如何获取TensorFlow中MobileNet V1的最后全连接/特征层以输入LSTM

Hey there! Let's figure out which layer to pick for your LSTM input—since you've already fine-tuned MobileNet V1 for 4-class classification, you're just one step away from getting the right feature representation.

Key Layers to Consider #

MobileNet V1 has two ideal candidates for feature extraction, depending on what kind of features you need for your LSTM:

1. Global Average Pooling Layer (MobilenetV1/Logits/AvgPool_1a/AvgPool)

   • This is the go-to choice if you want a compact, fixed-dimension global feature vector. It takes the spatial output from the last convolutional block and averages it down to a 1024-dimensional vector (default for MobileNet V1, regardless of input image size).
   • Perfect for sequence tasks where each frame/image needs to be represented as a single vector—you can stack these vectors into a sequence tensor (shape: (batch_size, sequence_length, 1024)) which LSTMs handle seamlessly.

2. Final Convolutional Block Output (MobilenetV1/Conv2d_13_pointwise/Relu6)

   • If you need richer spatial features (e.g., to model spatial patterns in your sequence), this layer is your best bet. It outputs a 7x7x1024 feature map (for 224x224 input) that retains spatial information.
   • You can reshape this into a sequence by flattening the spatial dimensions (e.g., (batch_size, 49, 1024) where 49 = 7*7) to feed into your LSTM, giving it access to fine-grained spatial details.

Avoid These Layers #

Skip the final classification layers like MobilenetV1/Logits/Conv2d_1c_1x1 (your retrained 4-class layer) or the Softmax output—these are already tailored to your 4-class task and won't provide the general-purpose features an LSTM needs.

How to Extract the Features #

Here's a quick code snippet to grab your desired layer's output:

import tensorflow as tf

# Load your retrained MobileNet V1 model
model = tf.keras.models.load_model('your_retrained_model.h5')

# First, confirm the layer names (you mentioned you already did this, but just to double-check)
for layer in model.layers:
    print(layer.name)

# Create a feature extractor for the global average pooling layer
feature_extractor = tf.keras.Model(
    inputs=model.input,
    outputs=model.get_layer('MobilenetV1/Logits/AvgPool_1a/AvgPool').output
)

# Or, if you want spatial features, use the final convolutional block:
# feature_extractor = tf.keras.Model(
#     inputs=model.input,
#     outputs=model.get_layer('MobilenetV1/Conv2d_13_pointwise/Relu6').output
# )

# Test with a sample input
sample_input = tf.random.normal([1, 224, 224, 3])  # Match your input size
features = feature_extractor(sample_input)
print(f"Feature shape: {features.shape}")
# For global pooling: (1, 1024)
# For conv block: (1, 7, 7, 1024)

Quick Tip #

If your LSTM expects a sequence of fixed-length vectors, the global pooling layer's output is the easiest to work with—no extra reshaping needed beyond stacking frames into a sequence. If you need spatial context, reshaping the convolutional block's output into (sequence_length, feature_dim) will do the trick.

内容的提问来源于stack exchange，提问作者Pototo