timbrooks commited on
Commit
2afcb7e
1 Parent(s): 658ad5e

Add InstructPix2Pix

Browse files

Former-commit-id: 3626d699482f2419961432bff2e1763ccf55f6e7

LICENSE ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
1
+ Copyright 2023 Timothy Brooks, Aleksander Holynski, Alexei A. Efros
2
+
3
+ Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
4
+
5
+ The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
6
+
7
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
8
+
9
+ Portions of code and models (such as pretrained checkpoints, which are fine-tuned starting from released Stable Diffusion checkpoints) are derived from the Stable Diffusion codebase (https://github.com/CompVis/stable-diffusion). Further restrictions may apply. Please consult the Stable Diffusion license `stable_diffusion/LICENSE`. Modified code is denoted as such in comments at the start of each file.
README.md ADDED
@@ -0,0 +1,158 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # InstructPix2Pix: Learning to Follow Image Editing Instructions
2
+ ### [Project Page](https://www.timothybrooks.com/instruct-pix2pix/) | [Paper](https://arxiv.org/abs/2211.09800) | [Data](http://instruct-pix2pix.eecs.berkeley.edu/)
3
+ PyTorch implementation of InstructPix2Pix, an instruction-based image editing model, based on the original [CompVis/stable_diffusion](https://github.com/CompVis/stable-diffusion) repo. <br>
4
+
5
+ [InstructPix2Pix: Learning to Follow Image Editing Instructions](https://www.timothybrooks.com/instruct-pix2pix/)
6
+ [Tim Brooks](https://www.timothybrooks.com/)\*,
7
+ [Aleksander Holynski](https://holynski.org/)\*,
8
+ [Alexei A. Efros](https://people.eecs.berkeley.edu/~efros/) <br>
9
+ UC Berkeley <br>
10
+ \*denotes equal contribution
11
+
12
+ <img src='https://instruct-pix2pix.timothybrooks.com/teaser.jpg'/>
13
+
14
+ ## TL;DR: quickstart
15
+
16
+ To setup a conda environment, download a pretrained model, and edit an image:
17
+ ```
18
+ conda env create -f environment.yaml
19
+ conda activate ip2p
20
+ bash scripts/download_checkpoints.sh
21
+ python edit_cli.py --input imgs/example.jpg --output imgs/output.jpg --edit "turn him into a cyborg"
22
+
23
+ # Optionally, you can specify parameters:
24
+ # python edit_cli.py --steps 100 --resolution 512 --seed 0 --cfg-text 7.5 --cfg-image 1.2 --input imgs/example.jpg --output imgs/output.jpg --edit "turn him into a cyborg"
25
+ ```
26
+
27
+ ## Setup
28
+
29
+ Install all dependencies with:
30
+ ```
31
+ conda env create -f environment.yaml
32
+ ```
33
+
34
+ Download the pretrained models by running:
35
+ ```
36
+ bash scripts/download_checkpoints.sh
37
+ ```
38
+
39
+ ## Generated Dataset
40
+
41
+ Our image editing model is trained on a generated dataset consisting of 454,445 examples. Each example contains (1) an input image, (2) an editing instruction, and (3) an output edited image. We provide two versions of the dataset, one in which each pair of edited images is generated 100 times, and the best examples are chosen based on CLIP metrics (Section 3.1.2 in the paper) (`clip-filtered-dataset`), and one in which examples are randomly chosen (`random-sample-dataset`).
42
+
43
+ For the released version of this dataset, we've additionally filtered prompts and images for NSFW content. After NSFW filtering, the GPT-3 generated dataset contains 451,990 examples. The final image-pair datasets contain:
44
+
45
+ | | # of image editing examples | Dataset size |
46
+ |--|-----------------------|----------------------- |
47
+ | `random-sample-dataset` |451990|727GB|
48
+ | `clip-filtered-dataset` |313010|436GB|
49
+
50
+ To download one of these datasets, along with the entire NSFW-filtered text data, run the following command with the appropriate dataset name:
51
+
52
+ ```
53
+ bash scripts/download_data.sh clip-filtered-dataset
54
+ ```
55
+
56
+
57
+ ## Training InstructPix2Pix
58
+
59
+ Need to modify configs/instruct-pix2pix/default.yaml to point to the dataset in the right location. Need to also download the Stable Diffusion checkpoint from which to finetune.
60
+
61
+ ```
62
+ python stable_diffusion/main.py --name default --base configs/train.yaml --train --gpus 0,1,2,3,4,5,6,7
63
+ ```
64
+
65
+
66
+ ## Creating your own dataset
67
+
68
+ Our generated dataset of paired images and editing instructions is made in two phases: First, we use GPT-3 to generate text triplets: (a) a caption describing an image, (b) an edit instruction, (c) a caption describing the image after the edit. Then, we turn pairs of captions (before/after the edit) into pairs of images using Stable Diffusion and Prompt-to-Prompt.
69
+
70
+ ### (1) Generate a dataset of captions and instructions
71
+
72
+ We provide our generated dataset of captions and edit instructions [here](https://instruct-pix2pix.eecs.berkeley.edu/gpt-generated-prompts.jsonl). If you plan to use our captions+instructions, skip to step (2). Otherwise, if you would like to create your own text dataset, please follow steps (1.1-1.3) below. Note that generating very large datasets using GPT-3 can be expensive.
73
+
74
+ #### (1.1) Manually write a dataset of instructions and captions
75
+
76
+ The first step of the process is fine-tuning GPT-3. To do this, we made a dataset of 700 examples broadly covering of edits that we might want our model to be able to perform. Our examples are available here [here](https://instruct-pix2pix.eecs.berkeley.edu/human_written_examples.jsonl). These should be diverse and cover a wide range of possible captions and types of edits. Ideally, they should avoid duplication or significant overlap of captions and instructions. It is also important to be mindful of limitations of Stable Diffusion and Prompt-to-Prompt in writing these examples, such as inability to perform large spatial transformations (e.g., moving the camera, zooming in, swapping object locations).
77
+
78
+ Input prompts should closely match the distribution of input prompts used to generate the larger dataset. We sampled the 700 input prompts from LAION Improves Aesthetics 6.5+ dataset and also use this dataset for generating examples. We found this dataset is quite noisy (many of the captions are overly long and contain irrelevant text). For this reason, we also considered MSCOCO and LAION-COCO datasets, but ultimately chose LAION Improves Aesthetics 6.5+ due to its diversity of content, proper nouns, and artistic mediums. If you choose to use another dataset or combination of datasets as input to GPT-3 when generating examples, we recomend you sample the input prompts from the same distribution when manually writing training examples.
79
+
80
+ #### (1.2) Finetune GPT-3
81
+
82
+ The next step is to finetune a large language model to generate an edit instruction and edited caption from a new input caption. We use GPT-3 Davinci via the OpenAI API, although other language models could be used.
83
+
84
+ To prepare training data for GPT-3, one must setup an OpenAI developer account to access the needed APIs. Run the `prompts/prepare_for_gpt.py` script, which forms the prompts into the correct format by concatenating instructions and captions and adding delimiters and stop sequences.
85
+
86
+ ```bash
87
+ python dataset_creation/prepare_for_gpt.py prompts/human_written_examples.jsonl prompts/human_written_examples_for_gpt.jsonl
88
+ ```
89
+
90
+ Next, finetune GPT-3 via the OpenAI CLI. We provide an example below, although please refer to the official documentation here as best practices may change. We trained the Davinci model for a single epoch. You could experiment with smaller less expensive GPT-3 variants or with open source language models, although this may negatively hurt performance.
91
+
92
+ ```bash
93
+ openai api fine_tunes.create -t prompts/human_written_examples_for_gpt.jsonl -m davinci --n_epochs 1 --suffix "instruct-pix2pix"
94
+ ```
95
+
96
+ You can test out the finetuned GPT-3 model by launching the provided Gradio app:
97
+
98
+ ```bash
99
+ python prompt_app.py OPENAI_MODEL_NAME
100
+ ```
101
+
102
+ #### (1.3) Generate a large dataset of captions and instructions
103
+
104
+ We now use the finetuned GPT-3 model to generate a large dataset. Our dataset cost thousands of dollars to create. See `prompts/gen_instructions_and_captions.py` for the script which generates these examples. We recommend first generating a small number of examples and gradually increasing the scale to ensure the results are working as desired before increasing scale.
105
+
106
+ ```bash
107
+ python dataset_creation/generate_txt_dataset.py OPENAI_MODEL_NAME
108
+ ```
109
+
110
+ If you are generating at a very large scale (e.g., 100K+), it will be noteably faster to generate the dataset with multiple processes running in parallel. This can be accomplished by setting `--partitions=N` to a higher number and running multiple processes, setting each `--partition` to the corresponding value.
111
+
112
+ ```bash
113
+ python dataset_creation/generate_txt_dataset.py OPENAI_MODEL_NAME --partitions=10 --partition=0
114
+ ```
115
+
116
+ ### (2) Turn paired captions into paired images
117
+
118
+ The next step is to turn pairs of text captions into pairs of images. For this, we need to copy a pre-trained Stable Diffusion model checkpoint to `stable_diffusion/models/ldm/stable-diffusion-v1/`. For our model, we used [checkpoint v1.5](https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned.ckpt), but other versions may also work. It is also necessary to download a checkpoint for the Stable Diffusion autoencoder. We used the [new autoencoder](https://huggingface.co/stabilityai/sd-vae-ft-mse-original/resolve/main/vae-ft-mse-840000-ema-pruned.ckpt), which should be put in the same directory. Once all checkpoints have been downloaded, we can generate the dataset with the following command:
119
+
120
+ ```
121
+ python dataset_creation/generate_img_dataset.py data/instruct-pix2pix-dataset-000 data/gpt_generated_prompts.jsonl
122
+ ```
123
+
124
+ This command operates on a single GPU (typically a V100 or A100). To parallelize over many GPUs/machines, set `--n-partitions` to the total number of parallel jobs and `--partition` to the index of each job.
125
+
126
+ ```
127
+ python dataset_creation/generate_img_dataset.py data/instruct-pix2pix-dataset-000 data/gpt_generated_prompts.jsonl --n-partitions 100 --partition 0
128
+ ```
129
+
130
+ The default parameters match that of our dataset, although in practice you can use a smaller number of steps (e.g., `--steps=25`) to generate high quality data faster. By default, we generate 100 samples per prompt and use CLIP filtering to keep a max of 4 per prompt. You can experiment with fewer samples by setting `--n-samples`. The command below turns off CLIP filtering entirely and is therefore faster:
131
+
132
+ ```
133
+ python dataset_creation/generate_img_dataset.py data/instruct-pix2pix-dataset-000 data/gpt_generated_prompts.jsonl --n-samples 4 --clip-threshold 0 --clip-dir-threshold 0 --clip-img-threshold 0 --n-partitions 100 --partition 0
134
+ ```
135
+
136
+ After generating all of the dataset examples, run the following command below to create a list of the examples. This is needed for the dataset onject to efficiently be able to sample examples without needing to iterate over the entire dataset directory at the start of each training run.
137
+
138
+ ```
139
+ python dataset_creation/prepare_dataset.py data/instruct-pix2pix-dataset-000
140
+ ```
141
+
142
+ ## Comments
143
+
144
+ - Our codebase is based on the [Stable Diffusion codebase](https://github.com/CompVis/stable-diffusion).
145
+
146
+ ## BibTeX
147
+
148
+ ```
149
+ @article{brooks2022instructpix2pix,
150
+ title={InstructPix2Pix: Learning to Follow Image Editing Instructions},
151
+ author={Brooks, Tim and Holynski, Aleksander and Efros, Alexei A},
152
+ journal={arXiv preprint arXiv:2211.09800},
153
+ year={2022}
154
+ }
155
+ ```
156
+
157
+
158
+
configs/generate.yaml ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # File modified by authors of InstructPix2Pix from original (https://github.com/CompVis/stable-diffusion).
2
+ # See more details in LICENSE.
3
+
4
+ model:
5
+ base_learning_rate: 1.0e-04
6
+ target: stable_diffusion.ldm.models.diffusion.ddpm_edit.LatentDiffusion
7
+ params:
8
+ linear_start: 0.00085
9
+ linear_end: 0.0120
10
+ num_timesteps_cond: 1
11
+ log_every_t: 200
12
+ timesteps: 1000
13
+ first_stage_key: edited
14
+ cond_stage_key: edit
15
+ # image_size: 64
16
+ # image_size: 32
17
+ image_size: 16
18
+ channels: 4
19
+ cond_stage_trainable: false # Note: different from the one we trained before
20
+ conditioning_key: hybrid
21
+ monitor: val/loss_simple_ema
22
+ scale_factor: 0.18215
23
+ use_ema: true
24
+ load_ema: true
25
+
26
+ scheduler_config: # 10000 warmup steps
27
+ target: stable_diffusion.ldm.lr_scheduler.LambdaLinearScheduler
28
+ params:
29
+ warm_up_steps: [ 0 ]
30
+ cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
31
+ f_start: [ 1.e-6 ]
32
+ f_max: [ 1. ]
33
+ f_min: [ 1. ]
34
+
35
+ unet_config:
36
+ target: stable_diffusion.ldm.modules.diffusionmodules.openaimodel.UNetModel
37
+ params:
38
+ image_size: 32 # unused
39
+ in_channels: 8
40
+ out_channels: 4
41
+ model_channels: 320
42
+ attention_resolutions: [ 4, 2, 1 ]
43
+ num_res_blocks: 2
44
+ channel_mult: [ 1, 2, 4, 4 ]
45
+ num_heads: 8
46
+ use_spatial_transformer: True
47
+ transformer_depth: 1
48
+ context_dim: 768
49
+ use_checkpoint: True
50
+ legacy: False
51
+
52
+ first_stage_config:
53
+ target: stable_diffusion.ldm.models.autoencoder.AutoencoderKL
54
+ params:
55
+ embed_dim: 4
56
+ monitor: val/rec_loss
57
+ ddconfig:
58
+ double_z: true
59
+ z_channels: 4
60
+ resolution: 256
61
+ in_channels: 3
62
+ out_ch: 3
63
+ ch: 128
64
+ ch_mult:
65
+ - 1
66
+ - 2
67
+ - 4
68
+ - 4
69
+ num_res_blocks: 2
70
+ attn_resolutions: []
71
+ dropout: 0.0
72
+ lossconfig:
73
+ target: torch.nn.Identity
74
+
75
+ cond_stage_config:
76
+ target: stable_diffusion.ldm.modules.encoders.modules.FrozenCLIPEmbedder
77
+
78
+ data:
79
+ target: main.DataModuleFromConfig
80
+ params:
81
+ batch_size: 128
82
+ num_workers: 1
83
+ wrap: false
84
+ validation:
85
+ target: edit_dataset.EditDataset
86
+ params:
87
+ path: /shared/holynski/laion-aesthetics-6.5_edit-model=davinci-laion700-1epoch_samples=10000/laion-aesthetics-6.5_edit-model=davinci-laion700-1epoch_samples=10000
88
+ cache_dir: /shared/timbrooks/image-edit-data/caches
89
+ cache_name: davinci10k
90
+ split: val
91
+ min_text_sim: 0.2
92
+ min_image_sim: 0.75
93
+ min_direction_sim: 0.2
94
+ max_samples_per_prompt: 1
95
+ min_resize_res: 512
96
+ max_resize_res: 512
97
+ crop_res: 512
98
+ output_as_edit: False
99
+ real_input: True
configs/train.yaml ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # File modified by authors of InstructPix2Pix from original (https://github.com/CompVis/stable-diffusion).
2
+ # See more details in LICENSE.
3
+
4
+ model:
5
+ base_learning_rate: 1.0e-04
6
+ target: stable_diffusion.ldm.models.diffusion.ddpm_edit.LatentDiffusion
7
+ params:
8
+ ckpt_path: stable_diffusion/models/ldm/stable-diffusion-v1/v1-5-pruned-emaonly.ckpt
9
+ linear_start: 0.00085
10
+ linear_end: 0.0120
11
+ num_timesteps_cond: 1
12
+ log_every_t: 200
13
+ timesteps: 1000
14
+ first_stage_key: edited
15
+ cond_stage_key: edit
16
+ image_size: 32
17
+ channels: 4
18
+ cond_stage_trainable: false # Note: different from the one we trained before
19
+ conditioning_key: hybrid
20
+ monitor: val/loss_simple_ema
21
+ scale_factor: 0.18215
22
+ use_ema: true
23
+ load_ema: false
24
+
25
+ scheduler_config: # 10000 warmup steps
26
+ target: stable_diffusion.ldm.lr_scheduler.LambdaLinearScheduler
27
+ params:
28
+ warm_up_steps: [ 0 ]
29
+ cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
30
+ f_start: [ 1.e-6 ]
31
+ f_max: [ 1. ]
32
+ f_min: [ 1. ]
33
+
34
+ unet_config:
35
+ target: stable_diffusion.ldm.modules.diffusionmodules.openaimodel.UNetModel
36
+ params:
37
+ image_size: 32 # unused
38
+ in_channels: 8
39
+ out_channels: 4
40
+ model_channels: 320
41
+ attention_resolutions: [ 4, 2, 1 ]
42
+ num_res_blocks: 2
43
+ channel_mult: [ 1, 2, 4, 4 ]
44
+ num_heads: 8
45
+ use_spatial_transformer: True
46
+ transformer_depth: 1
47
+ context_dim: 768
48
+ use_checkpoint: True
49
+ legacy: False
50
+
51
+ first_stage_config:
52
+ target: stable_diffusion.ldm.models.autoencoder.AutoencoderKL
53
+ params:
54
+ embed_dim: 4
55
+ monitor: val/rec_loss
56
+ ddconfig:
57
+ double_z: true
58
+ z_channels: 4
59
+ resolution: 256
60
+ in_channels: 3
61
+ out_ch: 3
62
+ ch: 128
63
+ ch_mult:
64
+ - 1
65
+ - 2
66
+ - 4
67
+ - 4
68
+ num_res_blocks: 2
69
+ attn_resolutions: []
70
+ dropout: 0.0
71
+ lossconfig:
72
+ target: torch.nn.Identity
73
+
74
+ cond_stage_config:
75
+ target: stable_diffusion.ldm.modules.encoders.modules.FrozenCLIPEmbedder
76
+
77
+ data:
78
+ target: main.DataModuleFromConfig
79
+ params:
80
+ batch_size: 32
81
+ num_workers: 2
82
+ train:
83
+ target: edit_dataset.EditDataset
84
+ params:
85
+ path: /home/timbrooks/instruct-pix2pix-datasets/20-20-75
86
+ split: train
87
+ min_resize_res: 256
88
+ max_resize_res: 256
89
+ crop_res: 256
90
+ flip_prob: 0.5
91
+ validation:
92
+ target: edit_dataset.EditDataset
93
+ params:
94
+ path: /home/timbrooks/instruct-pix2pix-datasets/20-20-75
95
+ split: val
96
+ min_resize_res: 256
97
+ max_resize_res: 256
98
+ crop_res: 256
99
+
100
+ lightning:
101
+ callbacks:
102
+ image_logger:
103
+ target: main.ImageLogger
104
+ params:
105
+ batch_frequency: 2000
106
+ max_images: 2
107
+ increase_log_steps: False
108
+
109
+ trainer:
110
+ max_epochs: 2000
111
+ benchmark: True
112
+ accumulate_grad_batches: 4
113
+ check_val_every_n_epoch: 4
dataset_creation/generate_img_dataset.py ADDED
@@ -0,0 +1,297 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import json
3
+ from pathlib import Path
4
+
5
+ import k_diffusion
6
+ import numpy as np
7
+ import torch
8
+ import torch.nn as nn
9
+ from einops import rearrange, repeat
10
+ from omegaconf import OmegaConf
11
+ from PIL import Image
12
+ from pytorch_lightning import seed_everything
13
+ from tqdm import tqdm
14
+
15
+ from stable_diffusion.ldm.modules.attention import CrossAttention
16
+ from stable_diffusion.ldm.util import instantiate_from_config
17
+ from metrics.clip_similarity import ClipSimilarity
18
+
19
+
20
+ ################################################################################
21
+ # Modified K-diffusion Euler ancestral sampler with prompt-to-prompt.
22
+ # https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py
23
+
24
+
25
+ def append_dims(x, target_dims):
26
+ """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
27
+ dims_to_append = target_dims - x.ndim
28
+ if dims_to_append < 0:
29
+ raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less")
30
+ return x[(...,) + (None,) * dims_to_append]
31
+
32
+
33
+ def to_d(x, sigma, denoised):
34
+ """Converts a denoiser output to a Karras ODE derivative."""
35
+ return (x - denoised) / append_dims(sigma, x.ndim)
36
+
37
+
38
+ def get_ancestral_step(sigma_from, sigma_to):
39
+ """Calculates the noise level (sigma_down) to step down to and the amount
40
+ of noise to add (sigma_up) when doing an ancestral sampling step."""
41
+ sigma_up = min(sigma_to, (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5)
42
+ sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5
43
+ return sigma_down, sigma_up
44
+
45
+
46
+ def sample_euler_ancestral(model, x, sigmas, prompt2prompt_threshold=0.0, **extra_args):
47
+ """Ancestral sampling with Euler method steps."""
48
+ s_in = x.new_ones([x.shape[0]])
49
+ for i in range(len(sigmas) - 1):
50
+ prompt_to_prompt = prompt2prompt_threshold > i / (len(sigmas) - 2)
51
+ for m in model.modules():
52
+ if isinstance(m, CrossAttention):
53
+ m.prompt_to_prompt = prompt_to_prompt
54
+ denoised = model(x, sigmas[i] * s_in, **extra_args)
55
+ sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1])
56
+ d = to_d(x, sigmas[i], denoised)
57
+ # Euler method
58
+ dt = sigma_down - sigmas[i]
59
+ x = x + d * dt
60
+ if sigmas[i + 1] > 0:
61
+ # Make noise the same across all samples in batch.
62
+ x = x + torch.randn_like(x[:1]) * sigma_up
63
+ return x
64
+
65
+
66
+ ################################################################################
67
+
68
+
69
+ def load_model_from_config(config, ckpt, vae_ckpt=None, verbose=False):
70
+ print(f"Loading model from {ckpt}")
71
+ pl_sd = torch.load(ckpt, map_location="cpu")
72
+ if "global_step" in pl_sd:
73
+ print(f"Global Step: {pl_sd['global_step']}")
74
+ sd = pl_sd["state_dict"]
75
+ if vae_ckpt is not None:
76
+ print(f"Loading VAE from {vae_ckpt}")
77
+ vae_sd = torch.load(vae_ckpt, map_location="cpu")["state_dict"]
78
+ sd = {
79
+ k: vae_sd[k[len("first_stage_model.") :]] if k.startswith("first_stage_model.") else v
80
+ for k, v in sd.items()
81
+ }
82
+ model = instantiate_from_config(config.model)
83
+ m, u = model.load_state_dict(sd, strict=False)
84
+ if len(m) > 0 and verbose:
85
+ print("missing keys:")
86
+ print(m)
87
+ if len(u) > 0 and verbose:
88
+ print("unexpected keys:")
89
+ print(u)
90
+ return model
91
+
92
+
93
+ class CFGDenoiser(nn.Module):
94
+ def __init__(self, model):
95
+ super().__init__()
96
+ self.inner_model = model
97
+
98
+ def forward(self, x, sigma, uncond, cond, cfg_scale):
99
+ x_in = torch.cat([x] * 2)
100
+ sigma_in = torch.cat([sigma] * 2)
101
+ cond_in = torch.cat([uncond, cond])
102
+ uncond, cond = self.inner_model(x_in, sigma_in, cond=cond_in).chunk(2)
103
+ return uncond + (cond - uncond) * cfg_scale
104
+
105
+
106
+ def to_pil(image: torch.Tensor) -> Image.Image:
107
+ image = 255.0 * rearrange(image.cpu().numpy(), "c h w -> h w c")
108
+ image = Image.fromarray(image.astype(np.uint8))
109
+ return image
110
+
111
+
112
+ def main():
113
+ parser = argparse.ArgumentParser()
114
+ parser.add_argument(
115
+ "out_dir",
116
+ type=str,
117
+ help="Path to output dataset directory.",
118
+ )
119
+ parser.add_argument(
120
+ "prompts_file",
121
+ type=str,
122
+ help="Path to prompts .jsonl file.",
123
+ )
124
+ parser.add_argument(
125
+ "--steps",
126
+ type=int,
127
+ default=100,
128
+ help="Number of sampling steps.",
129
+ )
130
+ parser.add_argument(
131
+ "--n-samples",
132
+ type=int,
133
+ default=100,
134
+ help="Number of samples to generate per prompt (before CLIP filtering).",
135
+ )
136
+ parser.add_argument(
137
+ "--max-out-samples",
138
+ type=int,
139
+ default=4,
140
+ help="Max number of output samples to save per prompt (after CLIP filtering).",
141
+ )
142
+ parser.add_argument(
143
+ "--n-partitions",
144
+ type=int,
145
+ default=1,
146
+ help="Number of total partitions.",
147
+ )
148
+ parser.add_argument(
149
+ "--partition",
150
+ type=int,
151
+ default=0,
152
+ help="Partition index.",
153
+ )
154
+ parser.add_argument(
155
+ "--min-p2p",
156
+ type=float,
157
+ default=0.1,
158
+ help="Min prompt2prompt threshold (portion of denoising for which to fix self attention maps).",
159
+ )
160
+ parser.add_argument(
161
+ "--max-p2p",
162
+ type=float,
163
+ default=0.9,
164
+ help="Max prompt2prompt threshold (portion of denoising for which to fix self attention maps).",
165
+ )
166
+ parser.add_argument(
167
+ "--min-cfg",
168
+ type=float,
169
+ default=7.5,
170
+ help="Min classifier free guidance scale.",
171
+ )
172
+ parser.add_argument(
173
+ "--max-cfg",
174
+ type=float,
175
+ default=15,
176
+ help="Max classifier free guidance scale.",
177
+ )
178
+ parser.add_argument(
179
+ "--clip-threshold",
180
+ type=float,
181
+ default=0.2,
182
+ help="CLIP threshold for text-image similarity of each image.",
183
+ )
184
+ parser.add_argument(
185
+ "--clip-dir-threshold",
186
+ type=float,
187
+ default=0.2,
188
+ help="Directional CLIP threshold for similarity of change between pairs of text and pairs of images.",
189
+ )
190
+ parser.add_argument(
191
+ "--clip-img-threshold",
192
+ type=float,
193
+ default=0.7,
194
+ help="CLIP threshold for image-image similarity.",
195
+ )
196
+ opt = parser.parse_args()
197
+
198
+ global_seed = torch.randint(1 << 32, ()).item()
199
+ print(f"Global seed: {global_seed}")
200
+ seed_everything(global_seed)
201
+
202
+ model = load_model_from_config(
203
+ OmegaConf.load("configs/stable-diffusion/v1-inference.yaml"),
204
+ ckpt="models/ldm/stable-diffusion-v1/v1-5-pruned-emaonly.ckpt",
205
+ vae_ckpt="models/ldm/stable-diffusion-v1/vae-ft-mse-840000-ema-pruned.ckpt",
206
+ )
207
+ model.cuda().eval()
208
+ model_wrap = k_diffusion.external.CompVisDenoiser(model)
209
+
210
+ clip_similarity = ClipSimilarity().cuda()
211
+
212
+ out_dir = Path(opt.out_dir)
213
+ out_dir.mkdir(exist_ok=True, parents=True)
214
+
215
+ with open(opt.prompts_file) as fp:
216
+ prompts = [json.loads(line) for line in fp]
217
+
218
+ print(f"Partition index {opt.partition} ({opt.partition + 1} / {opt.n_partitions})")
219
+ prompts = np.array_split(list(enumerate(prompts)), opt.n_partitions)[opt.partition]
220
+
221
+ with torch.no_grad(), torch.autocast("cuda"), model.ema_scope():
222
+ uncond = model.get_learned_conditioning(2 * [""])
223
+ sigmas = model_wrap.get_sigmas(opt.steps)
224
+
225
+ for i, prompt in tqdm(prompts, desc="Prompts"):
226
+ prompt_dir = out_dir.joinpath(f"{i:07d}")
227
+ prompt_dir.mkdir(exist_ok=True)
228
+
229
+ with open(prompt_dir.joinpath("prompt.json"), "w") as fp:
230
+ json.dump(prompt, fp)
231
+
232
+ cond = model.get_learned_conditioning([prompt["input"], prompt["output"]])
233
+ results = {}
234
+
235
+ with tqdm(total=opt.n_samples, desc="Samples") as progress_bar:
236
+
237
+ while len(results) < opt.n_samples:
238
+ seed = torch.randint(1 << 32, ()).item()
239
+ if seed in results:
240
+ continue
241
+ torch.manual_seed(seed)
242
+
243
+ x = torch.randn(1, 4, 512 // 8, 512 // 8, device="cuda") * sigmas[0]
244
+ x = repeat(x, "1 ... -> n ...", n=2)
245
+
246
+ model_wrap_cfg = CFGDenoiser(model_wrap)
247
+ p2p_threshold = opt.min_p2p + torch.rand(()).item() * (opt.max_p2p - opt.min_p2p)
248
+ cfg_scale = opt.min_cfg + torch.rand(()).item() * (opt.max_cfg - opt.min_cfg)
249
+ extra_args = {"cond": cond, "uncond": uncond, "cfg_scale": cfg_scale}
250
+ samples_ddim = sample_euler_ancestral(model_wrap_cfg, x, sigmas, p2p_threshold, **extra_args)
251
+ x_samples_ddim = model.decode_first_stage(samples_ddim)
252
+ x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
253
+
254
+ x0 = x_samples_ddim[0]
255
+ x1 = x_samples_ddim[1]
256
+
257
+ clip_sim_0, clip_sim_1, clip_sim_dir, clip_sim_image = clip_similarity(
258
+ x0[None], x1[None], [prompt["input"]], [prompt["output"]]
259
+ )
260
+
261
+ results[seed] = dict(
262
+ image_0=to_pil(x0),
263
+ image_1=to_pil(x1),
264
+ p2p_threshold=p2p_threshold,
265
+ cfg_scale=cfg_scale,
266
+ clip_sim_0=clip_sim_0[0].item(),
267
+ clip_sim_1=clip_sim_1[0].item(),
268
+ clip_sim_dir=clip_sim_dir[0].item(),
269
+ clip_sim_image=clip_sim_image[0].item(),
270
+ )
271
+
272
+ progress_bar.update()
273
+
274
+ # CLIP filter to get best samples for each prompt.
275
+ metadata = [
276
+ (result["clip_sim_dir"], seed)
277
+ for seed, result in results.items()
278
+ if result["clip_sim_image"] >= opt.clip_img_threshold
279
+ and result["clip_sim_dir"] >= opt.clip_dir_threshold
280
+ and result["clip_sim_0"] >= opt.clip_threshold
281
+ and result["clip_sim_1"] >= opt.clip_threshold
282
+ ]
283
+ metadata.sort(reverse=True)
284
+ for _, seed in metadata[: opt.max_out_samples]:
285
+ result = results[seed]
286
+ image_0 = result.pop("image_0")
287
+ image_1 = result.pop("image_1")
288
+ image_0.save(prompt_dir.joinpath(f"{seed}_0.jpg"), quality=100)
289
+ image_1.save(prompt_dir.joinpath(f"{seed}_1.jpg"), quality=100)
290
+ with open(prompt_dir.joinpath(f"metadata.jsonl"), "a") as fp:
291
+ fp.write(f"{json.dumps(dict(seed=seed, **result))}\n")
292
+
293
+ print("Done.")
294
+
295
+
296
+ if __name__ == "__main__":
297
+ main()
dataset_creation/generate_txt_dataset.py ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from __future__ import annotations
2
+
3
+ import json
4
+ import time
5
+ from argparse import ArgumentParser
6
+ from pathlib import Path
7
+ from typing import Optional
8
+
9
+ import datasets
10
+ import numpy as np
11
+ import openai
12
+ from tqdm.auto import tqdm
13
+
14
+
15
+ DELIMITER_0 = "\n##\n"
16
+ DELIMITER_1 = "\n%%\n"
17
+ STOP = "\nEND"
18
+
19
+
20
+ def generate(
21
+ openai_model: str,
22
+ caption: str,
23
+ num_retries: int = 3,
24
+ max_tokens: int = 256,
25
+ temperature: float = 0.7,
26
+ top_p: float = 1.0,
27
+ frequency_penalty: float = 0.1,
28
+ presence_penalty: float = 0.0,
29
+ sleep_on_error: float = 1.0,
30
+ ) -> Optional[tuple[str, str]]:
31
+ for _ in range(1 + num_retries):
32
+ try:
33
+ response = openai.Completion.create(
34
+ model=openai_model,
35
+ prompt=caption + DELIMITER_0,
36
+ temperature=temperature,
37
+ max_tokens=max_tokens,
38
+ top_p=top_p,
39
+ frequency_penalty=frequency_penalty,
40
+ presence_penalty=presence_penalty,
41
+ stop=[STOP],
42
+ )
43
+ except Exception as e:
44
+ print(e)
45
+ time.sleep(sleep_on_error)
46
+ continue
47
+ output = response["choices"][0]["text"].split(DELIMITER_1)
48
+ if len(output) == 2:
49
+ instruction, edited_caption = output
50
+ results = openai.Moderation.create([instruction, edited_caption])["results"]
51
+ if results[0]["flagged"] or results[1]["flagged"]:
52
+ continue
53
+ if caption.strip().strip(".!?").lower() != edited_caption.strip().strip(".!?").lower():
54
+ return instruction, edited_caption
55
+
56
+
57
+ def main(openai_model: str, num_samples: int, num_partitions: int, partition: int, seed: int):
58
+ dataset = datasets.load_dataset("ChristophSchuhmann/improved_aesthetics_6.5plus", split="train")
59
+ # Other datasets we considered that may be worth trying:
60
+ # dataset = datasets.load_dataset("ChristophSchuhmann/MS_COCO_2017_URL_TEXT", split="train")
61
+ # dataset = datasets.load_dataset("laion/laion-coco", split="train")
62
+
63
+ np.random.seed(seed)
64
+ permutation = np.array_split(np.random.permutation(len(dataset)), num_partitions)[partition]
65
+ dataset = dataset[permutation]
66
+ captions = dataset["TEXT"]
67
+ urls = dataset["URL"]
68
+ output_path = f"prompts/dataset=laion-aesthetics-6.5_model={openai_model}_samples={num_samples}_partition={partition}.jsonl" # fmt: skip
69
+ print(f"Prompt file path: {output_path}")
70
+
71
+ count = 0
72
+ caption_set = set()
73
+ url_set = set()
74
+
75
+ if Path(output_path).exists():
76
+ with open(output_path, "r") as f:
77
+ for line in tqdm(f, desc="Resuming from existing prompts"):
78
+ prompt = json.loads(line)
79
+ if prompt["caption"] not in caption_set and prompt["url"] not in url_set:
80
+ caption_set.add(prompt["caption"])
81
+ url_set.add(prompt["url"])
82
+ count += 1
83
+
84
+ with open(output_path, "a") as fp:
85
+ with tqdm(total=num_samples - count, desc="Generating instructions and edited captions") as progress_bar:
86
+ for caption, url in zip(captions, urls):
87
+ if caption in caption_set or url in url_set:
88
+ continue
89
+ if openai.Moderation.create(caption)["results"][0]["flagged"]:
90
+ continue
91
+ edit_output = generate(caption)
92
+ if edit_output is not None:
93
+ edit, output = edit_output
94
+ fp.write(f"{json.dumps(dict(caption=caption, edit=edit, output=output, url=url))}\n")
95
+ count += 1
96
+ progress_bar.update()
97
+ caption_set.add(caption)
98
+ url_set.add(url)
99
+ if count == num_samples:
100
+ break
101
+
102
+
103
+ if __name__ == "__main__":
104
+ parser = ArgumentParser()
105
+ parser.add_argument("openai-api-key", type=str)
106
+ parser.add_argument("openai-model", type=str)
107
+ parser.add_argument("--num-samples", default=10000, type=int)
108
+ parser.add_argument("--num-partitions", default=1, type=int)
109
+ parser.add_argument("--partition", default=0, type=int)
110
+ parser.add_argument("--seed", default=0, type=int)
111
+ args = parser.parse_args()
112
+ openai.api_key = args.openai_api_key
113
+ main(args.openai_model, args.num_samples, args.num_partitions, args.partition, args.seed)
dataset_creation/prepare_dataset.py ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import json
2
+ from argparse import ArgumentParser
3
+ from pathlib import Path
4
+
5
+ from tqdm.auto import tqdm
6
+
7
+
8
+ def main():
9
+ parser = ArgumentParser()
10
+ parser.add_argument("dataset_dir")
11
+ args = parser.parse_args()
12
+ dataset_dir = Path(args.dataset_dir)
13
+
14
+ seeds = []
15
+ with tqdm(desc="Listing dataset image seeds") as progress_bar:
16
+ for prompt_dir in dataset_dir.iterdir():
17
+ if prompt_dir.is_dir():
18
+ prompt_seeds = [image_path.name.split("_")[0] for image_path in sorted(prompt_dir.glob("*_0.jpg"))]
19
+ if len(prompt_seeds) > 0:
20
+ seeds.append((prompt_dir.name, prompt_seeds))
21
+ progress_bar.update()
22
+ seeds.sort()
23
+
24
+ with open(dataset_dir.joinpath("seeds.json"), "w") as f:
25
+ json.dump(seeds, f)
26
+
27
+
28
+ if __name__ == "__main__":
29
+ main()
dataset_creation/prepare_for_gpt.py ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import json
2
+ from argparse import ArgumentParser
3
+
4
+ from .generate_txt_dataset import DELIMITER_0, DELIMITER_1, STOP
5
+
6
+
7
+ def main(input_path: str, output_path: str):
8
+ with open(input_path) as f:
9
+ prompts = [json.loads(l) for l in f]
10
+
11
+ with open(output_path, "w") as f:
12
+ for prompt in prompts:
13
+ prompt_for_gpt = {
14
+ "prompt": f"{prompt['input']}{DELIMITER_0}",
15
+ "completion": f"{prompt['edit']}{DELIMITER_1}{prompt['output']}{STOP}",
16
+ }
17
+ f.write(f"{json.dumps(prompt_for_gpt)}\n")
18
+
19
+
20
+ if __name__ == "__main__":
21
+ parser = ArgumentParser()
22
+ parser.add_argument("input-path", type=str)
23
+ parser.add_argument("output-path", type=str)
24
+ args = parser.parse_args()
25
+ main(args.input_path, args.output_path)
edit_app.py ADDED
@@ -0,0 +1,269 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from __future__ import annotations
2
+
3
+ import math
4
+ import random
5
+ import sys
6
+ from argparse import ArgumentParser
7
+
8
+ import einops
9
+ import gradio as gr
10
+ import k_diffusion as K
11
+ import numpy as np
12
+ import torch
13
+ import torch.nn as nn
14
+ from einops import rearrange
15
+ from omegaconf import OmegaConf
16
+ from PIL import Image, ImageOps
17
+ from torch import autocast
18
+
19
+ sys.path.append("./stable_diffusion")
20
+
21
+ from stable_diffusion.ldm.util import instantiate_from_config
22
+
23
+
24
+ help_text = """
25
+ If you're not getting what you want, there may be a few reasons:
26
+ 1. Is the image not changing enough? Your Image CFG weight may be too high. This value dictates how similar the output should be to the input. It's possible your edit requires larger changes from the original image, and your Image CFG weight isn't allowing that. Alternatively, your Text CFG weight may be too low. This value dictates how much to listen to the text instruction. The default Image CFG of 1.5 and Text CFG of 7.5 are a good starting point, but aren't necessarily optimal for each edit. Try:
27
+ * Decreasing the Image CFG weight, or
28
+ * Incerasing the Text CFG weight, or
29
+ 2. Conversely, is the image changing too much, such that the details in the original image aren't preserved? Try:
30
+ * Increasing the Image CFG weight, or
31
+ * Decreasing the Text CFG weight
32
+ 3. Try generating results with different random seeds by setting "Randomize Seed" and running generation multiple times. You can also try setting "Randomize CFG" to sample new Text CFG and Image CFG values each time.
33
+ 4. Rephrasing the instruction sometimes improves results (e.g., "turn him into a dog" vs. "make him a dog" vs. "as a dog").
34
+ 5. Increasing the number of steps sometimes improves results.
35
+ 6. Do faces look weird? The Stable Diffusion autoencoder has a hard time with faces that are small in the image. Try:
36
+ * Cropping the image so the face takes up a larger portion of the frame.
37
+ """
38
+
39
+
40
+ example_instructions = [
41
+ "Make it a picasso painting",
42
+ "as if it were by modigliani",
43
+ "convert to a bronze statue",
44
+ "Turn it into an anime.",
45
+ "have it look like a graphic novel",
46
+ "make him gain weight",
47
+ "what would he look like bald?",
48
+ "Have him smile",
49
+ "Put him in a cocktail party.",
50
+ "move him at the beach.",
51
+ "add dramatic lighting",
52
+ "Convert to black and white",
53
+ "What if it were snowing?",
54
+ "Give him a leather jacket",
55
+ "Turn him into a cyborg!",
56
+ "make him wear a beanie",
57
+ ]
58
+
59
+
60
+ class CFGDenoiser(nn.Module):
61
+ def __init__(self, model):
62
+ super().__init__()
63
+ self.inner_model = model
64
+
65
+ def forward(self, z, sigma, cond, uncond, text_cfg_scale, image_cfg_scale):
66
+ cfg_z = einops.repeat(z, "1 ... -> n ...", n=3)
67
+ cfg_sigma = einops.repeat(sigma, "1 ... -> n ...", n=3)
68
+ cfg_cond = {
69
+ "c_crossattn": [torch.cat([cond["c_crossattn"][0], uncond["c_crossattn"][0], uncond["c_crossattn"][0]])],
70
+ "c_concat": [torch.cat([cond["c_concat"][0], cond["c_concat"][0], uncond["c_concat"][0]])],
71
+ }
72
+ out_cond, out_img_cond, out_uncond = self.inner_model(cfg_z, cfg_sigma, cond=cfg_cond).chunk(3)
73
+ return out_uncond + text_cfg_scale * (out_cond - out_img_cond) + image_cfg_scale * (out_img_cond - out_uncond)
74
+
75
+
76
+ def load_model_from_config(config, ckpt, vae_ckpt=None, verbose=False, cached=False):
77
+ print(f"Cache: {cached}")
78
+ print(f"Loading model from {ckpt}")
79
+ pl_sd = torch.load(ckpt, map_location="cpu")
80
+ if "global_step" in pl_sd:
81
+ print(f"Global Step: {pl_sd['global_step']}")
82
+ sd = pl_sd["state_dict"]
83
+ if vae_ckpt is not None:
84
+ print(f"Loading VAE from {vae_ckpt}")
85
+ vae_sd = torch.load(vae_ckpt, map_location="cpu")["state_dict"]
86
+ sd = {
87
+ k: vae_sd[k[len("first_stage_model.") :]] if k.startswith("first_stage_model.") else v
88
+ for k, v in sd.items()
89
+ }
90
+ model = instantiate_from_config(config.model, cached=cached)
91
+ m, u = model.load_state_dict(sd, strict=False)
92
+ if len(m) > 0 and verbose:
93
+ print("missing keys:")
94
+ print(m)
95
+ if len(u) > 0 and verbose:
96
+ print("unexpected keys:")
97
+ print(u)
98
+ return model
99
+
100
+
101
+ def main():
102
+ parser = ArgumentParser()
103
+ parser.add_argument("--resolution", default=512, type=int)
104
+ parser.add_argument("--config", default="configs/instruct-pix2pix/generate.yaml", type=str)
105
+ parser.add_argument("--ckpt", default="checkpoints/instruct-pix2pix-00-20000.ckpt", type=str)
106
+ parser.add_argument("--vae-ckpt", default=None, type=str)
107
+ args = parser.parse_args()
108
+
109
+ config = OmegaConf.load(args.config)
110
+ model = load_model_from_config(config, args.ckpt, args.vae_ckpt)
111
+ model.eval().cuda()
112
+ model_wrap = K.external.CompVisDenoiser(model)
113
+ model_wrap_cfg = CFGDenoiser(model_wrap)
114
+ null_token = model.get_learned_conditioning([""])
115
+ example_image = Image.open("imgs/example.jpg").convert("RGB")
116
+
117
+ def load_example(
118
+ steps: int,
119
+ randomize_seed: bool,
120
+ seed: int,
121
+ randomize_cfg: bool,
122
+ text_cfg_scale: float,
123
+ image_cfg_scale: float,
124
+ ):
125
+ example_instruction = random.choice(example_instructions)
126
+ return [example_image, example_instruction] + generate(
127
+ example_image,
128
+ example_instruction,
129
+ steps,
130
+ randomize_seed,
131
+ seed,
132
+ randomize_cfg,
133
+ text_cfg_scale,
134
+ image_cfg_scale,
135
+ )
136
+
137
+ def generate(
138
+ input_image: Image.Image,
139
+ instruction: str,
140
+ steps: int,
141
+ randomize_seed: bool,
142
+ seed: int,
143
+ randomize_cfg: bool,
144
+ text_cfg_scale: float,
145
+ image_cfg_scale: float,
146
+ ):
147
+ seed = random.randint(0, 100000) if randomize_seed else seed
148
+ text_cfg_scale = round(random.uniform(6.0, 9.0), ndigits=2) if randomize_cfg else text_cfg_scale
149
+ image_cfg_scale = round(random.uniform(1.2, 1.8), ndigits=2) if randomize_cfg else image_cfg_scale
150
+
151
+ width, height = input_image.size
152
+ factor = args.resolution / max(width, height)
153
+ factor = math.ceil(min(width, height) * factor / 64) * 64 / min(width, height)
154
+ width = int((width * factor) // 64) * 64
155
+ height = int((height * factor) // 64) * 64
156
+ input_image = ImageOps.fit(input_image, (width, height), method=Image.Resampling.LANCZOS)
157
+
158
+ if instruction == "":
159
+ return [input_image, seed]
160
+
161
+ with torch.no_grad(), autocast("cuda"), model.ema_scope():
162
+ cond = {}
163
+ cond["c_crossattn"] = [model.get_learned_conditioning([instruction])]
164
+ input_image = 2 * torch.tensor(np.array(input_image)).float() / 255 - 1
165
+ input_image = rearrange(input_image, "h w c -> 1 c h w").to(model.device)
166
+ cond["c_concat"] = [model.encode_first_stage(input_image).mode()]
167
+
168
+ uncond = {}
169
+ uncond["c_crossattn"] = [null_token]
170
+ uncond["c_concat"] = [torch.zeros_like(cond["c_concat"][0])]
171
+
172
+ sigmas = model_wrap.get_sigmas(steps)
173
+
174
+ extra_args = {
175
+ "cond": cond,
176
+ "uncond": uncond,
177
+ "text_cfg_scale": text_cfg_scale,
178
+ "image_cfg_scale": image_cfg_scale,
179
+ }
180
+ torch.manual_seed(seed)
181
+ z = torch.randn_like(cond["c_concat"][0]) * sigmas[0]
182
+ z = K.sampling.sample_euler_ancestral(model_wrap_cfg, z, sigmas, extra_args=extra_args)
183
+ x = model.decode_first_stage(z)
184
+ x = torch.clamp((x + 1.0) / 2.0, min=0.0, max=1.0)
185
+ x = 255.0 * rearrange(x, "1 c h w -> h w c")
186
+ edited_image = Image.fromarray(x.type(torch.uint8).cpu().numpy())
187
+
188
+ return [seed, text_cfg_scale, image_cfg_scale, edited_image]
189
+
190
+ def reset():
191
+ return [50, "Randomize Seed", random.randint(0, 100000), "Fix CFG", 7.5, 1.5, None]
192
+
193
+ with gr.Blocks(css="footer {visibility: hidden}") as demo:
194
+ with gr.Row():
195
+ with gr.Column(scale=1, min_width=100):
196
+ generate_button = gr.Button("Generate")
197
+ with gr.Column(scale=1, min_width=100):
198
+ load_button = gr.Button("Load Example")
199
+ with gr.Column(scale=1, min_width=100):
200
+ reset_button = gr.Button("Reset")
201
+ with gr.Column(scale=3):
202
+ instruction = gr.Textbox(lines=1, label="Edit Instruction", interactive=True)
203
+
204
+ with gr.Row():
205
+ input_image = gr.Image(label="Input Image", type="pil", interactive=True)
206
+ edited_image = gr.Image(label=f"Edited Image", type="pil", interactive=False)
207
+ input_image.style(height=512, width=512)
208
+ edited_image.style(height=512, width=512)
209
+
210
+ with gr.Row():
211
+ steps = gr.Number(value=50, precision=0, label="Steps", interactive=True)
212
+ randomize_seed = gr.Radio(
213
+ ["Fix Seed", "Randomize Seed"],
214
+ value="Randomize Seed",
215
+ type="index",
216
+ show_label=False,
217
+ interactive=True,
218
+ )
219
+ seed = gr.Number(value=random.randint(0, 100000), precision=0, label="Seed", interactive=True)
220
+ randomize_cfg = gr.Radio(
221
+ ["Fix CFG", "Randomize CFG"],
222
+ value="Fix CFG",
223
+ type="index",
224
+ show_label=False,
225
+ interactive=True,
226
+ )
227
+ text_cfg_scale = gr.Number(value=7.5, label=f"Text CFG", interactive=True)
228
+ image_cfg_scale = gr.Number(value=1.5, label=f"Image CFG", interactive=True)
229
+
230
+ gr.Markdown(help_text)
231
+
232
+ load_button.click(
233
+ fn=load_example,
234
+ inputs=[
235
+ steps,
236
+ randomize_seed,
237
+ seed,
238
+ randomize_cfg,
239
+ text_cfg_scale,
240
+ image_cfg_scale,
241
+ ],
242
+ outputs=[input_image, instruction, seed, text_cfg_scale, image_cfg_scale, edited_image],
243
+ )
244
+ generate_button.click(
245
+ fn=generate,
246
+ inputs=[
247
+ input_image,
248
+ instruction,
249
+ steps,
250
+ randomize_seed,
251
+ seed,
252
+ randomize_cfg,
253
+ text_cfg_scale,
254
+ image_cfg_scale,
255
+ ],
256
+ outputs=[seed, text_cfg_scale, image_cfg_scale, edited_image],
257
+ )
258
+ reset_button.click(
259
+ fn=reset,
260
+ inputs=[],
261
+ outputs=[steps, randomize_seed, seed, randomize_cfg, text_cfg_scale, image_cfg_scale, edited_image],
262
+ )
263
+
264
+ demo.queue(concurrency_count=1)
265
+ demo.launch(share=True)
266
+
267
+
268
+ if __name__ == "__main__":
269
+ main()
edit_cli.py ADDED
@@ -0,0 +1,128 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from __future__ import annotations
2
+
3
+ import math
4
+ import random
5
+ import sys
6
+ from argparse import ArgumentParser
7
+
8
+ import einops
9
+ import k_diffusion as K
10
+ import numpy as np
11
+ import torch
12
+ import torch.nn as nn
13
+ from einops import rearrange
14
+ from omegaconf import OmegaConf
15
+ from PIL import Image, ImageOps
16
+ from torch import autocast
17
+
18
+ sys.path.append("./stable_diffusion")
19
+
20
+ from stable_diffusion.ldm.util import instantiate_from_config
21
+
22
+
23
+ class CFGDenoiser(nn.Module):
24
+ def __init__(self, model):
25
+ super().__init__()
26
+ self.inner_model = model
27
+
28
+ def forward(self, z, sigma, cond, uncond, text_cfg_scale, image_cfg_scale):
29
+ cfg_z = einops.repeat(z, "1 ... -> n ...", n=3)
30
+ cfg_sigma = einops.repeat(sigma, "1 ... -> n ...", n=3)
31
+ cfg_cond = {
32
+ "c_crossattn": [torch.cat([cond["c_crossattn"][0], uncond["c_crossattn"][0], uncond["c_crossattn"][0]])],
33
+ "c_concat": [torch.cat([cond["c_concat"][0], cond["c_concat"][0], uncond["c_concat"][0]])],
34
+ }
35
+ out_cond, out_img_cond, out_uncond = self.inner_model(cfg_z, cfg_sigma, cond=cfg_cond).chunk(3)
36
+ return out_uncond + text_cfg_scale * (out_cond - out_img_cond) + image_cfg_scale * (out_img_cond - out_uncond)
37
+
38
+
39
+ def load_model_from_config(config, ckpt, vae_ckpt=None, verbose=False):
40
+ print(f"Loading model from {ckpt}")
41
+ pl_sd = torch.load(ckpt, map_location="cpu")
42
+ if "global_step" in pl_sd:
43
+ print(f"Global Step: {pl_sd['global_step']}")
44
+ sd = pl_sd["state_dict"]
45
+ if vae_ckpt is not None:
46
+ print(f"Loading VAE from {vae_ckpt}")
47
+ vae_sd = torch.load(vae_ckpt, map_location="cpu")["state_dict"]
48
+ sd = {
49
+ k: vae_sd[k[len("first_stage_model.") :]] if k.startswith("first_stage_model.") else v
50
+ for k, v in sd.items()
51
+ }
52
+ model = instantiate_from_config(config.model)
53
+ m, u = model.load_state_dict(sd, strict=False)
54
+ if len(m) > 0 and verbose:
55
+ print("missing keys:")
56
+ print(m)
57
+ if len(u) > 0 and verbose:
58
+ print("unexpected keys:")
59
+ print(u)
60
+ return model
61
+
62
+
63
+ def main():
64
+ parser = ArgumentParser()
65
+ parser.add_argument("--resolution", default=512, type=int)
66
+ parser.add_argument("--steps", default=100, type=int)
67
+ parser.add_argument("--config", default="configs/generate.yaml", type=str)
68
+ parser.add_argument("--ckpt", default="checkpoints/instruct-pix2pix-00-20000.ckpt", type=str)
69
+ parser.add_argument("--vae-ckpt", default=None, type=str)
70
+ parser.add_argument("--input", required=True, type=str)
71
+ parser.add_argument("--output", required=True, type=str)
72
+ parser.add_argument("--edit", required=True, type=str)
73
+ parser.add_argument("--cfg-text", default=7.5, type=float)
74
+ parser.add_argument("--cfg-image", default=1.2, type=float)
75
+ parser.add_argument("--seed", type=int)
76
+ args = parser.parse_args()
77
+
78
+ config = OmegaConf.load(args.config)
79
+ model = load_model_from_config(config, args.ckpt, args.vae_ckpt)
80
+ model.eval().cuda()
81
+ model_wrap = K.external.CompVisDenoiser(model)
82
+ model_wrap_cfg = CFGDenoiser(model_wrap)
83
+ null_token = model.get_learned_conditioning([""])
84
+
85
+ seed = random.randint(0, 100000) if args.seed is None else args.seed
86
+ input_image = Image.open(args.input).convert("RGB")
87
+ width, height = input_image.size
88
+ factor = args.resolution / max(width, height)
89
+ factor = math.ceil(min(width, height) * factor / 64) * 64 / min(width, height)
90
+ width = int((width * factor) // 64) * 64
91
+ height = int((height * factor) // 64) * 64
92
+ input_image = ImageOps.fit(input_image, (width, height), method=Image.Resampling.LANCZOS)
93
+
94
+ if args.edit == "":
95
+ input_image.save(args.output)
96
+ return
97
+
98
+ with torch.no_grad(), autocast("cuda"), model.ema_scope():
99
+ cond = {}
100
+ cond["c_crossattn"] = [model.get_learned_conditioning([args.edit])]
101
+ input_image = 2 * torch.tensor(np.array(input_image)).float() / 255 - 1
102
+ input_image = rearrange(input_image, "h w c -> 1 c h w").to(model.device)
103
+ cond["c_concat"] = [model.encode_first_stage(input_image).mode()]
104
+
105
+ uncond = {}
106
+ uncond["c_crossattn"] = [null_token]
107
+ uncond["c_concat"] = [torch.zeros_like(cond["c_concat"][0])]
108
+
109
+ sigmas = model_wrap.get_sigmas(args.steps)
110
+
111
+ extra_args = {
112
+ "cond": cond,
113
+ "uncond": uncond,
114
+ "text_cfg_scale": args.cfg_text,
115
+ "image_cfg_scale": args.cfg_image,
116
+ }
117
+ torch.manual_seed(seed)
118
+ z = torch.randn_like(cond["c_concat"][0]) * sigmas[0]
119
+ z = K.sampling.sample_euler_ancestral(model_wrap_cfg, z, sigmas, extra_args=extra_args)
120
+ x = model.decode_first_stage(z)
121
+ x = torch.clamp((x + 1.0) / 2.0, min=0.0, max=1.0)
122
+ x = 255.0 * rearrange(x, "1 c h w -> h w c")
123
+ edited_image = Image.fromarray(x.type(torch.uint8).cpu().numpy())
124
+ edited_image.save(args.output)
125
+
126
+
127
+ if __name__ == "__main__":
128
+ main()
edit_dataset.py ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from __future__ import annotations
2
+
3
+ import json
4
+ import math
5
+ from pathlib import Path
6
+ from typing import Any
7
+
8
+ import numpy as np
9
+ import torch
10
+ import torchvision
11
+ from einops import rearrange
12
+ from PIL import Image
13
+ from torch.utils.data import Dataset
14
+
15
+
16
+ class EditDataset(Dataset):
17
+ def __init__(
18
+ self,
19
+ path: str,
20
+ split: str = "train",
21
+ splits: tuple[float, float, float] = (0.9, 0.05, 0.05),
22
+ min_resize_res: int = 256,
23
+ max_resize_res: int = 256,
24
+ crop_res: int = 256,
25
+ flip_prob: float = 0.0,
26
+ ):
27
+ assert split in ("train", "val", "test")
28
+ assert sum(splits) == 1
29
+ self.path = path
30
+ self.min_resize_res = min_resize_res
31
+ self.max_resize_res = max_resize_res
32
+ self.crop_res = crop_res
33
+ self.flip_prob = flip_prob
34
+
35
+ with open(Path(self.path, "seeds.json")) as f:
36
+ self.seeds = json.load(f)
37
+
38
+ split_0, split_1 = {
39
+ "train": (0.0, splits[0]),
40
+ "val": (splits[0], splits[0] + splits[1]),
41
+ "test": (splits[0] + splits[1], 1.0),
42
+ }[split]
43
+
44
+ idx_0 = math.floor(split_0 * len(self.seeds))
45
+ idx_1 = math.floor(split_1 * len(self.seeds))
46
+ self.seeds = self.seeds[idx_0:idx_1]
47
+
48
+ def __len__(self) -> int:
49
+ return len(self.seeds)
50
+
51
+ def __getitem__(self, i: int) -> dict[str, Any]:
52
+ name, seeds = self.seeds[i]
53
+ propt_dir = Path(self.path, name)
54
+ seed = seeds[torch.randint(0, len(seeds), ()).item()]
55
+ with open(propt_dir.joinpath("prompt.json")) as fp:
56
+ prompt = json.load(fp)["edit"]
57
+
58
+ image_0 = Image.open(propt_dir.joinpath(f"{seed}_0.jpg"))
59
+ image_1 = Image.open(propt_dir.joinpath(f"{seed}_1.jpg"))
60
+
61
+ reize_res = torch.randint(self.min_resize_res, self.max_resize_res + 1, ()).item()
62
+ image_0 = image_0.resize((reize_res, reize_res), Image.Resampling.LANCZOS)
63
+ image_1 = image_1.resize((reize_res, reize_res), Image.Resampling.LANCZOS)
64
+
65
+ image_0 = rearrange(2 * torch.tensor(np.array(image_0)).float() / 255 - 1, "h w c -> c h w")
66
+ image_1 = rearrange(2 * torch.tensor(np.array(image_1)).float() / 255 - 1, "h w c -> c h w")
67
+
68
+ crop = torchvision.transforms.RandomCrop(self.crop_res)
69
+ flip = torchvision.transforms.RandomHorizontalFlip(float(self.flip_prob))
70
+ image_0, image_1 = flip(crop(torch.cat((image_0, image_1)))).chunk(2)
71
+
72
+ return dict(edited=image_1, edit=dict(c_concat=image_0, c_crossattn=prompt))
environment.yaml ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # File modified by authors of InstructPix2Pix from original (https://github.com/CompVis/stable-diffusion).
2
+ # See more details in LICENSE.
3
+
4
+ name: ip2p
5
+ channels:
6
+ - pytorch
7
+ - defaults
8
+ dependencies:
9
+ - python=3.8.5
10
+ - pip=20.3
11
+ - cudatoolkit=11.3
12
+ - pytorch=1.11.0
13
+ - torchvision=0.12.0
14
+ - numpy=1.19.2
15
+ - pip:
16
+ - albumentations==0.4.3
17
+ - diffusers
18
+ - opencv-python==4.1.2.30
19
+ - pudb==2019.2
20
+ - invisible-watermark
21
+ - imageio==2.9.0
22
+ - imageio-ffmpeg==0.4.2
23
+ - pytorch-lightning==1.4.2
24
+ - omegaconf==2.1.1
25
+ - test-tube>=0.7.5
26
+ - streamlit>=0.73.1
27
+ - einops==0.3.0
28
+ - torch-fidelity==0.3.0
29
+ - transformers==4.19.2
30
+ - torchmetrics==0.6.0
31
+ - kornia==0.6
32
+ - -e git+https://github.com/CompVis/taming-transformers.git@master#egg=taming-transformers
33
+ - -e git+https://github.com/openai/CLIP.git@main#egg=clip
34
+ - openai
35
+ - gradio
36
+ - seaborn
37
+ - git+https://github.com/crowsonkb/k-diffusion.git
imgs/example.jpg ADDED
main.py ADDED
@@ -0,0 +1,797 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # File modified by authors of InstructPix2Pix from original (https://github.com/CompVis/stable-diffusion).
2
+ # See more details in LICENSE.
3
+
4
+ import argparse, os, sys, datetime, glob
5
+ import numpy as np
6
+ import time
7
+ import torch
8
+ import torchvision
9
+ import pytorch_lightning as pl
10
+ import json
11
+ import pickle
12
+
13
+ from packaging import version
14
+ from omegaconf import OmegaConf
15
+ from torch.utils.data import DataLoader, Dataset
16
+ from functools import partial
17
+ from PIL import Image
18
+
19
+ import torch.distributed as dist
20
+ from pytorch_lightning import seed_everything
21
+ from pytorch_lightning.trainer import Trainer
22
+ from pytorch_lightning.callbacks import ModelCheckpoint, Callback, LearningRateMonitor
23
+ from pytorch_lightning.utilities.distributed import rank_zero_only
24
+ from pytorch_lightning.utilities import rank_zero_info
25
+ from pytorch_lightning.plugins import DDPPlugin
26
+
27
+ sys.path.append("./stable_diffusion")
28
+
29
+ from stable_diffusion.ldm.data.base import Txt2ImgIterableBaseDataset
30
+ from stable_diffusion.ldm.util import instantiate_from_config
31
+
32
+
33
+ def get_parser(**parser_kwargs):
34
+ def str2bool(v):
35
+ if isinstance(v, bool):
36
+ return v
37
+ if v.lower() in ("yes", "true", "t", "y", "1"):
38
+ return True
39
+ elif v.lower() in ("no", "false", "f", "n", "0"):
40
+ return False
41
+ else:
42
+ raise argparse.ArgumentTypeError("Boolean value expected.")
43
+
44
+ parser = argparse.ArgumentParser(**parser_kwargs)
45
+ parser.add_argument(
46
+ "-n",
47
+ "--name",
48
+ type=str,
49
+ const=True,
50
+ default="",
51
+ nargs="?",
52
+ help="postfix for logdir",
53
+ )
54
+ parser.add_argument(
55
+ "-r",
56
+ "--resume",
57
+ type=str,
58
+ const=True,
59
+ default="",
60
+ nargs="?",
61
+ help="resume from logdir or checkpoint in logdir",
62
+ )
63
+ parser.add_argument(
64
+ "-b",
65
+ "--base",
66
+ nargs="*",
67
+ metavar="base_config.yaml",
68
+ help="paths to base configs. Loaded from left-to-right. "
69
+ "Parameters can be overwritten or added with command-line options of the form `--key value`.",
70
+ default=list(),
71
+ )
72
+ parser.add_argument(
73
+ "-t",
74
+ "--train",
75
+ type=str2bool,
76
+ const=True,
77
+ default=False,
78
+ nargs="?",
79
+ help="train",
80
+ )
81
+ parser.add_argument(
82
+ "--no-test",
83
+ type=str2bool,
84
+ const=True,
85
+ default=False,
86
+ nargs="?",
87
+ help="disable test",
88
+ )
89
+ parser.add_argument(
90
+ "-p",
91
+ "--project",
92
+ help="name of new or path to existing project"
93
+ )
94
+ parser.add_argument(
95
+ "-d",
96
+ "--debug",
97
+ type=str2bool,
98
+ nargs="?",
99
+ const=True,
100
+ default=False,
101
+ help="enable post-mortem debugging",
102
+ )
103
+ parser.add_argument(
104
+ "-s",
105
+ "--seed",
106
+ type=int,
107
+ default=23,
108
+ help="seed for seed_everything",
109
+ )
110
+ parser.add_argument(
111
+ "-f",
112
+ "--postfix",
113
+ type=str,
114
+ default="",
115
+ help="post-postfix for default name",
116
+ )
117
+ parser.add_argument(
118
+ "-l",
119
+ "--logdir",
120
+ type=str,
121
+ default="logs",
122
+ help="directory for logging dat shit",
123
+ )
124
+ parser.add_argument(
125
+ "--scale_lr",
126
+ action="store_true",
127
+ default=False,
128
+ help="scale base-lr by ngpu * batch_size * n_accumulate",
129
+ )
130
+ return parser
131
+
132
+
133
+ def nondefault_trainer_args(opt):
134
+ parser = argparse.ArgumentParser()
135
+ parser = Trainer.add_argparse_args(parser)
136
+ args = parser.parse_args([])
137
+ return sorted(k for k in vars(args) if getattr(opt, k) != getattr(args, k))
138
+
139
+
140
+ class WrappedDataset(Dataset):
141
+ """Wraps an arbitrary object with __len__ and __getitem__ into a pytorch dataset"""
142
+
143
+ def __init__(self, dataset):
144
+ self.data = dataset
145
+
146
+ def __len__(self):
147
+ return len(self.data)
148
+
149
+ def __getitem__(self, idx):
150
+ return self.data[idx]
151
+
152
+
153
+ def worker_init_fn(_):
154
+ worker_info = torch.utils.data.get_worker_info()
155
+
156
+ dataset = worker_info.dataset
157
+ worker_id = worker_info.id
158
+
159
+ if isinstance(dataset, Txt2ImgIterableBaseDataset):
160
+ split_size = dataset.num_records // worker_info.num_workers
161
+ # reset num_records to the true number to retain reliable length information
162
+ dataset.sample_ids = dataset.valid_ids[worker_id * split_size:(worker_id + 1) * split_size]
163
+ current_id = np.random.choice(len(np.random.get_state()[1]), 1)
164
+ return np.random.seed(np.random.get_state()[1][current_id] + worker_id)
165
+ else:
166
+ return np.random.seed(np.random.get_state()[1][0] + worker_id)
167
+
168
+
169
+ class DataModuleFromConfig(pl.LightningDataModule):
170
+ def __init__(self, batch_size, train=None, validation=None, test=None, predict=None,
171
+ wrap=False, num_workers=None, shuffle_test_loader=False, use_worker_init_fn=False,
172
+ shuffle_val_dataloader=False):
173
+ super().__init__()
174
+ self.batch_size = batch_size
175
+ self.dataset_configs = dict()
176
+ self.num_workers = num_workers if num_workers is not None else batch_size * 2
177
+ self.use_worker_init_fn = use_worker_init_fn
178
+ if train is not None:
179
+ self.dataset_configs["train"] = train
180
+ self.train_dataloader = self._train_dataloader
181
+ if validation is not None:
182
+ self.dataset_configs["validation"] = validation
183
+ self.val_dataloader = partial(self._val_dataloader, shuffle=shuffle_val_dataloader)
184
+ if test is not None:
185
+ self.dataset_configs["test"] = test
186
+ self.test_dataloader = partial(self._test_dataloader, shuffle=shuffle_test_loader)
187
+ if predict is not None:
188
+ self.dataset_configs["predict"] = predict
189
+ self.predict_dataloader = self._predict_dataloader
190
+ self.wrap = wrap
191
+
192
+ def prepare_data(self):
193
+ for data_cfg in self.dataset_configs.values():
194
+ instantiate_from_config(data_cfg)
195
+
196
+ def setup(self, stage=None):
197
+ self.datasets = dict(
198
+ (k, instantiate_from_config(self.dataset_configs[k]))
199
+ for k in self.dataset_configs)
200
+ if self.wrap:
201
+ for k in self.datasets:
202
+ self.datasets[k] = WrappedDataset(self.datasets[k])
203
+
204
+ def _train_dataloader(self):
205
+ is_iterable_dataset = isinstance(self.datasets['train'], Txt2ImgIterableBaseDataset)
206
+ if is_iterable_dataset or self.use_worker_init_fn:
207
+ init_fn = worker_init_fn
208
+ else:
209
+ init_fn = None
210
+ return DataLoader(self.datasets["train"], batch_size=self.batch_size,
211
+ num_workers=self.num_workers, shuffle=False if is_iterable_dataset else True,
212
+ worker_init_fn=init_fn, persistent_workers=True)
213
+
214
+ def _val_dataloader(self, shuffle=False):
215
+ if isinstance(self.datasets['validation'], Txt2ImgIterableBaseDataset) or self.use_worker_init_fn:
216
+ init_fn = worker_init_fn
217
+ else:
218
+ init_fn = None
219
+ return DataLoader(self.datasets["validation"],
220
+ batch_size=self.batch_size,
221
+ num_workers=self.num_workers,
222
+ worker_init_fn=init_fn,
223
+ shuffle=shuffle, persistent_workers=True)
224
+
225
+ def _test_dataloader(self, shuffle=False):
226
+ is_iterable_dataset = isinstance(self.datasets['train'], Txt2ImgIterableBaseDataset)
227
+ if is_iterable_dataset or self.use_worker_init_fn:
228
+ init_fn = worker_init_fn
229
+ else:
230
+ init_fn = None
231
+
232
+ # do not shuffle dataloader for iterable dataset
233
+ shuffle = shuffle and (not is_iterable_dataset)
234
+
235
+ return DataLoader(self.datasets["test"], batch_size=self.batch_size,
236
+ num_workers=self.num_workers, worker_init_fn=init_fn, shuffle=shuffle, persistent_workers=True)
237
+
238
+ def _predict_dataloader(self, shuffle=False):
239
+ if isinstance(self.datasets['predict'], Txt2ImgIterableBaseDataset) or self.use_worker_init_fn:
240
+ init_fn = worker_init_fn
241
+ else:
242
+ init_fn = None
243
+ return DataLoader(self.datasets["predict"], batch_size=self.batch_size,
244
+ num_workers=self.num_workers, worker_init_fn=init_fn, persistent_workers=True)
245
+
246
+
247
+ class SetupCallback(Callback):
248
+ def __init__(self, resume, now, logdir, ckptdir, cfgdir, config, lightning_config):
249
+ super().__init__()
250
+ self.resume = resume
251
+ self.now = now
252
+ self.logdir = logdir
253
+ self.ckptdir = ckptdir
254
+ self.cfgdir = cfgdir
255
+ self.config = config
256
+ self.lightning_config = lightning_config
257
+
258
+ def on_keyboard_interrupt(self, trainer, pl_module):
259
+ if trainer.global_rank == 0:
260
+ print("Summoning checkpoint.")
261
+ ckpt_path = os.path.join(self.ckptdir, "last.ckpt")
262
+ trainer.save_checkpoint(ckpt_path)
263
+
264
+ def on_pretrain_routine_start(self, trainer, pl_module):
265
+ if trainer.global_rank == 0:
266
+ # Create logdirs and save configs
267
+ # os.makedirs(self.logdir, exist_ok=True)
268
+ # os.makedirs(self.ckptdir, exist_ok=True)
269
+ # os.makedirs(self.cfgdir, exist_ok=True)
270
+
271
+ if "callbacks" in self.lightning_config:
272
+ if 'metrics_over_trainsteps_checkpoint' in self.lightning_config['callbacks']:
273
+ os.makedirs(os.path.join(self.ckptdir, 'trainstep_checkpoints'), exist_ok=True)
274
+ print("Project config")
275
+ print(OmegaConf.to_yaml(self.config))
276
+ OmegaConf.save(self.config,
277
+ os.path.join(self.cfgdir, "{}-project.yaml".format(self.now)))
278
+
279
+ print("Lightning config")
280
+ print(OmegaConf.to_yaml(self.lightning_config))
281
+ OmegaConf.save(OmegaConf.create({"lightning": self.lightning_config}),
282
+ os.path.join(self.cfgdir, "{}-lightning.yaml".format(self.now)))
283
+
284
+ def get_world_size():
285
+ if not dist.is_available():
286
+ return 1
287
+ if not dist.is_initialized():
288
+ return 1
289
+ return dist.get_world_size()
290
+
291
+ def all_gather(data):
292
+ """
293
+ Run all_gather on arbitrary picklable data (not necessarily tensors)
294
+ Args:
295
+ data: any picklable object
296
+ Returns:
297
+ list[data]: list of data gathered from each rank
298
+ """
299
+ world_size = get_world_size()
300
+ if world_size == 1:
301
+ return [data]
302
+
303
+ # serialized to a Tensor
304
+ origin_size = None
305
+ if not isinstance(data, torch.Tensor):
306
+ buffer = pickle.dumps(data)
307
+ storage = torch.ByteStorage.from_buffer(buffer)
308
+ tensor = torch.ByteTensor(storage).to("cuda")
309
+ else:
310
+ origin_size = data.size()
311
+ tensor = data.reshape(-1)
312
+
313
+ tensor_type = tensor.dtype
314
+
315
+ # obtain Tensor size of each rank
316
+ local_size = torch.LongTensor([tensor.numel()]).to("cuda")
317
+ size_list = [torch.LongTensor([0]).to("cuda") for _ in range(world_size)]
318
+ dist.all_gather(size_list, local_size)
319
+ size_list = [int(size.item()) for size in size_list]
320
+ max_size = max(size_list)
321
+
322
+ # receiving Tensor from all ranks
323
+ # we pad the tensor because torch all_gather does not support
324
+ # gathering tensors of different shapes
325
+ tensor_list = []
326
+ for _ in size_list:
327
+ tensor_list.append(torch.FloatTensor(size=(max_size,)).cuda().to(tensor_type))
328
+ if local_size != max_size:
329
+ padding = torch.FloatTensor(size=(max_size - local_size,)).cuda().to(tensor_type)
330
+ tensor = torch.cat((tensor, padding), dim=0)
331
+ dist.all_gather(tensor_list, tensor)
332
+
333
+ data_list = []
334
+ for size, tensor in zip(size_list, tensor_list):
335
+ if origin_size is None:
336
+ buffer = tensor.cpu().numpy().tobytes()[:size]
337
+ data_list.append(pickle.loads(buffer))
338
+ else:
339
+ buffer = tensor[:size]
340
+ data_list.append(buffer)
341
+
342
+ if origin_size is not None:
343
+ new_shape = [-1] + list(origin_size[1:])
344
+ resized_list = []
345
+ for data in data_list:
346
+ # suppose the difference of tensor size exist in first dimension
347
+ data = data.reshape(new_shape)
348
+ resized_list.append(data)
349
+
350
+ return resized_list
351
+ else:
352
+ return data_list
353
+
354
+ class ImageLogger(Callback):
355
+ def __init__(self, batch_frequency, max_images, clamp=True, increase_log_steps=True,
356
+ rescale=True, disabled=False, log_on_batch_idx=False, log_first_step=False,
357
+ log_images_kwargs=None):
358
+ super().__init__()
359
+ self.rescale = rescale
360
+ self.batch_freq = batch_frequency
361
+ self.max_images = max_images
362
+ self.logger_log_images = {
363
+ pl.loggers.TestTubeLogger: self._testtube,
364
+ }
365
+ self.log_steps = [2 ** n for n in range(6, int(np.log2(self.batch_freq)) + 1)]
366
+ if not increase_log_steps:
367
+ self.log_steps = [self.batch_freq]
368
+ self.clamp = clamp
369
+ self.disabled = disabled
370
+ self.log_on_batch_idx = log_on_batch_idx
371
+ self.log_images_kwargs = log_images_kwargs if log_images_kwargs else {}
372
+ self.log_first_step = log_first_step
373
+
374
+ @rank_zero_only
375
+ def _testtube(self, pl_module, images, batch_idx, split):
376
+ for k in images:
377
+ grid = torchvision.utils.make_grid(images[k])
378
+ grid = (grid + 1.0) / 2.0 # -1,1 -> 0,1; c,h,w
379
+
380
+ tag = f"{split}/{k}"
381
+ pl_module.logger.experiment.add_image(
382
+ tag, grid,
383
+ global_step=pl_module.global_step)
384
+
385
+ @rank_zero_only
386
+ def log_local(self, save_dir, split, images, prompts,
387
+ global_step, current_epoch, batch_idx):
388
+ root = os.path.join(save_dir, "images", split)
389
+ names = {"reals": "before", "inputs": "after", "reconstruction": "before-vq", "samples": "after-gen"}
390
+ # print(root)
391
+ for k in images:
392
+ grid = torchvision.utils.make_grid(images[k], nrow=8)
393
+ if self.rescale:
394
+ grid = (grid + 1.0) / 2.0 # -1,1 -> 0,1; c,h,w
395
+ grid = grid.transpose(0, 1).transpose(1, 2).squeeze(-1)
396
+ grid = grid.numpy()
397
+ grid = (grid * 255).astype(np.uint8)
398
+ filename = "gs-{:06}_e-{:06}_b-{:06}_{}.png".format(
399
+ global_step,
400
+ current_epoch,
401
+ batch_idx,
402
+ names[k])
403
+ path = os.path.join(root, filename)
404
+ os.makedirs(os.path.split(path)[0], exist_ok=True)
405
+ # print(path)
406
+ Image.fromarray(grid).save(path)
407
+
408
+ filename = "gs-{:06}_e-{:06}_b-{:06}_prompt.json".format(
409
+ global_step,
410
+ current_epoch,
411
+ batch_idx)
412
+ path = os.path.join(root, filename)
413
+ with open(path, "w") as f:
414
+ for p in prompts:
415
+ f.write(f"{json.dumps(p)}\n")
416
+
417
+ def log_img(self, pl_module, batch, batch_idx, split="train"):
418
+ check_idx = batch_idx if self.log_on_batch_idx else pl_module.global_step
419
+ if (self.check_frequency(check_idx) and # batch_idx % self.batch_freq == 0
420
+ hasattr(pl_module, "log_images") and
421
+ callable(pl_module.log_images) and
422
+ self.max_images > 0) or (split == "val" and batch_idx == 0):
423
+ logger = type(pl_module.logger)
424
+
425
+ is_train = pl_module.training
426
+ if is_train:
427
+ pl_module.eval()
428
+
429
+ with torch.no_grad():
430
+ images = pl_module.log_images(batch, split=split, **self.log_images_kwargs)
431
+
432
+ prompts = batch["edit"]["c_crossattn"][:self.max_images]
433
+ prompts = [p for ps in all_gather(prompts) for p in ps]
434
+
435
+ for k in images:
436
+ N = min(images[k].shape[0], self.max_images)
437
+ images[k] = images[k][:N]
438
+ images[k] = torch.cat(all_gather(images[k][:N]))
439
+ if isinstance(images[k], torch.Tensor):
440
+ images[k] = images[k].detach().cpu()
441
+ if self.clamp:
442
+ images[k] = torch.clamp(images[k], -1., 1.)
443
+
444
+ self.log_local(pl_module.logger.save_dir, split, images, prompts,
445
+ pl_module.global_step, pl_module.current_epoch, batch_idx)
446
+
447
+ logger_log_images = self.logger_log_images.get(logger, lambda *args, **kwargs: None)
448
+ logger_log_images(pl_module, images, pl_module.global_step, split)
449
+
450
+ if is_train:
451
+ pl_module.train()
452
+
453
+ def check_frequency(self, check_idx):
454
+ if ((check_idx % self.batch_freq) == 0 or (check_idx in self.log_steps)) and (
455
+ check_idx > 0 or self.log_first_step):
456
+ if len(self.log_steps) > 0:
457
+ self.log_steps.pop(0)
458
+ return True
459
+ return False
460
+
461
+ def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx):
462
+ if not self.disabled and (pl_module.global_step > 0 or self.log_first_step):
463
+ self.log_img(pl_module, batch, batch_idx, split="train")
464
+
465
+ def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx):
466
+ if not self.disabled and pl_module.global_step > 0:
467
+ self.log_img(pl_module, batch, batch_idx, split="val")
468
+ if hasattr(pl_module, 'calibrate_grad_norm'):
469
+ if (pl_module.calibrate_grad_norm and batch_idx % 25 == 0) and batch_idx > 0:
470
+ self.log_gradients(trainer, pl_module, batch_idx=batch_idx)
471
+
472
+
473
+ class CUDACallback(Callback):
474
+ # see https://github.com/SeanNaren/minGPT/blob/master/mingpt/callback.py
475
+ def on_train_epoch_start(self, trainer, pl_module):
476
+ # Reset the memory use counter
477
+ torch.cuda.reset_peak_memory_stats(trainer.root_gpu)
478
+ torch.cuda.synchronize(trainer.root_gpu)
479
+ self.start_time = time.time()
480
+
481
+ def on_train_epoch_end(self, trainer, pl_module, outputs):
482
+ torch.cuda.synchronize(trainer.root_gpu)
483
+ max_memory = torch.cuda.max_memory_allocated(trainer.root_gpu) / 2 ** 20
484
+ epoch_time = time.time() - self.start_time
485
+
486
+ try:
487
+ max_memory = trainer.training_type_plugin.reduce(max_memory)
488
+ epoch_time = trainer.training_type_plugin.reduce(epoch_time)
489
+
490
+ rank_zero_info(f"Average Epoch time: {epoch_time:.2f} seconds")
491
+ rank_zero_info(f"Average Peak memory {max_memory:.2f}MiB")
492
+ except AttributeError:
493
+ pass
494
+
495
+
496
+ if __name__ == "__main__":
497
+ # custom parser to specify config files, train, test and debug mode,
498
+ # postfix, resume.
499
+ # `--key value` arguments are interpreted as arguments to the trainer.
500
+ # `nested.key=value` arguments are interpreted as config parameters.
501
+ # configs are merged from left-to-right followed by command line parameters.
502
+
503
+ # model:
504
+ # base_learning_rate: float
505
+ # target: path to lightning module
506
+ # params:
507
+ # key: value
508
+ # data:
509
+ # target: main.DataModuleFromConfig
510
+ # params:
511
+ # batch_size: int
512
+ # wrap: bool
513
+ # train:
514
+ # target: path to train dataset
515
+ # params:
516
+ # key: value
517
+ # validation:
518
+ # target: path to validation dataset
519
+ # params:
520
+ # key: value
521
+ # test:
522
+ # target: path to test dataset
523
+ # params:
524
+ # key: value
525
+ # lightning: (optional, has sane defaults and can be specified on cmdline)
526
+ # trainer:
527
+ # additional arguments to trainer
528
+ # logger:
529
+ # logger to instantiate
530
+ # modelcheckpoint:
531
+ # modelcheckpoint to instantiate
532
+ # callbacks:
533
+ # callback1:
534
+ # target: importpath
535
+ # params:
536
+ # key: value
537
+
538
+ now = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S")
539
+
540
+ # add cwd for convenience and to make classes in this file available when
541
+ # running as `python main.py`
542
+ # (in particular `main.DataModuleFromConfig`)
543
+ sys.path.append(os.getcwd())
544
+
545
+ parser = get_parser()
546
+ parser = Trainer.add_argparse_args(parser)
547
+
548
+ opt, unknown = parser.parse_known_args()
549
+
550
+ assert opt.name
551
+ cfg_fname = os.path.split(opt.base[0])[-1]
552
+ cfg_name = os.path.splitext(cfg_fname)[0]
553
+ nowname = f"{cfg_name}_{opt.name}"
554
+ logdir = os.path.join(opt.logdir, nowname)
555
+ ckpt = os.path.join(logdir, "checkpoints", "last.ckpt")
556
+
557
+ if os.path.isfile(ckpt):
558
+ opt.resume_from_checkpoint = ckpt
559
+ base_configs = sorted(glob.glob(os.path.join(logdir, "configs/*.yaml")))
560
+ opt.base = base_configs + opt.base
561
+ _tmp = logdir.split("/")
562
+ nowname = _tmp[-1]
563
+ # By default, when finetuning from Stable Diffusion, we load the EMA-only checkpoint to initialize all weights.
564
+ # If resuming InstructPix2Pix from a finetuning checkpoint, instead load both EMA and non-EMA weights.
565
+ opt.model.params.load_ema = True
566
+
567
+ ckptdir = os.path.join(logdir, "checkpoints")
568
+ cfgdir = os.path.join(logdir, "configs")
569
+
570
+ os.makedirs(logdir, exist_ok=True)
571
+ os.makedirs(ckptdir, exist_ok=True)
572
+ os.makedirs(cfgdir, exist_ok=True)
573
+
574
+ try:
575
+ # init and save configs
576
+ configs = [OmegaConf.load(cfg) for cfg in opt.base]
577
+ cli = OmegaConf.from_dotlist(unknown)
578
+ config = OmegaConf.merge(*configs, cli)
579
+ lightning_config = config.pop("lightning", OmegaConf.create())
580
+ # merge trainer cli with config
581
+ trainer_config = lightning_config.get("trainer", OmegaConf.create())
582
+ # default to ddp
583
+ trainer_config["accelerator"] = "ddp"
584
+ for k in nondefault_trainer_args(opt):
585
+ trainer_config[k] = getattr(opt, k)
586
+ if not "gpus" in trainer_config:
587
+ del trainer_config["accelerator"]
588
+ cpu = True
589
+ else:
590
+ gpuinfo = trainer_config["gpus"]
591
+ print(f"Running on GPUs {gpuinfo}")
592
+ cpu = False
593
+ trainer_opt = argparse.Namespace(**trainer_config)
594
+ lightning_config.trainer = trainer_config
595
+
596
+ # model
597
+ model = instantiate_from_config(config.model)
598
+
599
+ # trainer and callbacks
600
+ trainer_kwargs = dict()
601
+
602
+ # default logger configs
603
+ default_logger_cfgs = {
604
+ "wandb": {
605
+ "target": "pytorch_lightning.loggers.WandbLogger",
606
+ "params": {
607
+ "name": nowname,
608
+ "save_dir": logdir,
609
+ "id": nowname,
610
+ }
611
+ },
612
+ "testtube": {
613
+ "target": "pytorch_lightning.loggers.TestTubeLogger",
614
+ "params": {
615
+ "name": "testtube",
616
+ "save_dir": logdir,
617
+ }
618
+ },
619
+ }
620
+ default_logger_cfg = default_logger_cfgs["wandb"]
621
+ if "logger" in lightning_config:
622
+ logger_cfg = lightning_config.logger
623
+ else:
624
+ logger_cfg = OmegaConf.create()
625
+ logger_cfg = OmegaConf.merge(default_logger_cfg, logger_cfg)
626
+ trainer_kwargs["logger"] = instantiate_from_config(logger_cfg)
627
+
628
+ # modelcheckpoint - use TrainResult/EvalResult(checkpoint_on=metric) to
629
+ # specify which metric is used to determine best models
630
+ default_modelckpt_cfg = {
631
+ "target": "pytorch_lightning.callbacks.ModelCheckpoint",
632
+ "params": {
633
+ "dirpath": ckptdir,
634
+ "filename": "{epoch:06}",
635
+ "verbose": True,
636
+ "save_last": True,
637
+ }
638
+ }
639
+
640
+ if "modelcheckpoint" in lightning_config:
641
+ modelckpt_cfg = lightning_config.modelcheckpoint
642
+ else:
643
+ modelckpt_cfg = OmegaConf.create()
644
+ modelckpt_cfg = OmegaConf.merge(default_modelckpt_cfg, modelckpt_cfg)
645
+ print(f"Merged modelckpt-cfg: \n{modelckpt_cfg}")
646
+ if version.parse(pl.__version__) < version.parse('1.4.0'):
647
+ trainer_kwargs["checkpoint_callback"] = instantiate_from_config(modelckpt_cfg)
648
+
649
+ # add callback which sets up log directory
650
+ default_callbacks_cfg = {
651
+ "setup_callback": {
652
+ "target": "main.SetupCallback",
653
+ "params": {
654
+ "resume": opt.resume,
655
+ "now": now,
656
+ "logdir": logdir,
657
+ "ckptdir": ckptdir,
658
+ "cfgdir": cfgdir,
659
+ "config": config,
660
+ "lightning_config": lightning_config,
661
+ }
662
+ },
663
+ "image_logger": {
664
+ "target": "main.ImageLogger",
665
+ "params": {
666
+ "batch_frequency": 750,
667
+ "max_images": 4,
668
+ "clamp": True
669
+ }
670
+ },
671
+ "learning_rate_logger": {
672
+ "target": "main.LearningRateMonitor",
673
+ "params": {
674
+ "logging_interval": "step",
675
+ # "log_momentum": True
676
+ }
677
+ },
678
+ "cuda_callback": {
679
+ "target": "main.CUDACallback"
680
+ },
681
+ }
682
+ if version.parse(pl.__version__) >= version.parse('1.4.0'):
683
+ default_callbacks_cfg.update({'checkpoint_callback': modelckpt_cfg})
684
+
685
+ if "callbacks" in lightning_config:
686
+ callbacks_cfg = lightning_config.callbacks
687
+ else:
688
+ callbacks_cfg = OmegaConf.create()
689
+
690
+ print(
691
+ 'Caution: Saving checkpoints every n train steps without deleting. This might require some free space.')
692
+ default_metrics_over_trainsteps_ckpt_dict = {
693
+ 'metrics_over_trainsteps_checkpoint': {
694
+ "target": 'pytorch_lightning.callbacks.ModelCheckpoint',
695
+ 'params': {
696
+ "dirpath": os.path.join(ckptdir, 'trainstep_checkpoints'),
697
+ "filename": "{epoch:06}-{step:09}",
698
+ "verbose": True,
699
+ 'save_top_k': -1,
700
+ 'every_n_train_steps': 1000,
701
+ 'save_weights_only': True
702
+ }
703
+ }
704
+ }
705
+ default_callbacks_cfg.update(default_metrics_over_trainsteps_ckpt_dict)
706
+
707
+ callbacks_cfg = OmegaConf.merge(default_callbacks_cfg, callbacks_cfg)
708
+ if 'ignore_keys_callback' in callbacks_cfg and hasattr(trainer_opt, 'resume_from_checkpoint'):
709
+ callbacks_cfg.ignore_keys_callback.params['ckpt_path'] = trainer_opt.resume_from_checkpoint
710
+ elif 'ignore_keys_callback' in callbacks_cfg:
711
+ del callbacks_cfg['ignore_keys_callback']
712
+
713
+ trainer_kwargs["callbacks"] = [instantiate_from_config(callbacks_cfg[k]) for k in callbacks_cfg]
714
+
715
+ trainer = Trainer.from_argparse_args(trainer_opt, plugins=DDPPlugin(find_unused_parameters=False), **trainer_kwargs)
716
+ trainer.logdir = logdir ###
717
+
718
+ # data
719
+ data = instantiate_from_config(config.data)
720
+ # NOTE according to https://pytorch-lightning.readthedocs.io/en/latest/datamodules.html
721
+ # calling these ourselves should not be necessary but it is.
722
+ # lightning still takes care of proper multiprocessing though
723
+ data.prepare_data()
724
+ data.setup()
725
+ print("#### Data #####")
726
+ for k in data.datasets:
727
+ print(f"{k}, {data.datasets[k].__class__.__name__}, {len(data.datasets[k])}")
728
+
729
+ # configure learning rate
730
+ bs, base_lr = config.data.params.batch_size, config.model.base_learning_rate
731
+ if not cpu:
732
+ ngpu = len(lightning_config.trainer.gpus.strip(",").split(','))
733
+ else:
734
+ ngpu = 1
735
+ if 'accumulate_grad_batches' in lightning_config.trainer:
736
+ accumulate_grad_batches = lightning_config.trainer.accumulate_grad_batches
737
+ else:
738
+ accumulate_grad_batches = 1
739
+ print(f"accumulate_grad_batches = {accumulate_grad_batches}")
740
+ lightning_config.trainer.accumulate_grad_batches = accumulate_grad_batches
741
+ if opt.scale_lr:
742
+ model.learning_rate = accumulate_grad_batches * ngpu * bs * base_lr
743
+ print(
744
+ "Setting learning rate to {:.2e} = {} (accumulate_grad_batches) * {} (num_gpus) * {} (batchsize) * {:.2e} (base_lr)".format(
745
+ model.learning_rate, accumulate_grad_batches, ngpu, bs, base_lr))
746
+ else:
747
+ model.learning_rate = base_lr
748
+ print("++++ NOT USING LR SCALING ++++")
749
+ print(f"Setting learning rate to {model.learning_rate:.2e}")
750
+
751
+
752
+ # allow checkpointing via USR1
753
+ def melk(*args, **kwargs):
754
+ # run all checkpoint hooks
755
+ if trainer.global_rank == 0:
756
+ print("Summoning checkpoint.")
757
+ ckpt_path = os.path.join(ckptdir, "last.ckpt")
758
+ trainer.save_checkpoint(ckpt_path)
759
+
760
+
761
+ def divein(*args, **kwargs):
762
+ if trainer.global_rank == 0:
763
+ import pudb;
764
+ pudb.set_trace()
765
+
766
+
767
+ import signal
768
+
769
+ signal.signal(signal.SIGUSR1, melk)
770
+ signal.signal(signal.SIGUSR2, divein)
771
+
772
+ # run
773
+ if opt.train:
774
+ try:
775
+ trainer.fit(model, data)
776
+ except Exception:
777
+ melk()
778
+ raise
779
+ if not opt.no_test and not trainer.interrupted:
780
+ trainer.test(model, data)
781
+ except Exception:
782
+ if opt.debug and trainer.global_rank == 0:
783
+ try:
784
+ import pudb as debugger
785
+ except ImportError:
786
+ import pdb as debugger
787
+ debugger.post_mortem()
788
+ raise
789
+ finally:
790
+ # move newly created debug project to debug_runs
791
+ if opt.debug and not opt.resume and trainer.global_rank == 0:
792
+ dst, name = os.path.split(logdir)
793
+ dst = os.path.join(dst, "debug_runs", name)
794
+ os.makedirs(os.path.split(dst)[0], exist_ok=True)
795
+ os.rename(logdir, dst)
796
+ if trainer.global_rank == 0:
797
+ print(trainer.profiler.summary())
metrics/clip_similarity.py ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from __future__ import annotations
2
+
3
+ import clip
4
+ import torch
5
+ import torch.nn as nn
6
+ import torch.nn.functional as F
7
+ from einops import rearrange
8
+
9
+
10
+ class ClipSimilarity(nn.Module):
11
+ def __init__(self, name: str = "ViT-L/14"):
12
+ super().__init__()
13
+ assert name in ("RN50", "RN101", "RN50x4", "RN50x16", "RN50x64", "ViT-B/32", "ViT-B/16", "ViT-L/14", "ViT-L/14@336px") # fmt: skip
14
+ self.size = {"RN50x4": 288, "RN50x16": 384, "RN50x64": 448, "ViT-L/14@336px": 336}.get(name, 224)
15
+
16
+ self.model, _ = clip.load(name, device="cpu", download_root="./")
17
+ self.model.eval().requires_grad_(False)
18
+
19
+ self.register_buffer("mean", torch.tensor((0.48145466, 0.4578275, 0.40821073)))
20
+ self.register_buffer("std", torch.tensor((0.26862954, 0.26130258, 0.27577711)))
21
+
22
+ def encode_text(self, text: list[str]) -> torch.Tensor:
23
+ text = clip.tokenize(text, truncate=True).to(next(self.parameters()).device)
24
+ text_features = self.model.encode_text(text)
25
+ text_features = text_features / text_features.norm(dim=1, keepdim=True)
26
+ return text_features
27
+
28
+ def encode_image(self, image: torch.Tensor) -> torch.Tensor: # Input images in range [0, 1].
29
+ image = F.interpolate(image.float(), size=self.size, mode="bicubic", align_corners=False)
30
+ image = image - rearrange(self.mean, "c -> 1 c 1 1")
31
+ image = image / rearrange(self.std, "c -> 1 c 1 1")
32
+ image_features = self.model.encode_image(image)
33
+ image_features = image_features / image_features.norm(dim=1, keepdim=True)
34
+ return image_features
35
+
36
+ def forward(
37
+ self, image_0: torch.Tensor, image_1: torch.Tensor, text_0: list[str], text_1: list[str]
38
+ ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
39
+ image_features_0 = self.encode_image(image_0)
40
+ image_features_1 = self.encode_image(image_1)
41
+ text_features_0 = self.encode_text(text_0)
42
+ text_features_1 = self.encode_text(text_1)
43
+ sim_0 = F.cosine_similarity(image_features_0, text_features_0)
44
+ sim_1 = F.cosine_similarity(image_features_1, text_features_1)
45
+ sim_direction = F.cosine_similarity(image_features_1 - image_features_0, text_features_1 - text_features_0)
46
+ sim_image = F.cosine_similarity(image_features_0, image_features_1)
47
+ return sim_0, sim_1, sim_direction, sim_image
prompt_app.py ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from __future__ import annotations
2
+
3
+ from argparse import ArgumentParser
4
+
5
+ import datasets
6
+ import gradio as gr
7
+ import numpy as np
8
+ import openai
9
+
10
+ from dataset_creation.generate_txt_dataset import generate
11
+
12
+
13
+ def main(openai_model: str):
14
+ dataset = datasets.load_dataset("ChristophSchuhmann/improved_aesthetics_6.5plus", split="train")
15
+ captions = dataset[np.random.permutation(len(dataset))]["TEXT"]
16
+ index = 0
17
+
18
+ def click_random():
19
+ nonlocal index
20
+ output = captions[index]
21
+ index = (index + 1) % len(captions)
22
+ return output
23
+
24
+ def click_generate(input: str):
25
+ if input == "":
26
+ raise gr.Error("Input caption is missing!")
27
+ edit_output = generate(openai_model, input)
28
+ if edit_output is None:
29
+ return "Failed :(", "Failed :("
30
+ return edit_output
31
+
32
+ with gr.Blocks(css="footer {visibility: hidden}") as demo:
33
+ txt_input = gr.Textbox(lines=3, label="Input Caption", interactive=True, placeholder="Type image caption here...") # fmt: skip
34
+ txt_edit = gr.Textbox(lines=1, label="GPT-3 Instruction", interactive=False)
35
+ txt_output = gr.Textbox(lines=3, label="GPT3 Edited Caption", interactive=False)
36
+
37
+ with gr.Row():
38
+ clear_btn = gr.Button("Clear")
39
+ random_btn = gr.Button("Random Input")
40
+ generate_btn = gr.Button("Generate Instruction + Edited Caption")
41
+
42
+ clear_btn.click(fn=lambda: ("", "", ""), inputs=[], outputs=[txt_input, txt_edit, txt_output])
43
+ random_btn.click(fn=click_random, inputs=[], outputs=[txt_input])
44
+ generate_btn.click(fn=click_generate, inputs=[txt_input], outputs=[txt_edit, txt_output])
45
+
46
+ demo.launch(share=True)
47
+
48
+
49
+ if __name__ == "__main__":
50
+ parser = ArgumentParser()
51
+ parser.add_argument("openai-api-key", type=str)
52
+ parser.add_argument("openai-model", type=str)
53
+ args = parser.parse_args()
54
+ openai.api_key = args.openai_api_key
55
+ main(args.openai_model)
scripts/download_checkpoints.sh ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
4
+
5
+ mkdir -p $SCRIPT_DIR/../checkpoints
6
+
7
+ curl http://instruct-pix2pix.eecs.berkeley.edu/instruct-pix2pix-00-20000.ckpt -o $SCRIPT_DIR/../checkpoints/instruct-pix2pix-00-20000.ckpt
scripts/download_data.sh ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
4
+
5
+ mkdir -p $SCRIPT_DIR/../data
6
+
7
+ wget http://instruct-pix2pix.eecs.berkeley.edu/gpt-generated-prompts.jsonl -O $SCRIPT_DIR/../data/gpt-generated-prompts.jsonl
8
+ wget http://instruct-pix2pix.eecs.berkeley.edu/human-written-prompts.jsonl -O $SCRIPT_DIR/../data/human-written-prompts.jsonl
9
+
10
+ mkdir $SCRIPT_DIR/../data/$1
11
+ wget -A zip,json -r http://instruct-pix2pix.eecs.berkeley.edu/$1 -nd -P $SCRIPT_DIR/../data/$1
stable_diffusion/ldm/models/diffusion/ddpm_edit.py ADDED
@@ -0,0 +1,1459 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ wild mixture of
3
+ https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
4
+ https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
5
+ https://github.com/CompVis/taming-transformers
6
+ -- merci
7
+ """
8
+
9
+ # File modified by authors of InstructPix2Pix from original (https://github.com/CompVis/stable-diffusion).
10
+ # See more details in LICENSE.
11
+
12
+ import torch
13
+ import torch.nn as nn
14
+ import numpy as np
15
+ import pytorch_lightning as pl
16
+ from torch.optim.lr_scheduler import LambdaLR
17
+ from einops import rearrange, repeat
18
+ from contextlib import contextmanager
19
+ from functools import partial
20
+ from tqdm import tqdm
21
+ from torchvision.utils import make_grid
22
+ from pytorch_lightning.utilities.distributed import rank_zero_only
23
+
24
+ from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config
25
+ from ldm.modules.ema import LitEma
26
+ from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution
27
+ from ldm.models.autoencoder import VQModelInterface, IdentityFirstStage, AutoencoderKL
28
+ from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
29
+ from ldm.models.diffusion.ddim import DDIMSampler
30
+
31
+
32
+ __conditioning_keys__ = {'concat': 'c_concat',
33
+ 'crossattn': 'c_crossattn',
34
+ 'adm': 'y'}
35
+
36
+
37
+ def disabled_train(self, mode=True):
38
+ """Overwrite model.train with this function to make sure train/eval mode
39
+ does not change anymore."""
40
+ return self
41
+
42
+
43
+ def uniform_on_device(r1, r2, shape, device):
44
+ return (r1 - r2) * torch.rand(*shape, device=device) + r2
45
+
46
+
47
+ class DDPM(pl.LightningModule):
48
+ # classic DDPM with Gaussian diffusion, in image space
49
+ def __init__(self,
50
+ unet_config,
51
+ timesteps=1000,
52
+ beta_schedule="linear",
53
+ loss_type="l2",
54
+ ckpt_path=None,
55
+ ignore_keys=[],
56
+ load_only_unet=False,
57
+ monitor="val/loss",
58
+ use_ema=True,
59
+ first_stage_key="image",
60
+ image_size=256,
61
+ channels=3,
62
+ log_every_t=100,
63
+ clip_denoised=True,
64
+ linear_start=1e-4,
65
+ linear_end=2e-2,
66
+ cosine_s=8e-3,
67
+ given_betas=None,
68
+ original_elbo_weight=0.,
69
+ v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
70
+ l_simple_weight=1.,
71
+ conditioning_key=None,
72
+ parameterization="eps", # all assuming fixed variance schedules
73
+ scheduler_config=None,
74
+ use_positional_encodings=False,
75
+ learn_logvar=False,
76
+ logvar_init=0.,
77
+ load_ema=True,
78
+ ):
79
+ super().__init__()
80
+ assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"'
81
+ self.parameterization = parameterization
82
+ print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
83
+ self.cond_stage_model = None
84
+ self.clip_denoised = clip_denoised
85
+ self.log_every_t = log_every_t
86
+ self.first_stage_key = first_stage_key
87
+ self.image_size = image_size # try conv?
88
+ self.channels = channels
89
+ self.use_positional_encodings = use_positional_encodings
90
+ self.model = DiffusionWrapper(unet_config, conditioning_key)
91
+ count_params(self.model, verbose=True)
92
+ self.use_ema = use_ema
93
+
94
+ self.use_scheduler = scheduler_config is not None
95
+ if self.use_scheduler:
96
+ self.scheduler_config = scheduler_config
97
+
98
+ self.v_posterior = v_posterior
99
+ self.original_elbo_weight = original_elbo_weight
100
+ self.l_simple_weight = l_simple_weight
101
+
102
+ if monitor is not None:
103
+ self.monitor = monitor
104
+
105
+ if self.use_ema and load_ema:
106
+ self.model_ema = LitEma(self.model)
107
+ print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
108
+
109
+ if ckpt_path is not None:
110
+ self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet)
111
+
112
+ # If initialing from EMA-only checkpoint, create EMA model after loading.
113
+ if self.use_ema and not load_ema:
114
+ self.model_ema = LitEma(self.model)
115
+ print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
116
+
117
+ self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
118
+ linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
119
+
120
+ self.loss_type = loss_type
121
+
122
+ self.learn_logvar = learn_logvar
123
+ self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
124
+ if self.learn_logvar:
125
+ self.logvar = nn.Parameter(self.logvar, requires_grad=True)
126
+
127
+
128
+ def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
129
+ linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
130
+ if exists(given_betas):
131
+ betas = given_betas
132
+ else:
133
+ betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
134
+ cosine_s=cosine_s)
135
+ alphas = 1. - betas
136
+ alphas_cumprod = np.cumprod(alphas, axis=0)
137
+ alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
138
+
139
+ timesteps, = betas.shape
140
+ self.num_timesteps = int(timesteps)
141
+ self.linear_start = linear_start
142
+ self.linear_end = linear_end
143
+ assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
144
+
145
+ to_torch = partial(torch.tensor, dtype=torch.float32)
146
+
147
+ self.register_buffer('betas', to_torch(betas))
148
+ self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
149
+ self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
150
+
151
+ # calculations for diffusion q(x_t | x_{t-1}) and others
152
+ self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
153
+ self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
154
+ self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
155
+ self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
156
+ self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
157
+
158
+ # calculations for posterior q(x_{t-1} | x_t, x_0)
159
+ posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (
160
+ 1. - alphas_cumprod) + self.v_posterior * betas
161
+ # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
162
+ self.register_buffer('posterior_variance', to_torch(posterior_variance))
163
+ # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
164
+ self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
165
+ self.register_buffer('posterior_mean_coef1', to_torch(
166
+ betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
167
+ self.register_buffer('posterior_mean_coef2', to_torch(
168
+ (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
169
+
170
+ if self.parameterization == "eps":
171
+ lvlb_weights = self.betas ** 2 / (
172
+ 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
173
+ elif self.parameterization == "x0":
174
+ lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod))
175
+ else:
176
+ raise NotImplementedError("mu not supported")
177
+ # TODO how to choose this term
178
+ lvlb_weights[0] = lvlb_weights[1]
179
+ self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
180
+ assert not torch.isnan(self.lvlb_weights).all()
181
+
182
+ @contextmanager
183
+ def ema_scope(self, context=None):
184
+ if self.use_ema:
185
+ self.model_ema.store(self.model.parameters())
186
+ self.model_ema.copy_to(self.model)
187
+ if context is not None:
188
+ print(f"{context}: Switched to EMA weights")
189
+ try:
190
+ yield None
191
+ finally:
192
+ if self.use_ema:
193
+ self.model_ema.restore(self.model.parameters())
194
+ if context is not None:
195
+ print(f"{context}: Restored training weights")
196
+
197
+ def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
198
+ sd = torch.load(path, map_location="cpu")
199
+ if "state_dict" in list(sd.keys()):
200
+ sd = sd["state_dict"]
201
+ keys = list(sd.keys())
202
+
203
+ # Our model adds additional channels to the first layer to condition on an input image.
204
+ # For the first layer, copy existing channel weights and initialize new channel weights to zero.
205
+ input_keys = [
206
+ "model.diffusion_model.input_blocks.0.0.weight",
207
+ "model_ema.diffusion_modelinput_blocks00weight",
208
+ ]
209
+
210
+ self_sd = self.state_dict()
211
+ for input_key in input_keys:
212
+ if input_key not in sd or input_key not in self_sd:
213
+ continue
214
+
215
+ input_weight = self_sd[input_key]
216
+
217
+ if input_weight.size() != sd[input_key].size():
218
+ print(f"Manual init: {input_key}")
219
+ input_weight.zero_()
220
+ input_weight[:, :4, :, :].copy_(sd[input_key])
221
+ ignore_keys.append(input_key)
222
+
223
+ for k in keys:
224
+ for ik in ignore_keys:
225
+ if k.startswith(ik):
226
+ print("Deleting key {} from state_dict.".format(k))
227
+ del sd[k]
228
+ missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
229
+ sd, strict=False)
230
+ print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
231
+ if len(missing) > 0:
232
+ print(f"Missing Keys: {missing}")
233
+ if len(unexpected) > 0:
234
+ print(f"Unexpected Keys: {unexpected}")
235
+
236
+ def q_mean_variance(self, x_start, t):
237
+ """
238
+ Get the distribution q(x_t | x_0).
239
+ :param x_start: the [N x C x ...] tensor of noiseless inputs.
240
+ :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
241
+ :return: A tuple (mean, variance, log_variance), all of x_start's shape.
242
+ """
243
+ mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
244
+ variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
245
+ log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
246
+ return mean, variance, log_variance
247
+
248
+ def predict_start_from_noise(self, x_t, t, noise):
249
+ return (
250
+ extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
251
+ extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
252
+ )
253
+
254
+ def q_posterior(self, x_start, x_t, t):
255
+ posterior_mean = (
256
+ extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
257
+ extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
258
+ )
259
+ posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
260
+ posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
261
+ return posterior_mean, posterior_variance, posterior_log_variance_clipped
262
+
263
+ def p_mean_variance(self, x, t, clip_denoised: bool):
264
+ model_out = self.model(x, t)
265
+ if self.parameterization == "eps":
266
+ x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
267
+ elif self.parameterization == "x0":
268
+ x_recon = model_out
269
+ if clip_denoised:
270
+ x_recon.clamp_(-1., 1.)
271
+
272
+ model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
273
+ return model_mean, posterior_variance, posterior_log_variance
274
+
275
+ @torch.no_grad()
276
+ def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
277
+ b, *_, device = *x.shape, x.device
278
+ model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
279
+ noise = noise_like(x.shape, device, repeat_noise)
280
+ # no noise when t == 0
281
+ nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
282
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
283
+
284
+ @torch.no_grad()
285
+ def p_sample_loop(self, shape, return_intermediates=False):
286
+ device = self.betas.device
287
+ b = shape[0]
288
+ img = torch.randn(shape, device=device)
289
+ intermediates = [img]
290
+ for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
291
+ img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
292
+ clip_denoised=self.clip_denoised)
293
+ if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
294
+ intermediates.append(img)
295
+ if return_intermediates:
296
+ return img, intermediates
297
+ return img
298
+
299
+ @torch.no_grad()
300
+ def sample(self, batch_size=16, return_intermediates=False):
301
+ image_size = self.image_size
302
+ channels = self.channels
303
+ return self.p_sample_loop((batch_size, channels, image_size, image_size),
304
+ return_intermediates=return_intermediates)
305
+
306
+ def q_sample(self, x_start, t, noise=None):
307
+ noise = default(noise, lambda: torch.randn_like(x_start))
308
+ return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
309
+ extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
310
+
311
+ def get_loss(self, pred, target, mean=True):
312
+ if self.loss_type == 'l1':
313
+ loss = (target - pred).abs()
314
+ if mean:
315
+ loss = loss.mean()
316
+ elif self.loss_type == 'l2':
317
+ if mean:
318
+ loss = torch.nn.functional.mse_loss(target, pred)
319
+ else:
320
+ loss = torch.nn.functional.mse_loss(target, pred, reduction='none')
321
+ else:
322
+ raise NotImplementedError("unknown loss type '{loss_type}'")
323
+
324
+ return loss
325
+
326
+ def p_losses(self, x_start, t, noise=None):
327
+ noise = default(noise, lambda: torch.randn_like(x_start))
328
+ x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
329
+ model_out = self.model(x_noisy, t)
330
+
331
+ loss_dict = {}
332
+ if self.parameterization == "eps":
333
+ target = noise
334
+ elif self.parameterization == "x0":
335
+ target = x_start
336
+ else:
337
+ raise NotImplementedError(f"Paramterization {self.parameterization} not yet supported")
338
+
339
+ loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
340
+
341
+ log_prefix = 'train' if self.training else 'val'
342
+
343
+ loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()})
344
+ loss_simple = loss.mean() * self.l_simple_weight
345
+
346
+ loss_vlb = (self.lvlb_weights[t] * loss).mean()
347
+ loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb})
348
+
349
+ loss = loss_simple + self.original_elbo_weight * loss_vlb
350
+
351
+ loss_dict.update({f'{log_prefix}/loss': loss})
352
+
353
+ return loss, loss_dict
354
+
355
+ def forward(self, x, *args, **kwargs):
356
+ # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
357
+ # assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
358
+ t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
359
+ return self.p_losses(x, t, *args, **kwargs)
360
+
361
+ def get_input(self, batch, k):
362
+ return batch[k]
363
+
364
+ def shared_step(self, batch):
365
+ x = self.get_input(batch, self.first_stage_key)
366
+ loss, loss_dict = self(x)
367
+ return loss, loss_dict
368
+
369
+ def training_step(self, batch, batch_idx):
370
+ loss, loss_dict = self.shared_step(batch)
371
+
372
+ self.log_dict(loss_dict, prog_bar=True,
373
+ logger=True, on_step=True, on_epoch=True)
374
+
375
+ self.log("global_step", self.global_step,
376
+ prog_bar=True, logger=True, on_step=True, on_epoch=False)
377
+
378
+ if self.use_scheduler:
379
+ lr = self.optimizers().param_groups[0]['lr']
380
+ self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
381
+
382
+ return loss
383
+
384
+ @torch.no_grad()
385
+ def validation_step(self, batch, batch_idx):
386
+ _, loss_dict_no_ema = self.shared_step(batch)
387
+ with self.ema_scope():
388
+ _, loss_dict_ema = self.shared_step(batch)
389
+ loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema}
390
+ self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
391
+ self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
392
+
393
+ def on_train_batch_end(self, *args, **kwargs):
394
+ if self.use_ema:
395
+ self.model_ema(self.model)
396
+
397
+ def _get_rows_from_list(self, samples):
398
+ n_imgs_per_row = len(samples)
399
+ denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
400
+ denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
401
+ denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
402
+ return denoise_grid
403
+
404
+ @torch.no_grad()
405
+ def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
406
+ log = dict()
407
+ x = self.get_input(batch, self.first_stage_key)
408
+ N = min(x.shape[0], N)
409
+ n_row = min(x.shape[0], n_row)
410
+ x = x.to(self.device)[:N]
411
+ log["inputs"] = x
412
+
413
+ # get diffusion row
414
+ diffusion_row = list()
415
+ x_start = x[:n_row]
416
+
417
+ for t in range(self.num_timesteps):
418
+ if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
419
+ t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
420
+ t = t.to(self.device).long()
421
+ noise = torch.randn_like(x_start)
422
+ x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
423
+ diffusion_row.append(x_noisy)
424
+
425
+ log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
426
+
427
+ if sample:
428
+ # get denoise row
429
+ with self.ema_scope("Plotting"):
430
+ samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
431
+
432
+ log["samples"] = samples
433
+ log["denoise_row"] = self._get_rows_from_list(denoise_row)
434
+
435
+ if return_keys:
436
+ if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
437
+ return log
438
+ else:
439
+ return {key: log[key] for key in return_keys}
440
+ return log
441
+
442
+ def configure_optimizers(self):
443
+ lr = self.learning_rate
444
+ params = list(self.model.parameters())
445
+ if self.learn_logvar:
446
+ params = params + [self.logvar]
447
+ opt = torch.optim.AdamW(params, lr=lr)
448
+ return opt
449
+
450
+
451
+ class LatentDiffusion(DDPM):
452
+ """main class"""
453
+ def __init__(self,
454
+ first_stage_config,
455
+ cond_stage_config,
456
+ num_timesteps_cond=None,
457
+ cond_stage_key="image",
458
+ cond_stage_trainable=False,
459
+ concat_mode=True,
460
+ cond_stage_forward=None,
461
+ conditioning_key=None,
462
+ scale_factor=1.0,
463
+ scale_by_std=False,
464
+ load_ema=True,
465
+ *args, **kwargs):
466
+ self.num_timesteps_cond = default(num_timesteps_cond, 1)
467
+ self.scale_by_std = scale_by_std
468
+ assert self.num_timesteps_cond <= kwargs['timesteps']
469
+ # for backwards compatibility after implementation of DiffusionWrapper
470
+ if conditioning_key is None:
471
+ conditioning_key = 'concat' if concat_mode else 'crossattn'
472
+ if cond_stage_config == '__is_unconditional__':
473
+ conditioning_key = None
474
+ ckpt_path = kwargs.pop("ckpt_path", None)
475
+ ignore_keys = kwargs.pop("ignore_keys", [])
476
+ super().__init__(conditioning_key=conditioning_key, *args, load_ema=load_ema, **kwargs)
477
+ self.concat_mode = concat_mode
478
+ self.cond_stage_trainable = cond_stage_trainable
479
+ self.cond_stage_key = cond_stage_key
480
+ try:
481
+ self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
482
+ except:
483
+ self.num_downs = 0
484
+ if not scale_by_std:
485
+ self.scale_factor = scale_factor
486
+ else:
487
+ self.register_buffer('scale_factor', torch.tensor(scale_factor))
488
+ self.instantiate_first_stage(first_stage_config)
489
+ self.instantiate_cond_stage(cond_stage_config)
490
+ self.cond_stage_forward = cond_stage_forward
491
+ self.clip_denoised = False
492
+ self.bbox_tokenizer = None
493
+
494
+ self.restarted_from_ckpt = False
495
+ if ckpt_path is not None:
496
+ self.init_from_ckpt(ckpt_path, ignore_keys)
497
+ self.restarted_from_ckpt = True
498
+
499
+ if self.use_ema and not load_ema:
500
+ self.model_ema = LitEma(self.model)
501
+ print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
502
+
503
+ def make_cond_schedule(self, ):
504
+ self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
505
+ ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
506
+ self.cond_ids[:self.num_timesteps_cond] = ids
507
+
508
+ @rank_zero_only
509
+ @torch.no_grad()
510
+ def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
511
+ # only for very first batch
512
+ if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
513
+ assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
514
+ # set rescale weight to 1./std of encodings
515
+ print("### USING STD-RESCALING ###")
516
+ x = super().get_input(batch, self.first_stage_key)
517
+ x = x.to(self.device)
518
+ encoder_posterior = self.encode_first_stage(x)
519
+ z = self.get_first_stage_encoding(encoder_posterior).detach()
520
+ del self.scale_factor
521
+ self.register_buffer('scale_factor', 1. / z.flatten().std())
522
+ print(f"setting self.scale_factor to {self.scale_factor}")
523
+ print("### USING STD-RESCALING ###")
524
+
525
+ def register_schedule(self,
526
+ given_betas=None, beta_schedule="linear", timesteps=1000,
527
+ linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
528
+ super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
529
+
530
+ self.shorten_cond_schedule = self.num_timesteps_cond > 1
531
+ if self.shorten_cond_schedule:
532
+ self.make_cond_schedule()
533
+
534
+ def instantiate_first_stage(self, config):
535
+ model = instantiate_from_config(config)
536
+ self.first_stage_model = model.eval()
537
+ self.first_stage_model.train = disabled_train
538
+ for param in self.first_stage_model.parameters():
539
+ param.requires_grad = False
540
+
541
+ def instantiate_cond_stage(self, config):
542
+ if not self.cond_stage_trainable:
543
+ if config == "__is_first_stage__":
544
+ print("Using first stage also as cond stage.")
545
+ self.cond_stage_model = self.first_stage_model
546
+ elif config == "__is_unconditional__":
547
+ print(f"Training {self.__class__.__name__} as an unconditional model.")
548
+ self.cond_stage_model = None
549
+ # self.be_unconditional = True
550
+ else:
551
+ model = instantiate_from_config(config)
552
+ self.cond_stage_model = model.eval()
553
+ self.cond_stage_model.train = disabled_train
554
+ for param in self.cond_stage_model.parameters():
555
+ param.requires_grad = False
556
+ else:
557
+ assert config != '__is_first_stage__'
558
+ assert config != '__is_unconditional__'
559
+ model = instantiate_from_config(config)
560
+ self.cond_stage_model = model
561
+
562
+ def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False):
563
+ denoise_row = []
564
+ for zd in tqdm(samples, desc=desc):
565
+ denoise_row.append(self.decode_first_stage(zd.to(self.device),
566
+ force_not_quantize=force_no_decoder_quantization))
567
+ n_imgs_per_row = len(denoise_row)
568
+ denoise_row = torch.stack(denoise_row) # n_log_step, n_row, C, H, W
569
+ denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
570
+ denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
571
+ denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
572
+ return denoise_grid
573
+
574
+ def get_first_stage_encoding(self, encoder_posterior):
575
+ if isinstance(encoder_posterior, DiagonalGaussianDistribution):
576
+ z = encoder_posterior.sample()
577
+ elif isinstance(encoder_posterior, torch.Tensor):
578
+ z = encoder_posterior
579
+ else:
580
+ raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
581
+ return self.scale_factor * z
582
+
583
+ def get_learned_conditioning(self, c):
584
+ if self.cond_stage_forward is None:
585
+ if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
586
+ c = self.cond_stage_model.encode(c)
587
+ if isinstance(c, DiagonalGaussianDistribution):
588
+ c = c.mode()
589
+ else:
590
+ c = self.cond_stage_model(c)
591
+ else:
592
+ assert hasattr(self.cond_stage_model, self.cond_stage_forward)
593
+ c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
594
+ return c
595
+
596
+ def meshgrid(self, h, w):
597
+ y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
598
+ x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)
599
+
600
+ arr = torch.cat([y, x], dim=-1)
601
+ return arr
602
+
603
+ def delta_border(self, h, w):
604
+ """
605
+ :param h: height
606
+ :param w: width
607
+ :return: normalized distance to image border,
608
+ wtith min distance = 0 at border and max dist = 0.5 at image center
609
+ """
610
+ lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
611
+ arr = self.meshgrid(h, w) / lower_right_corner
612
+ dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
613
+ dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
614
+ edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0]
615
+ return edge_dist
616
+
617
+ def get_weighting(self, h, w, Ly, Lx, device):
618
+ weighting = self.delta_border(h, w)
619
+ weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"],
620
+ self.split_input_params["clip_max_weight"], )
621
+ weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)
622
+
623
+ if self.split_input_params["tie_braker"]:
624
+ L_weighting = self.delta_border(Ly, Lx)
625
+ L_weighting = torch.clip(L_weighting,
626
+ self.split_input_params["clip_min_tie_weight"],
627
+ self.split_input_params["clip_max_tie_weight"])
628
+
629
+ L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
630
+ weighting = weighting * L_weighting
631
+ return weighting
632
+
633
+ def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1): # todo load once not every time, shorten code
634
+ """
635
+ :param x: img of size (bs, c, h, w)
636
+ :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
637
+ """
638
+ bs, nc, h, w = x.shape
639
+
640
+ # number of crops in image
641
+ Ly = (h - kernel_size[0]) // stride[0] + 1
642
+ Lx = (w - kernel_size[1]) // stride[1] + 1
643
+
644
+ if uf == 1 and df == 1:
645
+ fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
646
+ unfold = torch.nn.Unfold(**fold_params)
647
+
648
+ fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)
649
+
650
+ weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype)
651
+ normalization = fold(weighting).view(1, 1, h, w) # normalizes the overlap
652
+ weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))
653
+
654
+ elif uf > 1 and df == 1:
655
+ fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
656
+ unfold = torch.nn.Unfold(**fold_params)
657
+
658
+ fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
659
+ dilation=1, padding=0,
660
+ stride=(stride[0] * uf, stride[1] * uf))
661
+ fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2)
662
+
663
+ weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype)
664
+ normalization = fold(weighting).view(1, 1, h * uf, w * uf) # normalizes the overlap
665
+ weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx))
666
+
667
+ elif df > 1 and uf == 1:
668
+ fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
669
+ unfold = torch.nn.Unfold(**fold_params)
670
+
671
+ fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
672
+ dilation=1, padding=0,
673
+ stride=(stride[0] // df, stride[1] // df))
674
+ fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2)
675
+
676
+ weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype)
677
+ normalization = fold(weighting).view(1, 1, h // df, w // df) # normalizes the overlap
678
+ weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx))
679
+
680
+ else:
681
+ raise NotImplementedError
682
+
683
+ return fold, unfold, normalization, weighting
684
+
685
+ @torch.no_grad()
686
+ def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False,
687
+ cond_key=None, return_original_cond=False, bs=None, uncond=0.05):
688
+ x = super().get_input(batch, k)
689
+ if bs is not None:
690
+ x = x[:bs]
691
+ x = x.to(self.device)
692
+ encoder_posterior = self.encode_first_stage(x)
693
+ z = self.get_first_stage_encoding(encoder_posterior).detach()
694
+ cond_key = cond_key or self.cond_stage_key
695
+ xc = super().get_input(batch, cond_key)
696
+ if bs is not None:
697
+ xc["c_crossattn"] = xc["c_crossattn"][:bs]
698
+ xc["c_concat"] = xc["c_concat"][:bs]
699
+ cond = {}
700
+
701
+ # To support classifier-free guidance, randomly drop out only text conditioning 5%, only image conditioning 5%, and both 5%.
702
+ random = torch.rand(x.size(0), device=x.device)
703
+ prompt_mask = rearrange(random < 2 * uncond, "n -> n 1 1")
704
+ input_mask = 1 - rearrange((random >= uncond).float() * (random < 3 * uncond).float(), "n -> n 1 1 1")
705
+
706
+ null_prompt = self.get_learned_conditioning([""])
707
+ cond["c_crossattn"] = [torch.where(prompt_mask, null_prompt, self.get_learned_conditioning(xc["c_crossattn"]).detach())]
708
+ cond["c_concat"] = [input_mask * self.encode_first_stage((xc["c_concat"].to(self.device))).mode().detach()]
709
+
710
+ out = [z, cond]
711
+ if return_first_stage_outputs:
712
+ xrec = self.decode_first_stage(z)
713
+ out.extend([x, xrec])
714
+ if return_original_cond:
715
+ out.append(xc)
716
+ return out
717
+
718
+ @torch.no_grad()
719
+ def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
720
+ if predict_cids:
721
+ if z.dim() == 4:
722
+ z = torch.argmax(z.exp(), dim=1).long()
723
+ z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
724
+ z = rearrange(z, 'b h w c -> b c h w').contiguous()
725
+
726
+ z = 1. / self.scale_factor * z
727
+
728
+ if hasattr(self, "split_input_params"):
729
+ if self.split_input_params["patch_distributed_vq"]:
730
+ ks = self.split_input_params["ks"] # eg. (128, 128)
731
+ stride = self.split_input_params["stride"] # eg. (64, 64)
732
+ uf = self.split_input_params["vqf"]
733
+ bs, nc, h, w = z.shape
734
+ if ks[0] > h or ks[1] > w:
735
+ ks = (min(ks[0], h), min(ks[1], w))
736
+ print("reducing Kernel")
737
+
738
+ if stride[0] > h or stride[1] > w:
739
+ stride = (min(stride[0], h), min(stride[1], w))
740
+ print("reducing stride")
741
+
742
+ fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
743
+
744
+ z = unfold(z) # (bn, nc * prod(**ks), L)
745
+ # 1. Reshape to img shape
746
+ z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
747
+
748
+ # 2. apply model loop over last dim
749
+ if isinstance(self.first_stage_model, VQModelInterface):
750
+ output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
751
+ force_not_quantize=predict_cids or force_not_quantize)
752
+ for i in range(z.shape[-1])]
753
+ else:
754
+
755
+ output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
756
+ for i in range(z.shape[-1])]
757
+
758
+ o = torch.stack(output_list, axis=-1) # # (bn, nc, ks[0], ks[1], L)
759
+ o = o * weighting
760
+ # Reverse 1. reshape to img shape
761
+ o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
762
+ # stitch crops together
763
+ decoded = fold(o)
764
+ decoded = decoded / normalization # norm is shape (1, 1, h, w)
765
+ return decoded
766
+ else:
767
+ if isinstance(self.first_stage_model, VQModelInterface):
768
+ return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
769
+ else:
770
+ return self.first_stage_model.decode(z)
771
+
772
+ else:
773
+ if isinstance(self.first_stage_model, VQModelInterface):
774
+ return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
775
+ else:
776
+ return self.first_stage_model.decode(z)
777
+
778
+ # same as above but without decorator
779
+ def differentiable_decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
780
+ if predict_cids:
781
+ if z.dim() == 4:
782
+ z = torch.argmax(z.exp(), dim=1).long()
783
+ z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
784
+ z = rearrange(z, 'b h w c -> b c h w').contiguous()
785
+
786
+ z = 1. / self.scale_factor * z
787
+
788
+ if hasattr(self, "split_input_params"):
789
+ if self.split_input_params["patch_distributed_vq"]:
790
+ ks = self.split_input_params["ks"] # eg. (128, 128)
791
+ stride = self.split_input_params["stride"] # eg. (64, 64)
792
+ uf = self.split_input_params["vqf"]
793
+ bs, nc, h, w = z.shape
794
+ if ks[0] > h or ks[1] > w:
795
+ ks = (min(ks[0], h), min(ks[1], w))
796
+ print("reducing Kernel")
797
+
798
+ if stride[0] > h or stride[1] > w:
799
+ stride = (min(stride[0], h), min(stride[1], w))
800
+ print("reducing stride")
801
+
802
+ fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
803
+
804
+ z = unfold(z) # (bn, nc * prod(**ks), L)
805
+ # 1. Reshape to img shape
806
+ z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
807
+
808
+ # 2. apply model loop over last dim
809
+ if isinstance(self.first_stage_model, VQModelInterface):
810
+ output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
811
+ force_not_quantize=predict_cids or force_not_quantize)
812
+ for i in range(z.shape[-1])]
813
+ else:
814
+
815
+ output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
816
+ for i in range(z.shape[-1])]
817
+
818
+ o = torch.stack(output_list, axis=-1) # # (bn, nc, ks[0], ks[1], L)
819
+ o = o * weighting
820
+ # Reverse 1. reshape to img shape
821
+ o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
822
+ # stitch crops together
823
+ decoded = fold(o)
824
+ decoded = decoded / normalization # norm is shape (1, 1, h, w)
825
+ return decoded
826
+ else:
827
+ if isinstance(self.first_stage_model, VQModelInterface):
828
+ return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
829
+ else:
830
+ return self.first_stage_model.decode(z)
831
+
832
+ else:
833
+ if isinstance(self.first_stage_model, VQModelInterface):
834
+ return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
835
+ else:
836
+ return self.first_stage_model.decode(z)
837
+
838
+ @torch.no_grad()
839
+ def encode_first_stage(self, x):
840
+ if hasattr(self, "split_input_params"):
841
+ if self.split_input_params["patch_distributed_vq"]:
842
+ ks = self.split_input_params["ks"] # eg. (128, 128)
843
+ stride = self.split_input_params["stride"] # eg. (64, 64)
844
+ df = self.split_input_params["vqf"]
845
+ self.split_input_params['original_image_size'] = x.shape[-2:]
846
+ bs, nc, h, w = x.shape
847
+ if ks[0] > h or ks[1] > w:
848
+ ks = (min(ks[0], h), min(ks[1], w))
849
+ print("reducing Kernel")
850
+
851
+ if stride[0] > h or stride[1] > w:
852
+ stride = (min(stride[0], h), min(stride[1], w))
853
+ print("reducing stride")
854
+
855
+ fold, unfold, normalization, weighting = self.get_fold_unfold(x, ks, stride, df=df)
856
+ z = unfold(x) # (bn, nc * prod(**ks), L)
857
+ # Reshape to img shape
858
+ z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
859
+
860
+ output_list = [self.first_stage_model.encode(z[:, :, :, :, i])
861
+ for i in range(z.shape[-1])]
862
+
863
+ o = torch.stack(output_list, axis=-1)
864
+ o = o * weighting
865
+
866
+ # Reverse reshape to img shape
867
+ o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
868
+ # stitch crops together
869
+ decoded = fold(o)
870
+ decoded = decoded / normalization
871
+ return decoded
872
+
873
+ else:
874
+ return self.first_stage_model.encode(x)
875
+ else:
876
+ return self.first_stage_model.encode(x)
877
+
878
+ def shared_step(self, batch, **kwargs):
879
+ x, c = self.get_input(batch, self.first_stage_key)
880
+ loss = self(x, c)
881
+ return loss
882
+
883
+ def forward(self, x, c, *args, **kwargs):
884
+ t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
885
+ if self.model.conditioning_key is not None:
886
+ assert c is not None
887
+ if self.cond_stage_trainable:
888
+ c = self.get_learned_conditioning(c)
889
+ if self.shorten_cond_schedule: # TODO: drop this option
890
+ tc = self.cond_ids[t].to(self.device)
891
+ c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float()))
892
+ return self.p_losses(x, c, t, *args, **kwargs)
893
+
894
+ def _rescale_annotations(self, bboxes, crop_coordinates): # TODO: move to dataset
895
+ def rescale_bbox(bbox):
896
+ x0 = clamp((bbox[0] - crop_coordinates[0]) / crop_coordinates[2])
897
+ y0 = clamp((bbox[1] - crop_coordinates[1]) / crop_coordinates[3])
898
+ w = min(bbox[2] / crop_coordinates[2], 1 - x0)
899
+ h = min(bbox[3] / crop_coordinates[3], 1 - y0)
900
+ return x0, y0, w, h
901
+
902
+ return [rescale_bbox(b) for b in bboxes]
903
+
904
+ def apply_model(self, x_noisy, t, cond, return_ids=False):
905
+
906
+ if isinstance(cond, dict):
907
+ # hybrid case, cond is exptected to be a dict
908
+ pass
909
+ else:
910
+ if not isinstance(cond, list):
911
+ cond = [cond]
912
+ key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
913
+ cond = {key: cond}
914
+
915
+ if hasattr(self, "split_input_params"):
916
+ assert len(cond) == 1 # todo can only deal with one conditioning atm
917
+ assert not return_ids
918
+ ks = self.split_input_params["ks"] # eg. (128, 128)
919
+ stride = self.split_input_params["stride"] # eg. (64, 64)
920
+
921
+ h, w = x_noisy.shape[-2:]
922
+
923
+ fold, unfold, normalization, weighting = self.get_fold_unfold(x_noisy, ks, stride)
924
+
925
+ z = unfold(x_noisy) # (bn, nc * prod(**ks), L)
926
+ # Reshape to img shape
927
+ z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
928
+ z_list = [z[:, :, :, :, i] for i in range(z.shape[-1])]
929
+
930
+ if self.cond_stage_key in ["image", "LR_image", "segmentation",
931
+ 'bbox_img'] and self.model.conditioning_key: # todo check for completeness
932
+ c_key = next(iter(cond.keys())) # get key
933
+ c = next(iter(cond.values())) # get value
934
+ assert (len(c) == 1) # todo extend to list with more than one elem
935
+ c = c[0] # get element
936
+
937
+ c = unfold(c)
938
+ c = c.view((c.shape[0], -1, ks[0], ks[1], c.shape[-1])) # (bn, nc, ks[0], ks[1], L )
939
+
940
+ cond_list = [{c_key: [c[:, :, :, :, i]]} for i in range(c.shape[-1])]
941
+
942
+ elif self.cond_stage_key == 'coordinates_bbox':
943
+ assert 'original_image_size' in self.split_input_params, 'BoudingBoxRescaling is missing original_image_size'
944
+
945
+ # assuming padding of unfold is always 0 and its dilation is always 1
946
+ n_patches_per_row = int((w - ks[0]) / stride[0] + 1)
947
+ full_img_h, full_img_w = self.split_input_params['original_image_size']
948
+ # as we are operating on latents, we need the factor from the original image size to the
949
+ # spatial latent size to properly rescale the crops for regenerating the bbox annotations
950
+ num_downs = self.first_stage_model.encoder.num_resolutions - 1
951
+ rescale_latent = 2 ** (num_downs)
952
+
953
+ # get top left postions of patches as conforming for the bbbox tokenizer, therefore we
954
+ # need to rescale the tl patch coordinates to be in between (0,1)
955
+ tl_patch_coordinates = [(rescale_latent * stride[0] * (patch_nr % n_patches_per_row) / full_img_w,
956
+ rescale_latent * stride[1] * (patch_nr // n_patches_per_row) / full_img_h)
957
+ for patch_nr in range(z.shape[-1])]
958
+
959
+ # patch_limits are tl_coord, width and height coordinates as (x_tl, y_tl, h, w)
960
+ patch_limits = [(x_tl, y_tl,
961
+ rescale_latent * ks[0] / full_img_w,
962
+ rescale_latent * ks[1] / full_img_h) for x_tl, y_tl in tl_patch_coordinates]
963
+ # patch_values = [(np.arange(x_tl,min(x_tl+ks, 1.)),np.arange(y_tl,min(y_tl+ks, 1.))) for x_tl, y_tl in tl_patch_coordinates]
964
+
965
+ # tokenize crop coordinates for the bounding boxes of the respective patches
966
+ patch_limits_tknzd = [torch.LongTensor(self.bbox_tokenizer._crop_encoder(bbox))[None].to(self.device)
967
+ for bbox in patch_limits] # list of length l with tensors of shape (1, 2)
968
+ print(patch_limits_tknzd[0].shape)
969
+ # cut tknzd crop position from conditioning
970
+ assert isinstance(cond, dict), 'cond must be dict to be fed into model'
971
+ cut_cond = cond['c_crossattn'][0][..., :-2].to(self.device)
972
+ print(cut_cond.shape)
973
+
974
+ adapted_cond = torch.stack([torch.cat([cut_cond, p], dim=1) for p in patch_limits_tknzd])
975
+ adapted_cond = rearrange(adapted_cond, 'l b n -> (l b) n')
976
+ print(adapted_cond.shape)
977
+ adapted_cond = self.get_learned_conditioning(adapted_cond)
978
+ print(adapted_cond.shape)
979
+ adapted_cond = rearrange(adapted_cond, '(l b) n d -> l b n d', l=z.shape[-1])
980
+ print(adapted_cond.shape)
981
+
982
+ cond_list = [{'c_crossattn': [e]} for e in adapted_cond]
983
+
984
+ else:
985
+ cond_list = [cond for i in range(z.shape[-1])] # Todo make this more efficient
986
+
987
+ # apply model by loop over crops
988
+ output_list = [self.model(z_list[i], t, **cond_list[i]) for i in range(z.shape[-1])]
989
+ assert not isinstance(output_list[0],
990
+ tuple) # todo cant deal with multiple model outputs check this never happens
991
+
992
+ o = torch.stack(output_list, axis=-1)
993
+ o = o * weighting
994
+ # Reverse reshape to img shape
995
+ o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
996
+ # stitch crops together
997
+ x_recon = fold(o) / normalization
998
+
999
+ else:
1000
+ x_recon = self.model(x_noisy, t, **cond)
1001
+
1002
+ if isinstance(x_recon, tuple) and not return_ids:
1003
+ return x_recon[0]
1004
+ else:
1005
+ return x_recon
1006
+
1007
+ def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
1008
+ return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \
1009
+ extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
1010
+
1011
+ def _prior_bpd(self, x_start):
1012
+ """
1013
+ Get the prior KL term for the variational lower-bound, measured in
1014
+ bits-per-dim.
1015
+ This term can't be optimized, as it only depends on the encoder.
1016
+ :param x_start: the [N x C x ...] tensor of inputs.
1017
+ :return: a batch of [N] KL values (in bits), one per batch element.
1018
+ """
1019
+ batch_size = x_start.shape[0]
1020
+ t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
1021
+ qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
1022
+ kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0)
1023
+ return mean_flat(kl_prior) / np.log(2.0)
1024
+
1025
+ def p_losses(self, x_start, cond, t, noise=None):
1026
+ noise = default(noise, lambda: torch.randn_like(x_start))
1027
+ x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
1028
+ model_output = self.apply_model(x_noisy, t, cond)
1029
+
1030
+ loss_dict = {}
1031
+ prefix = 'train' if self.training else 'val'
1032
+
1033
+ if self.parameterization == "x0":
1034
+ target = x_start
1035
+ elif self.parameterization == "eps":
1036
+ target = noise
1037
+ else:
1038
+ raise NotImplementedError()
1039
+
1040
+ loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3])
1041
+ loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
1042
+
1043
+ logvar_t = self.logvar[t].to(self.device)
1044
+ loss = loss_simple / torch.exp(logvar_t) + logvar_t
1045
+ # loss = loss_simple / torch.exp(self.logvar) + self.logvar
1046
+ if self.learn_logvar:
1047
+ loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
1048
+ loss_dict.update({'logvar': self.logvar.data.mean()})
1049
+
1050
+ loss = self.l_simple_weight * loss.mean()
1051
+
1052
+ loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3))
1053
+ loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
1054
+ loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
1055
+ loss += (self.original_elbo_weight * loss_vlb)
1056
+ loss_dict.update({f'{prefix}/loss': loss})
1057
+
1058
+ return loss, loss_dict
1059
+
1060
+ def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False,
1061
+ return_x0=False, score_corrector=None, corrector_kwargs=None):
1062
+ t_in = t
1063
+ model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
1064
+
1065
+ if score_corrector is not None:
1066
+ assert self.parameterization == "eps"
1067
+ model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
1068
+
1069
+ if return_codebook_ids:
1070
+ model_out, logits = model_out
1071
+
1072
+ if self.parameterization == "eps":
1073
+ x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
1074
+ elif self.parameterization == "x0":
1075
+ x_recon = model_out
1076
+ else:
1077
+ raise NotImplementedError()
1078
+
1079
+ if clip_denoised:
1080
+ x_recon.clamp_(-1., 1.)
1081
+ if quantize_denoised:
1082
+ x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
1083
+ model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
1084
+ if return_codebook_ids:
1085
+ return model_mean, posterior_variance, posterior_log_variance, logits
1086
+ elif return_x0:
1087
+ return model_mean, posterior_variance, posterior_log_variance, x_recon
1088
+ else:
1089
+ return model_mean, posterior_variance, posterior_log_variance
1090
+
1091
+ @torch.no_grad()
1092
+ def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False,
1093
+ return_codebook_ids=False, quantize_denoised=False, return_x0=False,
1094
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None):
1095
+ b, *_, device = *x.shape, x.device
1096
+ outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised,
1097
+ return_codebook_ids=return_codebook_ids,
1098
+ quantize_denoised=quantize_denoised,
1099
+ return_x0=return_x0,
1100
+ score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
1101
+ if return_codebook_ids:
1102
+ raise DeprecationWarning("Support dropped.")
1103
+ model_mean, _, model_log_variance, logits = outputs
1104
+ elif return_x0:
1105
+ model_mean, _, model_log_variance, x0 = outputs
1106
+ else:
1107
+ model_mean, _, model_log_variance = outputs
1108
+
1109
+ noise = noise_like(x.shape, device, repeat_noise) * temperature
1110
+ if noise_dropout > 0.:
1111
+ noise = torch.nn.functional.dropout(noise, p=noise_dropout)
1112
+ # no noise when t == 0
1113
+ nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
1114
+
1115
+ if return_codebook_ids:
1116
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1)
1117
+ if return_x0:
1118
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
1119
+ else:
1120
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
1121
+
1122
+ @torch.no_grad()
1123
+ def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False,
1124
+ img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0.,
1125
+ score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None,
1126
+ log_every_t=None):
1127
+ if not log_every_t:
1128
+ log_every_t = self.log_every_t
1129
+ timesteps = self.num_timesteps
1130
+ if batch_size is not None:
1131
+ b = batch_size if batch_size is not None else shape[0]
1132
+ shape = [batch_size] + list(shape)
1133
+ else:
1134
+ b = batch_size = shape[0]
1135
+ if x_T is None:
1136
+ img = torch.randn(shape, device=self.device)
1137
+ else:
1138
+ img = x_T
1139
+ intermediates = []
1140
+ if cond is not None:
1141
+ if isinstance(cond, dict):
1142
+ cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
1143
+ list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
1144
+ else:
1145
+ cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
1146
+
1147
+ if start_T is not None:
1148
+ timesteps = min(timesteps, start_T)
1149
+ iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation',
1150
+ total=timesteps) if verbose else reversed(
1151
+ range(0, timesteps))
1152
+ if type(temperature) == float:
1153
+ temperature = [temperature] * timesteps
1154
+
1155
+ for i in iterator:
1156
+ ts = torch.full((b,), i, device=self.device, dtype=torch.long)
1157
+ if self.shorten_cond_schedule:
1158
+ assert self.model.conditioning_key != 'hybrid'
1159
+ tc = self.cond_ids[ts].to(cond.device)
1160
+ cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
1161
+
1162
+ img, x0_partial = self.p_sample(img, cond, ts,
1163
+ clip_denoised=self.clip_denoised,
1164
+ quantize_denoised=quantize_denoised, return_x0=True,
1165
+ temperature=temperature[i], noise_dropout=noise_dropout,
1166
+ score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
1167
+ if mask is not None:
1168
+ assert x0 is not None
1169
+ img_orig = self.q_sample(x0, ts)
1170
+ img = img_orig * mask + (1. - mask) * img
1171
+
1172
+ if i % log_every_t == 0 or i == timesteps - 1:
1173
+ intermediates.append(x0_partial)
1174
+ if callback: callback(i)
1175
+ if img_callback: img_callback(img, i)
1176
+ return img, intermediates
1177
+
1178
+ @torch.no_grad()
1179
+ def p_sample_loop(self, cond, shape, return_intermediates=False,
1180
+ x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False,
1181
+ mask=None, x0=None, img_callback=None, start_T=None,
1182
+ log_every_t=None):
1183
+
1184
+ if not log_every_t:
1185
+ log_every_t = self.log_every_t
1186
+ device = self.betas.device
1187
+ b = shape[0]
1188
+ if x_T is None:
1189
+ img = torch.randn(shape, device=device)
1190
+ else:
1191
+ img = x_T
1192
+
1193
+ intermediates = [img]
1194
+ if timesteps is None:
1195
+ timesteps = self.num_timesteps
1196
+
1197
+ if start_T is not None:
1198
+ timesteps = min(timesteps, start_T)
1199
+ iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(
1200
+ range(0, timesteps))
1201
+
1202
+ if mask is not None:
1203
+ assert x0 is not None
1204
+ assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match
1205
+
1206
+ for i in iterator:
1207
+ ts = torch.full((b,), i, device=device, dtype=torch.long)
1208
+ if self.shorten_cond_schedule:
1209
+ assert self.model.conditioning_key != 'hybrid'
1210
+ tc = self.cond_ids[ts].to(cond.device)
1211
+ cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
1212
+
1213
+ img = self.p_sample(img, cond, ts,
1214
+ clip_denoised=self.clip_denoised,
1215
+ quantize_denoised=quantize_denoised)
1216
+ if mask is not None:
1217
+ img_orig = self.q_sample(x0, ts)
1218
+ img = img_orig * mask + (1. - mask) * img
1219
+
1220
+ if i % log_every_t == 0 or i == timesteps - 1:
1221
+ intermediates.append(img)
1222
+ if callback: callback(i)
1223
+ if img_callback: img_callback(img, i)
1224
+
1225
+ if return_intermediates:
1226
+ return img, intermediates
1227
+ return img
1228
+
1229
+ @torch.no_grad()
1230
+ def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None,
1231
+ verbose=True, timesteps=None, quantize_denoised=False,
1232
+ mask=None, x0=None, shape=None,**kwargs):
1233
+ if shape is None:
1234
+ shape = (batch_size, self.channels, self.image_size, self.image_size)
1235
+ if cond is not None:
1236
+ if isinstance(cond, dict):
1237
+ cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
1238
+ list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
1239
+ else:
1240
+ cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
1241
+ return self.p_sample_loop(cond,
1242
+ shape,
1243
+ return_intermediates=return_intermediates, x_T=x_T,
1244
+ verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised,
1245
+ mask=mask, x0=x0)
1246
+
1247
+ @torch.no_grad()
1248
+ def sample_log(self,cond,batch_size,ddim, ddim_steps,**kwargs):
1249
+
1250
+ if ddim:
1251
+ ddim_sampler = DDIMSampler(self)
1252
+ shape = (self.channels, self.image_size, self.image_size)
1253
+ samples, intermediates =ddim_sampler.sample(ddim_steps,batch_size,
1254
+ shape,cond,verbose=False,**kwargs)
1255
+
1256
+ else:
1257
+ samples, intermediates = self.sample(cond=cond, batch_size=batch_size,
1258
+ return_intermediates=True,**kwargs)
1259
+
1260
+ return samples, intermediates
1261
+
1262
+
1263
+ @torch.no_grad()
1264
+ def log_images(self, batch, N=4, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
1265
+ quantize_denoised=True, inpaint=False, plot_denoise_rows=False, plot_progressive_rows=False,
1266
+ plot_diffusion_rows=False, **kwargs):
1267
+
1268
+ use_ddim = False
1269
+
1270
+ log = dict()
1271
+ z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key,
1272
+ return_first_stage_outputs=True,
1273
+ force_c_encode=True,
1274
+ return_original_cond=True,
1275
+ bs=N, uncond=0)
1276
+ N = min(x.shape[0], N)
1277
+ n_row = min(x.shape[0], n_row)
1278
+ log["inputs"] = x
1279
+ log["reals"] = xc["c_concat"]
1280
+ log["reconstruction"] = xrec
1281
+ if self.model.conditioning_key is not None:
1282
+ if hasattr(self.cond_stage_model, "decode"):
1283
+ xc = self.cond_stage_model.decode(c)
1284
+ log["conditioning"] = xc
1285
+ elif self.cond_stage_key in ["caption"]:
1286
+ xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["caption"])
1287
+ log["conditioning"] = xc
1288
+ elif self.cond_stage_key == 'class_label':
1289
+ xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
1290
+ log['conditioning'] = xc
1291
+ elif isimage(xc):
1292
+ log["conditioning"] = xc
1293
+ if ismap(xc):
1294
+ log["original_conditioning"] = self.to_rgb(xc)
1295
+
1296
+ if plot_diffusion_rows:
1297
+ # get diffusion row
1298
+ diffusion_row = list()
1299
+ z_start = z[:n_row]
1300
+ for t in range(self.num_timesteps):
1301
+ if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
1302
+ t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
1303
+ t = t.to(self.device).long()
1304
+ noise = torch.randn_like(z_start)
1305
+ z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
1306
+ diffusion_row.append(self.decode_first_stage(z_noisy))
1307
+
1308
+ diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W
1309
+ diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
1310
+ diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
1311
+ diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
1312
+ log["diffusion_row"] = diffusion_grid
1313
+
1314
+ if sample:
1315
+ # get denoise row
1316
+ with self.ema_scope("Plotting"):
1317
+ samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
1318
+ ddim_steps=ddim_steps,eta=ddim_eta)
1319
+ # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
1320
+ x_samples = self.decode_first_stage(samples)
1321
+ log["samples"] = x_samples
1322
+ if plot_denoise_rows:
1323
+ denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
1324
+ log["denoise_row"] = denoise_grid
1325
+
1326
+ if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance(
1327
+ self.first_stage_model, IdentityFirstStage):
1328
+ # also display when quantizing x0 while sampling
1329
+ with self.ema_scope("Plotting Quantized Denoised"):
1330
+ samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
1331
+ ddim_steps=ddim_steps,eta=ddim_eta,
1332
+ quantize_denoised=True)
1333
+ # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True,
1334
+ # quantize_denoised=True)
1335
+ x_samples = self.decode_first_stage(samples.to(self.device))
1336
+ log["samples_x0_quantized"] = x_samples
1337
+
1338
+ if inpaint:
1339
+ # make a simple center square
1340
+ b, h, w = z.shape[0], z.shape[2], z.shape[3]
1341
+ mask = torch.ones(N, h, w).to(self.device)
1342
+ # zeros will be filled in
1343
+ mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0.
1344
+ mask = mask[:, None, ...]
1345
+ with self.ema_scope("Plotting Inpaint"):
1346
+
1347
+ samples, _ = self.sample_log(cond=c,batch_size=N,ddim=use_ddim, eta=ddim_eta,
1348
+ ddim_steps=ddim_steps, x0=z[:N], mask=mask)
1349
+ x_samples = self.decode_first_stage(samples.to(self.device))
1350
+ log["samples_inpainting"] = x_samples
1351
+ log["mask"] = mask
1352
+
1353
+ # outpaint
1354
+ with self.ema_scope("Plotting Outpaint"):
1355
+ samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,eta=ddim_eta,
1356
+ ddim_steps=ddim_steps, x0=z[:N], mask=mask)
1357
+ x_samples = self.decode_first_stage(samples.to(self.device))
1358
+ log["samples_outpainting"] = x_samples
1359
+
1360
+ if plot_progressive_rows:
1361
+ with self.ema_scope("Plotting Progressives"):
1362
+ img, progressives = self.progressive_denoising(c,
1363
+ shape=(self.channels, self.image_size, self.image_size),
1364
+ batch_size=N)
1365
+ prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
1366
+ log["progressive_row"] = prog_row
1367
+
1368
+ if return_keys:
1369
+ if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
1370
+ return log
1371
+ else:
1372
+ return {key: log[key] for key in return_keys}
1373
+ return log
1374
+
1375
+ def configure_optimizers(self):
1376
+ lr = self.learning_rate
1377
+ params = list(self.model.parameters())
1378
+ if self.cond_stage_trainable:
1379
+ print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
1380
+ params = params + list(self.cond_stage_model.parameters())
1381
+ if self.learn_logvar:
1382
+ print('Diffusion model optimizing logvar')
1383
+ params.append(self.logvar)
1384
+ opt = torch.optim.AdamW(params, lr=lr)
1385
+ if self.use_scheduler:
1386
+ assert 'target' in self.scheduler_config
1387
+ scheduler = instantiate_from_config(self.scheduler_config)
1388
+
1389
+ print("Setting up LambdaLR scheduler...")
1390
+ scheduler = [
1391
+ {
1392
+ 'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule),
1393
+ 'interval': 'step',
1394
+ 'frequency': 1
1395
+ }]
1396
+ return [opt], scheduler
1397
+ return opt
1398
+
1399
+ @torch.no_grad()
1400
+ def to_rgb(self, x):
1401
+ x = x.float()
1402
+ if not hasattr(self, "colorize"):
1403
+ self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x)
1404
+ x = nn.functional.conv2d(x, weight=self.colorize)
1405
+ x = 2. * (x - x.min()) / (x.max() - x.min()) - 1.
1406
+ return x
1407
+
1408
+
1409
+ class DiffusionWrapper(pl.LightningModule):
1410
+ def __init__(self, diff_model_config, conditioning_key):
1411
+ super().__init__()
1412
+ self.diffusion_model = instantiate_from_config(diff_model_config)
1413
+ self.conditioning_key = conditioning_key
1414
+ assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm']
1415
+
1416
+ def forward(self, x, t, c_concat: list = None, c_crossattn: list = None):
1417
+ if self.conditioning_key is None:
1418
+ out = self.diffusion_model(x, t)
1419
+ elif self.conditioning_key == 'concat':
1420
+ xc = torch.cat([x] + c_concat, dim=1)
1421
+ out = self.diffusion_model(xc, t)
1422
+ elif self.conditioning_key == 'crossattn':
1423
+ cc = torch.cat(c_crossattn, 1)
1424
+ out = self.diffusion_model(x, t, context=cc)
1425
+ elif self.conditioning_key == 'hybrid':
1426
+ xc = torch.cat([x] + c_concat, dim=1)
1427
+ cc = torch.cat(c_crossattn, 1)
1428
+ out = self.diffusion_model(xc, t, context=cc)
1429
+ elif self.conditioning_key == 'adm':
1430
+ cc = c_crossattn[0]
1431
+ out = self.diffusion_model(x, t, y=cc)
1432
+ else:
1433
+ raise NotImplementedError()
1434
+
1435
+ return out
1436
+
1437
+
1438
+ class Layout2ImgDiffusion(LatentDiffusion):
1439
+ # TODO: move all layout-specific hacks to this class
1440
+ def __init__(self, cond_stage_key, *args, **kwargs):
1441
+ assert cond_stage_key == 'coordinates_bbox', 'Layout2ImgDiffusion only for cond_stage_key="coordinates_bbox"'
1442
+ super().__init__(cond_stage_key=cond_stage_key, *args, **kwargs)
1443
+
1444
+ def log_images(self, batch, N=8, *args, **kwargs):
1445
+ logs = super().log_images(batch=batch, N=N, *args, **kwargs)
1446
+
1447
+ key = 'train' if self.training else 'validation'
1448
+ dset = self.trainer.datamodule.datasets[key]
1449
+ mapper = dset.conditional_builders[self.cond_stage_key]
1450
+
1451
+ bbox_imgs = []
1452
+ map_fn = lambda catno: dset.get_textual_label(dset.get_category_id(catno))
1453
+ for tknzd_bbox in batch[self.cond_stage_key][:N]:
1454
+ bboximg = mapper.plot(tknzd_bbox.detach().cpu(), map_fn, (256, 256))
1455
+ bbox_imgs.append(bboximg)
1456
+
1457
+ cond_img = torch.stack(bbox_imgs, dim=0)
1458
+ logs['bbox_image'] = cond_img
1459
+ return logs
stable_diffusion/ldm/modules/attention.py CHANGED
@@ -1,3 +1,6 @@
 
 
 
1
  from inspect import isfunction
2
  import math
3
  import torch
@@ -89,7 +92,7 @@ class LinearAttention(nn.Module):
89
  b, c, h, w = x.shape
90
  qkv = self.to_qkv(x)
91
  q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
92
- k = k.softmax(dim=-1)
93
  context = torch.einsum('bhdn,bhen->bhde', k, v)
94
  out = torch.einsum('bhde,bhdn->bhen', context, q)
95
  out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
@@ -167,7 +170,11 @@ class CrossAttention(nn.Module):
167
  nn.Dropout(dropout)
168
  )
169
 
 
 
170
  def forward(self, x, context=None, mask=None):
 
 
171
  h = self.heads
172
 
173
  q = self.to_q(x)
@@ -179,6 +186,13 @@ class CrossAttention(nn.Module):
179
 
180
  sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
181
 
 
 
 
 
 
 
 
182
  if exists(mask):
183
  mask = rearrange(mask, 'b ... -> b (...)')
184
  max_neg_value = -torch.finfo(sim.dtype).max
@@ -258,4 +272,4 @@ class SpatialTransformer(nn.Module):
258
  x = block(x, context=context)
259
  x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w)
260
  x = self.proj_out(x)
261
- return x + x_in
1
+ # File modified by authors of InstructPix2Pix from original (https://github.com/CompVis/stable-diffusion).
2
+ # See more details in LICENSE.
3
+
4
  from inspect import isfunction
5
  import math
6
  import torch
92
  b, c, h, w = x.shape
93
  qkv = self.to_qkv(x)
94
  q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
95
+ k = k.softmax(dim=-1)
96
  context = torch.einsum('bhdn,bhen->bhde', k, v)
97
  out = torch.einsum('bhde,bhdn->bhen', context, q)
98
  out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
170
  nn.Dropout(dropout)
171
  )
172
 
173
+ self.prompt_to_prompt = False
174
+
175
  def forward(self, x, context=None, mask=None):
176
+ is_self_attn = context is None
177
+
178
  h = self.heads
179
 
180
  q = self.to_q(x)
186
 
187
  sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
188
 
189
+ if self.prompt_to_prompt and is_self_attn:
190
+ # Unlike the original Prompt-to-Prompt which uses cross-attention layers, we copy attention maps for self-attention layers.
191
+ # There must be 4 elements in the batch: {conditional, unconditional} x {prompt 1, prompt 2}
192
+ assert x.size(0) == 4
193
+ sims = sim.chunk(4)
194
+ sim = torch.cat((sims[0], sims[0], sims[2], sims[2]))
195
+
196
  if exists(mask):
197
  mask = rearrange(mask, 'b ... -> b (...)')
198
  max_neg_value = -torch.finfo(sim.dtype).max
272
  x = block(x, context=context)
273
  x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w)
274
  x = self.proj_out(x)
275
+ return x + x_in
stable_diffusion/main.py CHANGED
@@ -738,4 +738,4 @@ if __name__ == "__main__":
738
  os.makedirs(os.path.split(dst)[0], exist_ok=True)
739
  os.rename(logdir, dst)
740
  if trainer.global_rank == 0:
741
- print(trainer.profiler.summary())
738
  os.makedirs(os.path.split(dst)[0], exist_ok=True)
739
  os.rename(logdir, dst)
740
  if trainer.global_rank == 0:
741
+ print(trainer.profiler.summary())