Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,61 +1,118 @@
|
|
| 1 |
-
import
|
| 2 |
-
from tensorflow.keras import layers, models, regularizers
|
| 3 |
import json
|
| 4 |
-
from PIL import Image
|
| 5 |
import numpy as np
|
| 6 |
-
import
|
|
|
|
|
|
|
|
|
|
| 7 |
import gradio as gr
|
| 8 |
|
| 9 |
-
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
| 11 |
if os.path.exists("config.json"):
|
| 12 |
with open("config.json") as f:
|
| 13 |
cfg = json.load(f)
|
| 14 |
|
| 15 |
-
#
|
| 16 |
-
|
|
|
|
|
|
|
|
|
|
| 17 |
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
base_model
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
# ----
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
def predict(image: Image.Image):
|
|
|
|
|
|
|
|
|
|
| 52 |
x = preprocess(image)
|
| 53 |
-
preds = model.predict(x)
|
| 54 |
-
class_idx = int(np.argmax(preds
|
| 55 |
-
confidence = float(
|
| 56 |
return class_names[class_idx], confidence
|
| 57 |
|
| 58 |
-
|
|
|
|
|
|
|
|
|
|
| 59 |
interface = gr.Interface(
|
| 60 |
fn=predict,
|
| 61 |
inputs=gr.Image(type="pil", label="Upload an image"),
|
|
@@ -64,9 +121,9 @@ interface = gr.Interface(
|
|
| 64 |
gr.Number(label="Confidence")
|
| 65 |
],
|
| 66 |
title="NeuRecAI - Material Classifier",
|
| 67 |
-
description="Upload an image to classify it as battery, glass, metal, organic,
|
|
|
|
| 68 |
)
|
| 69 |
|
| 70 |
-
# ---- Launch ----
|
| 71 |
if __name__ == "__main__":
|
| 72 |
interface.launch()
|
|
|
|
| 1 |
+
import os
|
|
|
|
| 2 |
import json
|
|
|
|
| 3 |
import numpy as np
|
| 4 |
+
from PIL import Image
|
| 5 |
+
import tensorflow as tf
|
| 6 |
+
from tensorflow.keras import layers, models, regularizers
|
| 7 |
+
from tensorflow.keras.applications.efficientnet import preprocess_input
|
| 8 |
import gradio as gr
|
| 9 |
|
| 10 |
+
|
| 11 |
+
# -----------------------
|
| 12 |
+
# Config and class labels
|
| 13 |
+
# -----------------------
|
| 14 |
+
cfg = {}
|
| 15 |
if os.path.exists("config.json"):
|
| 16 |
with open("config.json") as f:
|
| 17 |
cfg = json.load(f)
|
| 18 |
|
| 19 |
+
# Default labels; override via config.json if provided
|
| 20 |
+
default_class_names = ['battery', 'glass', 'metal', 'organic', 'paper', 'plastic']
|
| 21 |
+
class_names = cfg.get("class_names", default_class_names)
|
| 22 |
+
num_classes = int(cfg.get("num_classes", len(class_names)))
|
| 23 |
+
assert num_classes == len(class_names), "num_classes must equal len(class_names)"
|
| 24 |
|
| 25 |
+
|
| 26 |
+
# -----------------------
|
| 27 |
+
# Model construction
|
| 28 |
+
# -----------------------
|
| 29 |
+
def build_model(num_classes: int) -> tf.keras.Model:
|
| 30 |
+
"""
|
| 31 |
+
Rebuild the exact training head to load .h5 weights reliably.
|
| 32 |
+
Uses EfficientNetB7 backbone with a small dense head.
|
| 33 |
+
"""
|
| 34 |
+
base_model = tf.keras.applications.EfficientNetB7(
|
| 35 |
+
weights="imagenet", # keep as used during training
|
| 36 |
+
include_top=False,
|
| 37 |
+
input_shape=(224, 224, 3)
|
| 38 |
+
)
|
| 39 |
+
base_model.trainable = False
|
| 40 |
+
|
| 41 |
+
model = models.Sequential([
|
| 42 |
+
base_model,
|
| 43 |
+
layers.GlobalAveragePooling2D(),
|
| 44 |
+
layers.Dense(256, activation="relu",
|
| 45 |
+
kernel_regularizer=regularizers.l2(0.001)),
|
| 46 |
+
layers.Dropout(0.3),
|
| 47 |
+
layers.Dense(128, activation="relu",
|
| 48 |
+
kernel_regularizer=regularizers.l2(0.001)),
|
| 49 |
+
layers.Dropout(0.3),
|
| 50 |
+
layers.Dense(num_classes, activation="softmax")
|
| 51 |
+
])
|
| 52 |
+
return model
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
# -----------------------
|
| 56 |
+
# Load model or weights
|
| 57 |
+
# -----------------------
|
| 58 |
+
model = None
|
| 59 |
+
_loaded_full = False
|
| 60 |
+
|
| 61 |
+
# Prefer a full saved model if available
|
| 62 |
+
for fname in ("model.keras", "model.h5"):
|
| 63 |
+
if os.path.exists(fname):
|
| 64 |
+
model = tf.keras.models.load_model(fname, compile=False)
|
| 65 |
+
_loaded_full = True
|
| 66 |
+
print(f"Loaded full model from {fname}")
|
| 67 |
+
break
|
| 68 |
+
|
| 69 |
+
if model is None:
|
| 70 |
+
# Fallback: rebuild and load weights
|
| 71 |
+
model = build_model(num_classes)
|
| 72 |
+
weights_path = "model_weights.h5" if os.path.exists("model_weights.h5") else "model.weights.h5"
|
| 73 |
+
if os.path.exists(weights_path):
|
| 74 |
+
model.load_weights(weights_path)
|
| 75 |
+
print(f"Weights loaded from {weights_path}.")
|
| 76 |
+
else:
|
| 77 |
+
print("No model file found: expected model.keras/model.h5 or model_weights.h5/model.weights.h5.")
|
| 78 |
+
print("Current directory:", os.listdir())
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
# -----------------------
|
| 82 |
+
# Preprocessing
|
| 83 |
+
# -----------------------
|
| 84 |
+
def _to_rgb(img: Image.Image) -> Image.Image:
|
| 85 |
+
"""Force 3-channel RGB to avoid grayscale/RGBA issues."""
|
| 86 |
+
return img.convert("RGB")
|
| 87 |
+
|
| 88 |
+
def preprocess(image: Image.Image) -> np.ndarray:
|
| 89 |
+
"""
|
| 90 |
+
Resize to 224, EfficientNet preprocess, add batch dimension.
|
| 91 |
+
"""
|
| 92 |
+
image = _to_rgb(image).resize((224, 224), Image.BILINEAR)
|
| 93 |
+
x = np.asarray(image, dtype=np.float32)
|
| 94 |
+
x = preprocess_input(x) # correct scaling for EfficientNet
|
| 95 |
+
x = np.expand_dims(x, axis=0) # (1, 224, 224, 3)
|
| 96 |
+
return x
|
| 97 |
+
|
| 98 |
+
|
| 99 |
+
# -----------------------
|
| 100 |
+
# Inference
|
| 101 |
+
# -----------------------
|
| 102 |
def predict(image: Image.Image):
|
| 103 |
+
"""
|
| 104 |
+
Returns (label, confidence) for Gradio.
|
| 105 |
+
"""
|
| 106 |
x = preprocess(image)
|
| 107 |
+
preds = model.predict(x, verbose=0)[0] # shape: (num_classes,)
|
| 108 |
+
class_idx = int(np.argmax(preds))
|
| 109 |
+
confidence = float(preds[class_idx])
|
| 110 |
return class_names[class_idx], confidence
|
| 111 |
|
| 112 |
+
|
| 113 |
+
# -----------------------
|
| 114 |
+
# Gradio interface
|
| 115 |
+
# -----------------------
|
| 116 |
interface = gr.Interface(
|
| 117 |
fn=predict,
|
| 118 |
inputs=gr.Image(type="pil", label="Upload an image"),
|
|
|
|
| 121 |
gr.Number(label="Confidence")
|
| 122 |
],
|
| 123 |
title="NeuRecAI - Material Classifier",
|
| 124 |
+
description=("Upload an image to classify it as battery, glass, metal, organic, "
|
| 125 |
+
"paper, or plastic.")
|
| 126 |
)
|
| 127 |
|
|
|
|
| 128 |
if __name__ == "__main__":
|
| 129 |
interface.launch()
|