Spaces:
Runtime error
Runtime error
Update inference.py
Browse files- inference.py +100 -1
inference.py
CHANGED
@@ -17,7 +17,8 @@ class PaliGemmaModel:
|
|
17 |
|
18 |
@spaces.GPU
|
19 |
def infer(self, image: PIL.Image.Image, text: str, max_new_tokens: int) -> str:
|
20 |
-
inputs = self.processor(text=text, images=image, return_tensors="pt")
|
|
|
21 |
with torch.inference_mode():
|
22 |
generated_ids = self.model.generate(
|
23 |
**inputs,
|
@@ -26,3 +27,101 @@ class PaliGemmaModel:
|
|
26 |
)
|
27 |
result = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
|
28 |
return result[0][len(text):].lstrip("\n")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
|
18 |
@spaces.GPU
|
19 |
def infer(self, image: PIL.Image.Image, text: str, max_new_tokens: int) -> str:
|
20 |
+
inputs = self.processor(text=text, images=image, return_tensors="pt")
|
21 |
+
inputs = {k: v.to(self.device) for k, v in inputs.items()} # Move inputs to the correct device
|
22 |
with torch.inference_mode():
|
23 |
generated_ids = self.model.generate(
|
24 |
**inputs,
|
|
|
27 |
)
|
28 |
result = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
|
29 |
return result[0][len(text):].lstrip("\n")
|
30 |
+
|
31 |
+
class VAEModel:
|
32 |
+
def __init__(self, model_path: str):
|
33 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
34 |
+
self.params = self._get_params(model_path)
|
35 |
+
|
36 |
+
def _get_params(self, checkpoint_path):
|
37 |
+
"""Converts PyTorch checkpoint to Flax params."""
|
38 |
+
checkpoint = dict(np.load(checkpoint_path))
|
39 |
+
|
40 |
+
def transp(kernel):
|
41 |
+
return np.transpose(kernel, (2, 3, 1, 0))
|
42 |
+
|
43 |
+
def conv(name):
|
44 |
+
return {
|
45 |
+
'bias': checkpoint[name + '.bias'],
|
46 |
+
'kernel': transp(checkpoint[name + '.weight']),
|
47 |
+
}
|
48 |
+
|
49 |
+
def resblock(name):
|
50 |
+
return {
|
51 |
+
'Conv_0': conv(name + '.0'),
|
52 |
+
'Conv_1': conv(name + '.2'),
|
53 |
+
'Conv_2': conv(name + '.4'),
|
54 |
+
}
|
55 |
+
|
56 |
+
return {
|
57 |
+
'_embeddings': checkpoint['_vq_vae._embedding'],
|
58 |
+
'Conv_0': conv('decoder.0'),
|
59 |
+
'ResBlock_0': resblock('decoder.2.net'),
|
60 |
+
'ResBlock_1': resblock('decoder.3.net'),
|
61 |
+
'ConvTranspose_0': conv('decoder.4'),
|
62 |
+
'ConvTranspose_1': conv('decoder.6'),
|
63 |
+
'ConvTranspose_2': conv('decoder.8'),
|
64 |
+
'ConvTranspose_3': conv('decoder.10'),
|
65 |
+
'Conv_1': conv('decoder.12'),
|
66 |
+
}
|
67 |
+
|
68 |
+
def reconstruct_masks(self, codebook_indices):
|
69 |
+
quantized = self._quantized_values_from_codebook_indices(codebook_indices)
|
70 |
+
return self._decoder().apply({'params': self.params}, quantized)
|
71 |
+
|
72 |
+
def _quantized_values_from_codebook_indices(self, codebook_indices):
|
73 |
+
batch_size, num_tokens = codebook_indices.shape
|
74 |
+
assert num_tokens == 16, codebook_indices.shape
|
75 |
+
unused_num_embeddings, embedding_dim = self.params['_embeddings'].shape
|
76 |
+
|
77 |
+
encodings = jnp.take(self.params['_embeddings'], codebook_indices.reshape((-1)), axis=0)
|
78 |
+
encodings = encodings.reshape((batch_size, 4, 4, embedding_dim))
|
79 |
+
return encodings
|
80 |
+
|
81 |
+
@functools.cache
|
82 |
+
def _decoder(self):
|
83 |
+
class ResBlock(nn.Module):
|
84 |
+
features: int
|
85 |
+
|
86 |
+
@nn.compact
|
87 |
+
def __call__(self, x):
|
88 |
+
original_x = x
|
89 |
+
x = nn.Conv(features=self.features, kernel_size=(3, 3), padding=1)(x)
|
90 |
+
x = nn.relu(x)
|
91 |
+
x = nn.Conv(features=self.features, kernel_size=(3, 3), padding=1)(x)
|
92 |
+
x = nn.relu(x)
|
93 |
+
x = nn.Conv(features=self.features, kernel_size=(1, 1), padding=0)(x)
|
94 |
+
return x + original_x
|
95 |
+
|
96 |
+
class Decoder(nn.Module):
|
97 |
+
"""Upscales quantized vectors to mask."""
|
98 |
+
|
99 |
+
@nn.compact
|
100 |
+
def __call__(self, x):
|
101 |
+
num_res_blocks = 2
|
102 |
+
dim = 128
|
103 |
+
num_upsample_layers = 4
|
104 |
+
|
105 |
+
x = nn.Conv(features=dim, kernel_size=(1, 1), padding=0)(x)
|
106 |
+
x = nn.relu(x)
|
107 |
+
|
108 |
+
for _ in range(num_res_blocks):
|
109 |
+
x = ResBlock(features=dim)(x)
|
110 |
+
|
111 |
+
for _ in range(num_upsample_layers):
|
112 |
+
x = nn.ConvTranspose(
|
113 |
+
features=dim,
|
114 |
+
kernel_size=(4, 4),
|
115 |
+
strides=(2, 2),
|
116 |
+
padding=2,
|
117 |
+
transpose_kernel=True,
|
118 |
+
)(x)
|
119 |
+
x = nn.relu(x)
|
120 |
+
dim //= 2
|
121 |
+
|
122 |
+
x = nn.Conv(features=1, kernel_size=(1, 1), padding=0)(x)
|
123 |
+
|
124 |
+
return x
|
125 |
+
|
126 |
+
return jax.jit(Decoder().apply, backend='cpu')
|
127 |
+
|