File size: 3,880 Bytes
9e14838 | 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 | import wandb
import torch
import pickle
import torch.nn as nn
from typing import List
from src.model.base import ODBinaryMetricClassifier
from src.model.clip import VideoAttrExtractor
from src.utility.loss import focal_loss
class BinaryLinearClassifier(nn.Module):
def __init__(
self,
*args,
**kargs,
):
super().__init__()
self.encoder = VideoAttrExtractor(
*args,
**kargs
)
self.projs = self.make_linear(self.encoder.embed_dim)
def make_linear(self, embed_dim):
linear = nn.Linear(
embed_dim,
2
)
nn.init.normal_(linear.weight, std=0.001)
nn.init.normal_(linear.bias, std=0.001)
return linear
@property
def transform(self):
return self.encoder.transform
@property
def n_px(self):
return self.encoder.model.input_resolution
def forward(self, x, *args, **kargs):
results = self.encoder(x)
embeds = results["embeds"]
logits = self.projs(embeds.mean(1))
return dict(
logits=logits,
** results
)
class LinearVideoLearner(ODBinaryMetricClassifier):
def __init__(
self,
architecture: str = 'ViT-B/16',
text_embed: bool = False,
attn_record: bool = False,
pretrain: str = None,
label_weights: List[float] = [1, 1],
cls_weight: float = 10.0,
store_attrs: List[str] = [],
):
super().__init__()
self.save_hyperparameters()
params = dict(
architecture=architecture,
text_embed=text_embed,
attn_record=attn_record,
pretrain=pretrain,
store_attrs=store_attrs,
)
self.model = BinaryLinearClassifier(**params)
self.label_weights = torch.tensor(label_weights)
self.cls_weight = cls_weight
@property
def transform(self):
return self.model.transform
@property
def n_px(self):
return self.model.n_px
def shared_step(self, batch, stage):
x, y, z = batch["xyz"]
indices = batch["indices"]
dts_name = batch["dts_name"]
names = batch["names"]
output = self(x, **z)
logits = output["logits"]
loss = 0
# classification loss
if (stage == "train"):
cls_loss = focal_loss(
logits,
y,
gamma=4,
weight=(
self.label_weights.to(y.device)
if stage == "train" else
None
)
)
loss += cls_loss.mean() * self.cls_weight
self.log(
f"{stage}/{dts_name}/loss",
cls_loss.mean(),
batch_size=logits.shape[0]
)
else:
# classification loss
cls_loss = nn.functional.cross_entropy(
logits,
y,
reduction="none",
weight=(
self.label_weights.to(y.device)
if stage == "train" else
None
)
)
loss += cls_loss.mean()
return {
"logits": logits,
"labels": y,
"loss": loss,
"dts_name": dts_name,
"indices": indices,
"output": output
}
if __name__ == "__main__":
frames = 5
model = BinaryLinearClassifier(
architecture="ViT-L/14",
attn_record=True,
text_embed=False
)
model.to("cuda")
result = model(torch.randn(9, frames, 3, 224, 224).to("cuda"))
logit = result["logits"]
logit.sum().backward()
print([m for m, v in model.named_parameters() if v.requires_grad])
print("done")
|