holynski commited on
Commit
926ff6c
1 Parent(s): 2afcb7e

Updates for v1 release

Browse files
README.md CHANGED
@@ -13,16 +13,28 @@ PyTorch implementation of InstructPix2Pix, an instruction-based image editing mo
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
 
@@ -56,10 +68,18 @@ bash scripts/download_data.sh clip-filtered-dataset
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
 
@@ -73,64 +93,72 @@ We provide our generated dataset of captions and edit instructions [here](https:
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.
@@ -139,6 +167,28 @@ After generating all of the dataset examples, run the following command below to
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).
13
 
14
  ## TL;DR: quickstart
15
 
16
+ Set up a conda environment, and download a pretrained model:
17
  ```
18
  conda env create -f environment.yaml
19
  conda activate ip2p
20
  bash scripts/download_checkpoints.sh
21
+ ```
22
+
23
+ Edit a single image:
24
+ ```
25
  python edit_cli.py --input imgs/example.jpg --output imgs/output.jpg --edit "turn him into a cyborg"
26
 
27
+ # Optionally, you can specify parameters to tune your result:
28
+ # python edit_cli.py --steps 100 --resolution 512 --seed 1371 --cfg-text 7.5 --cfg-image 1.2 --input imgs/example.jpg --output imgs/output.jpg --edit "turn him into a cyborg"
29
+ ```
30
+
31
+ Or launch your own interactive editing Gradio app:
32
+ ```
33
+ python edit_app.py
34
  ```
35
+ ![Edit app](https://github.com/timothybrooks/instruct-pix2pix/blob/main/imgs/edit_app.jpg?raw=true)
36
+
37
+ _(For advice on how to get the best results by tuning parameters, see the [Tips](https://github.com/timothybrooks/instruct-pix2pix#tips) section)._
38
 
39
  ## Setup
40
 
68
 
69
  ## Training InstructPix2Pix
70
 
71
+ InstructPix2Pix is trained by fine-tuning from an initial StableDiffusion checkpoint. The first step is to download a Stable Diffusion checkpoint. For our trained models, we used the v1.5 checkpoint as the starting point. To download the same ones we used, you can run the following script:
72
+ ```
73
+ bash scripts/download_pretrained_sd.sh
74
+ ```
75
+ If you'd like to use a different checkpoint, point to it in the config file `configs/train.yaml`, on line 8, after `ckpt_path:`.
76
+
77
+ Next, we need to change the config to point to our downloaded (or generated) dataset. If you're using the `clip-filtered-dataset` from above, you can skip this. Otherwise, you may need to edit lines 85 and 94 of the config (`data.params.train.params.path`, `data.params.validation.params.path`).
78
+
79
+ Finally, start a training job with the following command:
80
 
81
  ```
82
+ python main.py --name default --base configs/train.yaml --train --gpus 0,1,2,3,4,5,6,7
83
  ```
84
 
85
 
93
 
94
  #### (1.1) Manually write a dataset of instructions and captions
95
 
96
+ 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](https://instruct-pix2pix.eecs.berkeley.edu/human-written-prompts.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).
97
 
98
+ Input prompts should closely match the distribution of input prompts used to generate the larger dataset. We sampled the 700 input prompts from the _LAION Improved 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 Improved 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 recommend you sample the input prompts from the same distribution when manually writing training examples.
99
 
100
  #### (1.2) Finetune GPT-3
101
 
102
+ The next step is to finetune a large language model on the manually written instructions/outputs to generate edit instructions and edited caption from a new input caption. For this, we finetune GPT-3's Davinci model via the OpenAI API, although other language models could be used.
103
 
104
+ To prepare training data for GPT-3, one must first create an OpenAI developer account to access the needed APIs, and [set up the API keys on your local device](https://beta.openai.com/docs/api-reference/introduction). Also, 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.
105
 
106
  ```bash
107
+ python dataset_creation/prepare_for_gpt.py --input-path data/human-written-prompts.jsonl --output-path data/human-written-prompts-for-gpt.jsonl
108
  ```
109
 
110
+ Next, finetune GPT-3 via the OpenAI CLI. We provide an example below, although please refer to OpenAI's official documentation for this, as best practices may change. We trained the Davinci model for a single epoch. You can experiment with smaller less expensive GPT-3 variants or with open source language models, although this may negatively affect performance.
111
 
112
  ```bash
113
+ openai api fine_tunes.create -t data/human-written-prompts-for-gpt.jsonl -m davinci --n_epochs 1 --suffix "instruct-pix2pix"
114
  ```
115
 
116
  You can test out the finetuned GPT-3 model by launching the provided Gradio app:
117
 
118
  ```bash
119
+ python prompt_app.py --openai-api-key OPENAI_KEY --openai-model OPENAI_MODEL_NAME
120
  ```
121
 
122
+ ![Prompt app](https://github.com/timothybrooks/instruct-pix2pix/blob/main/imgs/prompt_app.jpg?raw=true)
123
+
124
  #### (1.3) Generate a large dataset of captions and instructions
125
 
126
+ 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 (by setting a low value of `--num-samples`) and gradually increasing the scale to ensure the results are working as desired before increasing scale.
127
 
128
  ```bash
129
+ python dataset_creation/generate_txt_dataset.py --openai-api-key OPENAI_KEY --openai-model OPENAI_MODEL_NAME
130
  ```
131
 
132
  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.
133
 
134
  ```bash
135
+ python dataset_creation/generate_txt_dataset.py --openai-api-key OPENAI_KEY --openai-model OPENAI_MODEL_NAME --partitions=10 --partition=0
136
  ```
137
 
138
  ### (2) Turn paired captions into paired images
139
 
140
+ The next step is to turn pairs of text captions into pairs of images. For this, we need to copy some pre-trained Stable Diffusion checkpoints to `stable_diffusion/models/ldm/stable-diffusion-v1/`. You may have already done this if you followed the instructions above for training with our provided data, but if not, you can do this by running:
141
+
142
+ ```bash
143
+ bash scripts/download_pretrained_sd.sh
144
+ ```
145
+
146
+ For our model, we used [checkpoint v1.5](https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned.ckpt), and the [new autoencoder](https://huggingface.co/stabilityai/sd-vae-ft-mse-original/resolve/main/vae-ft-mse-840000-ema-pruned.ckpt), but other models may work as well. If you choose to use other models, make sure to change point to the corresponding checkpoints by passing in the `--ckpt` and `--vae-ckpt` arguments. Once all checkpoints have been downloaded, we can generate the dataset with the following command:
147
 
148
  ```
149
+ python dataset_creation/generate_img_dataset.py --out_dir data/instruct-pix2pix-dataset-000 --prompts_file path/to/generated_prompts.jsonl
150
  ```
151
 
152
  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.
153
 
154
  ```
155
+ python dataset_creation/generate_img_dataset.py --out_dir data/instruct-pix2pix-dataset-000 --prompts_file path/to/generated_prompts.jsonl --n-partitions 100 --partition 0
156
  ```
157
 
158
  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:
159
 
160
  ```
161
+ python dataset_creation/generate_img_dataset.py --out_dir data/instruct-pix2pix-dataset-000 --prompts_file path/to/generated_prompts.jsonl --n-samples 4 --clip-threshold 0 --clip-dir-threshold 0 --clip-img-threshold 0 --n-partitions 100 --partition 0
162
  ```
163
 
164
  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.
167
  python dataset_creation/prepare_dataset.py data/instruct-pix2pix-dataset-000
168
  ```
169
 
170
+ ## Evaluation
171
+
172
+ To generate plots like the ones in Figures 8 and 10 in the paper, run the following command:
173
+
174
+ ```
175
+ python metrics/compute_metrics.py --ckpt /path/to/your/model.ckpt
176
+ ```
177
+
178
+ ## Tips
179
+
180
+ If you're not getting the quality result you want, there may be a few reasons:
181
+ 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:
182
+ * Decreasing the Image CFG weight, or
183
+ * Incerasing the Text CFG weight, or
184
+ 2. Conversely, **is the image changing too much**, such that the details in the original image aren't preserved? Try:
185
+ * Increasing the Image CFG weight, or
186
+ * Decreasing the Text CFG weight
187
+ 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.
188
+ 4. Rephrasing the instruction sometimes improves results (e.g., "turn him into a dog" vs. "make him a dog" vs. "as a dog").
189
+ 5. Increasing the number of steps sometimes improves results.
190
+ 6. Do faces look weird? The Stable Diffusion autoencoder has a hard time with faces that are small in the image. Try cropping the image so the face takes up a larger portion of the frame.
191
+
192
  ## Comments
193
 
194
  - Our codebase is based on the [Stable Diffusion codebase](https://github.com/CompVis/stable-diffusion).
configs/generate.yaml CHANGED
@@ -3,7 +3,7 @@
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
@@ -24,7 +24,7 @@ model:
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
@@ -33,7 +33,7 @@ model:
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
@@ -50,7 +50,7 @@ model:
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
@@ -73,7 +73,7 @@ model:
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
@@ -84,9 +84,9 @@ data:
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
3
 
4
  model:
5
  base_learning_rate: 1.0e-04
6
+ target: ldm.models.diffusion.ddpm_edit.LatentDiffusion
7
  params:
8
  linear_start: 0.00085
9
  linear_end: 0.0120
24
  load_ema: true
25
 
26
  scheduler_config: # 10000 warmup steps
27
+ target: ldm.lr_scheduler.LambdaLinearScheduler
28
  params:
29
  warm_up_steps: [ 0 ]
30
  cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
33
  f_min: [ 1. ]
34
 
35
  unet_config:
36
+ target: ldm.modules.diffusionmodules.openaimodel.UNetModel
37
  params:
38
  image_size: 32 # unused
39
  in_channels: 8
50
  legacy: False
51
 
52
  first_stage_config:
53
+ target: ldm.models.autoencoder.AutoencoderKL
54
  params:
55
  embed_dim: 4
56
  monitor: val/rec_loss
73
  target: torch.nn.Identity
74
 
75
  cond_stage_config:
76
+ target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
77
 
78
  data:
79
  target: main.DataModuleFromConfig
84
  validation:
85
  target: edit_dataset.EditDataset
86
  params:
87
+ path: data/clip-filtered-dataset
88
+ cache_dir: data/
89
+ cache_name: data_10k
90
  split: val
91
  min_text_sim: 0.2
92
  min_image_sim: 0.75
configs/train.yaml CHANGED
@@ -3,7 +3,7 @@
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
@@ -23,7 +23,7 @@ model:
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
@@ -32,7 +32,7 @@ model:
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
@@ -49,7 +49,7 @@ model:
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
@@ -72,7 +72,7 @@ model:
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
@@ -82,7 +82,7 @@ data:
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
@@ -91,7 +91,7 @@ data:
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
3
 
4
  model:
5
  base_learning_rate: 1.0e-04
6
+ target: 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
23
  load_ema: false
24
 
25
  scheduler_config: # 10000 warmup steps
26
+ target: ldm.lr_scheduler.LambdaLinearScheduler
27
  params:
28
  warm_up_steps: [ 0 ]
29
  cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
32
  f_min: [ 1. ]
33
 
34
  unet_config:
35
+ target: ldm.modules.diffusionmodules.openaimodel.UNetModel
36
  params:
37
  image_size: 32 # unused
38
  in_channels: 8
49
  legacy: False
50
 
51
  first_stage_config:
52
+ target: ldm.models.autoencoder.AutoencoderKL
53
  params:
54
  embed_dim: 4
55
  monitor: val/rec_loss
72
  target: torch.nn.Identity
73
 
74
  cond_stage_config:
75
+ target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
76
 
77
  data:
78
  target: main.DataModuleFromConfig
82
  train:
83
  target: edit_dataset.EditDataset
84
  params:
85
+ path: data/clip-filtered-dataset
86
  split: train
87
  min_resize_res: 256
88
  max_resize_res: 256
91
  validation:
92
  target: edit_dataset.EditDataset
93
  params:
94
+ path: data/clip-filtered-dataset
95
  split: val
96
  min_resize_res: 256
97
  max_resize_res: 256
dataset_creation/generate_img_dataset.py CHANGED
@@ -1,5 +1,6 @@
1
  import argparse
2
  import json
 
3
  from pathlib import Path
4
 
5
  import k_diffusion
@@ -12,8 +13,11 @@ 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
 
@@ -112,15 +116,29 @@ def to_pil(image: torch.Tensor) -> Image.Image:
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,
@@ -200,9 +218,9 @@ def main():
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)
@@ -229,7 +247,7 @@ def main():
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:
@@ -255,7 +273,7 @@ def main():
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(
1
  import argparse
2
  import json
3
+ import sys
4
  from pathlib import Path
5
 
6
  import k_diffusion
13
  from pytorch_lightning import seed_everything
14
  from tqdm import tqdm
15
 
16
+ sys.path.append("./")
17
+ sys.path.append("./stable_diffusion")
18
+
19
+ from ldm.modules.attention import CrossAttention
20
+ from ldm.util import instantiate_from_config
21
  from metrics.clip_similarity import ClipSimilarity
22
 
23
 
116
  def main():
117
  parser = argparse.ArgumentParser()
118
  parser.add_argument(
119
+ "--out_dir",
120
  type=str,
121
+ required=True,
122
  help="Path to output dataset directory.",
123
  )
124
  parser.add_argument(
125
+ "--prompts_file",
126
  type=str,
127
+ required=True,
128
  help="Path to prompts .jsonl file.",
129
  )
130
+ parser.add_argument(
131
+ "--ckpt",
132
+ type=str,
133
+ default="stable_diffusion/models/ldm/stable-diffusion-v1/v1-5-pruned-emaonly.ckpt",
134
+ help="Path to stable diffusion checkpoint.",
135
+ )
136
+ parser.add_argument(
137
+ "--vae-ckpt",
138
+ type=str,
139
+ default="stable_diffusion/models/ldm/stable-diffusion-v1/vae-ft-mse-840000-ema-pruned.ckpt",
140
+ help="Path to vae checkpoint.",
141
+ )
142
  parser.add_argument(
143
  "--steps",
144
  type=int,
218
  seed_everything(global_seed)
219
 
220
  model = load_model_from_config(
221
+ OmegaConf.load("stable_diffusion/configs/stable-diffusion/v1-inference.yaml"),
222
+ ckpt=opt.ckpt,
223
+ vae_ckpt=opt.vae_ckpt,
224
  )
225
  model.cuda().eval()
226
  model_wrap = k_diffusion.external.CompVisDenoiser(model)
247
  with open(prompt_dir.joinpath("prompt.json"), "w") as fp:
248
  json.dump(prompt, fp)
249
 
250
+ cond = model.get_learned_conditioning([prompt["caption"], prompt["output"]])
251
  results = {}
252
 
253
  with tqdm(total=opt.n_samples, desc="Samples") as progress_bar:
273
  x1 = x_samples_ddim[1]
274
 
275
  clip_sim_0, clip_sim_1, clip_sim_dir, clip_sim_image = clip_similarity(
276
+ x0[None], x1[None], [prompt["caption"]], [prompt["output"]]
277
  )
278
 
279
  results[seed] = dict(
dataset_creation/generate_txt_dataset.py CHANGED
@@ -65,7 +65,7 @@ def main(openai_model: str, num_samples: int, num_partitions: int, partition: in
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
@@ -88,7 +88,7 @@ def main(openai_model: str, num_samples: int, num_partitions: int, partition: in
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")
@@ -102,8 +102,8 @@ def main(openai_model: str, num_samples: int, num_partitions: int, partition: in
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)
65
  dataset = dataset[permutation]
66
  captions = dataset["TEXT"]
67
  urls = dataset["URL"]
68
+ output_path = f"data/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
88
  continue
89
  if openai.Moderation.create(caption)["results"][0]["flagged"]:
90
  continue
91
+ edit_output = generate(openai_model, 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")
102
 
103
  if __name__ == "__main__":
104
  parser = ArgumentParser()
105
+ parser.add_argument("--openai-api-key", required=True, type=str)
106
+ parser.add_argument("--openai-model", required=True, 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)
dataset_creation/prepare_for_gpt.py CHANGED
@@ -1,7 +1,7 @@
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):
@@ -19,7 +19,7 @@ def main(input_path: str, output_path: str):
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)
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):
19
 
20
  if __name__ == "__main__":
21
  parser = ArgumentParser()
22
+ parser.add_argument("--input-path", required=True, type=str)
23
+ parser.add_argument("--output-path", required=True, type=str)
24
  args = parser.parse_args()
25
  main(args.input_path, args.output_path)
edit_app.py CHANGED
@@ -73,8 +73,7 @@ class CFGDenoiser(nn.Module):
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:
@@ -87,7 +86,7 @@ def load_model_from_config(config, ckpt, vae_ckpt=None, verbose=False, cached=Fa
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:")
@@ -101,8 +100,8 @@ def load_model_from_config(config, ckpt, vae_ckpt=None, verbose=False, cached=Fa
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
 
@@ -188,7 +187,7 @@ def main():
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():
@@ -208,7 +207,7 @@ def main():
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",
@@ -216,7 +215,7 @@ def main():
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",
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):
 
77
  print(f"Loading model from {ckpt}")
78
  pl_sd = torch.load(ckpt, map_location="cpu")
79
  if "global_step" in pl_sd:
86
  k: vae_sd[k[len("first_stage_model.") :]] if k.startswith("first_stage_model.") else v
87
  for k, v in sd.items()
88
  }
89
+ model = instantiate_from_config(config.model)
90
  m, u = model.load_state_dict(sd, strict=False)
91
  if len(m) > 0 and verbose:
92
  print("missing keys:")
100
  def main():
101
  parser = ArgumentParser()
102
  parser.add_argument("--resolution", default=512, type=int)
103
+ parser.add_argument("--config", default="configs/generate.yaml", type=str)
104
+ parser.add_argument("--ckpt", default="checkpoints/instruct-pix2pix-00-22000.ckpt", type=str)
105
  parser.add_argument("--vae-ckpt", default=None, type=str)
106
  args = parser.parse_args()
107
 
187
  return [seed, text_cfg_scale, image_cfg_scale, edited_image]
188
 
189
  def reset():
190
+ return [0, "Randomize Seed", 1371, "Fix CFG", 7.5, 1.5, None]
191
 
192
  with gr.Blocks(css="footer {visibility: hidden}") as demo:
193
  with gr.Row():
207
  edited_image.style(height=512, width=512)
208
 
209
  with gr.Row():
210
+ steps = gr.Number(value=100, precision=0, label="Steps", interactive=True)
211
  randomize_seed = gr.Radio(
212
  ["Fix Seed", "Randomize Seed"],
213
  value="Randomize Seed",
215
  show_label=False,
216
  interactive=True,
217
  )
218
+ seed = gr.Number(value=1371, precision=0, label="Seed", interactive=True)
219
  randomize_cfg = gr.Radio(
220
  ["Fix CFG", "Randomize CFG"],
221
  value="Fix CFG",
edit_cli.py CHANGED
@@ -65,13 +65,13 @@ def main():
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
 
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-00-22000.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=1.5, type=float)
75
  parser.add_argument("--seed", type=int)
76
  args = parser.parse_args()
77
 
edit_dataset.py CHANGED
@@ -70,3 +70,52 @@ class EditDataset(Dataset):
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))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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))
73
+
74
+
75
+ class EditDatasetEval(Dataset):
76
+ def __init__(
77
+ self,
78
+ path: str,
79
+ split: str = "train",
80
+ splits: tuple[float, float, float] = (0.9, 0.05, 0.05),
81
+ res: int = 256,
82
+ ):
83
+ assert split in ("train", "val", "test")
84
+ assert sum(splits) == 1
85
+ self.path = path
86
+ self.res = res
87
+
88
+ with open(Path(self.path, "seeds.json")) as f:
89
+ self.seeds = json.load(f)
90
+
91
+ split_0, split_1 = {
92
+ "train": (0.0, splits[0]),
93
+ "val": (splits[0], splits[0] + splits[1]),
94
+ "test": (splits[0] + splits[1], 1.0),
95
+ }[split]
96
+
97
+ idx_0 = math.floor(split_0 * len(self.seeds))
98
+ idx_1 = math.floor(split_1 * len(self.seeds))
99
+ self.seeds = self.seeds[idx_0:idx_1]
100
+
101
+ def __len__(self) -> int:
102
+ return len(self.seeds)
103
+
104
+ def __getitem__(self, i: int) -> dict[str, Any]:
105
+ name, seeds = self.seeds[i]
106
+ propt_dir = Path(self.path, name)
107
+ seed = seeds[torch.randint(0, len(seeds), ()).item()]
108
+ with open(propt_dir.joinpath("prompt.json")) as fp:
109
+ prompt = json.load(fp)
110
+ edit = prompt["edit"]
111
+ input_prompt = prompt["input"]
112
+ output_prompt = prompt["output"]
113
+
114
+ image_0 = Image.open(propt_dir.joinpath(f"{seed}_0.jpg"))
115
+
116
+ reize_res = torch.randint(self.res, self.res + 1, ()).item()
117
+ image_0 = image_0.resize((reize_res, reize_res), Image.Resampling.LANCZOS)
118
+
119
+ image_0 = rearrange(2 * torch.tensor(np.array(image_0)).float() / 255 - 1, "h w c -> c h w")
120
+
121
+ return dict(image_0=image_0, input_prompt=input_prompt, edit=edit, output_prompt=output_prompt)
environment.yaml CHANGED
@@ -14,6 +14,7 @@ dependencies:
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
14
  - numpy=1.19.2
15
  - pip:
16
  - albumentations==0.4.3
17
+ - datasets==2.8.0
18
  - diffusers
19
  - opencv-python==4.1.2.30
20
  - pudb==2019.2
imgs/dataset.jpg ADDED
imgs/edit_app.jpg ADDED
imgs/prompt_app.jpg ADDED
main.py CHANGED
@@ -1,6 +1,3 @@
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
@@ -26,8 +23,8 @@ 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):
@@ -553,6 +550,7 @@ if __name__ == "__main__":
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
@@ -560,9 +558,7 @@ if __name__ == "__main__":
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")
@@ -576,6 +572,12 @@ if __name__ == "__main__":
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())
 
 
 
1
  import argparse, os, sys, datetime, glob
2
  import numpy as np
3
  import time
23
 
24
  sys.path.append("./stable_diffusion")
25
 
26
+ from ldm.data.base import Txt2ImgIterableBaseDataset
27
+ from ldm.util import instantiate_from_config
28
 
29
 
30
  def get_parser(**parser_kwargs):
550
  nowname = f"{cfg_name}_{opt.name}"
551
  logdir = os.path.join(opt.logdir, nowname)
552
  ckpt = os.path.join(logdir, "checkpoints", "last.ckpt")
553
+ resume = False
554
 
555
  if os.path.isfile(ckpt):
556
  opt.resume_from_checkpoint = ckpt
558
  opt.base = base_configs + opt.base
559
  _tmp = logdir.split("/")
560
  nowname = _tmp[-1]
561
+ resume = True
 
 
562
 
563
  ckptdir = os.path.join(logdir, "checkpoints")
564
  cfgdir = os.path.join(logdir, "configs")
572
  configs = [OmegaConf.load(cfg) for cfg in opt.base]
573
  cli = OmegaConf.from_dotlist(unknown)
574
  config = OmegaConf.merge(*configs, cli)
575
+
576
+ if resume:
577
+ # By default, when finetuning from Stable Diffusion, we load the EMA-only checkpoint to initialize all weights.
578
+ # If resuming InstructPix2Pix from a finetuning checkpoint, instead load both EMA and non-EMA weights.
579
+ config.model.params.load_ema = True
580
+
581
  lightning_config = config.pop("lightning", OmegaConf.create())
582
  # merge trainer cli with config
583
  trainer_config = lightning_config.get("trainer", OmegaConf.create())
metrics/compute_metrics.py ADDED
@@ -0,0 +1,235 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 tqdm.auto import tqdm
14
+ from einops import rearrange
15
+ from omegaconf import OmegaConf
16
+ from PIL import Image, ImageOps
17
+ from torch import autocast
18
+
19
+ import json
20
+ import matplotlib.pyplot as plt
21
+ import seaborn
22
+ from pathlib import Path
23
+
24
+ sys.path.append("./")
25
+
26
+ from clip_similarity import ClipSimilarity
27
+ from edit_dataset import EditDatasetEval
28
+
29
+ sys.path.append("./stable_diffusion")
30
+
31
+ from ldm.util import instantiate_from_config
32
+
33
+
34
+ class CFGDenoiser(nn.Module):
35
+ def __init__(self, model):
36
+ super().__init__()
37
+ self.inner_model = model
38
+
39
+ def forward(self, z, sigma, cond, uncond, text_cfg_scale, image_cfg_scale):
40
+ cfg_z = einops.repeat(z, "1 ... -> n ...", n=3)
41
+ cfg_sigma = einops.repeat(sigma, "1 ... -> n ...", n=3)
42
+ cfg_cond = {
43
+ "c_crossattn": [torch.cat([cond["c_crossattn"][0], uncond["c_crossattn"][0], uncond["c_crossattn"][0]])],
44
+ "c_concat": [torch.cat([cond["c_concat"][0], cond["c_concat"][0], uncond["c_concat"][0]])],
45
+ }
46
+ out_cond, out_img_cond, out_uncond = self.inner_model(cfg_z, cfg_sigma, cond=cfg_cond).chunk(3)
47
+ return out_uncond + text_cfg_scale * (out_cond - out_img_cond) + image_cfg_scale * (out_img_cond - out_uncond)
48
+
49
+
50
+ def load_model_from_config(config, ckpt, vae_ckpt=None, verbose=False):
51
+ print(f"Loading model from {ckpt}")
52
+ pl_sd = torch.load(ckpt, map_location="cpu")
53
+ if "global_step" in pl_sd:
54
+ print(f"Global Step: {pl_sd['global_step']}")
55
+ sd = pl_sd["state_dict"]
56
+ if vae_ckpt is not None:
57
+ print(f"Loading VAE from {vae_ckpt}")
58
+ vae_sd = torch.load(vae_ckpt, map_location="cpu")["state_dict"]
59
+ sd = {
60
+ k: vae_sd[k[len("first_stage_model.") :]] if k.startswith("first_stage_model.") else v
61
+ for k, v in sd.items()
62
+ }
63
+ model = instantiate_from_config(config.model)
64
+ m, u = model.load_state_dict(sd, strict=False)
65
+ if len(m) > 0 and verbose:
66
+ print("missing keys:")
67
+ print(m)
68
+ if len(u) > 0 and verbose:
69
+ print("unexpected keys:")
70
+ print(u)
71
+ return model
72
+
73
+ class ImageEditor(nn.Module):
74
+ def __init__(self, config, ckpt, vae_ckpt=None):
75
+ super().__init__()
76
+
77
+ config = OmegaConf.load(config)
78
+ self.model = load_model_from_config(config, ckpt, vae_ckpt)
79
+ self.model.eval().cuda()
80
+ self.model_wrap = K.external.CompVisDenoiser(self.model)
81
+ self.model_wrap_cfg = CFGDenoiser(self.model_wrap)
82
+ self.null_token = self.model.get_learned_conditioning([""])
83
+
84
+ def forward(
85
+ self,
86
+ image: torch.Tensor,
87
+ edit: str,
88
+ scale_txt: float = 7.5,
89
+ scale_img: float = 1.0,
90
+ steps: int = 100,
91
+ ) -> torch.Tensor:
92
+ assert image.dim() == 3
93
+ assert image.size(1) % 64 == 0
94
+ assert image.size(2) % 64 == 0
95
+ with torch.no_grad(), autocast("cuda"), self.model.ema_scope():
96
+ cond = {
97
+ "c_crossattn": [self.model.get_learned_conditioning([edit])],
98
+ "c_concat": [self.model.encode_first_stage(image[None]).mode()],
99
+ }
100
+ uncond = {
101
+ "c_crossattn": [self.model.get_learned_conditioning([""])],
102
+ "c_concat": [torch.zeros_like(cond["c_concat"][0])],
103
+ }
104
+ extra_args = {
105
+ "uncond": uncond,
106
+ "cond": cond,
107
+ "image_cfg_scale": scale_img,
108
+ "text_cfg_scale": scale_txt,
109
+ }
110
+ sigmas = self.model_wrap.get_sigmas(steps)
111
+ x = torch.randn_like(cond["c_concat"][0]) * sigmas[0]
112
+ x = K.sampling.sample_euler_ancestral(self.model_wrap_cfg, x, sigmas, extra_args=extra_args)
113
+ x = self.model.decode_first_stage(x)[0]
114
+ return x
115
+
116
+
117
+ def compute_metrics(config,
118
+ model_path,
119
+ vae_ckpt,
120
+ data_path,
121
+ output_path,
122
+ scales_img,
123
+ scales_txt,
124
+ num_samples = 5000,
125
+ split = "test",
126
+ steps = 50,
127
+ res = 512,
128
+ seed = 0):
129
+ editor = ImageEditor(config, model_path, vae_ckpt).cuda()
130
+ clip_similarity = ClipSimilarity().cuda()
131
+
132
+
133
+
134
+ outpath = Path(output_path, f"n={num_samples}_p={split}_s={steps}_r={res}_e={seed}.jsonl")
135
+ Path(output_path).mkdir(parents=True, exist_ok=True)
136
+
137
+ for scale_txt in scales_txt:
138
+ for scale_img in scales_img:
139
+ dataset = EditDatasetEval(
140
+ path=data_path,
141
+ split=split,
142
+ res=res
143
+ )
144
+ assert num_samples <= len(dataset)
145
+ print(f'Processing t={scale_txt}, i={scale_img}')
146
+ torch.manual_seed(seed)
147
+ perm = torch.randperm(len(dataset))
148
+ count = 0
149
+ i = 0
150
+
151
+ sim_0_avg = 0
152
+ sim_1_avg = 0
153
+ sim_direction_avg = 0
154
+ sim_image_avg = 0
155
+ count = 0
156
+
157
+ pbar = tqdm(total=num_samples)
158
+ while count < num_samples:
159
+
160
+ idx = perm[i].item()
161
+ sample = dataset[idx]
162
+ i += 1
163
+
164
+ gen = editor(sample["image_0"].cuda(), sample["edit"], scale_txt=scale_txt, scale_img=scale_img, steps=steps)
165
+
166
+ sim_0, sim_1, sim_direction, sim_image = clip_similarity(
167
+ sample["image_0"][None].cuda(), gen[None].cuda(), [sample["input_prompt"]], [sample["output_prompt"]]
168
+ )
169
+ sim_0_avg += sim_0.item()
170
+ sim_1_avg += sim_1.item()
171
+ sim_direction_avg += sim_direction.item()
172
+ sim_image_avg += sim_image.item()
173
+ count += 1
174
+ pbar.update(count)
175
+ pbar.close()
176
+
177
+ sim_0_avg /= count
178
+ sim_1_avg /= count
179
+ sim_direction_avg /= count
180
+ sim_image_avg /= count
181
+
182
+ with open(outpath, "a") as f:
183
+ f.write(f"{json.dumps(dict(sim_0=sim_0_avg, sim_1=sim_1_avg, sim_direction=sim_direction_avg, sim_image=sim_image_avg, num_samples=num_samples, split=split, scale_txt=scale_txt, scale_img=scale_img, steps=steps, res=res, seed=seed))}\n")
184
+ return outpath
185
+
186
+ def plot_metrics(metrics_file, output_path):
187
+
188
+ with open(metrics_file, 'r') as f:
189
+ data = [json.loads(line) for line in f]
190
+
191
+ plt.rcParams.update({'font.size': 11.5})
192
+ seaborn.set_style("darkgrid")
193
+ plt.figure(figsize=(20.5* 0.7, 10.8* 0.7), dpi=200)
194
+
195
+ x = [d["sim_direction"] for d in data]
196
+ y = [d["sim_image"] for d in data]
197
+
198
+ plt.plot(x, y, marker='o', linewidth=2, markersize=4)
199
+
200
+ plt.xlabel("CLIP Text-Image Direction Similarity", labelpad=10)
201
+ plt.ylabel("CLIP Image Similarity", labelpad=10)
202
+
203
+ plt.savefig(Path(output_path) / Path("plot.pdf"), bbox_inches="tight")
204
+
205
+ def main():
206
+ parser = ArgumentParser()
207
+ parser.add_argument("--resolution", default=512, type=int)
208
+ parser.add_argument("--steps", default=100, type=int)
209
+ parser.add_argument("--config", default="configs/generate.yaml", type=str)
210
+ parser.add_argument("--output_path", default="analysis/", type=str)
211
+ parser.add_argument("--ckpt", default="checkpoints/instruct-pix2pix-00-22000.ckpt", type=str)
212
+ parser.add_argument("--dataset", default="data/clip-filtered-dataset/", type=str)
213
+ parser.add_argument("--vae-ckpt", default=None, type=str)
214
+ args = parser.parse_args()
215
+
216
+ scales_img = [1.0, 1.2, 1.4, 1.6, 1.8, 2.0, 2.2]
217
+ scales_txt = [7.5]
218
+
219
+ metrics_file = compute_metrics(
220
+ args.config,
221
+ args.ckpt,
222
+ args.vae_ckpt,
223
+ args.dataset,
224
+ args.output_path,
225
+ scales_img,
226
+ scales_txt
227
+ steps = args.steps
228
+ )
229
+
230
+ plot_metrics(metrics_file, args.output_path)
231
+
232
+
233
+
234
+ if __name__ == "__main__":
235
+ main()
prompt_app.py CHANGED
@@ -48,8 +48,8 @@ def main(openai_model: str):
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)
48
 
49
  if __name__ == "__main__":
50
  parser = ArgumentParser()
51
+ parser.add_argument("--openai-api-key", required=True, type=str)
52
+ parser.add_argument("--openai-model", required=True, type=str)
53
  args = parser.parse_args()
54
  openai.api_key = args.openai_api_key
55
  main(args.openai_model)
scripts/download_checkpoints.sh CHANGED
@@ -4,4 +4,4 @@ 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
4
 
5
  mkdir -p $SCRIPT_DIR/../checkpoints
6
 
7
+ curl http://instruct-pix2pix.eecs.berkeley.edu/instruct-pix2pix-00-22000.ckpt -o $SCRIPT_DIR/../checkpoints/instruct-pix2pix-00-22000.ckpt
scripts/download_data.sh CHANGED
@@ -1,11 +1,27 @@
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
 
 
 
 
 
 
 
1
  #!/bin/bash
2
 
3
+ # Make data folder relative to script location
4
  SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
5
 
6
  mkdir -p $SCRIPT_DIR/../data
7
 
8
+ # Copy text datasets
9
+ wget -q --show-progress http://instruct-pix2pix.eecs.berkeley.edu/gpt-generated-prompts.jsonl -O $SCRIPT_DIR/../data/gpt-generated-prompts.jsonl
10
+ wget -q --show-progress http://instruct-pix2pix.eecs.berkeley.edu/human-written-prompts.jsonl -O $SCRIPT_DIR/../data/human-written-prompts.jsonl
11
 
12
+ # If dataset name isn't provided, exit.
13
+ if [ -z $1 ]
14
+ then
15
+ exit 0
16
+ fi
17
+
18
+ # Copy dataset files
19
  mkdir $SCRIPT_DIR/../data/$1
20
+ wget -A zip,json -R "index.html*" -q --show-progress -r --no-parent http://instruct-pix2pix.eecs.berkeley.edu/$1/ -nd -P $SCRIPT_DIR/../data/$1/
21
+
22
+ # Unzip to folders
23
+ unzip $SCRIPT_DIR/../data/$1/\*.zip -d $SCRIPT_DIR/../data/$1/
24
+
25
+ # Cleanup
26
+ rm -f $SCRIPT_DIR/../data/$1/*.zip
27
+ rm -f $SCRIPT_DIR/../data/$1/*.html
scripts/download_pretrained_sd.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/../stable_diffusion/models/ldm/stable-diffusion-v1
6
+ curl -L https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.ckpt -o $SCRIPT_DIR/../stable_diffusion/models/ldm/stable-diffusion-v1/v1-5-pruned-emaonly.ckpt
7
+ curl -L https://huggingface.co/stabilityai/sd-vae-ft-mse-original/resolve/main/vae-ft-mse-840000-ema-pruned.ckpt -o $SCRIPT_DIR/../stable_diffusion/models/ldm/stable-diffusion-v1/vae-ft-mse-840000-ema-pruned.ckpt
stable_diffusion/main.py CHANGED
@@ -737,5 +737,8 @@ if __name__ == "__main__":
737
  dst = os.path.join(dst, "debug_runs", name)
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())
 
 
 
737
  dst = os.path.join(dst, "debug_runs", name)
738
  os.makedirs(os.path.split(dst)[0], exist_ok=True)
739
  os.rename(logdir, dst)
740
+ try:
741
+ if trainer.global_rank == 0:
742
+ print(trainer.profiler.summary())
743
+ except:
744
+ pass