保存一个Tensorflow模型并将其加载到Tensorflow.js中

Question

In my regular python code, I have a CNN implemented. I save it using model.save and produce some four files (checkpoint, meta, index, and some other file). However, I cannot load these four files directly to tensorflow.js. Here is the sample CNN:

import tflearn
from tflearn.layers.conv import conv_2d, max_pool_2d
from tflearn.layers.core import input_data, dropout, fully_connected
from tflearn.layers.estimator import regression

convnet = input_data(shape=[None, IMG_SIZE, IMG_SIZE, 1], name='input')

convnet = conv_2d(convnet, FIRST_NUM_CHANNEL, FILTER_SIZE, activation='relu')
convnet = max_pool_2d(convnet, 2)

convnet = conv_2d(convnet, FIRST_NUM_CHANNEL*2, FILTER_SIZE, activation='relu')
convnet = max_pool_2d(convnet, 2)

convnet = conv_2d(convnet, FIRST_NUM_CHANNEL*4, FILTER_SIZE, activation='relu')
convnet = max_pool_2d(convnet, 2)

convnet = fully_connected(convnet, FIRST_NUM_CHANNEL*8, activation='relu')
convnet = dropout(convnet, 0.7)

convnet = fully_connected(convnet, NUM_OUTPUT, activation='softmax')
convnet = regression(convnet, optimizer='adam', learning_rate=LR, loss='categorical_crossentropy', name='targets')

model = tflearn.DNN(convnet, tensorboard_dir='log')

train = train_data[:7000]
test = train_data[-1000:]


X = np.array([i[0] for i in train]).reshape(-1,IMG_SIZE,IMG_SIZE,1)
Y = [i[1] for i in train]

test_x = np.array([i[0] for i in test]).reshape(-1,IMG_SIZE,IMG_SIZE,1)
test_y = [i[1] for i in test]


model.fit({'input': X}, {'targets': Y}, n_epoch=NUM_EPOCHS, validation_set=({'input': test_x}, {'targets': test_y}), 
    snapshot_step=500, show_metric=True, run_id=MODEL_NAME)

model.save(MODEL_NAME)
print('MODEL SAVED:', MODEL_NAME)

The last two lines of this code snippet is for saving the model. I can just load the model in a flask app but I want to port it to tensorflow.js. Can anyone give me a tutorial on how to do this?

Answer 1

tensorflowjs_converted outputs among other files a weight file weights_manifest.json and a model topology file tensorflowjs_model.pb . To load this model in tensorflow.js, follow the steps outllined below.

• serve the folder containing the files using a server

// cd to the directory containing the files

// then launch the python server
python3 -m http-server

// or install and launch npm module http-server
npm install -g http-server
http-server --cors -c1 .

• create a js script to load the model

(async () => {
   ...
   const model = await tf.loadFrozenModel('http://localhost:8080/tensorflowjs_model.pb', 'http://localhost:8080/weights_manifest.json')
 })()

---

There is a difference between loadModel and loadFrozenModel .

• loadModel is for loading model that were saved locally. The model could either be retrieved from the indexDB of the browser or from the localStorage. Maybe it can be used to retrieve a model saved by another tensorflow API different than Js, but the user will be required to select the file by using tf.io.browserFiles ( I haven't tried it)

• loadFrozenModel is for loading model that are served by a server