Closed paulds8 closed 2 years ago
paulds8
commented
2 years ago It's also worth noting that I don't need to use CLIP specifically, if there is a more efficient alternative you know of that achieves the same outcome I can 100% consider this instead.
PINTO0309
commented
2 years ago Can the batch size be fixed at 1?
paulds8
commented
2 years ago The batch size can be fixed in the context I need to apply this. Perhaps not necessarily at 1, but at a fixed, number, yes. I will attempt to generate an ONNX graph with a few fixed batch sizes and let you know the outcome.
paulds8
commented
2 years ago Doesn't seem like it made any significant difference in this case. Perhaps I'll get the speed up I'm looking for once pruned models are released.
PINTO0309
commented
2 years ago I am not really interested in talking about your performance. But I am at a loss to understand why you would ignore the error message. I am not your handyman.
I will delete the file attached here within a few days, as it is taking up space on my Google Drive. Whether I get a reply from you or not.
ArgMax Float64
from collections import OrderedDict
from typing import Tuple, Unionimport numpy as np import torch import torch.nn.functional as F from torch import nn
class Bottleneck(nn.Module): expansion = 4
def __init__(self, inplanes, planes, stride=1):
super().__init__()
# all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()
self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
self.bn3 = nn.BatchNorm2d(planes * self.expansion)
self.relu = nn.ReLU(inplace=True)
self.downsample = None
self.stride = stride
if stride > 1 or inplanes != planes * Bottleneck.expansion:
# downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
self.downsample = nn.Sequential(OrderedDict([
("-1", nn.AvgPool2d(stride)),
("0", nn.Conv2d(inplanes, planes * self.expansion, 1, stride=1, bias=False)),
("1", nn.BatchNorm2d(planes * self.expansion))
]))
def forward(self, x: torch.Tensor):
identity = x
out = self.relu(self.bn1(self.conv1(x)))
out = self.relu(self.bn2(self.conv2(out)))
out = self.avgpool(out)
out = self.bn3(self.conv3(out))
if self.downsample is not None:
identity = self.downsample(x)
out += identity
out = self.relu(out)
return outclass AttentionPool2d(nn.Module): def init(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None): super().init() self.positional_embedding = nn.Parameter(torch.randn(spacial_dim 2 + 1, embed_dim) / embed_dim 0.5) self.k_proj = nn.Linear(embed_dim, embed_dim) self.q_proj = nn.Linear(embed_dim, embed_dim) self.v_proj = nn.Linear(embed_dim, embed_dim) self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim) self.num_heads = num_heads
def forward(self, x):
x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(2, 0, 1) # NCHW -> (HW)NC
x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0) # (HW+1)NC
x = x + self.positional_embedding[:, None, :].to(x.dtype) # (HW+1)NC
x, _ = F.multi_head_attention_forward(
query=x, key=x, value=x,
embed_dim_to_check=x.shape[-1],
num_heads=self.num_heads,
q_proj_weight=self.q_proj.weight,
k_proj_weight=self.k_proj.weight,
v_proj_weight=self.v_proj.weight,
in_proj_weight=None,
in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]),
bias_k=None,
bias_v=None,
add_zero_attn=False,
dropout_p=0,
out_proj_weight=self.c_proj.weight,
out_proj_bias=self.c_proj.bias,
use_separate_proj_weight=True,
training=self.training,
need_weights=False
)
return x[0]class ModifiedResNet(nn.Module): """ A ResNet class that is similar to torchvision's but contains the following changes:
The final pooling layer is a QKV attention instead of an average pool """
def init(self, layers, output_dim, heads, input_resolution=224, width=64): super().init() self.output_dim = output_dim self.input_resolution = input_resolution
# the 3-layer stem
self.conv1 = nn.Conv2d(3, width // 2, kernel_size=3, stride=2, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(width // 2)
self.conv2 = nn.Conv2d(width // 2, width // 2, kernel_size=3, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(width // 2)
self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
self.bn3 = nn.BatchNorm2d(width)
self.avgpool = nn.AvgPool2d(2)
self.relu = nn.ReLU(inplace=True)
# residual layers
self._inplanes = width # this is a *mutable* variable used during construction
self.layer1 = self._make_layer(width, layers[0])
self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
self.layer4 = self._make_layer(width * 8, layers[3], stride=2)
embed_dim = width * 32 # the ResNet feature dimension
self.attnpool = AttentionPool2d(input_resolution // 32, embed_dim, heads, output_dim)def _make_layer(self, planes, blocks, stride=1): layers = [Bottleneck(self._inplanes, planes, stride)]
self._inplanes = planes * Bottleneck.expansion
for _ in range(1, blocks):
layers.append(Bottleneck(self._inplanes, planes))
return nn.Sequential(*layers)def forward(self, x): def stem(x): for conv, bn in [(self.conv1, self.bn1), (self.conv2, self.bn2), (self.conv3, self.bn3)]: x = self.relu(bn(conv(x))) x = self.avgpool(x) return x
x = x.type(self.conv1.weight.dtype)
x = stem(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.attnpool(x)
return xclass LayerNorm(nn.LayerNorm): """Subclass torch's LayerNorm to handle fp16."""
def forward(self, x: torch.Tensor):
orig_type = x.dtype
ret = super().forward(x.type(torch.float32))
return ret.type(orig_type)class QuickGELU(nn.Module): def forward(self, x: torch.Tensor): return x torch.sigmoid(1.702 x)
class ResidualAttentionBlock(nn.Module): def init(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None): super().init()
self.attn = nn.MultiheadAttention(d_model, n_head)
self.ln_1 = LayerNorm(d_model)
self.mlp = nn.Sequential(OrderedDict([
("c_fc", nn.Linear(d_model, d_model * 4)),
("gelu", QuickGELU()),
("c_proj", nn.Linear(d_model * 4, d_model))
]))
self.ln_2 = LayerNorm(d_model)
self.attn_mask = attn_mask
def attention(self, x: torch.Tensor):
self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
def forward(self, x: torch.Tensor):
x = x + self.attention(self.ln_1(x))
x = x + self.mlp(self.ln_2(x))
return xclass Transformer(nn.Module): def init(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None): super().init() self.width = width self.layers = layers self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attnmask) for in range(layers)])
def forward(self, x: torch.Tensor):
return self.resblocks(x)class VisionTransformer(nn.Module): def init(self, input_resolution: int, patch_size: int, width: int, layers: int, heads: int, output_dim: int): super().init() self.input_resolution = input_resolution self.output_dim = output_dim self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False)
scale = width ** -0.5
self.class_embedding = nn.Parameter(scale * torch.randn(width))
self.positional_embedding = nn.Parameter(scale * torch.randn((input_resolution // patch_size) ** 2 + 1, width))
self.ln_pre = LayerNorm(width)
self.transformer = Transformer(width, layers, heads)
self.ln_post = LayerNorm(width)
self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
def forward(self, x: torch.Tensor):
x = self.conv1(x) # shape = [*, width, grid, grid]
x = x.reshape(x.shape[0], x.shape[1], -1) # shape = [*, width, grid ** 2]
x = x.permute(0, 2, 1) # shape = [*, grid ** 2, width]
x = torch.cat([self.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1) # shape = [*, grid ** 2 + 1, width]
x = x + self.positional_embedding.to(x.dtype)
x = self.ln_pre(x)
x = x.permute(1, 0, 2) # NLD -> LND
x = self.transformer(x)
x = x.permute(1, 0, 2) # LND -> NLD
x = self.ln_post(x[:, 0, :])
if self.proj is not None:
x = x @ self.proj
return xclass CLIP(nn.Module): def init(self, embed_dim: int,
image_resolution: int,
vision_layers: Union[Tuple[int, int, int, int], int],
vision_width: int,
vision_patch_size: int,
# text
context_length: int,
vocab_size: int,
transformer_width: int,
transformer_heads: int,
transformer_layers: int
):
super().__init__()
self.context_length = context_length
if isinstance(vision_layers, (tuple, list)):
vision_heads = vision_width * 32 // 64
self.visual = ModifiedResNet(
layers=vision_layers,
output_dim=embed_dim,
heads=vision_heads,
input_resolution=image_resolution,
width=vision_width
)
else:
vision_heads = vision_width // 64
self.visual = VisionTransformer(
input_resolution=image_resolution,
patch_size=vision_patch_size,
width=vision_width,
layers=vision_layers,
heads=vision_heads,
output_dim=embed_dim
)
self.transformer = Transformer(
width=transformer_width,
layers=transformer_layers,
heads=transformer_heads,
attn_mask=self.build_attention_mask()
)
self.vocab_size = vocab_size
self.token_embedding = nn.Embedding(vocab_size, transformer_width)
self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width))
self.ln_final = LayerNorm(transformer_width)
self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim))
self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
self.initialize_parameters()
def initialize_parameters(self):
nn.init.normal_(self.token_embedding.weight, std=0.02)
nn.init.normal_(self.positional_embedding, std=0.01)
if isinstance(self.visual, ModifiedResNet):
if self.visual.attnpool is not None:
std = self.visual.attnpool.c_proj.in_features ** -0.5
nn.init.normal_(self.visual.attnpool.q_proj.weight, std=std)
nn.init.normal_(self.visual.attnpool.k_proj.weight, std=std)
nn.init.normal_(self.visual.attnpool.v_proj.weight, std=std)
nn.init.normal_(self.visual.attnpool.c_proj.weight, std=std)
for resnet_block in [self.visual.layer1, self.visual.layer2, self.visual.layer3, self.visual.layer4]:
for name, param in resnet_block.named_parameters():
if name.endswith("bn3.weight"):
nn.init.zeros_(param)
proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
attn_std = self.transformer.width ** -0.5
fc_std = (2 * self.transformer.width) ** -0.5
for block in self.transformer.resblocks:
nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
if self.text_projection is not None:
nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)
def build_attention_mask(self):
# lazily create causal attention mask, with full attention between the vision tokens
# pytorch uses additive attention mask; fill with -inf
mask = torch.empty(self.context_length, self.context_length)
mask.fill_(float("-inf"))
mask.triu_(1) # zero out the lower diagonal
return mask
@property
def dtype(self):
return self.visual.conv1.weight.dtype
def encode_image(self, image):
return self.visual(image.type(self.dtype))
def encode_text(self, text):
x = self.token_embedding(text).type(self.dtype) # [batch_size, n_ctx, d_model]
x = x + self.positional_embedding.type(self.dtype)
x = x.permute(1, 0, 2) # NLD -> LND
x = self.transformer(x)
x = x.permute(1, 0, 2) # LND -> NLD
x = self.ln_final(x).type(self.dtype)
# x.shape = [batch_size, n_ctx, transformer.width]
# take features from the eot embedding (eot_token is the highest number in each sequence)
x = x[torch.arange(x.shape[0]), text.to(torch.float64).argmax(dim=-1)] @ self.text_projection
return x
def forward(self, image, text):
image_features = self.encode_image(image).cpu()
text_features = self.encode_text(text).cpu()
# normalized features
image_features = image_features / image_features.norm(dim=1, keepdim=True)
text_features = text_features / text_features.norm(dim=1, keepdim=True)
# cosine similarity as logits
logit_scale = self.logit_scale.cpu().exp()
logits_per_image = logit_scale * image_features @ text_features.t()
logits_per_text = logits_per_image.t()
# shape = [global_batch_size, global_batch_size]
return logits_per_image, logits_per_textdef convert_weights(model: nn.Module): """Convert applicable model parameters to fp16"""
def _convert_weights_to_fp16(l):
if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
l.weight.data = l.weight.data.half()
if l.bias is not None:
l.bias.data = l.bias.data.half()
if isinstance(l, nn.MultiheadAttention):
for attr in [*[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]], "in_proj_bias", "bias_k", "bias_v"]:
tensor = getattr(l, attr)
if tensor is not None:
tensor.data = tensor.data.half()
for name in ["text_projection", "proj"]:
if hasattr(l, name):
attr = getattr(l, name)
if attr is not None:
attr.data = attr.data.half()
model.apply(_convert_weights_to_fp16)def build_model(state_dict: dict): vit = "visual.proj" in state_dict
if vit:
vision_width = state_dict["visual.conv1.weight"].shape[0]
vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")])
vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5)
image_resolution = vision_patch_size * grid_size
else:
counts: list = [len(set(k.split(".")[2] for k in state_dict if k.startswith(f"visual.layer{b}"))) for b in [1, 2, 3, 4]]
vision_layers = tuple(counts)
vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0]
output_width = round((state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5)
vision_patch_size = None
assert output_width ** 2 + 1 == state_dict["visual.attnpool.positional_embedding"].shape[0]
image_resolution = output_width * 32
embed_dim = state_dict["text_projection"].shape[1]
context_length = state_dict["positional_embedding"].shape[0]
vocab_size = state_dict["token_embedding.weight"].shape[0]
transformer_width = state_dict["ln_final.weight"].shape[0]
transformer_heads = transformer_width // 64
transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith(f"transformer.resblocks")))
model = CLIP(
embed_dim,
image_resolution, vision_layers, vision_width, vision_patch_size,
context_length, vocab_size, transformer_width, transformer_heads, transformer_layers
)
for key in ["input_resolution", "context_length", "vocab_size"]:
if key in state_dict:
del state_dict[key]
convert_weights(model)
model.load_state_dict(state_dict)
return model.eval()
- export.py
```python
import os
import skimage
from PIL import Image
import numpy as np
import torch
import onnx
from onnxsim import simplify
import clip
clip.available_models()
load_model = 'ViT-B/32'
model, preprocess = clip.load(load_model)
model.cuda().eval()
model.float()
input_resolution = model.visual.input_resolution
context_length = model.context_length
vocab_size = model.vocab_size
print("Model parameters:", f"{np.sum([int(np.prod(p.shape)) for p in model.parameters()]):,}")
print("Input resolution:", input_resolution)
print("Context length:", context_length)
print("Vocab size:", vocab_size)
clip.tokenize("Hello World!")
# images in skimage to use and their textual descriptions
descriptions = {
"astronaut": "a portrait of an astronaut with the American flag",
}
original_images = []
images = []
texts = []
for filename in [filename for filename in os.listdir(skimage.data_dir) if filename.endswith(".png") or filename.endswith(".jpg")]:
name = os.path.splitext(filename)[0]
if name not in descriptions:
continue
image = Image.open(os.path.join(skimage.data_dir, filename)).convert("RGB")
original_images.append(image)
images.append(preprocess(image))
texts.append(descriptions[name])
image_input = torch.tensor(np.stack(images)).cpu()#.cuda()
text_tokens = clip.tokenize(["This is " + desc for desc in texts]).cuda()
model.visual.cpu()
model.visual(image_input)[0] # astronaut pic embedding
model(image_input, text_tokens.to(torch.int32))[0] # astronaut text embedding
onnx_file = f"clip_{load_model.lower().replace('-','_').replace('/','_')}.onnx"
torch.onnx.export(
model,
(image_input, text_tokens),
onnx_file,
opset_version=11,
)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)python export.pyonnxsim clip_vit_b_32.onnx clip_vit_b_32.onnxonnx2trt clip_vit_b_32.onnx -o clip_vit_b_32.trt -b 1 -d 16 -v[2022-04-02 13:07:15 WARNING] TensorRT was linked against cuBLAS/cuBLAS LT 11.6.5 but loaded cuBLAS/cuBLAS LT 11.6.1
[2022-04-02 13:07:15 INFO] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +8, now: CPU 4041, GPU 2430 (MiB)
[2022-04-02 13:07:15 INFO] [MemUsageChange] Init cuDNN: CPU +0, GPU +8, now: CPU 4041, GPU 2438 (MiB)
[2022-04-02 13:07:15 INFO] [MemUsageChange] TensorRT-managed allocation in building engine: CPU +4, GPU +289, now: CPU 4, GPU 289 (MiB)
Writing TensorRT engine to clip_vit_b_32.trt
All done
javiabellan
commented
1 year ago @PINTO0309 How did you get such a neat and clean ONNX visualization of your CLIP model? When I use Netron, (even simplifying my model onnxsim) I get a messy graph like this:
PINTO0309
commented
1 year ago @javiabellan Do not start another discussion on a closed issue.
Issue Type
Feature Request
OS
Ubuntu
OS architecture
aarch64
Programming Language
Python
Framework
ONNX
Model name and Weights/Checkpoints URL
OpenAI CLIP
https://github.com/openai/CLIP
I used the following notebook with opset=14 targeting this PR (https://github.com/openai/CLIP/pull/219) to convert both the image & text "branches" of ViT-B/32: https://colab.research.google.com/github/josephrocca/openai-clip-js/blob/main/Export_CLIP_to_ONNX_tflite_tfjs_tf_saved_model.ipynb#scrollTo=kDmmi0vMI9WY
I then ensure all shapes are properly defined so I can use the onnxruntime with the TensorRT backend:
python3 -m onnxruntime.tools.symbolic_shape_infer --input onnx_models/clip-text-vit-32.onnx --output onnx_models/clip-text-vit-32-shaped.onnxpython3 -m onnxruntime.tools.symbolic_shape_infer --input onnx_models/clip-image-vit-32.onnx --output onnx_models/clip-image-vit-32-shaped.onnxDescription
I am working on a real-time AI in an embedded context (Jetson TX2 is the target device). I need the latency to be as low as possible for all elements.
I stumbled across this repo today and there are definitely some models here that will speed up my dev process considerably.
Thank you!!
I see you haven't yet added CLIP to the zoo. As you have significant experience optimizing models in constrained environments, I was hoping for some assistance.
The first challenge is getting the latency of using a CLIP model as low as possible.
Using the code below, on a Jetson TX2 I am able to get a result in ~0.24s on average for the model described above. I was hoping you'd have some techniques to help significantly reduce this. My ideal goal is to reduce this to roughly 0.05s on a TX2.
CLIP already used FP16 in many places which the TX2's GPU can take advantage of. I haven't been able to successfully simplify the graph further or force more of the graph to FP16.
I was hoping you'd be able to help. I am happy to test and help in any way I can!
Code:
Relevant Log Output
No response
URL or source code for simple inference testing code
No response