This repo is for fine-tuning Long-CLIP in the command line. It does not add custom nodes to ComfyUI; however, you can easily use your fine-tune with ComfyUI:
- First, fine-tune with exp-ft-B-LongGmP-finetune-LongCLIP-L.py (recommended)
- Or with ft-B-train-LongCLIP-ViT-L-14.py (likely inferior) (deprecated)
- If you used "exp-ft-B-LongGmP", use this to convert the model: exp-ft-C-convert-Long-GmP-back-to-weight.py
- Then, for both fine-tune scripts, use ft-C-convert-for-SDXL-comfyUI-longCLIP.py
- Now you have a state_dict you can plug into ComfyUI for use with SD / SDXL!
Added folder Convert-for-HuggingFace-Spaces-etc
- Includes the convert_clip_original_pytorch_to_hf.py script from HuggingFace + configuration .json files (modified for Long-CLIP 248 tokens).
- Includes optional code to add metadata
{"format": "pt"}- use it in case you get an error about 'pt'! - Please check the included
how-to-use.txt& code comments for details
Added a-loss-to-penalize-overfit-via-entropy.py
- A custom loss with an
entropy penaltyterm that penalizes over-confidence (overfit) - For a diverse and general (!) dataset, the results of fine-tuning are good, but slightly worse than without entropy penalty (fine-tune on COCO-SPRIGHT):
- (Note: This test was done using ViT-L/14 - 77 tokens, not Long-CLIP - but applies just the same)
- ImageNet/ObjectNet accuracy without entropy penalty: 0.845 -> 0.914
- ImageNet/ObjectNet accuracy with entropy penalty: 0.845 -> 0.908
- Alas, I don't want to integrate this loss into the 'best' working code; whether or not it is useful depends entirely on your dataset. If you have a very narrow dataset (e.g. just sneakers), however, and you find CLIP to begin overfitting (val loss increasing) before it has converged to a good (low) loss, then this entropy penalty could be very useful. Simply replace the loss in the actual training script if you observe overfitting, and tinker with the
lambda_entropyfactor. Actual example fromlog_train.txtof1.:
Epoch n:
Validation Acc: 0.9012, Validation F1: 0.8749
Training Loss: 0.6043, Validation Loss: 1.1853
Epoch n+1:
Validation Acc: 0.8942, Validation F1: 0.8652 <- decrease
Training Loss: 0.6018, Validation Loss: 1.1894 <- increase
Now, for the diverse dataset, this was overtraining, not overfitting; the model had already converged (good Acc/F1, low loss). In this case, early stopping (or saving checkpoints every epoch, then hand-selecting the best one - an earlier one, in this case) is recommended. However, I did not observe such an uptick with entropy penalty for a few epochs of overtraining (albeit the model converged at less ideal Validation Acc: 0.8902, Validation F1: 0.8600). So, give it a try if you see CLIP do this with your dataset (very extreme example; better to check log_train.txt to catch it early!):
- This is the same as
exp-ft-M-Long-CLIP-L-GmP-plus-manipulate-activations.py - BUT it introduces label smoothing and a custom loss function for Long-CLIP (CLIP doesn't have discrete 'classes').
- In general: Label smoothing is a regularization technique that softens the target labels by assigning a small probability to incorrect classes, rather than using hard one-hot labels. This approach prevents the model from becoming overly confident in its predictions, encouraging better generalization and reducing the risk of overfitting, especially in cases where the data might be noisy or limited in diversity.
- Read more in the paper When Does Label Smoothing Help? / Google Brain.
- Is it for me? 🤔
- If you have a small dataset or suboptimal labels or are GPU-poor (24/48 GB VRAM), or all of that:
- YES! This may improve your results quite dramatically! 🤓💡
It further improved the results for the high-quality COCO-40K-SPRIGHT dataset with short labels <77 tokens (!) by ~0.2% accuracy, trained on 1x RTX4090; it's likely to show much greater relative benefits with long labels and a small / less diverse dataset (such as your fine-tune of 'sneakers' or whatever floats your boat):
-
You can download this model on my HuggingFace if you don't want to reproduce the fine-tune with the provided code. :-)
-
Technical / code summary of changes:
class ContrastiveLoss(nn.Module):
def __init__(self, temperature=0.07):
super(ContrastiveLoss, self).__init__()
self.temperature = temperature
self.criterion = nn.CrossEntropyLoss()
def forward(self, logits_per_image, logits_per_text):
# Normalize the features to avoid overflow or underflow
logits_per_image = F.normalize(logits_per_image, p=2, dim=1)
logits_per_text = F.normalize(logits_per_text, p=2, dim=1)
# Calculate logits
logits = torch.matmul(logits_per_image, logits_per_text.t()) / self.temperature
labels = torch.arange(logits.size(0), device=logits.device)
# Calculate loss as the mean of the two cross-entropy losses
loss_img = self.criterion(logits, labels)
loss_txt = self.criterion(logits.t(), labels)
return (loss_img + loss_txt) / 2
- Shout out to GPT-4o, my team-mate, for implementation help! 🤓🤝🤖
class ContrastiveLoss(nn.Module):
def __init__(self, temperature=0.07, smoothing=0.1):
super(ContrastiveLoss, self).__init__()
self.temperature = temperature
self.smoothing = smoothing
def forward(self, logits_per_image, logits_per_text):
# Normalize the features to avoid overflow or underflow
logits_per_image = F.normalize(logits_per_image, p=2, dim=1)
logits_per_text = F.normalize(logits_per_text, p=2, dim=1)
# Calculate logits
logits = torch.matmul(logits_per_image, logits_per_text.t()) / self.temperature
labels = torch.arange(logits.size(0), device=logits.device)
# Apply label smoothing
N = logits.size(0)
smoothed_labels = torch.full_like(logits, self.smoothing / (N - 1))
smoothed_labels.scatter_(1, labels.unsqueeze(1), 1.0 - self.smoothing)
# Calculate loss manually using log-softmax and smoothed labels
log_probs = F.log_softmax(logits, dim=1)
loss_img = -(smoothed_labels * log_probs).sum(dim=1).mean()
log_probs = F.log_softmax(logits.t(), dim=1)
loss_txt = -(smoothed_labels * log_probs).sum(dim=1).mean()
return (loss_img + loss_txt) / 2
⬇️ Download my best-performing (ImageNet/ObjectNet accuracy of 0.89) GmP fine-tune here:
- ⬇️ https://huggingface.co/zer0int/LongCLIP-GmP-ViT-L-14
- It's a state-dict; use with ComfyUI as-is, or load it as the state_dict of the original LongCLIP-L for inference, to fine-tune, etc.
- Added exp-ft-M-Long-CLIP-L-GmP-plus-manipulate-activations.py
- Allows scaling / manipulating activation value of specified neurons @ layers during fine-tuning
- You can also just download the resulting model here.
- Eval: Best model so far!
👉 ...Eliminates typographic attack vulnerability in Long-CLIP-L ✅🤯
- Add COCO-SPRIGHT dataset labels + training code for reproduction of my results
- Exact code I used for fine-tune (10 Epochs, 3 hours or so, RTX4090); see code comments
- Attention visualization
- Use exp-ft-B-reproduce-LongGmP-finetune-LongCLIP-L-on-long-labels.py to reproduce my results with GmP-CoCo-SPRIGHT-40k
- Use ft-D-Long-eval-single-image-multi-labels.py to classify single image with multi-labels, e.g. typographic attack probing
- Simply use runall-longclip-explain.py to batch processes all images in "IMG_IN"; will do:
- Getting a Long-CLIP opinion about them / gradient ascent, longclipga_AMP-finetune.py
- Using Long-CLIP opinion + image -> attention heatmaps / what CLIP was 'looking at': longclip-explain-attention-visualization.py - adaptation of hila-chefer/Transformer-MM-Explainability for Long-CLIP
- Note: CLIP often predicts emojis. This can lead to [unicode] errors on certain OS. If this happens, edit the files in the "TOK" folder and remove CLIP's predicted emojis and other special characters, then run "python longclip-explain-attention-visualization.py" + the below scripts one after another manually. Emojis are important predictions in CLIP for salient features (lol - but it's true!), so I didn't want to strip them by default.
- Write CLIP opinion word into image as text, make group mosaics / runmakenames.py + runmakemosaic.py
- --> As always, check the code for comments / details and usage info! <--
Results of the above fine-tuning and evaluation / visualization:
Limitation: Non-English typographic attack vulnerability; here: German, "attention eavesdropping" on an Alexa: Microphone, Satire: Remarkably correct. "lichmachpuck": Nonsensical token-smasher word "light-make-puck". "schtenberg" - made-up "last-name-sound-alike" because "just prepend 'sch-' and add '-berg' or '-ung' and it appears as if it was German", right?! Well; CLIP found that pattern during its pre-training, too. =)
Would likely need to train CLIP on German (and other languages) to stop that, as it seems to be a "biased rant due to minority representation in training data". Maybe I'll spam GPT-4o @ OpenAI API with 40k labels for translating. =)
- Added exp-ft-**.py and eval-*.py scripts
⚠️ Extremely experimental Geometric Parameterization (GmP) inspired by this paper.
- Introduces sophisticated per-parameter learning needed for GmP fine-tune.
- Introduces gradient accumulation. Check "finetune" code for details.
- Otherwise, this mainly changes the model architecture (!) via custom Long-CLIP model code.
⚠️ It is normal / inevitable to get large or even 'inf' gradients in Epoch 0. But 'inf' should NOT happen in later epochs!- Optional: Use "exp-ft-X-visualize-theta-r-barplots.py" to visualize distribution of 'theta' and 'r' components (only works with GmP fine-tuned model).
- Use "exp-ft-C-convert-Long-GmP-back-to-weight.py" to convert fine-tune to normal model object. Otherwise, the model won't be compatible with any third party code at all!
- Once converted back to ".weight", you can use the full model object "as normal" and e.g. convert to state_dict with "ft-C-convert-for-SDXL-comfyUI-longCLIP.py".
- See the code comments in the eval-*.py files for info how to use them. Check model accuracy by evaluating against a dataset.
My GmP-CLIP, fine-tuned on https://huggingface.co/datasets/SPRIGHT-T2I/spright_coco Initially with short (<77 tokens) labels so I can compare it to CLIP (not-long, original-CLIP). You can find the short labels (if you need them for whatever reason) here: CLIP-fine-tune/tree/CLIP-vision/COCO
- GmP replaces linear layer ".weight" with GeometricLinear() for c_fc and c_proj in the MLP (multi-layer perceptron):
"Normal" CLIP MLP (multi-layer perceptron):
(mlp): Sequential(
|-(c_fc): Linear(in_features=1024, out_features=4096, bias=True)
| (gelu): QuickGELU()
|-}-(c_proj): Linear(in_features=4096, out_features=1024, bias=True)
| |
| |-- visual.transformer.resblocks.0.mlp.c_fc.weight
| |-- visual.transformer.resblocks.0.mlp.c_fc.bias
|
|---- visual.transformer.resblocks.0.mlp.c_proj.weight
|---- visual.transformer.resblocks.0.mlp.c_proj.bias
GmP CLIP MLP:
Weight decomposition into:
- radial component 'r' as norm of pre-trained weights
- angular component 'theta' as normalized direction
-> preserves weight vectors' directionality and magnitude
(mlp): Sequential(
|-(c_fc): GeometricLinear()
| (gelu): QuickGELU()
|-}-(c_proj): GeometricLinear()
| |
| |-- visual.transformer.resblocks.0.mlp.c_fc.r
| |-- visual.transformer.resblocks.0.mlp.c_fc.theta
| |-- visual.transformer.resblocks.0.mlp.c_fc.bias
|
|---- visual.transformer.resblocks.0.mlp.c_proj.r
|---- visual.transformer.resblocks.0.mlp.c_proj.theta
|---- visual.transformer.resblocks.0.mlp.c_proj.bias
(Same thing for [text] transformer.resblocks)
-
Refactored GA = Gradient Ascent, gets a CLIP "opinion" (text) about an image
-
(optimizes for cosine similarity of text embeddings with image embeddings)
-
Long-CLIP ViT-L/14 (the model guiding stable diffusion) now fits in <24 GB memory!
-
Approx. 1.5 minutes / image (RTX 4090) / uses torch.cuda.amp / autocast + GradScaler
-
To use: python longclipga_AMP.py --image_path "IMG_IN/catpiz.png"
-
Likewise, longclipga_AMP_anti.py gets the cosine "DIS-similarity" ("opposite of") an image
-
There is no antonym to "cat" in real life - but in CLIP's embeddings, there is!
-
Use run_longclipga_AMP_opposites.py for both (batch) -> "What's most ALIKE to the image?" + "What's most UNLIKE the image?"
-
Saves output (all + best words) to "TOK" folder / txt files. -- Pro Tip: Use "best" to prompt SDXL. =)
-
⚠️ Highly recommended: Use "Sysmem Fallback" (NVIDIA Control Panel). It should fit in <24 GB VRAM - BUT that depends on what else is running on your box. Plus, you wouldn't want a CUDA OOM crash just because you opened your browser to a video. You can also lower the batch_size in the code, but that degrades CLIP's "opinion" quality (but try e.g. "8" if you absolutely must). -
You won't win benchmarks with throwing small batch_sizes at a big model such as ViT-L/14; but using a finetune as the text encoder for e.g. Stable Diffusion SDXL, this CLIP will win some hearts! 💙🤖
- Uses AMP (automatic mixed precision) + AdaBelief optimizer (optional: fall back to AdamW) + OneCycleLR scheduler with warmup
- Gradually unfreeze CLIP (optional) or train whole model (default) + set Learning Rate for individual parameters (optional)
- Debug print when exploding or vanishing gradients occur + Many fancy logs and plots with live training updates
- Converts a "desc.csv" from CLIP Interrogator to dataset labels .json.
- Example: ft-X-example-my-dataset-labels.json is the expected format for my fine-tuning script; if you have a different format - e.g. single text files next to images - explain that to GPT-4, Claude 3, or any other AI assistant + "and I need to convert them to be labels in a single .json file that should look like so:" copy-paste the content of ft-X-example-my-dataset-labels.json into prompt as a one-shot example
- If you load your dataset: dataset1 = ImageTextDataset("path/to/image/folder", "path/to/my-text-labels.json", transform=preprocess), and inside the .json images are: "subpath/to/0001.jpg" -> then the script dataloader will look for the image in "path/to/image/folder/subpath/to/0001.jpg".
- Data augmentation: If your dataset is ~1000 images, consider augmenting the images by flipping them horizontally etc.
- The script example will create a copy of your images with color jitter, which prevents CLIP from overfitting on specific colors.
- Use augmented images with .json labels and randomly select from multiple labels for a given image. See code in (3) for details.
- Fine-tune CLIP. Insert dataset .json and path to images as per previous step. See code # comments for details.
- 10,000 text-image pairs can archive good fine-tuning results within 1-2 hours (RTX 4090).
- Convert the torch.save model .pt into a state_dict you can then just plug into SDXL as the text encoder.
- Easy as Pi with ComfyUI, see SeaArtLab/ComfyUI-Long-CLIP for details!
- Same random seed etc., just swapping out the original longCLIP-L model for my fine-tune. CFG scale 14 = high CLIP influence / guidance.
- Please note: The U-Net of SDXL was also trained on the same dataset, with a frozen CLIP (independent of CLIP).
- For fine-tuning the SDXL U-Net Diffusion Model to complement CLIP, please refer to kohya-ss/sd-scripts
Added run_visualization.py / 'vitvis' for LongCLIP feature activation max visualization
- Check run_visualization.py code # comments for instructions
- Based on hamidkazemi22/vit-visualization
- Added longclipga.py -> Get 'opinion' text from model about an image
- (Optimize cosine similarity of text embeddings for image embeddings)
- Check the code, I left comments.
- Original CLIP Gradient Ascent Script: Used with permission by Twitter / X: @advadnoun
- Added longclip-token-to-ID.py -> Get token <-> ID mapping
This repository is the official implementation of Long-CLIP
Long-CLIP: Unlocking the Long-Text Capability of CLIP
Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Jiaqi Wang
- 🔥 Long Input length Increase the maximum input length of CLIP from 77 to 248.
- 🔥 Strong Performace Improve the R@5 of long-caption text-image retrieval by 20% and traditional text-image retrieval by 6%.
- 🔥 Plug-in and play Can be directly applied in any work that requires long-text capability.
🚀 [2024/4/1] The training code is released!
🚀 [2024/3/25] The Inference code and models (LongCLIP-B and LongCLIP-L) are released!
🚀 [2024/3/25] The paper is released!
- Training code for Long-CLIP based on OpenAI-CLIP
- Evaluation code for Long-CLIP
- evaluation code for zero-shot classification and text-image retrieval tasks.
- Usage example of Long-CLIP
- Checkpoints of Long-CLIP
Our model is based on CLIP, please prepare environment for CLIP.
Please first clone our repo from github by running the following command.
git clone https://github.com/beichenzbc/Long-CLIP.git
cd Long-CLIPThen, download the checkpoints of our model LongCLIP-B and/or LongCLIP-L and place it under ./checkpoints
from model import longclip
import torch
from PIL import Image
device = "cuda" if torch.cuda.is_available() else "cpu"
model, preprocess = longclip.load("./checkpoints/longclip-B.pt", device=device)
text = longclip.tokenize(["A man is crossing the street with a red car parked nearby.", "A man is driving a car in an urban scene."]).to(device)
image = preprocess(Image.open("./img/demo.png")).unsqueeze(0).to(device)
with torch.no_grad():
image_features = model.encode_image(image)
text_features = model.encode_text(text)
logits_per_image, logits_per_text = model(image, text)
probs = logits_per_image.softmax(dim=-1).cpu().numpy()
print("Label probs:", probs) # prints: [[0.982 0.01799]]To run zero-shot classification on imagenet dataset, run the following command after preparing the data
cd eval/classification/imagenet
python imagenet.pySimilarly, run the following command for cifar datset
cd eval/classification/cifar
python cifar10.py #cifar10
python cifar100.py #cifar100To run text-image retrieval on COCO2017 or Flickr30k, run the following command after preparing the data
cd eval/retrieval
python coco.py #COCO2017
python flickr30k.py #Flickr30kPlease refer to train/train.md for training details.
