Spaces:
Build error
Build error
danseith
commited on
Commit
·
2ce1788
1
Parent(s):
ca69fee
Implemented temperature scaling and changed output sampling to uniform.
Browse files
app.py
CHANGED
|
@@ -71,7 +71,7 @@ class TempScalePipe(FillMaskPipeline):
|
|
| 71 |
masked_index = torch.nonzero(input_ids == self.tokenizer.mask_token_id, as_tuple=False).squeeze(-1)
|
| 72 |
# Fill mask pipeline supports only one ${mask_token} per sample
|
| 73 |
|
| 74 |
-
logits = outputs[0, masked_index, :] /
|
| 75 |
probs = logits.softmax(dim=-1)
|
| 76 |
sampling = False
|
| 77 |
if sampling:
|
|
@@ -114,17 +114,18 @@ PIPELINE_REGISTRY.register_pipeline(
|
|
| 114 |
)
|
| 115 |
scrambler = pipeline("temp-scale", model="anferico/bert-for-patents")
|
| 116 |
|
| 117 |
-
def unmask(text, temp):
|
| 118 |
# text = add_mask(text)
|
| 119 |
split_text = text.split()
|
| 120 |
-
res = scrambler(text)
|
| 121 |
mask_pos = [i for i, t in enumerate(split_text) if 'MASK' in t][0]
|
| 122 |
out = {item["token_str"]: item["score"] for item in res}
|
| 123 |
score_to_str = {out[k]:k for k in out.keys()}
|
| 124 |
-
print(score_to_str)
|
| 125 |
-
print(out)
|
| 126 |
score_list = list(score_to_str.keys())
|
| 127 |
-
|
|
|
|
|
|
|
|
|
|
| 128 |
score = score_list[idx]
|
| 129 |
new_token = score_to_str[score]
|
| 130 |
split_text[mask_pos] = new_token
|
|
@@ -132,7 +133,7 @@ def unmask(text, temp):
|
|
| 132 |
|
| 133 |
textbox = gr.Textbox(label="Type language here", lines=5)
|
| 134 |
textbox2 = gr.Textbox(placeholder="Type here...", lines=4)
|
| 135 |
-
temp_slider = gr.Slider(1.0,
|
| 136 |
|
| 137 |
demo = gr.Interface(
|
| 138 |
fn=unmask,
|
|
|
|
| 71 |
masked_index = torch.nonzero(input_ids == self.tokenizer.mask_token_id, as_tuple=False).squeeze(-1)
|
| 72 |
# Fill mask pipeline supports only one ${mask_token} per sample
|
| 73 |
|
| 74 |
+
logits = outputs[0, masked_index, :] / temp
|
| 75 |
probs = logits.softmax(dim=-1)
|
| 76 |
sampling = False
|
| 77 |
if sampling:
|
|
|
|
| 114 |
)
|
| 115 |
scrambler = pipeline("temp-scale", model="anferico/bert-for-patents")
|
| 116 |
|
| 117 |
+
def unmask(text, temp, sampling='uniform'):
|
| 118 |
# text = add_mask(text)
|
| 119 |
split_text = text.split()
|
| 120 |
+
res = scrambler(text, temp=temp, top_k=3)
|
| 121 |
mask_pos = [i for i, t in enumerate(split_text) if 'MASK' in t][0]
|
| 122 |
out = {item["token_str"]: item["score"] for item in res}
|
| 123 |
score_to_str = {out[k]:k for k in out.keys()}
|
|
|
|
|
|
|
| 124 |
score_list = list(score_to_str.keys())
|
| 125 |
+
if sampling == 'multi':
|
| 126 |
+
idx = np.argmax(np.random.multinomial(1, score_list, 1))
|
| 127 |
+
else:
|
| 128 |
+
idx = np.random.randint(0, len(score_list))
|
| 129 |
score = score_list[idx]
|
| 130 |
new_token = score_to_str[score]
|
| 131 |
split_text[mask_pos] = new_token
|
|
|
|
| 133 |
|
| 134 |
textbox = gr.Textbox(label="Type language here", lines=5)
|
| 135 |
textbox2 = gr.Textbox(placeholder="Type here...", lines=4)
|
| 136 |
+
temp_slider = gr.Slider(1.0, 2.0, value=1.0, label='Temperature Scale')
|
| 137 |
|
| 138 |
demo = gr.Interface(
|
| 139 |
fn=unmask,
|