Spaces:
Running
on
Zero
Running
on
Zero
File size: 8,817 Bytes
e428df4 000c55e e428df4 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 |
# Copyright 2024 Alibaba DAMO Academy
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import re
import einops
import torch
import torch.nn as nn
import torch.nn.functional as F
from timm.models.regnet import RegStage
from timm.models.layers import LayerNorm2d
from transformers import TRANSFORMERS_CACHE
def parse_snapshot_folder(repo_id, cache_dir=None, repo_type="model"):
revision = "main"
# 1. parse the downloaded cache folder
if cache_dir is None:
cache_dir = TRANSFORMERS_CACHE
else:
cache_dir = cache_dir
object_id = repo_id.replace("/", "--")
repo_cache = os.path.join(cache_dir, f"{repo_type}s--{object_id}")
# 2. resolve refs (for instance to convert main to the associated commit sha)
refs_dir = os.path.join(repo_cache, "refs")
if os.path.isdir(refs_dir):
revision_file = os.path.join(refs_dir, revision)
if os.path.isfile(revision_file):
with open(revision_file) as f:
revision = f.read()
# 3. acquire the snapshot folder
folder = os.path.join(repo_cache, "snapshots", revision)
return folder
def load_mm_projector(model_path, cache_dir=None, token=None):
if os.path.exists(os.path.join(model_path, 'mm_projector.bin')):
is_local = True
folder = model_path
else:
is_local = False
folder = parse_snapshot_folder(model_path, cache_dir=cache_dir, repo_type="model")
if not os.path.exists(os.path.join(folder, 'mm_projector.bin')):
# downloading from remote repo
from huggingface_hub import snapshot_download
snapshot_download(repo_id=model_path, cache_dir=cache_dir, token=token)
mm_projector_weights = torch.load(os.path.join(folder, 'mm_projector.bin'), map_location='cpu')
mm_projector_weights = {k: v.to(torch.float16) for k, v in mm_projector_weights.items()}
return mm_projector_weights
class IdentityMap(nn.Module):
def __init__(self):
super().__init__()
def forward(self, x, *args, **kwargs):
return x
@property
def config(self):
return {"mm_projector_type": 'identity'}
class SimpleResBlock(nn.Module):
def __init__(self, channels):
super().__init__()
self.pre_norm = nn.LayerNorm(channels)
self.proj = nn.Sequential(
nn.Linear(channels, channels),
nn.GELU(),
nn.Linear(channels, channels)
)
def forward(self, x):
x = self.pre_norm(x)
return x + self.proj(x)
def build_vision_projector(config, delay_load=False, **kwargs):
projector_type = getattr(config, 'mm_projector_type', 'linear')
mlp_gelu_match = re.match(r'^mlp(\d+)x_gelu$', projector_type)
if mlp_gelu_match:
mlp_depth = int(mlp_gelu_match.group(1))
modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
for _ in range(1, mlp_depth):
modules.append(nn.GELU())
modules.append(nn.Linear(config.hidden_size, config.hidden_size))
return nn.Sequential(*modules)
if projector_type == "linear":
# NOTE: for both linear and mlp2x_gelu projector type, mean pooling is adopted to aggreate video features
return nn.Linear(config.mm_hidden_size, config.hidden_size)
elif projector_type == "stc_connector":
return STCConnector(config)
elif projector_type == "stp_connector":
return STPConnector(config)
elif projector_type == "stc_connector_v35":
return STCConnectorV35(config)
elif projector_type == "spatial_conv":
return SpatialConv(config)
elif projector_type == "spatial_pool":
return SpatialPool(config)
if projector_type == 'identity':
return IdentityMap()
raise ValueError(f'Unknown projector type: {projector_type}')
def build_mlp(depth, hidden_size, output_hidden_size):
modules = [nn.Linear(hidden_size, output_hidden_size)]
for _ in range(1, depth):
modules.append(nn.GELU())
modules.append(nn.Linear(output_hidden_size, output_hidden_size))
return nn.Sequential(*modules)
class STCConnector(nn.Module):
def __init__(self, config, downsample=(2, 2, 2), depth=4, mlp_depth=2):
"""Temporal Convolutional Vision-Language Connector.
Args:
config: config object.
downsample: (temporal, height, width) downsample rate.
depth: depth of the spatial interaction blocks.
mlp_depth: depth of the vision-language projector layers.
"""
super().__init__()
self.encoder_hidden_size = encoder_hidden_size = config.mm_hidden_size
self.hidden_size = hidden_size = config.hidden_size
self.output_hidden_size = output_hidden_size = config.hidden_size
# TODO: make these as config arguments
self.depth = depth
self.mlp_depth = mlp_depth
self.downsample = downsample
if depth != 0:
self.s1 = RegStage(
depth=depth,
in_chs=encoder_hidden_size,
out_chs=hidden_size,
stride=1,
dilation=1,
act_layer=nn.SiLU,
norm_layer=LayerNorm2d,
)
else:
self.s1 = nn.Identity()
self.sampler = nn.Sequential(
nn.Conv3d(
in_channels=hidden_size,
out_channels=hidden_size,
kernel_size=downsample,
stride=downsample,
padding=1,
bias=True
),
nn.SiLU()
)
if depth != 0:
self.s2 = RegStage(
depth=depth,
in_chs=hidden_size,
out_chs=hidden_size,
stride=1,
dilation=1,
act_layer=nn.SiLU,
norm_layer=LayerNorm2d,
)
else:
self.s2 = nn.Identity()
self.readout = build_mlp(mlp_depth, hidden_size, output_hidden_size)
def forward(self, x):
"""Aggregate tokens on the temporal and spatial dimensions.
Args:
x: input tokens [b, t, h, w, d] / [b, t, l, d]
Returns:
aggregated tokens [b, l, d]
"""
t = x.size(1)
if x.ndim == 4:
hw = int(x.size(2) ** 0.5)
x = einops.rearrange(x, "b t (h w) d -> b d t h w", h=hw, w=hw)
elif x.ndim == 5:
x = einops.rearrange(x, "b t h w d -> b d t h w")
x = einops.rearrange(x, "b d t h w -> (b t) d h w")
# 1. the first stage of the adapter
x = self.s1(x)
x = einops.rearrange(x, "(b t) d h w -> b d t h w", t=t)
# 2. downsampler
x = self.sampler(x)
new_t = x.size(2)
# 3. the second stage of the adapter
x = einops.rearrange(x, "b d t h w -> (b t) d h w")
x = self.s2(x)
x = einops.rearrange(x, "(b t) d h w -> b (t h w) d", t=new_t)
x = self.readout(x)
return x
class STPConnector(STCConnector):
def __init__(self, config, downsample=(2, 2, 2), depth=4, mlp_depth=2):
super().__init__(config=config, downsample=downsample, depth=depth, mlp_depth=mlp_depth)
self.sampler = nn.Sequential(nn.AvgPool3d(downsample), nn.SiLU())
class STCConnectorV35(STCConnector):
def __init__(self, config, downsample=(2, 2, 2), depth=4, mlp_depth=2):
super().__init__(config=config, downsample=downsample, depth=depth, mlp_depth=mlp_depth)
self.sampler = nn.Sequential(
nn.Conv3d(
in_channels=self.hidden_size,
out_channels=self.hidden_size,
kernel_size=downsample,
stride=downsample,
padding=0,
bias=True
),
nn.SiLU())
class SpatialConv(STCConnector):
def __init__(self, config, downsample=(1, 2, 2), depth=0, mlp_depth=2):
super().__init__(config=config, downsample=downsample, depth=depth, mlp_depth=mlp_depth)
class SpatialPool(STPConnector):
def __init__(self, config, downsample=(1, 2, 2), depth=0, mlp_depth=2):
super().__init__(config=config, downsample=downsample, depth=depth, mlp_depth=mlp_depth)
|