Dear All,
I have an imagery with a single fish species within each image along with a list of morphometric measurements of the fish (length, width, length of tail, etc). I would like to train a CNN model that will predict these measurements having as input only the images. Any ideas what kind of architecture is ideal for this task? I read about multioutput learning, but I haven't found a practical implementation in Python.
Thank you for your time.
For multioutput implementation you have two approaches:
1.- Train a model for each output. This is the most common approach and it lets you known the performance of the model for each output (at least for TF 2.10, I have tested it in TF 2.12 and I only got the sum loss. Well, haven't tried to change it so far).
2.- Train a single model for all outputs. This is valid only if the outputs are related. I don't know the details of your idea, but I think this approach is not valid for your case.
Since you will aim different fish metrics with this model, the best approach is to train a model for each metric. Also, standarize the y values, this will help the model to converge faster.
The implementation in Python can be done with Tensorflow/Keras. For example, a simple CNN model with the first approach (sorry, but this site is not ready for posting code):
def build_model(input_1, input_2, input_3):
cnn_input = Input(shape=(input_1, input_2, input_3), name='CNN Input')
cnn1 = Conv2D(filters=32, kernel_size=(3,3), activation='relu')(cnn_input)
cnn1 = BatchNormalization()(cnn1)
cnn1 = MaxPooling2D(pool_size=2)(cnn1)
cnn2 = Conv1D(filters=16, kernel_size=(3,3), activation='relu')(cnn1)
cnn2 = BatchNormalization()(cnn2)
cnn2 = MaxPooling2D(pool_size=2)(cnn2)
ft = Flatten()(cnn2)
dns1 = Dense(64, activation='relu')(ft)
dns2 = Dense(32, activation='relu')(dns1)
dns3 = Dense(32, activation='relu')(dns2)
output1 = Dense(1, name='length', activation = 'relu')(dns3)
output2 = Dense(1, name='width', activation = 'relu')(dns3)
output3 = Dense(1, name='taillength', activation = 'relu')(dns3)
model = Model(inputs=cnn_input, outputs=[output1, output2, output3])
optimizer = Adam(0.0001)
model.compile(optimizer=optimizer, loss={'length': 'mae', 'width': 'mae', 'taillength': 'mae'})
return model
Then you can call build_model using height, width and channel lenght of your input images.
The training with be something like model.fit(X, [y_0, y_1, y_2], epochs = your choice, batch_size = your choice, validation_data = ([Xval], [yval_0, yval_1, yval_2]))
Thank you Aldo for your suggestion. I can see the general framework.
Cheers!
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA