Spaces:
Sleeping
Sleeping
File size: 5,910 Bytes
d9272c6 |
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 |
"""
Copyright (c) 2022, salesforce.com, inc.
All rights reserved.
SPDX-License-Identifier: BSD-3-Clause
For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
"""
import warnings
from copy import deepcopy
import torch
import torch.nn.functional as F
from lavis.common.registry import registry
from lavis.models.albef_models import AlbefBase
from lavis.models.albef_models.albef_outputs import (
AlbefIntermediateOutput,
AlbefOutputWithLogits,
)
from lavis.models.base_model import MomentumDistilationMixin
from lavis.models.med import XBertEncoder
from lavis.models.vit import VisionTransformerEncoder
from torch import nn
@registry.register_model("albef_classification")
class AlbefClassification(AlbefBase, MomentumDistilationMixin):
PRETRAINED_MODEL_CONFIG_DICT = {
"ve": "configs/models/albef_classification_ve.yaml",
}
def __init__(
self,
image_encoder,
text_encoder,
num_classes,
momentum=0.995,
alpha=0.4,
use_distill=True,
max_txt_len=40,
):
super().__init__()
self.tokenizer = self.init_tokenizer()
self.max_txt_len = max_txt_len
self.use_distill = use_distill
self.visual_encoder = image_encoder
self.text_encoder = text_encoder
hidden_size = text_encoder.config.hidden_size
if num_classes > 0:
self.cls_head = nn.Sequential(
nn.Linear(hidden_size, hidden_size),
nn.ReLU(),
nn.Linear(hidden_size, num_classes),
)
else:
warnings.warn(
f"Found num_classes=0, initializing {type(self)} without classifier."
)
if self.use_distill:
self.visual_encoder_m = deepcopy(self.visual_encoder)
self.text_encoder_m = deepcopy(self.text_encoder)
self.cls_head_m = deepcopy(self.cls_head)
self.momentum = momentum
self.alpha = alpha
self.model_pairs = [
[self.visual_encoder, self.visual_encoder_m],
[self.text_encoder, self.text_encoder_m],
[self.cls_head, self.cls_head_m],
]
self.copy_params()
def _rampup_factor(self, epoch, iters, num_iters_per_epoch):
return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)
def forward(self, samples, is_train=True):
sentences = samples["text_input"]
sentences = self.tokenizer(
sentences,
padding="longest",
truncation=True,
max_length=self.max_txt_len,
return_tensors="pt",
).to(self.device)
samples.update({"tokenized_text": sentences})
targets = samples["label"]
image_embeds = self.visual_encoder.forward_features(samples["image"])
encoder_output = self.text_encoder.forward_automask(
samples["tokenized_text"], image_embeds
)
prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])
if is_train:
if self.use_distill:
with torch.no_grad():
self._momentum_update()
image_embeds_m = self.visual_encoder_m(samples["image"])
encoder_output_m = self.text_encoder_m.forward_automask(
samples["tokenized_text"], image_embeds_m
)
prediction_m = self.cls_head_m(
encoder_output_m.last_hidden_state[:, 0, :]
)
alpha = self.alpha * self._rampup_factor(
epoch=samples["epoch"],
iters=samples["iters"],
num_iters_per_epoch=samples["num_iters_per_epoch"],
)
loss = (1 - alpha) * F.cross_entropy(
prediction, targets
) - alpha * torch.sum(
F.log_softmax(prediction, dim=1) * F.softmax(prediction_m, dim=1),
dim=1,
).mean()
else:
loss = F.cross_entropy(prediction, targets)
image_embeds_m, encoder_output_m, prediction_m = None, None, None
# return {"loss": loss}
return AlbefOutputWithLogits(
loss=loss,
intermediate_output=AlbefIntermediateOutput(
image_embeds=image_embeds,
image_embeds_m=image_embeds_m,
encoder_output=encoder_output,
encoder_output_m=encoder_output_m,
),
logits=prediction,
logits_m=prediction_m,
)
else:
return {"predictions": prediction, "targets": targets}
def predict(self, samples):
output = self.forward(samples, is_train=False)
return output
@classmethod
def from_config(cls, cfg=None):
image_encoder = VisionTransformerEncoder.from_config(cfg)
# text encoder + multimodal encoder
text_encoder = XBertEncoder.from_config(cfg)
alpha = cfg.get("alpha", 0.4)
momentum = cfg.get("momentum", 0.995)
use_distill = cfg.get("use_distill", True)
num_classes = cfg.get("num_classes", -1)
max_txt_len = cfg.get("max_txt_len", 40)
assert num_classes > 1, "Invalid number of classes provided, found {}".format(
num_classes
)
model = cls(
image_encoder=image_encoder,
text_encoder=text_encoder,
use_distill=use_distill,
alpha=alpha,
num_classes=num_classes,
momentum=momentum,
max_txt_len=max_txt_len,
)
model.load_checkpoint_from_config(cfg)
return model
|