public/model-deepseek-ocr-xw
62
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

DESC: fix infer

Browse Source
This commit is contained in:
xiawei 2025-11-14 15:36:52 +08:00
parent 6f0e9b4195
commit 30aa84f057
1 changed files with 10 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
modeling_deepseekocr.py
View File

@ -944,17 +944,6 @@ class DeepseekOCRForCausalLM(DeepseekV2ForCausalLM):
no_repeat_ngram_size = 35, no_repeat_ngram_size = 35,
use_cache = True use_cache = True
) )
if '<image>' in conversation[0]['content'] and eval_mode:
outputs = tokenizer.decode(output_ids[0, input_ids.unsqueeze(0).cuda().shape[1]:])
stop_str = '<｜end▁of▁sentence｜>'
if outputs.endswith(stop_str):
outputs = outputs[:-len(stop_str)]
# re_match
outputs = outputs.strip()
return outputs
if '<image>' in conversation[0]['content'] and test_compress: if '<image>' in conversation[0]['content'] and test_compress:
outputs = tokenizer.decode(output_ids[0, input_ids.unsqueeze(0).cuda().shape[1]:]) outputs = tokenizer.decode(output_ids[0, input_ids.unsqueeze(0).cuda().shape[1]:])
@ -1035,3 +1024,13 @@ class DeepseekOCRForCausalLM(DeepseekV2ForCausalLM):
plt.close() plt.close()
result.save(f"{output_path}/result_with_boxes.jpg") result.save(f"{output_path}/result_with_boxes.jpg")
if '<image>' in conversation[0]['content'] and eval_mode:
outputs = tokenizer.decode(output_ids[0, input_ids.unsqueeze(0).cuda().shape[1]:])
stop_str = '<｜end▁of▁sentence｜>'
if outputs.endswith(stop_str):
outputs = outputs[:-len(stop_str)]
# re_match
outputs = outputs.strip()
return outputs