Model in memory
Browse files
app.py
CHANGED
@@ -66,27 +66,33 @@ st.image(image, caption='Your target document')
|
|
66 |
with st.spinner(f'Processing the document ...'):
|
67 |
pre_trained_model = "unstructuredio/chipper-fast-fine-tuning"
|
68 |
processor = DonutProcessor.from_pretrained(pre_trained_model)
|
69 |
-
|
|
|
70 |
|
71 |
-
|
|
|
|
|
|
|
72 |
|
73 |
-
|
74 |
-
repo_id=pre_trained_model, filename="lm_head.pth"
|
75 |
-
)
|
76 |
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
nn.Linear(rank, rank, bias=False),
|
81 |
-
nn.Linear(rank, model.decoder.lm_head.weight.shape[0], bias=True),
|
82 |
-
)
|
83 |
|
84 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
85 |
|
86 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
87 |
|
88 |
-
|
89 |
-
|
|
|
90 |
|
91 |
st.info(f'Parsing document')
|
92 |
parsed_info = run_prediction(image.convert("RGB"), model, processor, prompt)
|
|
|
66 |
with st.spinner(f'Processing the document ...'):
|
67 |
pre_trained_model = "unstructuredio/chipper-fast-fine-tuning"
|
68 |
processor = DonutProcessor.from_pretrained(pre_trained_model)
|
69 |
+
|
70 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
71 |
|
72 |
+
if 'model' in st.session_state:
|
73 |
+
model = st.session_state['model']
|
74 |
+
else:
|
75 |
+
model = VisionEncoderDecoderModel.from_pretrained(pre_trained_model)
|
76 |
|
77 |
+
from huggingface_hub import hf_hub_download
|
|
|
|
|
78 |
|
79 |
+
lm_head_file = hf_hub_download(
|
80 |
+
repo_id=pre_trained_model, filename="lm_head.pth"
|
81 |
+
)
|
|
|
|
|
|
|
82 |
|
83 |
+
rank = 128
|
84 |
+
model.decoder.lm_head = nn.Sequential(
|
85 |
+
nn.Linear(model.decoder.lm_head.weight.shape[1], rank, bias=False),
|
86 |
+
nn.Linear(rank, rank, bias=False),
|
87 |
+
nn.Linear(rank, model.decoder.lm_head.weight.shape[0], bias=True),
|
88 |
+
)
|
89 |
+
|
90 |
+
model.decoder.lm_head.load_state_dict(torch.load(lm_head_file))
|
91 |
|
|
|
92 |
|
93 |
+
model.eval()
|
94 |
+
model.to(device)
|
95 |
+
st.session_state['model'] = model
|
96 |
|
97 |
st.info(f'Parsing document')
|
98 |
parsed_info = run_prediction(image.convert("RGB"), model, processor, prompt)
|