Spaces:
Runtime error
Runtime error
Add expression model
Browse files
app.py
CHANGED
@@ -12,8 +12,9 @@ import audresample
|
|
12 |
|
13 |
|
14 |
device = 0 if torch.cuda.is_available() else "cpu"
|
15 |
-
model_name = "audeering/wav2vec2-large-robust-24-ft-age-gender"
|
16 |
duration = 1 # limit processing of audio
|
|
|
|
|
17 |
|
18 |
|
19 |
class AgeGenderHead(nn.Module):
|
@@ -66,10 +67,55 @@ class AgeGenderModel(Wav2Vec2PreTrainedModel):
|
|
66 |
return hidden_states, logits_age, logits_gender
|
67 |
|
68 |
|
|
|
|
|
69 |
|
70 |
-
|
71 |
-
|
72 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
|
74 |
|
75 |
def process_func(x: np.ndarray, sampling_rate: int) -> dict:
|
@@ -77,28 +123,38 @@ def process_func(x: np.ndarray, sampling_rate: int) -> dict:
|
|
77 |
# run through processor to normalize signal
|
78 |
# always returns a batch, so we just get the first entry
|
79 |
# then we put it on the device
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
y =
|
88 |
-
y = torch.
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
102 |
|
103 |
|
104 |
@spaces.GPU
|
@@ -117,17 +173,17 @@ def recognize(input_file):
|
|
117 |
target_rate = 16000
|
118 |
signal = audresample.resample(signal, sampling_rate, target_rate)
|
119 |
|
120 |
-
|
121 |
-
age = f"{round(age_gender['age'])} years"
|
122 |
-
gender = {k: v for k, v in age_gender.items() if k != "age"}
|
123 |
-
return age, gender
|
124 |
|
125 |
|
126 |
outputs = gr.Label()
|
127 |
title = "audEERING age and gender recognition"
|
128 |
description = (
|
129 |
-
"
|
130 |
-
f"
|
|
|
|
|
|
|
131 |
)
|
132 |
allow_flagging = "never"
|
133 |
|
@@ -159,8 +215,8 @@ with gr.Blocks() as demo:
|
|
159 |
gr.Markdown(description)
|
160 |
with gr.Tab(label="Speech analysis"):
|
161 |
with gr.Row():
|
162 |
-
gr.Markdown("Only the first second of the audio is processed.")
|
163 |
with gr.Column():
|
|
|
164 |
input = gr.Audio(
|
165 |
sources=["upload", "microphone"],
|
166 |
type="filepath",
|
@@ -170,8 +226,10 @@ with gr.Blocks() as demo:
|
|
170 |
with gr.Column():
|
171 |
output_age = gr.Textbox(label="Age")
|
172 |
output_gender = gr.Label(label="Gender")
|
|
|
173 |
|
174 |
-
|
|
|
175 |
|
176 |
|
177 |
demo.launch(debug=True)
|
|
|
12 |
|
13 |
|
14 |
device = 0 if torch.cuda.is_available() else "cpu"
|
|
|
15 |
duration = 1 # limit processing of audio
|
16 |
+
age_gender_model_name = "audeering/wav2vec2-large-robust-24-ft-age-gender"
|
17 |
+
expression_model_name = "audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim"
|
18 |
|
19 |
|
20 |
class AgeGenderHead(nn.Module):
|
|
|
67 |
return hidden_states, logits_age, logits_gender
|
68 |
|
69 |
|
70 |
+
class ExpressionHead(nn.Module):
|
71 |
+
r"""Expression model head."""
|
72 |
|
73 |
+
def __init__(self, config):
|
74 |
+
|
75 |
+
super().__init__()
|
76 |
+
|
77 |
+
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
|
78 |
+
self.dropout = nn.Dropout(config.final_dropout)
|
79 |
+
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
|
80 |
+
|
81 |
+
def forward(self, features, **kwargs):
|
82 |
+
|
83 |
+
x = features
|
84 |
+
x = self.dropout(x)
|
85 |
+
x = self.dense(x)
|
86 |
+
x = torch.tanh(x)
|
87 |
+
x = self.dropout(x)
|
88 |
+
x = self.out_proj(x)
|
89 |
+
|
90 |
+
return x
|
91 |
+
|
92 |
+
|
93 |
+
class ExpressionModel(Wav2Vec2PreTrainedModel):
|
94 |
+
r"""speech expression model."""
|
95 |
+
|
96 |
+
def __init__(self, config):
|
97 |
+
|
98 |
+
super().__init__(config)
|
99 |
+
|
100 |
+
self.config = config
|
101 |
+
self.wav2vec2 = Wav2Vec2Model(config)
|
102 |
+
self.classifier = ExpressionHead(config)
|
103 |
+
self.init_weights()
|
104 |
+
|
105 |
+
def forward(self, input_values):
|
106 |
+
outputs = self.wav2vec2(input_values)
|
107 |
+
hidden_states = outputs[0]
|
108 |
+
hidden_states = torch.mean(hidden_states, dim=1)
|
109 |
+
logits = self.classifier(hidden_states)
|
110 |
+
|
111 |
+
return hidden_states, logits
|
112 |
+
|
113 |
+
|
114 |
+
# Load models from hub
|
115 |
+
age_gender_processor = Wav2Vec2Processor.from_pretrained(age_gender_model_name)
|
116 |
+
age_gender_model = AgeGenderModel.from_pretrained(age_gender_model_name)
|
117 |
+
expression_processor = Wav2Vec2Processor.from_pretrained(expression_model_name)
|
118 |
+
expression_model = ExpressionModel.from_pretrained(expression_model_name)
|
119 |
|
120 |
|
121 |
def process_func(x: np.ndarray, sampling_rate: int) -> dict:
|
|
|
123 |
# run through processor to normalize signal
|
124 |
# always returns a batch, so we just get the first entry
|
125 |
# then we put it on the device
|
126 |
+
results = []
|
127 |
+
for processor, model in zip(
|
128 |
+
[age_gender_processor, expression_processor],
|
129 |
+
[age_gender_model, expression_model],
|
130 |
+
):
|
131 |
+
y = processor(x, sampling_rate=sampling_rate)
|
132 |
+
y = y['input_values'][0]
|
133 |
+
y = y.reshape(1, -1)
|
134 |
+
y = torch.from_numpy(y).to(device)
|
135 |
+
|
136 |
+
# run through model
|
137 |
+
with torch.no_grad():
|
138 |
+
y = model(y)
|
139 |
+
y = torch.hstack([y[1], y[2]])
|
140 |
+
|
141 |
+
# convert to numpy
|
142 |
+
y = y.detach().cpu().numpy()
|
143 |
+
results.append(y[0])
|
144 |
+
|
145 |
+
return (
|
146 |
+
100 * results[0][0], # age
|
147 |
+
{
|
148 |
+
"female": results[0][1],
|
149 |
+
"male": results[0][2],
|
150 |
+
"child": results[0][3],
|
151 |
+
},
|
152 |
+
{
|
153 |
+
"arousal": results[1][0],
|
154 |
+
"dominance": results[1][1],
|
155 |
+
"valence": results[1][2],
|
156 |
+
}
|
157 |
+
)
|
158 |
|
159 |
|
160 |
@spaces.GPU
|
|
|
173 |
target_rate = 16000
|
174 |
signal = audresample.resample(signal, sampling_rate, target_rate)
|
175 |
|
176 |
+
return process_func(signal, target_rate)
|
|
|
|
|
|
|
177 |
|
178 |
|
179 |
outputs = gr.Label()
|
180 |
title = "audEERING age and gender recognition"
|
181 |
description = (
|
182 |
+
"Speech analysis of an audio file or microphone recording. \n"
|
183 |
+
f"[{age_gender_model_name}](https://huggingface.co/{age_gender_model_name}) "
|
184 |
+
"is used for age and gender recognition, "
|
185 |
+
f"[{expression_model_name}](https://huggingface.co/{expression_model_name}) "
|
186 |
+
"is used for expression recognition."
|
187 |
)
|
188 |
allow_flagging = "never"
|
189 |
|
|
|
215 |
gr.Markdown(description)
|
216 |
with gr.Tab(label="Speech analysis"):
|
217 |
with gr.Row():
|
|
|
218 |
with gr.Column():
|
219 |
+
gr.Markdown("Only the first second of the audio is processed.")
|
220 |
input = gr.Audio(
|
221 |
sources=["upload", "microphone"],
|
222 |
type="filepath",
|
|
|
226 |
with gr.Column():
|
227 |
output_age = gr.Textbox(label="Age")
|
228 |
output_gender = gr.Label(label="Gender")
|
229 |
+
output_expression = gr.Label(label="Expression")
|
230 |
|
231 |
+
outputs = [output_age, output_gender, output_expression]
|
232 |
+
submit_btn.click(recognize, input, outputs)
|
233 |
|
234 |
|
235 |
demo.launch(debug=True)
|