and width (Hp, Wp).
        hw (Tuple): original height and width (H, W) before padding.
    Returns:
        x: unpartitioned sequences with [B, H, W, C].
    """
    Hp, Wp = pad_hw
    H, W = hw
    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
    x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)

    if Hp > H or Wp > W:
        x = x[:, :H, :W, :].contiguous()
    return x


def get_rel_pos(q_size, k_size, rel_pos):
    """
    Get relative positional embeddings according to the relative positions of
        query and key sizes.
    Args:
        q_size (int): size of query q.
        k_size (int): size of key k.
        rel_pos (Tensor): relative position embeddings (L, C).
    Returns:
        Extracted positional embeddings according to relative positions.
    """
    max_rel_dist = int(2 * max(q_size, k_size) - 1)
    use_log_interpolation = True

    # Interpolate rel pos if needed.
    if rel_pos.shape[0] != max_rel_dist:
        if not use_log_interpolation:
            # Interpolate rel pos.
            rel_pos_resized = F.interpolate(
                rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
                size=max_rel_dist,
                mode="linear",
            )
            rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
        else:
            src_size = rel_pos.shape[0]
            dst_size = max_rel_dist

            # q = 1.13492
            q = 1.0903078
            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
            t = dst_size // 2.0
            dx = np.arange(-t, t + 0.1, 1.0)
            # print("x = %s" % str(x))
            # print("dx = %s" % str(dx))
            all_rel_pos_bias = []
            for i in range(rel_pos.shape[1]):
                z = rel_pos[:, i].view(src_size).cpu().float().numpy()
                f = interpolate.interp1d(x, z, kind='cubic', fill_value="extrapolate")
                all_rel_pos_bias.append(
                    torch.Tensor(f(dx)).contiguous().view(-1, 1).to(rel_pos.device))
            rel_pos_resized = torch.cat(all_rel_pos_bias, dim=-1)
    else:
        rel_pos_resized = rel_pos

    # Scale the coords with short length if shapes for q and k are different.
    q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
    k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
    relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)

    return rel_pos_resized[relative_coords.long()]


def add_decomposed_rel_pos(attn, q, rel_pos_h, rel_pos_w, q_size, k_size):
    """
    Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.
    https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py   # noqa B950
    Args:
        attn (Tensor): attention map.
        q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).
        rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.
        rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.
        q_size (Tuple): spatial sequence size of query q with (q_h, q_w).
        k_size (Tuple): spatial sequence size of key k with (k_h, k_w).
    Returns:
        attn (Tensor): attention map with added relative positional embeddings.
    """
    q_h, q_w = q_size
    k_h, k_w = k_size
    Rh = get_rel_pos(q_h, k_h, rel_pos_h)
    Rw = get_rel_pos(q_w, k_w, rel_pos_w)

    B, _, dim = q.shape
    r_q = q.reshape(B, q_h, q_w, dim)
    rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh)
    rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw)

    attn = (
            attn.view(B, q_h, q_w, k_h, k_w) + rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]
    ).view(B, q_h * q_w, k_h * k_w)

    return attn


def get_abs_pos(abs_pos, has_cls_token, hw):
    """
    Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token
        dimension for the original embeddings.
    Args:
        abs_pos (Tensor): absolute positional embeddings with (1, num_position, C).
        has_cls_token (bool): If true, has 1 embedding in abs_pos for cls token.
        hw (Tuple): size of input image tokens.
    Returns:
        Absolute positional embeddings after processing with shape (1, H, W, C)
    """
    h, w = hw
    if has_cls_token:
        abs_pos = abs_pos[:, 1:]
    xy_num = abs_pos.shape[1]
    size = int(math.sqrt(xy_num))
    assert size * size == xy_num

    if size != h or size != w:
        original_datatype = abs_pos.dtype
        new_abs_pos = F.interpolate(
            abs_pos.reshape(1, size, size, -1).permute(0, 3, 1, 2).float(),  # bf16 is not implemented
            size=(h, w),
            mode="bicubic",
            align_corners=False,
        ).to(original_datatype)

        return new_abs_pos.permute(0, 2, 3, 1)
    else:
        return abs_pos.reshape(1, h, w, -1)


class PatchEmbed(nn.Module):
    """
    Image to Patch Embedding.
    """

    def __init__(
            self, kernel_size=(16, 16), stride=(16, 16), padding=(0, 0), in_chans=3, embed_dim=768
    ):
        """
        Args:
            kernel_size (Tuple): kernel size of the projection layer.
            stride (Tuple): stride of the projection layer.
            padding (Tuple): padding size of the projection layer.
            in_chans (int): Number of input image channels.
            embed_dim (int):  embed_dim (int): Patch embedding dimension.
        """
        super().__init__()

        self.proj = nn.Conv2d(
            in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
        )

    def forward(self, x):
        x = self.proj(x)
        # B C H W -> B H W C
        x = x.permute(0, 2, 3, 1)
        return x


def broadcat(tensors, dim=-1):
    num_tensors = len(tensors)
    shape_lens = set(list(map(lambda t: len(t.shape), tensors)))
    assert len(shape_lens) == 1, 'tensors must all have the same number of dimensions'
    shape_len = list(shape_lens)[0]
    dim = (dim + shape_len) if dim < 0 else dim
    dims = list(zip(*map(lambda t: list(t.shape), tensors)))
    expandable_dims = [(i, val) for i, val in enumerate(dims) if i != dim]
    assert all(
        [*map(lambda t: len(set(t[1])) <= 2, expandable_dims)]), 'invalid dimensions for broadcastable concatentation'
    max_dims = list(map(lambda t: (t[0], max(t[1])), expandable_dims))
    expanded_dims = list(map(lambda t: (t[0], (t[1],) * num_tensors), max_dims))
    expanded_dims.insert(dim, (dim, dims[dim]))
    expandable_shapes = list(zip(*map(lambda t: t[1], expanded_dims)))
    tensors = list(map(lambda t: t[0].expand(*t[1]), zip(tensors, expandable_shapes)))
    return torch.cat(tensors, dim=dim)


def rotate_half(x):
    x = rearrange(x, '... (d r) -> ... d r', r=2)
    x1, x2 = x.unbind(dim=-1)
    x = torch.stack((-x2, x1), dim=-1)
    return rearrange(x, '... d r -> ... (d r)')


class VisionRotaryEmbedding(nn.Module):
    def __init__(
            self,
            dim,
            pt_seq_len,
            ft_seq_len=None,
            custom_freqs=None,
            freqs_for='lang',
            theta=10000,
            max_freq=10,
            num_freqs=1,
    ):
        super().__init__()
        if custom_freqs:
            freqs = custom_freqs
        elif freqs_for == 'lang':
            freqs = 1. / (theta ** (torch.arange(0, dim, 2)[:(dim // 2)].float() / dim))
        elif freqs_for == 'pixel':
            freqs = torch.linspace(1., max_freq / 2, dim // 2) * pi
        elif freqs_for == 'constant':
            freqs = torch.ones(num_freqs).float()
        else:
            raise ValueError(f'unknown modality {freqs_for}')

        if ft_seq_len is None: ft_seq_len = pt_seq_len
        t = torch.arange(ft_seq_len) / ft_seq_len * pt_seq_len

        freqs_h = torch.einsum('..., f -> ... f', t, freqs)
        freqs_h = repeat(freqs_h, '... n -> ... (n r)', r=2)

        freqs_w = torch.einsum('..., f -> ... f', t, freqs)
        freqs_w = repeat(freqs_w, '... n -> ... (n r)', r=2)

        freqs = broadcat((freqs_h[:, None, :], freqs_w[None, :, :]), dim=-1)

        self.register_buffer("freqs_cos", freqs.cos())
        self.register_buffer("freqs_sin", freqs.sin())

        print('======== shape of rope freq', self.freqs_cos.shape, '========')

    def forward(self, t, start_index=0):
        rot_dim = self.freqs_cos.shape[-1]
        end_index = start_index + rot_dim
        assert rot_dim <= t.shape[
            -1], f'feature dimension {t.shape[-1]} is not of sufficient size to rotate in all the positions {rot_dim}'
        t_left, t, t_right = t[..., :start_index], t[..., start_index:end_index], t[..., end_index:]
        t = (t * self.freqs_cos) + (rotate_half(t) * self.freqs_sin)
        return torch.cat((t_left, t, t_right), dim=-1)


class VisionRotaryEmbeddingFast(nn.Module):
    def __init__(
            self,
            dim,
            pt_seq_len=16,
            ft_seq_len=None,
            custom_freqs=None,
            freqs_for='lang',
            theta=10000,
            max_freq=10,
            num_freqs=1,
    ):
        super().__init__()
        if custom_freqs:
            freqs = custom_freqs
        elif freqs_for == 'lang':
            freqs = 1. / (theta ** (torch.arange(0, dim, 2)[:(dim // 2)].float() / dim))
        elif freqs_for == 'pixel':
            freqs = torch.linspace(1., max_freq / 2, dim // 2) * pi
        elif freqs_for == 'constant':
            freqs = torch.ones(num_freqs).float()
        else:
            raise ValueError(f'unknown modality {freqs_for}')

        if ft_seq_len is None: ft_seq_len = pt_seq_len
        t = torch.arange(ft_seq_len) / ft_seq_len * pt_seq_len

        freqs = torch.einsum('..., f -> ... f', t, freqs)
        freqs = repeat(freqs, '... n -> ... (n r)', r=2)
        freqs = broadcat((freqs[:, None, :], freqs[None, :, :]), dim=-1)

        freqs_cos = freqs.cos().view(-1, freqs.shape[-1])
        freqs_sin = freqs.sin().view(-1, freqs.shape[-1])

        self.register_buffer("freqs_cos", freqs_cos)
        self.register_buffer("freqs_sin", freqs_sin)

        print('======== shape of rope freq', self.freqs_cos.shape, '========')

    def forward(self, t):
        return t * self.freqs_cos + rotate_half(t) * self.freqs_sin


class FrozenBatchNorm2d(nn.Module):
    """
    BatchNorm2d where the batch statistics and the affine parameters are fixed.
    It contains non-trainable buffers called
    "weight" and "bias", "running_mean", "running_var",
    initialized to perform identity transformation.
    The pre-trained backbone models from Caffe2 only contain "weight" and "bias",
    which are computed from the original four parameters of BN.
    The affine transform `x * weight + bias` will perform the equivalent
    computation of `(x - running_mean) / sqrt(running_var) * weight + bias`.
    When loading a backbone model from Caffe2, "running_mean" and "running_var"
    will be left unchanged as identity transformation.
    Other pre-trained backbone models may contain all 4 parameters.
    The forward is implemented by `F.batch_norm(..., training=False)`.
    """

    _version = 3

    def __init__(self, num_features, eps=1e-5):
        super().__init__()
        self.num_features = num_features
        self.eps = eps
        self.register_buffer("weight", torch.ones(num_features))
        self.register_buffer("bias", torch.zeros(num_features))
        self.register_buffer("running_mean", torch.zeros(num_features))
        self.register_buffer("running_var", torch.ones(num_features) - eps)

    def forward(self, x):
        if x.requires_grad:
            # When gradients are needed, F.batch_norm will use extra memory
            # because its backward op computes gradients for weight/bias as well.
            scale = self.weight * (self.running_var + self.eps).rsqrt()
            bias = self.bias - self.running_mean * scale
            scale = scale.reshape(1, -1, 1, 1)
            bias = bias.reshape(1, -1, 1, 1)
            out_dtype = x.dtype  # may be half
            return x * scale.to(out_dtype) + bias.to(out_dtype)
        else:
            # When gradients are not needed, F.batch_norm is a single fused op
            # and provide more optimization opportunities.
            return F.batch_norm(
                x,
                self.running_mean,
                self.running_var,
                self.weight,
                self.bias,
                training=False,
                eps=self.eps,
            )

    def _load_from_state_dict(
            self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
    ):
        version = local_metadata.get("version", None)

        if version is None or version < 2:
            # No running_mean/var in early versions
            # This will silent the warnings
            if prefix + "running_mean" not in state_dict:
                state_dict[prefix + "running_mean"] = torch.zeros_like(self.running_mean)
            if prefix + "running_var" not in state_dict:
                state_dict[prefix + "running_var"] = torch.ones_like(self.running_var)

        super()._load_from_state_dict(
            state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
        )

    def __repr__(self):
        return "FrozenBatchNorm2d(num_features={}, eps={})".format(self.num_features, self.eps)

    @classmethod
    def convert_frozen_batchnorm(cls, module):
        """
        Convert all BatchNorm/SyncBatchNorm in module into FrozenBatchNorm.
        Args:
            module (torch.nn.Module):
        Returns:
            If module is BatchNorm/SyncBatchNorm, returns a new module.
            Otherwise, in-place convert module and return it.
        Similar to convert_sync_batchnorm in
        https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/batchnorm.py
        """
        bn_module = nn.modules.batchnorm
        bn_module = (bn_module.BatchNorm2d, bn_module.SyncBatchNorm)
        res = module
        if isinstance(module, bn_module):
            res = cls(module.num_features)
            if module.affine:
                res.weight.data = module.weight.data.clone().detach()
                res.bias.data = module.bias.data.clone().detach()
            res.running_mean.data = module.running_mean.data
            res.running_var.data = module.running_var.data
            res.eps = module.eps
        else:
            for name, child in module.named_children():
                new_child = cls.convert_frozen_batchnorm(child)
                if new_child is not child:
                    res.add_module(name, new_child)
        return res


class LayerNorm(nn.Module):
    """
    A LayerNorm variant, popularized by Transformers, that performs point-wise mean and
    variance normalization over the channel dimension for inputs that have shape
    (batch_size, channels, height, width).
    https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119  # noqa B950
    """

    def __init__(self, normalized_shape, eps=1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(normalized_shape))
        self.bias = nn.Parameter(torch.zeros(normalized_shape))
        self.eps = eps
        self.normalized_shape = (normalized_shape,)

    def forward(self, x):
        u = x.mean(1, keepdim=True)
        s = (x - u).pow(2).mean(1, keepdim=True)
        x = (x - u) / torch.sqrt(s + self.eps)
        x = self.weight[:, None, None] * x + self.bias[:, None, None]
        return x


class CNNBlockBase(nn.Module):
    """
    A CNN block is assumed to have input channels, output channels and a stride.
    The input and output of `forward()` method must be NCHW tensors.
    The method can perform arbitrary computation but must match the given
    channels and stride specification.
    Attribute:
        in_channels (int):
        out_channels (int):
        stride (int):
    """

    def __init__(self, in_channels, out_channels, stride):
        """
        The `__init__` method of any subclass should also contain these arguments.
        Args:
            in_channels (int):
            out_channels (int):
            stride (int):
        """
        super().__init__()
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.stride = stride

    def freeze(self):
        """
        Make this block not trainable.
        This method sets all parameters to `requires_grad=False`,
        and convert all BatchNorm layers to FrozenBatchNorm
        Returns:
            the block itself
        """
        for p in self.parameters():
            p.requires_grad = False
        FrozenBatchNorm2d.convert_frozen_batchnorm(self)
        return self


def get_norm(norm, out_channels):
    """
    Args:
        norm (str or callable): either one of BN, SyncBN, FrozenBN, GN;
            or a callable that takes a channel number and returns
            the normalization layer as a nn.Module.
    Returns:
        nn.Module or None: the normalization layer
    """
    if norm is None:
        return None
    if isinstance(norm, str):
        if len(norm) == 0:
            return None
        norm = {
            "BN": BatchNorm2d,
            # Fixed in https://github.com/pytorch/pytorch/pull/36382
            "SyncBN": nn.SyncBatchNorm,
            "FrozenBN": FrozenBatchNorm2d,
            "GN": lambda channels: nn.GroupNorm(32, channels),
            # for debugging:
            "nnSyncBN": nn.SyncBatchNorm,
            "LN": lambda channels: LayerNorm(channels)
        }[norm]
    return norm(out_channels)


class DropPath(nn.Module):
    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
    """

    def __init__(self, drop_prob=None):
        super(DropPath, self).__init__()
        self.drop_prob = drop_prob

    def forward(self, x):
        if self.drop_prob == 0. or not self.training:
            return x
        keep_prob = 1 - self.drop_prob
        # work with diff dim tensors, not just 2D ConvNets
        shape = (x.shape[0],) + (1,) * (x.ndim - 1)
        random_tensor = keep_prob + \
                        torch.rand(shape, dtype=x.dtype, device=x.device)
        random_tensor.floor_()  # binarize
        output = x.div(keep_prob) * random_tensor
        return output


class SwiGLU(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.SiLU, drop=0.,
                 norm_layer=nn.LayerNorm, subln=False
                 ):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features

        self.w1 = nn.Linear(in_features, hidden_features)
        self.w2 = nn.Linear(in_features, hidden_features)

        self.act = act_layer()
        self.ffn_ln = norm_layer(hidden_features) if subln else nn.Identity()
        self.w3 = nn.Linear(hidden_features, out_features)

        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x1 = self.w1(x)
        x2 = self.w2(x)
        hidden = self.act(x1) * x2
        x = self.ffn_ln(hidden)
        x = self.w3(x)
        x = self.drop(x)
        return x


class Attention(nn.Module):
    def __init__(
            self,
            dim,
            num_heads=8,
            qkv_bias=True,
            qk_scale=None,
            attn_head_dim=None,
            norm_layer=nn.LayerNorm,
            rope=None,
            flash_attn=True,
            xformers_attn=True,
            subln=False
    ):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        if attn_head_dim is not None:
            head_dim = attn_head_dim
        all_head_dim = head_dim * self.num_heads
        self.scale = qk_scale or head_dim ** -0.5

        self.subln = subln
        self.q_proj = nn.Linear(dim, all_head_dim, bias=False)
        self.k_proj = nn.Linear(dim, all_head_dim, bias=False)
        self.v_proj = nn.Linear(dim, all_head_dim, bias=False)

        if qkv_bias:
            self.q_bias = nn.Parameter(torch.zeros(all_head_dim))
            self.v_bias = nn.Parameter(torch.zeros(all_head_dim))
        else:
            self.q_bias = None
            self.v_bias = None

        self.rope = rope
        # self.flash_attn = flash_attn
        self.proj = nn.Linear(all_head_dim, dim)
        self.inner_attn_ln = norm_layer(all_head_dim) if subln else nn.Identity()
        self.xformers_attn = xformers_attn
        # if self.flash_attn:
        #     factory_kwargs = {'device': 'cuda', 'dtype': torch.float16}
        #     self.inner_attn = FlashAttention(attention_dropout=0.0, **factory_kwargs)

    def forward(self, x):
        B, H, W, C = x.shape
        x = x.view(B, -1, C)
        N = H * W

        q = F.linear(input=x, weight=self.q_proj.weight, bias=self.q_bias)
        k = F.linear(input=x, weight=self.k_proj.weight, bias=None)
        v = F.linear(input=x, weight=self.v_proj.weight, bias=self.v_bias)

        q = q.reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3)  # B, num_heads, N, C
        k = k.reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3)
        v = v.reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3)

        ## rope
        q = self.rope(q).type_as(v)
        k = self.rope(k).type_as(v)

        if self.xformers_attn:
            q = q.permute(0, 2, 1, 3)  # B, num_heads, N, C -> B, N, num_heads, C
            k = k.permute(0, 2, 1, 3)
            v = v.permute(0, 2, 1, 3)

            # kv = torch.stack([k, v], dim=2)
            # x, attn_weights = self.inner_attn(q, kv, key_padding_mask=None, causal=False)
            x = memory_efficient_attention(q, k, v)
            x = x.reshape(B, N, -1)
            x = self.inner_attn_ln(x)
        else:
            q = q * self.scale
            attn = (q @ k.transpose(-2, -1))
            attn = attn.softmax(dim=-1).type_as(x)
            x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
            x = self.inner_attn_ln(x)

        x = self.proj(x)
        x = x.view(B, H, W, C)

        return x


class ResBottleneckBlock(CNNBlockBase):
    """
    The standard bottleneck residual block without the last activation layer.
    It contains 3 conv layers with kernels 1x1, 3x3, 1x1.
    """

    def __init__(
            self,
            in_channels,
            out_channels,
            bottleneck_channels,
            norm="LN",
            act_layer=nn.GELU,
    ):
        """
        Args:
            in_channels (int): Number of input channels.
            out_channels (int): Number of output channels.
            bottleneck_channels (int): number of output channels for the 3x3
                "bottleneck" conv layers.
            norm (str or callable): normalization for all conv layers.
                See :func:`layers.get_norm` for supported format.
            act_layer (callable): activation for all conv layers.
        """
        super().__init__(in_channels, out_channels, 1)

        self.conv1 = Conv2d(in_channels, bottleneck_channels, 1, bias=False)
        self.norm1 = get_norm(norm, bottleneck_channels)
        self.act1 = act_layer()

        self.conv2 = Conv2d(
            bottleneck_channels,
            bottleneck_channels,
            3,
            padding=1,
            bias=False,
        )
        self.norm2 = get_norm(norm, bottleneck_channels)
        self.act2 = act_layer()

        self.conv3 = Conv2d(bottleneck_channels, out_channels, 1, bias=False)
        self.norm3 = get_norm(norm, out_channels)

        for layer in [self.norm1, self.norm2]:
            layer.weight.data.fill_(1.0)
            layer.bias.data.zero_()
        # zero init last norm layer.
        self.norm3.weight.data.zero_()
        self.norm3.bias.data.zero_()

    def forward(self, x):
        out = x
        for layer in self.children():
            out = layer(out)

        out = x + out
        return out


class Block(nn.Module):
    """Transformer blocks with support of window attention and residual propagation blocks"""

    def __init__(
            self,
            dim,
            num_heads,
            mlp_ratio=4 * 2 / 3,
            qkv_bias=True,
            drop_path=0.0,
            norm_layer=partial(nn.LayerNorm, eps=1e-6),
            window_size=0,
            use_residual_block=False,
            rope=None,
            flash_attn=False,
            xformers_attn=True,
            subln=False,
            with_cp=True,
    ):
        """
        Args:
            dim (int): Number of input channels.
            num_heads (int): Number of attention heads in each ViT block.
            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
            qkv_bias (bool): If True, add a learnable bias to query, key, value.
            drop_path (float): Stochastic depth rate.
            norm_layer (nn.Module): Normalization layer.
            act_layer (nn.Module): Activation layer.
            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
            window_size (int): Window size for window attention blocks. If it equals 0, then not
                use window attention.
            use_residual_block (bool): If True, use a residual block after the MLP block.
            input_size (int or None): Input resolution for calculating the relative positional
                parameter size.
        """
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention(
            dim,
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            rope=rope,
            flash_attn=flash_attn,
            xformers_attn=xformers_attn,
            subln=subln
        )

        self.with_cp = with_cp
        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
        self.norm2 = norm_layer(dim)
        self.mlp = SwiGLU(
            in_features=dim,
            hidden_features=int(dim * mlp_ratio),
            subln=True,
            norm_layer=norm_layer,
        )

        self.window_size = window_size

        self.use_residual_block = use_residual_block
        if use_residual_block:
            # Use a residual block with bottleneck channel as dim // 2
            self.residual = ResBottleneckBlock(
                in_channels=dim,
                out_channels=dim,
                bottleneck_channels=dim // 2,
                norm="LN",
            )

    def _forward(self, x):
        shortcut = x
        x = self.norm1(x)

        # Window partition
        if self.window_size > 0:
            H, W = x.shape[1], x.shape[2]
            x, pad_hw = window_partition(x, self.window_size)

        x = self.attn(x)

        # Reverse window partition
        if self.window_size > 0:
            x = window_unpartition(x, self.window_size, pad_hw, (H, W))

        x = shortcut + self.drop_path(x)
        x = x + self.drop_path(self.mlp(self.norm2(x)))

        if self.use_residual_block:
            x = self.residual(x.permute(0, 3, 1, 2)).permute(0, 2, 3, 1)

        return x

    def forward(self, x):
        if self.with_cp and self.training:
            x = cp.checkpoint(self._forward, x)
        else:
            x = self._forward(x)
        return x


@BACKBONES.register_module()
class EVAViT(nn.Module):
    """
    This module implements Vision Transformer (ViT) backbone in :paper:`vitdet`.
    "Exploring Plain Vision Transformer Backbones for Object Detection",
    https://arxiv.org/abs/2203.16527
    """

    def __init__(
            self,
            img_size=1024,
            patch_size=16,
            in_chans=3,
            embed_dim=768,
            depth=12,
            num_heads=12,
            mlp_ratio=4 * 2 / 3,
            qkv_bias=True,
            drop_path_rate=0.0,
            norm_layer=partial(nn.LayerNorm, eps=1e-6),
            use_abs_pos=True,
            pt_hw_seq_len=16,
            intp_freq=True,
            window_size=0,
            global_window_size=0,
            window_block_indexes=(),
            residual_block_indexes=(),
            pretrain_img_size=224,
            pretrain_use_cls_token=True,
            out_feature="last_feat",
            subln=False,
            flash_attn=False,
            xformers_attn=True,
            with_cp=True,
            frozen=False,
    ):
        """
        Args:
            img_size (int): Input image size.
            patch_size (int): Patch size.
            in_chans (int): Number of input image channels.
            embed_dim (int): Patch embedding dimension.
            depth (int): Depth of ViT.
            num_heads (int): Number of attention heads in each ViT block.
            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
            qkv_bias (bool): If True, add a learnable bias to query, key, value.
            drop_path_rate (float): Stochastic depth rate.
            norm_layer (nn.Module): Normalization layer.
            act_layer (nn.Module): Activation layer.
            use_abs_pos (bool): If True, use absolute positional embeddings.
            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
            window_size (int): Window size for window attention blocks.
            window_block_indexes (list): Indexes for blocks using window attention.
            residual_block_indexes (list): Indexes for blocks using conv propagation.
            use_act_checkpoint (bool): If True, use activation checkpointing.
            pretrain_img_size (int): input image size for pretraining models.
            pretrain_use_cls_token (bool): If True, pretrainig models use class token.
            out_feature (str): name of the feature from the last block.
        """
        super().__init__()
        self.pretrain_use_cls_token = pretrain_use_cls_token
        self.patch_embed = PatchEmbed(
            kernel_size=(patch_size, patch_size),
            stride=(patch_size, patch_size),
            in_chans=in_chans,
            embed_dim=embed_dim,
        )
        self.frozen = frozen

        if use_abs_pos:
            # Initialize absolute positional embedding with pretrain image size.
            num_patches = (pretrain_img_size // patch_size) * (pretrain_img_size // patch_size)
            num_positions = (num_patches + 1) if pretrain_use_cls_token else num_patches
            self.pos_embed = nn.Parameter(torch.zeros(1, num_positions, embed_dim))
        else:
            self.pos_embed = None

        half_head_dim = embed_dim // num_heads // 2
        hw_seq_len = img_size // patch_size

        self.rope_win = VisionRotaryEmbeddingFast(
            dim=half_head_dim,
            pt_seq_len=pt_hw_seq_len,
            ft_seq_len=window_size if intp_freq else None,
        )
        self.rope_glb = VisionRotaryEmbeddingFast(
            dim=half_head_dim,
            pt_seq_len=pt_hw_seq_len,
            ft_seq_len=hw_seq_len if intp_freq else None,
        )

        # stochastic depth decay rule
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]

        self.blocks = nn.ModuleList()
        for i in range(depth):
            block = Block(
                dim=embed_dim,
                num_heads=num_heads,
                mlp_ratio=mlp_ratio,
                qkv_bias=qkv_bias,
                drop_path=dpr[i],
                norm_layer=norm_layer,
                window_size=window_size if i in window_block_indexes else global_window_size,
                use_residual_block=i in residual_block_indexes,
                rope=self.rope_win if i in window_block_indexes else self.rope_glb,
                flash_attn=flash_attn,
                xformers_attn=xformers_attn,
                subln=subln,
                with_cp=with_cp,
            )

            self.blocks.append(block)

        self._out_feature_channels = {out_feature: embed_dim}
        self._out_feature_strides = {out_feature: patch_size}
        self._out_features = [out_feature]

        self.adapter = None

        if self.pos_embed is not None:
            nn.init.normal_(self.pos_embed, std=0.02)

        # MIN SHI: I disable the weight initialization since they will be automatically loaded
        # **However, they will cause problems (deepspeed + bf16)**
        # self.apply(self._init_weights)
        self._freeze_stages()

    def _freeze_stages(self):
        if self.frozen:
            self.eval()
            for m in self.parameters():
                m.requires_grad = False

    def init_weights(self, ckpt_path):
        ckpt = _load_checkpoint(ckpt_path,
                                logger=logger,
                                map_location='cpu')
        if 'state_dict' in ckpt:
            _state_dict = ckpt['state_dict']
        elif 'model' in ckpt:
            _state_dict = ckpt['model']
        else:
            _state_dict = ckpt

        state_dict = OrderedDict()
        rope_keys = []
        for k, v in _state_dict.items():
            if 'rope' in k:
                rope_keys.append(k[13:])
            if k.startswith('backbone.'):
                state_dict[k[9:]] = v
            elif k.startswith('img_backbone.'):
                state_dict[k[13:]] = v
            else:
                state_dict[k] = v
        print(f'Before deleting rope keys, {len(state_dict)}')
        for rope_k in rope_keys:
            del state_dict[rope_k]
        print(f'After deleting rope keys, {len(state_dict)}')

        # strip prefix of state_dict
        if list(state_dict.keys())[0].startswith('module.'):
            state_dict = {k[7:]: v for k, v in state_dict.items()}

        # load state_dict
        meg = self.load_state_dict(state_dict, False)
        print(meg)

    def forward(self, x,):
        x = self.patch_embed(x)
        if self.pos_embed is not None:
            x = x + get_abs_pos(
                self.pos_embed, self.pretrain_use_cls_token, (x.shape[1], x.shape[2])
            )

        for blk in self.blocks:
            x = blk(x)  # b, h, w, c
        x = x.permute(0, 3, 1, 2)  # b, c, h, w
        # print(x.shape)
        return [x, ]