Upload Moondream
Browse files- model.safetensors +1 -1
- moondream.py +73 -2
- vision_encoder.py +8 -6
model.safetensors
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 3715037856
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:493ac8972766b8e4b9005bfab11454b93aab4987b44a01debebec3fa96773105
|
3 |
size 3715037856
|
moondream.py
CHANGED
@@ -9,13 +9,16 @@ from .configuration_moondream import PhiConfig
|
|
9 |
|
10 |
class Moondream(PreTrainedModel):
|
11 |
config_class = MoondreamConfig
|
|
|
12 |
|
13 |
def __init__(self, config):
|
14 |
super().__init__(config)
|
15 |
self.vision_encoder = VisionEncoder()
|
16 |
|
17 |
if type(config.phi_config) == dict:
|
18 |
-
phi_config = PhiConfig(
|
|
|
|
|
19 |
else:
|
20 |
phi_config = config.phi_config
|
21 |
self.text_model = PhiForCausalLM(phi_config)
|
@@ -94,7 +97,7 @@ class Moondream(PreTrainedModel):
|
|
94 |
prompt,
|
95 |
eos_text="<END>",
|
96 |
tokenizer=tokenizer,
|
97 |
-
max_new_tokens=
|
98 |
**kwargs,
|
99 |
)[0]
|
100 |
cleaned_answer = re.sub("<$|<END$", "", answer).strip()
|
@@ -104,3 +107,71 @@ class Moondream(PreTrainedModel):
|
|
104 |
result_queue.put(cleaned_answer)
|
105 |
else:
|
106 |
return cleaned_answer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
|
10 |
class Moondream(PreTrainedModel):
|
11 |
config_class = MoondreamConfig
|
12 |
+
_supports_flash_attn_2 = True
|
13 |
|
14 |
def __init__(self, config):
|
15 |
super().__init__(config)
|
16 |
self.vision_encoder = VisionEncoder()
|
17 |
|
18 |
if type(config.phi_config) == dict:
|
19 |
+
phi_config = PhiConfig(
|
20 |
+
**config.phi_config, attn_implementation=config._attn_implementation
|
21 |
+
)
|
22 |
else:
|
23 |
phi_config = config.phi_config
|
24 |
self.text_model = PhiForCausalLM(phi_config)
|
|
|
97 |
prompt,
|
98 |
eos_text="<END>",
|
99 |
tokenizer=tokenizer,
|
100 |
+
max_new_tokens=512,
|
101 |
**kwargs,
|
102 |
)[0]
|
103 |
cleaned_answer = re.sub("<$|<END$", "", answer).strip()
|
|
|
107 |
result_queue.put(cleaned_answer)
|
108 |
else:
|
109 |
return cleaned_answer
|
110 |
+
|
111 |
+
def batch_answer(
|
112 |
+
self,
|
113 |
+
images,
|
114 |
+
prompts,
|
115 |
+
tokenizer,
|
116 |
+
**kwargs,
|
117 |
+
):
|
118 |
+
eos_tokens = tokenizer("<END>", add_special_tokens=False)[0].ids
|
119 |
+
|
120 |
+
image_embeds = self.encode_image(images)
|
121 |
+
|
122 |
+
templated_prompts = [
|
123 |
+
f"<image>\n\nQuestion: {prompt}\n\nAnswer: " for prompt in prompts
|
124 |
+
]
|
125 |
+
prompt_embs = [
|
126 |
+
self.input_embeds(prompt, image_embed.unsqueeze(0), tokenizer)[0]
|
127 |
+
for prompt, image_embed in zip(templated_prompts, image_embeds)
|
128 |
+
]
|
129 |
+
|
130 |
+
bos_emb = prompt_embs[0][0]
|
131 |
+
max_len = max([p.shape[0] for p in prompt_embs])
|
132 |
+
|
133 |
+
inputs_embeds = torch.cat(
|
134 |
+
[
|
135 |
+
torch.cat([bos_emb.repeat(max_len - p.shape[0], 1), p]).unsqueeze(0)
|
136 |
+
for p in prompt_embs
|
137 |
+
],
|
138 |
+
dim=0,
|
139 |
+
)
|
140 |
+
attention_mask = torch.cat(
|
141 |
+
[
|
142 |
+
torch.cat(
|
143 |
+
[
|
144 |
+
torch.zeros(
|
145 |
+
1,
|
146 |
+
max_len - p.shape[0],
|
147 |
+
device=self.device,
|
148 |
+
dtype=torch.long,
|
149 |
+
),
|
150 |
+
torch.ones(1, p.shape[0], device=self.device, dtype=torch.long),
|
151 |
+
],
|
152 |
+
dim=1,
|
153 |
+
)
|
154 |
+
for p in prompt_embs
|
155 |
+
],
|
156 |
+
dim=0,
|
157 |
+
)
|
158 |
+
|
159 |
+
generate_config = {
|
160 |
+
"eos_token_id": eos_tokens,
|
161 |
+
"bos_token_id": tokenizer.bos_token_id,
|
162 |
+
"pad_token_id": tokenizer.eos_token_id,
|
163 |
+
"max_new_tokens": 512,
|
164 |
+
**kwargs,
|
165 |
+
}
|
166 |
+
|
167 |
+
with torch.no_grad():
|
168 |
+
output_ids = self.text_model.generate(
|
169 |
+
inputs_embeds=inputs_embeds,
|
170 |
+
attention_mask=attention_mask,
|
171 |
+
**generate_config,
|
172 |
+
)
|
173 |
+
|
174 |
+
return [
|
175 |
+
re.sub("<$|<END$", "", x).strip()
|
176 |
+
for x in tokenizer.batch_decode(output_ids, skip_special_tokens=True)
|
177 |
+
]
|
vision_encoder.py
CHANGED
@@ -121,13 +121,15 @@ class VisionEncoder(nn.Module):
|
|
121 |
def dtype(self):
|
122 |
return self.projection.mlp.fc1.weight.dtype
|
123 |
|
124 |
-
def __call__(self,
|
|
|
|
|
|
|
125 |
with torch.no_grad():
|
126 |
-
x = (
|
127 |
-
self.preprocess(image.convert("RGB"))
|
128 |
-
|
129 |
-
|
130 |
-
)
|
131 |
x = rearrange(x, "b c (h p1) (w p2) -> b (h w) (c p1 p2)", p1=14, p2=14)
|
132 |
|
133 |
x = self.encoder(x)
|
|
|
121 |
def dtype(self):
|
122 |
return self.projection.mlp.fc1.weight.dtype
|
123 |
|
124 |
+
def __call__(self, images) -> torch.Tensor:
|
125 |
+
if not isinstance(images, list):
|
126 |
+
images = [images]
|
127 |
+
|
128 |
with torch.no_grad():
|
129 |
+
x = torch.stack(
|
130 |
+
[self.preprocess(image.convert("RGB")) for image in images]
|
131 |
+
).to(self.device, dtype=self.dtype)
|
132 |
+
|
|
|
133 |
x = rearrange(x, "b c (h p1) (w p2) -> b (h w) (c p1 p2)", p1=14, p2=14)
|
134 |
|
135 |
x = self.encoder(x)
|