rwightman HF staff commited on
Commit
eadbfb0
1 Parent(s): 254c041

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +233 -0
README.md CHANGED
@@ -6,3 +6,236 @@ library_tag: open_clip
6
  license: mit
7
  ---
8
  # Model card for CLIP-convnext_large_d_320.laion2B-s29B-b131K-ft-soup
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6
  license: mit
7
  ---
8
  # Model card for CLIP-convnext_large_d_320.laion2B-s29B-b131K-ft-soup
9
+
10
+ # Table of Contents
11
+
12
+ 1. [Model Details](#model-details)
13
+ 2. [Uses](#uses)
14
+ 3. [Training Details](#training-details)
15
+ 4. [Evaluation](#evaluation)
16
+ 5. [Acknowledgements](#acknowledgements)
17
+ 6. [Citation](#citation)
18
+
19
+ # Model Details
20
+
21
+ ## Model Description
22
+
23
+ A series of CLIP [ConvNeXt-Large](https://arxiv.org/abs/2201.03545) (w/ extra text depth, vision MLP head) models trained on the LAION-2B (english) subset of [LAION-5B](https://arxiv.org/abs/2210.08402) using [OpenCLIP](https://github.com/mlfoundations/open_clip).
24
+
25
+ The models utilize:
26
+ * the [timm](https://github.com/rwightman/pytorch-image-models) ConvNeXt-Large model (`convnext_large`) as the image tower
27
+ * a MLP (`fc - gelu - drop - fc`) head in vision tower instead of the single projection of other CLIP models
28
+ * a text tower with same width but 4 layers more depth than ViT-L / RN50x16 models (depth 16, embed dim 768).
29
+
30
+ This 320x320 resolution model is a soup (weight average) of 3 fine-tunes of [CLIP-convnext_large_d.laion2B-s26B-b102K-augreg](https://huggingface.co/laion/CLIP-convnext_large_d.laion2B-s26B-b102K-augreg) at a higher resolution. It is an average of 3 fine-tunes from the final checkpoint of the original 256x256 training run w/ an additional ~2-3B samples for each fine-tune and a lower learning rate. Each fine-tune was a different learning rate (1e-4, 6e-5, 5e-5), and diff # of samples (3.2B, 2B, 2.5B).
31
+
32
+ At 320x320, the ConvNext-Large-D is significantly more efficient than the L/14 model at 336x336 that OpenAI fine-tuned. L/14-336 model is 2.5x more GMAC, 2.8x more activations, and 1.22x more parameters.
33
+
34
+ | Model | Dataset | Resolution | AugReg | Top-1 ImageNet Zero-Shot (%) |
35
+ | ----- | ------- | ---------- | ------------ | --------- |
36
+ | [convnext_large_d.laion2b_s26b_b102k-augreg](https://huggingface.co/laion/CLIP-convnext_large_d.laion2B-s26B-b102K-augreg) | LAION-2B | 256x256 | RRC (0.33, 1.0), RE (0.35), SD (0.1), D(0.1) | 75.9 |
37
+ | [convnext_large_d_320.laion2b_s29b_b131k-augreg-ft](https://huggingface.co/laion/CLIP-convnext_large_d_320.laion2B-s29B-b131K-augreg-ft) | LAION-2B | 320x320 | RRC (0.5, 1.0), RE (0.4), SD (0.1), D(0.0) | 76.6 |
38
+ | [convnext_large_d_320.laion2b_s29b_b131k-augreg-ft-soup](https://huggingface.co/laion/CLIP-convnext_large_d_320.laion2B-s29B-b131K-augreg-ft-soup) | LAION-2B | 320x320 | RRC (0.5, 1.0), RE (0.4), SD (0.1), D(0.0) | 76.9 |
39
+
40
+ RRC = Random Resize Crop (crop pcts), RE = Random Erasing (prob), SD = Stochastic Depth (prob) -- image tower only, D = Dropout (prob) -- image tower head only
41
+
42
+ LAION-A = LAION Aesthetic, an ~900M sample subset of LAION-2B with pHash dedupe and asthetic score filtering.
43
+
44
+ Model training done by Ross Wightman on the [stability.ai](https://stability.ai/) cluster.
45
+
46
+ # Uses
47
+
48
+ As per the original [OpenAI CLIP model card](https://github.com/openai/CLIP/blob/d50d76daa670286dd6cacf3bcd80b5e4823fc8e1/model-card.md), this model is intended as a research output for research communities. We hope that this model will enable researchers to better understand and explore zero-shot, arbitrary image classification. We also hope it can be used for interdisciplinary studies of the potential impact of such model.
49
+
50
+ The OpenAI CLIP paper includes a discussion of potential downstream impacts to provide an example for this sort of analysis. Additionally, the LAION-5B blog (https://laion.ai/blog/laion-5b/) and upcoming paper include additional discussion as it relates specifically to the training dataset.
51
+
52
+ ## Direct Use
53
+
54
+ Zero-shot image classification, image and text retrieval, among others.
55
+
56
+ ## Downstream Use
57
+
58
+ Image classification and other image task fine-tuning, linear probe image classification, image generation guiding and conditioning, among others.
59
+
60
+ ## Out-of-Scope Use
61
+
62
+ As per the OpenAI models,
63
+
64
+ **Any** deployed use case of the model - whether commercial or not - is currently out of scope. Non-deployed use cases such as image search in a constrained environment, are also not recommended unless there is thorough in-domain testing of the model with a specific, fixed class taxonomy. This is because our safety assessment demonstrated a high need for task specific testing especially given the variability of CLIP’s performance with different class taxonomies. This makes untested and unconstrained deployment of the model in any use case currently potentially harmful.
65
+
66
+ Certain use cases which would fall under the domain of surveillance and facial recognition are always out-of-scope regardless of performance of the model. This is because the use of artificial intelligence for tasks such as these can be premature currently given the lack of testing norms and checks to ensure its fair use.
67
+
68
+ Since the model has not been purposefully trained in or evaluated on any languages other than English, its use should be limited to English language use cases.
69
+
70
+ Further the above notice, the LAION-5B dataset used in training of these models has additional considerations, see below.
71
+
72
+ # Training Details
73
+
74
+ ## Training Data
75
+
76
+ This model was trained with one of (see table in intro):
77
+ * LAION-2B - A 2 billion sample English subset of LAION-5B (https://laion.ai/blog/laion-5b/).
78
+ * LAION-Aesthetic - A 900M sample subset of LAION-2B with pHash dedupe and asthetic score filtering
79
+
80
+ **IMPORTANT NOTE:** The motivation behind dataset creation is to democratize research and experimentation around large-scale multi-modal model training and handling of uncurated, large-scale datasets crawled from publically available internet. Our recommendation is therefore to use the dataset for research purposes. Be aware that this large-scale dataset is uncurated. Keep in mind that the uncurated nature of the dataset means that collected links may lead to strongly discomforting and disturbing content for a human viewer. Therefore, please use the demo links with caution and at your own risk. It is possible to extract a “safe” subset by filtering out samples based on the safety tags (using a customized trained NSFW classifier that we built). While this strongly reduces the chance for encountering potentially harmful content when viewing, we cannot entirely exclude the possibility for harmful content being still present in safe mode, so that the warning holds also there. We think that providing the dataset openly to broad research and other interested communities will allow for transparent investigation of benefits that come along with training large-scale models as well as pitfalls and dangers that may stay unreported or unnoticed when working with closed large datasets that remain restricted to a small community. Providing our dataset openly, we however do not recommend using it for creating ready-to-go industrial products, as the basic research about general properties and safety of such large-scale models, which we would like to encourage with this release, is still in progress.
81
+
82
+ ## Training Procedure
83
+
84
+ All 320x320 model fine-tunes were trained with a global batch size of 131072 for 10-16 checkpoint intervals of 203.7M samples for a total of ~2-3B samples seen over fine-tune.
85
+
86
+ For 320x320 models, a slurm script w/ srun below was used on 64 8-GPU (A100 40GB) nodes (Stability).
87
+
88
+ ```
89
+ /opt/slurm/sbin/srun --cpu_bind=v --accel-bind=gn python -m training.main \
90
+ --save-frequency 1 \
91
+ --name "convnext_large_320" \
92
+ --pretrained ""/runs/convnext_large_256/epoch_128.pt" \
93
+ --resume 'latest' \
94
+ --train-data="pipe:aws s3 cp s3://mybucket/path/{laion{00000..xxxxx}.tar -" \
95
+ --train-num-samples 203666042 \
96
+ --dataset-type webdataset \
97
+ --precision amp_bfloat16 \
98
+ --beta2 0.98 \
99
+ --warmup 2000 \
100
+ --batch-size=256 \
101
+ --epochs=12 \
102
+ --dataset-resampled \
103
+ --aug-cfg use_timm=True scale='(0.5, 1.0)' re_prob=0.4 \
104
+ --clip-grad-norm 5.0 \
105
+ --lr 5e-5 \
106
+ --workers=6 \
107
+ --model "convnext_large_d_320" \
108
+ --seed 0 \
109
+ --ddp-static-graph \
110
+ --local-loss \
111
+ --gather-with-grad \
112
+ --grad-checkpointing
113
+ ```
114
+
115
+ # Evaluation
116
+
117
+ Evaluation done with code in the [LAION CLIP Benchmark suite](https://github.com/LAION-AI/CLIP_benchmark).
118
+
119
+ ## Testing Data, Factors & Metrics
120
+
121
+ ### Testing Data
122
+
123
+ The testing is performed with VTAB+ (A combination of VTAB (https://arxiv.org/abs/1910.04867) w/ additional robustness datasets) for classification and COCO and Flickr for retrieval.
124
+
125
+ ## Results
126
+
127
+ The models achieve between 75.9 and 76.9 top-1 zero-shot accuracy on ImageNet-1k.
128
+
129
+ Zero-shot curve of origina from-scratch 256x256 training:
130
+ ![](convnext_large_zero_shot.png)
131
+
132
+ An initial round of benchmarks have been performed on a wider range of datasets, to be viewable at https://github.com/LAION-AI/CLIP_benchmark/blob/main/benchmark/results.ipynb
133
+
134
+ # Acknowledgements
135
+
136
+ Acknowledging [stability.ai](https://stability.ai/) for compute used to train this model.
137
+
138
+ # Citation
139
+
140
+ **BibTeX:**
141
+
142
+ LAION-5B
143
+ ```bibtex
144
+ @inproceedings{schuhmann2022laionb,
145
+ title={{LAION}-5B: An open large-scale dataset for training next generation image-text models},
146
+ author={Christoph Schuhmann and
147
+ Romain Beaumont and
148
+ Richard Vencu and
149
+ Cade W Gordon and
150
+ Ross Wightman and
151
+ Mehdi Cherti and
152
+ Theo Coombes and
153
+ Aarush Katta and
154
+ Clayton Mullis and
155
+ Mitchell Wortsman and
156
+ Patrick Schramowski and
157
+ Srivatsa R Kundurthy and
158
+ Katherine Crowson and
159
+ Ludwig Schmidt and
160
+ Robert Kaczmarczyk and
161
+ Jenia Jitsev},
162
+ booktitle={Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
163
+ year={2022},
164
+ url={https://openreview.net/forum?id=M3Y74vmsMcY}
165
+ }
166
+ ```
167
+
168
+ OpenCLIP software
169
+ ```bibtex
170
+ @software{ilharco_gabriel_2021_5143773,
171
+ author = {Ilharco, Gabriel and
172
+ Wortsman, Mitchell and
173
+ Wightman, Ross and
174
+ Gordon, Cade and
175
+ Carlini, Nicholas and
176
+ Taori, Rohan and
177
+ Dave, Achal and
178
+ Shankar, Vaishaal and
179
+ Namkoong, Hongseok and
180
+ Miller, John and
181
+ Hajishirzi, Hannaneh and
182
+ Farhadi, Ali and
183
+ Schmidt, Ludwig},
184
+ title = {OpenCLIP},
185
+ month = jul,
186
+ year = 2021,
187
+ note = {If you use this software, please cite it as below.},
188
+ publisher = {Zenodo},
189
+ version = {0.1},
190
+ doi = {10.5281/zenodo.5143773},
191
+ url = {https://doi.org/10.5281/zenodo.5143773}
192
+ }
193
+ ```
194
+
195
+ ```
196
+ @InProceedings{pmlr-v162-wortsman22a,
197
+ title = {Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time},
198
+ author = {Wortsman, Mitchell and Ilharco, Gabriel and Gadre, Samir Ya and Roelofs, Rebecca and Gontijo-Lopes, Raphael and Morcos, Ari S and Namkoong, Hongseok and Farhadi, Ali and Carmon, Yair and Kornblith, Simon and Schmidt, Ludwig},
199
+ booktitle = {Proceedings of the 39th International Conference on Machine Learning},
200
+ pages = {23965--23998},
201
+ year = {2022},
202
+ editor = {Chaudhuri, Kamalika and Jegelka, Stefanie and Song, Le and Szepesvari, Csaba and Niu, Gang and Sabato, Sivan},
203
+ volume = {162},
204
+ series = {Proceedings of Machine Learning Research},
205
+ month = {17--23 Jul},
206
+ publisher = {PMLR},
207
+ pdf = {https://proceedings.mlr.press/v162/wortsman22a/wortsman22a.pdf},
208
+ url = {https://proceedings.mlr.press/v162/wortsman22a.html}
209
+ }
210
+ ```
211
+
212
+ OpenAI CLIP paper
213
+ ```bibtex
214
+ @inproceedings{Radford2021LearningTV,
215
+ title={Learning Transferable Visual Models From Natural Language Supervision},
216
+ author={Alec Radford and Jong Wook Kim and Chris Hallacy and A. Ramesh and Gabriel Goh and Sandhini Agarwal and Girish Sastry and Amanda Askell and Pamela Mishkin and Jack Clark and Gretchen Krueger and Ilya Sutskever},
217
+ booktitle={ICML},
218
+ year={2021}
219
+ }
220
+ ```
221
+
222
+ ```bibtex
223
+ @Article{liu2022convnet,
224
+ author = {Zhuang Liu and Hanzi Mao and Chao-Yuan Wu and Christoph Feichtenhofer and Trevor Darrell and Saining Xie},
225
+ title = {A ConvNet for the 2020s},
226
+ journal = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
227
+ year = {2022},
228
+ }
229
+ ```
230
+
231
+ ```bibtex
232
+ @misc{rw2019timm,
233
+ author = {Ross Wightman},
234
+ title = {PyTorch Image Models},
235
+ year = {2019},
236
+ publisher = {GitHub},
237
+ journal = {GitHub repository},
238
+ doi = {10.5281/zenodo.4414861},
239
+ howpublished = {\url{https://github.com/rwightman/pytorch-image-models}}
240
+ }
241
+ ```