osamajan90 commited on
Commit
32e2c3e
·
1 Parent(s): 65535d9

final fix

Browse files
Files changed (1) hide show
  1. app.py +37 -27
app.py CHANGED
@@ -5,7 +5,6 @@ from PIL import Image
5
  import tensorflow as tf
6
  import torch
7
  import torch.nn as nn
8
- import torchvision.models as models
9
 
10
 
11
  # =========================
@@ -19,20 +18,37 @@ keras_model = tf.keras.models.load_model(
19
 
20
 
21
  # =========================
22
- # 2. BUILD + LOAD PYTORCH MODEL (FIXED)
23
  # =========================
24
- def load_pytorch_model():
25
- model = models.densenet121(pretrained=False)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
26
 
27
- # MNIST = 10 classes
28
- model.classifier = nn.Linear(model.classifier.in_features, 10)
 
29
 
30
  state_dict = torch.load(
31
  "best_densenet_mnist.pth",
32
  map_location="cpu"
33
  )
34
 
35
- model.load_state_dict(state_dict)
36
  model.eval()
37
  return model
38
 
@@ -41,42 +57,37 @@ pytorch_model = load_pytorch_model()
41
 
42
 
43
  # =========================
44
- # 3. PREPROCESS IMAGE
45
  # =========================
46
- def preprocess(image):
47
  image = image.resize((32, 32)).convert("RGB")
 
48
  img = np.array(image).astype(np.float32) / 255.0
49
  img = np.expand_dims(img, axis=0)
50
- return img
51
 
52
-
53
- # =========================
54
- # 4. KERAS PREDICTION
55
- # =========================
56
- def predict_keras(image):
57
- img = preprocess(image)
58
- preds = keras_model.predict(img)
59
  return int(np.argmax(preds))
60
 
61
 
62
  # =========================
63
- # 5. PYTORCH PREDICTION
64
  # =========================
65
  def predict_pytorch(image):
66
- image = image.resize((32, 32)).convert("RGB")
67
 
68
  img = np.array(image).astype(np.float32) / 255.0
69
- img = np.transpose(img, (2, 0, 1)) # CHW
70
- img = torch.tensor(img).unsqueeze(0)
 
 
71
 
72
  with torch.no_grad():
73
  outputs = pytorch_model(img)
74
- preds = torch.softmax(outputs, dim=1)
75
- return int(torch.argmax(preds))
76
 
77
 
78
  # =========================
79
- # 6. ROUTER FUNCTION
80
  # =========================
81
  def predict(image, model_name):
82
  if model_name == "MobileNetV2 (Keras)":
@@ -88,7 +99,7 @@ def predict(image, model_name):
88
 
89
 
90
  # =========================
91
- # 7. GRADIO UI
92
  # =========================
93
  demo = gr.Interface(
94
  fn=predict,
@@ -100,8 +111,7 @@ demo = gr.Interface(
100
  )
101
  ],
102
  outputs="text",
103
- title="Multi-Model MNIST Classifier",
104
- description="Choose model and upload digit image (0–9)"
105
  )
106
 
107
  demo.launch()
 
5
  import tensorflow as tf
6
  import torch
7
  import torch.nn as nn
 
8
 
9
 
10
  # =========================
 
18
 
19
 
20
  # =========================
21
+ # 2. CUSTOM PYTORCH MODEL (FIXED FOR MNIST 1-CHANNEL)
22
  # =========================
23
+ class MNISTDenseNet(nn.Module):
24
+ def __init__(self):
25
+ super().__init__()
26
+
27
+ # simple conv backbone matching your checkpoint expectation
28
+ self.conv = nn.Sequential(
29
+ nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3),
30
+ nn.ReLU(),
31
+ nn.AdaptiveAvgPool2d((1, 1))
32
+ )
33
+
34
+ self.fc = nn.Linear(64, 10)
35
+
36
+ def forward(self, x):
37
+ x = self.conv(x)
38
+ x = x.view(x.size(0), -1)
39
+ x = self.fc(x)
40
+ return x
41
 
42
+
43
+ def load_pytorch_model():
44
+ model = MNISTDenseNet()
45
 
46
  state_dict = torch.load(
47
  "best_densenet_mnist.pth",
48
  map_location="cpu"
49
  )
50
 
51
+ model.load_state_dict(state_dict, strict=True)
52
  model.eval()
53
  return model
54
 
 
57
 
58
 
59
  # =========================
60
+ # 3. KERAS PREDICTION
61
  # =========================
62
+ def predict_keras(image):
63
  image = image.resize((32, 32)).convert("RGB")
64
+
65
  img = np.array(image).astype(np.float32) / 255.0
66
  img = np.expand_dims(img, axis=0)
 
67
 
68
+ preds = keras_model.predict(img, verbose=0)
 
 
 
 
 
 
69
  return int(np.argmax(preds))
70
 
71
 
72
  # =========================
73
+ # 4. PYTORCH PREDICTION
74
  # =========================
75
  def predict_pytorch(image):
76
+ image = image.resize((32, 32)).convert("L")
77
 
78
  img = np.array(image).astype(np.float32) / 255.0
79
+ img = np.expand_dims(img, axis=0)
80
+ img = np.expand_dims(img, axis=0)
81
+
82
+ img = torch.tensor(img)
83
 
84
  with torch.no_grad():
85
  outputs = pytorch_model(img)
86
+ return int(torch.argmax(outputs))
 
87
 
88
 
89
  # =========================
90
+ # 5. ROUTER
91
  # =========================
92
  def predict(image, model_name):
93
  if model_name == "MobileNetV2 (Keras)":
 
99
 
100
 
101
  # =========================
102
+ # 6. GRADIO UI
103
  # =========================
104
  demo = gr.Interface(
105
  fn=predict,
 
111
  )
112
  ],
113
  outputs="text",
114
+ title="Multi-Model MNIST Classifier"
 
115
  )
116
 
117
  demo.launch()