Spaces:
Runtime error
Runtime error
Update inference_2.py
Browse files- inference_2.py +40 -9
inference_2.py
CHANGED
@@ -1,22 +1,32 @@
|
|
|
|
1 |
import cv2
|
2 |
import onnx
|
3 |
import torch
|
4 |
import argparse
|
5 |
import numpy as np
|
|
|
|
|
6 |
from models import image
|
7 |
|
8 |
-
import warnings
|
9 |
from onnx2pytorch import ConvertModel
|
10 |
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
onnx_model = onnx.load('models/efficientnet.onnx')
|
15 |
-
pytorch_model = ConvertModel(onnx_model)
|
16 |
torch.manual_seed(42)
|
17 |
|
18 |
|
19 |
-
audio_args = {
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
|
21 |
|
22 |
def get_args(parser):
|
@@ -55,14 +65,14 @@ def get_args(parser):
|
|
55 |
|
56 |
def load_img_modality_model(args):
|
57 |
rgb_encoder = pytorch_model
|
58 |
-
ckpt = torch.load('
|
59 |
rgb_encoder.load_state_dict(ckpt['rgb_encoder'], strict = True)
|
60 |
rgb_encoder.eval()
|
61 |
return rgb_encoder
|
62 |
|
63 |
def load_spec_modality_model(args):
|
64 |
spec_encoder = image.RawNet(args)
|
65 |
-
ckpt = torch.load('
|
66 |
spec_encoder.load_state_dict(ckpt['spec_encoder'], strict = True)
|
67 |
spec_encoder.eval()
|
68 |
return spec_encoder
|
@@ -71,9 +81,12 @@ parser = argparse.ArgumentParser(description="Inference models")
|
|
71 |
get_args(parser)
|
72 |
args, remaining_args = parser.parse_known_args()
|
73 |
assert remaining_args == [], remaining_args
|
|
|
74 |
spec_model = load_spec_modality_model(args)
|
|
|
75 |
img_model = load_img_modality_model(args)
|
76 |
|
|
|
77 |
def preprocess_img(face):
|
78 |
face = face / 255
|
79 |
face = cv2.resize(face, (256, 256))
|
@@ -90,12 +103,15 @@ def df_spec_pred(input_audio):
|
|
90 |
spec_grads = spec_model.forward(audio)
|
91 |
spec_grads_inv = np.exp(spec_grads.cpu().detach().numpy().squeeze())
|
92 |
max_value = np.argmax(spec_grads_inv)
|
|
|
93 |
if max_value > 0.5:
|
94 |
preds = round(100 - (max_value*100), 3)
|
95 |
text2 = f"The audio is REAL."
|
|
|
96 |
else:
|
97 |
preds = round(max_value*100, 3)
|
98 |
text2 = f"The audio is FAKE."
|
|
|
99 |
return text2
|
100 |
|
101 |
def df_img_pred(input_image):
|
@@ -104,25 +120,34 @@ def df_img_pred(input_image):
|
|
104 |
img_grads = img_model.forward(face)
|
105 |
img_grads = img_grads.cpu().detach().numpy()
|
106 |
img_grads_np = np.squeeze(img_grads)
|
|
|
107 |
if img_grads_np[0] > 0.5:
|
108 |
preds = round(img_grads_np[0] * 100, 3)
|
109 |
text2 = f"The image is REAL. \nConfidence score is: {preds}"
|
|
|
110 |
else:
|
111 |
preds = round(img_grads_np[1] * 100, 3)
|
112 |
text2 = f"The image is FAKE. \nConfidence score is: {preds}"
|
|
|
113 |
return text2
|
114 |
|
|
|
115 |
def preprocess_video(input_video, n_frames = 3):
|
116 |
v_cap = cv2.VideoCapture(input_video)
|
117 |
v_len = int(v_cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
|
|
|
|
118 |
if n_frames is None:
|
119 |
sample = np.arange(0, v_len)
|
120 |
else:
|
121 |
sample = np.linspace(0, v_len - 1, n_frames).astype(int)
|
|
|
|
|
122 |
frames = []
|
123 |
for j in range(v_len):
|
124 |
success = v_cap.grab()
|
125 |
if j in sample:
|
|
|
126 |
success, frame = v_cap.retrieve()
|
127 |
if not success:
|
128 |
continue
|
@@ -132,22 +157,28 @@ def preprocess_video(input_video, n_frames = 3):
|
|
132 |
v_cap.release()
|
133 |
return frames
|
134 |
|
|
|
135 |
def df_video_pred(input_video):
|
136 |
video_frames = preprocess_video(input_video)
|
137 |
real_faces_list = []
|
138 |
fake_faces_list = []
|
|
|
139 |
for face in video_frames:
|
140 |
img_grads = img_model.forward(face)
|
141 |
img_grads = img_grads.cpu().detach().numpy()
|
142 |
img_grads_np = np.squeeze(img_grads)
|
143 |
real_faces_list.append(img_grads_np[0])
|
144 |
fake_faces_list.append(img_grads_np[1])
|
|
|
145 |
real_faces_mean = np.mean(real_faces_list)
|
146 |
fake_faces_mean = np.mean(fake_faces_list)
|
|
|
147 |
if real_faces_mean > 0.5:
|
148 |
preds = round(real_faces_mean * 100, 3)
|
149 |
text2 = f"The video is REAL. \nConfidence score is: {preds}%"
|
|
|
150 |
else:
|
151 |
preds = round(fake_faces_mean * 100, 3)
|
152 |
text2 = f"The video is FAKE. \nConfidence score is: {preds}%"
|
|
|
153 |
return text2
|
|
|
1 |
+
import os
|
2 |
import cv2
|
3 |
import onnx
|
4 |
import torch
|
5 |
import argparse
|
6 |
import numpy as np
|
7 |
+
import torch.nn as nn
|
8 |
+
from models.TMC import ETMC
|
9 |
from models import image
|
10 |
|
|
|
11 |
from onnx2pytorch import ConvertModel
|
12 |
|
13 |
+
onnx_model = onnx.load('checkpoints/efficientnet.onnx')
|
14 |
+
pytorch_model = ConvertModel(onnx_model)
|
15 |
+
|
|
|
|
|
16 |
torch.manual_seed(42)
|
17 |
|
18 |
|
19 |
+
audio_args = {
|
20 |
+
'nb_samp': 64600,
|
21 |
+
'first_conv': 1024,
|
22 |
+
'in_channels': 1,
|
23 |
+
'filts': [20, [20, 20], [20, 128], [128, 128]],
|
24 |
+
'blocks': [2, 4],
|
25 |
+
'nb_fc_node': 1024,
|
26 |
+
'gru_node': 1024,
|
27 |
+
'nb_gru_layer': 3,
|
28 |
+
'nb_classes': 2
|
29 |
+
}
|
30 |
|
31 |
|
32 |
def get_args(parser):
|
|
|
65 |
|
66 |
def load_img_modality_model(args):
|
67 |
rgb_encoder = pytorch_model
|
68 |
+
ckpt = torch.load('checkpoints/model.pth', map_location = torch.device('cpu'))
|
69 |
rgb_encoder.load_state_dict(ckpt['rgb_encoder'], strict = True)
|
70 |
rgb_encoder.eval()
|
71 |
return rgb_encoder
|
72 |
|
73 |
def load_spec_modality_model(args):
|
74 |
spec_encoder = image.RawNet(args)
|
75 |
+
ckpt = torch.load('checkpoints/model.pth', map_location = torch.device('cpu'))
|
76 |
spec_encoder.load_state_dict(ckpt['spec_encoder'], strict = True)
|
77 |
spec_encoder.eval()
|
78 |
return spec_encoder
|
|
|
81 |
get_args(parser)
|
82 |
args, remaining_args = parser.parse_known_args()
|
83 |
assert remaining_args == [], remaining_args
|
84 |
+
|
85 |
spec_model = load_spec_modality_model(args)
|
86 |
+
|
87 |
img_model = load_img_modality_model(args)
|
88 |
|
89 |
+
|
90 |
def preprocess_img(face):
|
91 |
face = face / 255
|
92 |
face = cv2.resize(face, (256, 256))
|
|
|
103 |
spec_grads = spec_model.forward(audio)
|
104 |
spec_grads_inv = np.exp(spec_grads.cpu().detach().numpy().squeeze())
|
105 |
max_value = np.argmax(spec_grads_inv)
|
106 |
+
|
107 |
if max_value > 0.5:
|
108 |
preds = round(100 - (max_value*100), 3)
|
109 |
text2 = f"The audio is REAL."
|
110 |
+
|
111 |
else:
|
112 |
preds = round(max_value*100, 3)
|
113 |
text2 = f"The audio is FAKE."
|
114 |
+
|
115 |
return text2
|
116 |
|
117 |
def df_img_pred(input_image):
|
|
|
120 |
img_grads = img_model.forward(face)
|
121 |
img_grads = img_grads.cpu().detach().numpy()
|
122 |
img_grads_np = np.squeeze(img_grads)
|
123 |
+
|
124 |
if img_grads_np[0] > 0.5:
|
125 |
preds = round(img_grads_np[0] * 100, 3)
|
126 |
text2 = f"The image is REAL. \nConfidence score is: {preds}"
|
127 |
+
|
128 |
else:
|
129 |
preds = round(img_grads_np[1] * 100, 3)
|
130 |
text2 = f"The image is FAKE. \nConfidence score is: {preds}"
|
131 |
+
|
132 |
return text2
|
133 |
|
134 |
+
|
135 |
def preprocess_video(input_video, n_frames = 3):
|
136 |
v_cap = cv2.VideoCapture(input_video)
|
137 |
v_len = int(v_cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
138 |
+
|
139 |
+
# Pick 'n_frames' evenly spaced frames to sample
|
140 |
if n_frames is None:
|
141 |
sample = np.arange(0, v_len)
|
142 |
else:
|
143 |
sample = np.linspace(0, v_len - 1, n_frames).astype(int)
|
144 |
+
|
145 |
+
#Loop through frames.
|
146 |
frames = []
|
147 |
for j in range(v_len):
|
148 |
success = v_cap.grab()
|
149 |
if j in sample:
|
150 |
+
# Load frame
|
151 |
success, frame = v_cap.retrieve()
|
152 |
if not success:
|
153 |
continue
|
|
|
157 |
v_cap.release()
|
158 |
return frames
|
159 |
|
160 |
+
|
161 |
def df_video_pred(input_video):
|
162 |
video_frames = preprocess_video(input_video)
|
163 |
real_faces_list = []
|
164 |
fake_faces_list = []
|
165 |
+
|
166 |
for face in video_frames:
|
167 |
img_grads = img_model.forward(face)
|
168 |
img_grads = img_grads.cpu().detach().numpy()
|
169 |
img_grads_np = np.squeeze(img_grads)
|
170 |
real_faces_list.append(img_grads_np[0])
|
171 |
fake_faces_list.append(img_grads_np[1])
|
172 |
+
|
173 |
real_faces_mean = np.mean(real_faces_list)
|
174 |
fake_faces_mean = np.mean(fake_faces_list)
|
175 |
+
|
176 |
if real_faces_mean > 0.5:
|
177 |
preds = round(real_faces_mean * 100, 3)
|
178 |
text2 = f"The video is REAL. \nConfidence score is: {preds}%"
|
179 |
+
|
180 |
else:
|
181 |
preds = round(fake_faces_mean * 100, 3)
|
182 |
text2 = f"The video is FAKE. \nConfidence score is: {preds}%"
|
183 |
+
|
184 |
return text2
|