LiruiZhao commited on
Commit
bc59621
1 Parent(s): cbbb276
Files changed (2) hide show
  1. app.py +1 -3
  2. configs/generate_diffree.yaml +119 -0
app.py CHANGED
@@ -388,6 +388,4 @@ with gr.Blocks(css="footer {visibility: hidden}") as demo:
388
  # demo.launch(share=True)
389
 
390
 
391
- # demo.queue().launch()
392
- demo.queue(concurrency_count=1)
393
- demo.launch(share=False)
 
388
  # demo.launch(share=True)
389
 
390
 
391
+ demo.queue().launch()
 
 
configs/generate_diffree.yaml ADDED
@@ -0,0 +1,119 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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: 5.0e-05
6
+ target: ldm.models.diffusion.ddpm_diffree.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
+ first_stage_downsample: True
16
+ # image_size: 64
17
+ # image_size: 32
18
+ image_size: 16
19
+ channels: 4
20
+ cond_stage_trainable: false # Note: different from the one we trained before
21
+ conditioning_key: hybrid
22
+ monitor: val/loss_simple_ema
23
+ scale_factor: 0.18215
24
+ use_ema: true
25
+ load_ema: true
26
+
27
+ scheduler_config: # 10000 warmup steps
28
+ target: ldm.lr_scheduler.LambdaLinearScheduler
29
+ params:
30
+ warm_up_steps: [ 0 ]
31
+ cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
32
+ f_start: [ 1.e-6 ]
33
+ f_max: [ 1. ]
34
+ f_min: [ 1. ]
35
+
36
+ unet_config:
37
+ target: ldm.modules.diffusionmodules.openaimodel_diffree.UNetModel
38
+ params:
39
+ image_size: 32 # unused
40
+ in_channels: 8
41
+ # in_mask_channels: 12
42
+ out_channels: 4
43
+ model_channels: 320
44
+ attention_resolutions: [ 4, 2, 1 ]
45
+ num_res_blocks: 2
46
+ channel_mult: [ 1, 2, 4, 4 ]
47
+ num_heads: 8
48
+ use_spatial_transformer: True
49
+ transformer_depth: 1
50
+ context_dim: 768
51
+ use_checkpoint: True
52
+ legacy: False
53
+
54
+ omp_config:
55
+ target: ldm.modules.diffusionmodules.openaimodel_diffree.OMPModule
56
+ params:
57
+ image_size: 32 # unused
58
+ in_channels: 8
59
+ # in_mask_channels: 12
60
+ out_channels: 4
61
+ model_channels: 320
62
+ attention_resolutions: [ 4, 2, 1 ]
63
+ num_res_blocks: 2
64
+ channel_mult: [ 1, 2, 4, 4 ]
65
+ num_heads: 8
66
+ use_spatial_transformer: True
67
+ transformer_depth: 1
68
+ context_dim: 768
69
+ use_checkpoint: True
70
+ legacy: False
71
+
72
+ first_stage_config:
73
+ target: ldm.models.autoencoder.AutoencoderKL
74
+ params:
75
+ embed_dim: 4
76
+ monitor: val/rec_loss
77
+ ddconfig:
78
+ double_z: true
79
+ z_channels: 4
80
+ resolution: 256
81
+ in_channels: 3
82
+ out_ch: 3
83
+ ch: 128
84
+ ch_mult:
85
+ - 1
86
+ - 2
87
+ - 4
88
+ - 4
89
+ num_res_blocks: 2
90
+ attn_resolutions: []
91
+ dropout: 0.0
92
+ lossconfig:
93
+ target: torch.nn.Identity
94
+
95
+ cond_stage_config:
96
+ target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
97
+
98
+ data:
99
+ target: main.DataModuleFromConfig
100
+ params:
101
+ batch_size: 128
102
+ num_workers: 1
103
+ wrap: false
104
+ validation:
105
+ target: edit_dataset_pam.EditDatasetMask
106
+ params:
107
+ path: data/clip-filtered-dataset
108
+ cache_dir: data/
109
+ cache_name: data_10k
110
+ split: val
111
+ min_text_sim: 0.2
112
+ min_image_sim: 0.75
113
+ min_direction_sim: 0.2
114
+ max_samples_per_prompt: 1
115
+ min_resize_res: 512
116
+ max_resize_res: 512
117
+ crop_res: 512
118
+ output_as_edit: False
119
+ real_input: True