Mashaan blog
Self-Supervised Learning with I-JEPA
Acknowledgment
Thanks to the authors for making their code available. If I had any misunderstandings while reading the paper, I had to check the code to confirm it.
References
@InProceedings{Assran_2023_CVPR,
author = {Assran, Mahmoud and Duval, Quentin and Misra, Ishan and Bojanowski, Piotr and Vincent, Pascal and Rabbat, Michael and LeCun, Yann and Ballas, Nicolas},
title = {Self-Supervised Learning From Images With a Joint-Embedding Predictive Architecture},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2023},
pages = {15619-15629}
}
Main Architecture
Target encoder
Here is a printout of the architecture for the target encoder:
VisionTransformer(
(patch_embed): PatchEmbed(
(proj): Conv2d(3, 768, kernel_size=(14, 14), stride=(14, 14))
)
(blocks): ModuleList(
(0-11): 12 x Block(
(norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
(attn): Attention(
(qkv): Linear(in_features=768, out_features=2304, bias=True)
(attn_drop): Dropout(p=0.0, inplace=False)
(proj): Linear(in_features=768, out_features=768, bias=True)
(proj_drop): Dropout(p=0.0, inplace=False)
)
(drop_path): Identity()
(norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
(mlp): MLP(
(fc1): Linear(in_features=768, out_features=3072, bias=True)
(act): GELU(approximate='none')
(fc2): Linear(in_features=3072, out_features=768, bias=True)
(drop): Dropout(p=0.0, inplace=False)
)
)
)
(norm): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
)
Context encoder
Here is a printout of the architecture for the context encoder:
VisionTransformer(
(patch_embed): PatchEmbed(
(proj): Conv2d(3, 768, kernel_size=(14, 14), stride=(14, 14))
)
(blocks): ModuleList(
(0-11): 12 x Block(
(norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
(attn): Attention(
(qkv): Linear(in_features=768, out_features=2304, bias=True)
(attn_drop): Dropout(p=0.0, inplace=False)
(proj): Linear(in_features=768, out_features=768, bias=True)
(proj_drop): Dropout(p=0.0, inplace=False)
)
(drop_path): Identity()
(norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
(mlp): MLP(
(fc1): Linear(in_features=768, out_features=3072, bias=True)
(act): GELU(approximate='none')
(fc2): Linear(in_features=3072, out_features=768, bias=True)
(drop): Dropout(p=0.0, inplace=False)
)
)
)
(norm): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
)
Predictor
Here is a printout of the architecture for the predictor:
VisionTransformerPredictor(
(predictor_embed): Linear(in_features=768, out_features=384, bias=True)
(predictor_blocks): ModuleList(
(0-11): 12 x Block(
(norm1): LayerNorm((384,), eps=1e-06, elementwise_affine=True)
(attn): Attention(
(qkv): Linear(in_features=384, out_features=1152, bias=True)
(attn_drop): Dropout(p=0.0, inplace=False)
(proj): Linear(in_features=384, out_features=384, bias=True)
(proj_drop): Dropout(p=0.0, inplace=False)
)
(drop_path): Identity()
(norm2): LayerNorm((384,), eps=1e-06, elementwise_affine=True)
(mlp): MLP(
(fc1): Linear(in_features=384, out_features=1536, bias=True)
(act): GELU(approximate='none')
(fc2): Linear(in_features=1536, out_features=384, bias=True)
(drop): Dropout(p=0.0, inplace=False)
)
)
)
(predictor_norm): LayerNorm((384,), eps=1e-06, elementwise_affine=True)
(predictor_proj): Linear(in_features=384, out_features=768, bias=True)
)
Loss
A printout of the selected patches for the masks:
INFO:root:Epoch 1
itr: 0
unsupervised_loader length 8
imgs.shape torch.Size([128, 3, 224, 224])
masks_enc[0].shape torch.Size([128, 87])
masks_enc length 1
CurrentMask[0][:] tensor([ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 17, 18,
19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 32, 33, 34, 35, 36, 37,
38, 39, 40, 41, 42, 43, 44, 45, 46, 48, 49, 50, 51, 52, 53, 54, 55, 56,
57, 58, 59, 60, 61, 62, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75,
76, 77, 78, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91])
masks_pred[0].shape torch.Size([128, 35])
masks_pred length 4
CurrentMask[0][:] tensor([151, 152, 153, 154, 155, 156, 157, 167, 168, 169, 170, 171, 172, 173,
183, 184, 185, 186, 187, 188, 189, 199, 200, 201, 202, 203, 204, 205,
215, 216, 217, 218, 219, 220, 221])
CurrentMask[0][:] tensor([133, 134, 135, 136, 137, 138, 139, 149, 150, 151, 152, 153, 154, 155,
165, 166, 167, 168, 169, 170, 171, 181, 182, 183, 184, 185, 186, 187,
197, 198, 199, 200, 201, 202, 203])
CurrentMask[0][:] tensor([118, 119, 120, 121, 122, 123, 124, 134, 135, 136, 137, 138, 139, 140,
150, 151, 152, 153, 154, 155, 156, 166, 167, 168, 169, 170, 171, 172,
182, 183, 184, 185, 186, 187, 188])
CurrentMask[0][:] tensor([160, 161, 162, 163, 164, 165, 166, 176, 177, 178, 179, 180, 181, 182,
192, 193, 194, 195, 196, 197, 198, 208, 209, 210, 211, 212, 213, 214,
224, 225, 226, 227, 228, 229, 230])
A printout for running one iteration with displaying tensor shapes:
INFO:root:Epoch 1
itr: 0
unsupervised_loader length 8
imgs.shape torch.Size([128, 3, 224, 224])
masks_enc[0].shape torch.Size([128, 57])
masks_enc length 1
masks_pred[0].shape torch.Size([128, 42])
masks_pred length 4
inside forward_target
---------------encoder start-----------------
x.shape: torch.Size([128, 3, 224, 224])
x.shape: torch.Size([128, 256, 768])
pos_embed.shape: torch.Size([1, 256, 768])
x.shape after pos_embed: torch.Size([128, 256, 768])
x.shape after mask: torch.Size([128, 256, 768])
x.shape after blocks: torch.Size([128, 256, 768])
---------------encoder end-------------------
inside forward_context
---------------encoder start-----------------
x.shape: torch.Size([128, 3, 224, 224])
x.shape: torch.Size([128, 256, 768])
pos_embed.shape: torch.Size([1, 256, 768])
x.shape after pos_embed: torch.Size([128, 256, 768])
x.shape after mask: torch.Size([128, 57, 768])
x.shape after blocks: torch.Size([128, 57, 768])
---------------encoder end-------------------
---------------predictor start---------------
x.shape: torch.Size([128, 57, 768])
masks_x[0].shape torch.Size([128, 57])
masks_x length 1
masks[0].shape torch.Size([128, 42])
masks length 4
Batch Size: 128
x.shape after predictor_embed: torch.Size([128, 57, 384])
x_pos_embed.shape: torch.Size([128, 256, 384])
x.shape after adding positional embedding: torch.Size([128, 57, 384])
N_ctxt: 57
pos_embs.shape: torch.Size([512, 42, 384])
pred_tokens.shape: torch.Size([512, 42, 384])
x.shape after concat mask tokens: torch.Size([512, 99, 384])
x.shape after predictor_blocks: torch.Size([512, 99, 384])
x.shape after pulling mask tokens: torch.Size([512, 42, 384])
x.shape after predictor_proj: torch.Size([512, 42, 768])
---------------predictor end-----------------
INFO:root:[1, 0] loss: 0.468 masks: 57.0 42.0 [wd: 4.00e-02] [lr: 2.03e-04] [mem: 0.00e+00] (-1.0 ms)
INFO:root:[1, 0] grad_stats: [2.91e-02 1.51e-02] (1.46e-02, 3.33e-02)
itr: 1