Update README.md
Browse files
README.md
CHANGED
@@ -27,21 +27,23 @@ from matplotlib import pyplot as plt
|
|
27 |
import torch
|
28 |
import torchvision.transforms as T
|
29 |
|
30 |
-
sample = PIL.Image.open("
|
31 |
sample = T.ToTensor()(sample)[None,:] # add batch dimension
|
32 |
-
sample = T.
|
33 |
|
34 |
-
vqvae=torch.jit.load("
|
35 |
|
36 |
-
# rec is
|
37 |
-
#
|
38 |
-
#
|
39 |
-
#
|
|
|
|
|
|
|
40 |
|
41 |
-
|
42 |
-
rec_ind = vqvae.decode_from_ind(ind)
|
43 |
-
rec=
|
44 |
-
rec_ind=rec_ind.sigmoid()
|
45 |
|
46 |
print("Original image shape",list(sample.shape[1:]))
|
47 |
print("ind shapes",[list(v.shape[1:]) for v in ind])
|
@@ -52,6 +54,7 @@ plt.imshow(T.ToPILImage()(sample[0]).resize((256,256)))
|
|
52 |
plt.title("original")
|
53 |
plt.axis('off')
|
54 |
|
|
|
55 |
plt.subplot(1,3,2)
|
56 |
plt.imshow(T.ToPILImage()(rec[0]).resize((256,256)))
|
57 |
plt.title("reconstruction")
|
@@ -64,6 +67,7 @@ plt.title("reconstruction from ind")
|
|
64 |
plt.axis('off')
|
65 |
plt.show()
|
66 |
|
|
|
67 |
plt.figure(figsize=(18,6))
|
68 |
plt.subplot(1,3,1)
|
69 |
plt.imshow(T.ToPILImage()(ind[0]/512).resize((256,256)))
|
@@ -82,7 +86,7 @@ plt.axis('off')
|
|
82 |
plt.show()
|
83 |
|
84 |
print("latent space render")
|
85 |
-
for z_ in
|
86 |
dims = len(z_[0])
|
87 |
dims_sqrt = int(dims**0.5)
|
88 |
plt.figure(figsize=(10,10))
|
@@ -98,20 +102,12 @@ for z_ in z:
|
|
98 |
```
|
99 |
|
100 |
```
|
101 |
-
Original image shape [3,
|
102 |
-
ind shapes [[
|
103 |
```
|
104 |
|
105 |
-
|
106 |
-

|
112 |
-
Mid
|
113 |
-

|
114 |
-
Top
|
115 |
-

|
116 |
-
|
117 |
-
As you can see, it properly handles different image aspects at different scales
|
|
|
27 |
import torch
|
28 |
import torchvision.transforms as T
|
29 |
|
30 |
+
sample = PIL.Image.open("image.png") # you sample image
|
31 |
sample = T.ToTensor()(sample)[None,:] # add batch dimension
|
32 |
+
sample = T.RandomCrop((256,256))(sample) # this vqvae works fine with any input image size that is divisible by 8
|
33 |
|
34 |
+
vqvae=torch.jit.load("model_v3.pt")
|
35 |
|
36 |
+
# rec, rec_ind is reconstructions
|
37 |
+
# rec is reconstruction from latent space values z
|
38 |
+
# rec_ind is reconstruction from model predicted vector indices
|
39 |
+
# z latent space tensor with 64 channels and 4x smaller than input image
|
40 |
+
# z_layers is list of latent space tensors at different scales
|
41 |
+
# z_q_layers is quantized list of latent space tensors
|
42 |
+
# ind is list of encoded indices of quantized elements in latent space for each scale
|
43 |
|
44 |
+
z, z_layers,z_q_layers, ind = vqvae.encode(sample)
|
45 |
+
rec_ind = vqvae.decode_from_ind(ind).sigmoid()
|
46 |
+
rec = vqvae.decode(z).sigmoid()
|
|
|
47 |
|
48 |
print("Original image shape",list(sample.shape[1:]))
|
49 |
print("ind shapes",[list(v.shape[1:]) for v in ind])
|
|
|
54 |
plt.title("original")
|
55 |
plt.axis('off')
|
56 |
|
57 |
+
# these two must look the same
|
58 |
plt.subplot(1,3,2)
|
59 |
plt.imshow(T.ToPILImage()(rec[0]).resize((256,256)))
|
60 |
plt.title("reconstruction")
|
|
|
67 |
plt.axis('off')
|
68 |
plt.show()
|
69 |
|
70 |
+
# this must look like a pile of mess
|
71 |
plt.figure(figsize=(18,6))
|
72 |
plt.subplot(1,3,1)
|
73 |
plt.imshow(T.ToPILImage()(ind[0]/512).resize((256,256)))
|
|
|
86 |
plt.show()
|
87 |
|
88 |
print("latent space render")
|
89 |
+
for z_ in z_layers:
|
90 |
dims = len(z_[0])
|
91 |
dims_sqrt = int(dims**0.5)
|
92 |
plt.figure(figsize=(10,10))
|
|
|
102 |
```
|
103 |
|
104 |
```
|
105 |
+
Original image shape [3, 256, 256]
|
106 |
+
ind shapes [[64, 64], [32, 32], [16, 16]]
|
107 |
```
|
108 |
|
109 |
+
Here is some examples at 256x256 resolution
|
110 |
+

|
111 |
+

|
112 |
+

|
113 |
+

|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|