how to freeze some layers when fine tune resnet50

Question

I am trying to fine tune resnet 50 with keras. When I freeze all the layers in resnet50, everything works OK. However, I want to freeze some layers of resnet50, not all of them. But when I do this, I get some errors. Here is my code:

base_model = ResNet50(include_top=False, weights="imagenet", input_shape=(input_size, input_size, input_channels))
model = Sequential()
model.add(base_model)
model.add(Flatten())
model.add(Dense(80, activation="softmax"))

#this is where the error happens. The commented code works fine
"""
for layer in base_model.layers:
    layer.trainable = False
"""
for layer in base_model.layers[:-26]:
    layer.trainable = False
model.summary()
optimizer = Adam(lr=1e-4)
model.compile(loss="categorical_crossentropy", optimizer=optimizer, metrics=["accuracy"])

callbacks = [
    EarlyStopping(monitor='val_loss', patience=4, verbose=1, min_delta=1e-4),
    ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=2, cooldown=2, verbose=1),
    ModelCheckpoint(filepath='weights/renet50_best_weight.fold_' + str(fold_count) + '.hdf5', save_best_only=True,
                    save_weights_only=True)
    ]

model.load_weights(filepath="weights/renet50_best_weight.fold_1.hdf5")
model.fit_generator(generator=train_generator(), steps_per_epoch=len(df_train) // batch_size,  epochs=epochs, verbose=1,
                  callbacks=callbacks, validation_data=valid_generator(), validation_steps = len(df_valid) // batch_size) 

the error is as follows:

Traceback (most recent call last):
File "/home/jamesben/ai_challenger/src/train.py", line 184, in <module> model.load_weights(filepath="weights/renet50_best_weight.fold_" + str(fold_count) + '.hdf5')
File "/usr/local/lib/python3.5/dist-packages/keras/models.py", line 719, in load_weights topology.load_weights_from_hdf5_group(f, layers)
File "/usr/local/lib/python3.5/dist-packages/keras/engine/topology.py", line 3095, in load_weights_from_hdf5_group K.batch_set_value(weight_value_tuples)
File "/usr/local/lib/python3.5/dist-packages/keras/backend/tensorflow_backend.py", line 2193, in batch_set_value get_session().run(assign_ops, feed_dict=feed_dict)
File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py", line 767, in run run_metadata_ptr)
File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py", line 944, in _run % (np_val.shape, subfeed_t.name, str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (128,) for Tensor 'Placeholder_72:0', which has shape '(3, 3, 128, 128)'

can anyone give me some help on how many layer should I freeze with resnet50?

Shardashare

I loss one line between the callbacks and model.fit_generator model.load_weights(filepath="weights/renet50_best_weight.fold_1.hdf5') – Shardashare

This error is not about freezing layers. It's about trying to fit with data that don't have the same shape as your model inputs. – Unkindly

however, when I freeze all the layers of the resnet50, there will not be any errors, does this means the input shape is OK? – Shardashare

Something else in the code changed. – Unkindly

Thanks for your answer. Does this means if I want to do fine-tuning on resnet, I have to frezze all the layers in resnet? – Shardashare

No. freezing layers should not change anything. The problem is very probably in another place. – Unkindly

if I change the freezing code as follows, everything works fine. for layer in base_model.layers: layer.trainable = False – Shardashare

Answer 1

When using load_weights() and save_weights() with a nested model, it's very easy to get an error if the trainable settings are not the same.

To solve the error, make sure you freeze the same layers before calling model.load_weights(). That is, if the weight file is saved with all layers frozen, the procedure will be:

1. Recreate the model
2. Freeze all layers in base_model
3. Load the weights
4. Unfreeze those layers you want to train (in this case, base_model.layers[-26:])

For example,

base_model = ResNet50(include_top=False, input_shape=(224, 224, 3))
model = Sequential()
model.add(base_model)
model.add(Flatten())
model.add(Dense(80, activation="softmax"))

for layer in base_model.layers:
    layer.trainable = False
model.load_weights('all_layers_freezed.h5')

for layer in base_model.layers[-26:]:
    layer.trainable = True

---

The underlying reason:

When you call model.load_weights(), (roughly) the weight for each layer is loaded by the following steps (in the function load_weights_from_hdf5_group() in topology.py):

1. Call layer.weights to get the weight tensors
2. Match each weight tensor with its corresponding weight value in the hdf5 file
3. Call K.batch_set_value() to assign the weight values to the weight tensors

If your model is a nested model, you have to be careful about trainable because of Step 1.

I'll use an example to explain it. For the same model as above, model.summary() gives:

_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
resnet50 (Model)             (None, 1, 1, 2048)        23587712
_________________________________________________________________
flatten_10 (Flatten)         (None, 2048)              0
_________________________________________________________________
dense_5 (Dense)              (None, 80)                163920
=================================================================
Total params: 23,751,632
Trainable params: 11,202,640
Non-trainable params: 12,548,992
_________________________________________________________________

The inner ResNet50 model is treated as a layer of model during weight loading. When loading the layer resnet50, in Step 1, calling layer.weights is equivalent to calling base_model.weights. The list of weight tensors for all layers in the ResNet50 model will be collected and returned.

Now the problem is that, when constructing the list of weight tensors, trainable weights will come before non-trainable weights. In the definition of Layer class:

@property
def weights(self):
    return self.trainable_weights + self.non_trainable_weights

If all layers in base_model are frozen, the weight tensors will be in the following order:

for layer in base_model.layers:
    layer.trainable = False
print(base_model.weights)

[<tf.Variable 'conv1/kernel:0' shape=(7, 7, 3, 64) dtype=float32_ref>,
 <tf.Variable 'conv1/bias:0' shape=(64,) dtype=float32_ref>,
 <tf.Variable 'bn_conv1/gamma:0' shape=(64,) dtype=float32_ref>,
 <tf.Variable 'bn_conv1/beta:0' shape=(64,) dtype=float32_ref>,
 <tf.Variable 'bn_conv1/moving_mean:0' shape=(64,) dtype=float32_ref>,
 <tf.Variable 'bn_conv1/moving_variance:0' shape=(64,) dtype=float32_ref>,
 <tf.Variable 'res2a_branch2a/kernel:0' shape=(1, 1, 64, 64) dtype=float32_ref>,
 <tf.Variable 'res2a_branch2a/bias:0' shape=(64,) dtype=float32_ref>,
 ...
 <tf.Variable 'res5c_branch2c/kernel:0' shape=(1, 1, 512, 2048) dtype=float32_ref>,
 <tf.Variable 'res5c_branch2c/bias:0' shape=(2048,) dtype=float32_ref>,
 <tf.Variable 'bn5c_branch2c/gamma:0' shape=(2048,) dtype=float32_ref>,
 <tf.Variable 'bn5c_branch2c/beta:0' shape=(2048,) dtype=float32_ref>,
 <tf.Variable 'bn5c_branch2c/moving_mean:0' shape=(2048,) dtype=float32_ref>,
 <tf.Variable 'bn5c_branch2c/moving_variance:0' shape=(2048,) dtype=float32_ref>]

However, if some layers are trainable, the weight tensors of the trainable layers will come before that of the frozen ones:

for layer in base_model.layers[-5:]:
    layer.trainable = True
print(base_model.weights)

[<tf.Variable 'res5c_branch2c/kernel:0' shape=(1, 1, 512, 2048) dtype=float32_ref>,
 <tf.Variable 'res5c_branch2c/bias:0' shape=(2048,) dtype=float32_ref>,
 <tf.Variable 'bn5c_branch2c/gamma:0' shape=(2048,) dtype=float32_ref>,
 <tf.Variable 'bn5c_branch2c/beta:0' shape=(2048,) dtype=float32_ref>,
 <tf.Variable 'conv1/kernel:0' shape=(7, 7, 3, 64) dtype=float32_ref>,
 <tf.Variable 'conv1/bias:0' shape=(64,) dtype=float32_ref>,
 <tf.Variable 'bn_conv1/gamma:0' shape=(64,) dtype=float32_ref>,
 <tf.Variable 'bn_conv1/beta:0' shape=(64,) dtype=float32_ref>,
 <tf.Variable 'bn_conv1/moving_mean:0' shape=(64,) dtype=float32_ref>,
 <tf.Variable 'bn_conv1/moving_variance:0' shape=(64,) dtype=float32_ref>,
 <tf.Variable 'res2a_branch2a/kernel:0' shape=(1, 1, 64, 64) dtype=float32_ref>,
 <tf.Variable 'res2a_branch2a/bias:0' shape=(64,) dtype=float32_ref>,
 ...
 <tf.Variable 'bn5c_branch2b/moving_mean:0' shape=(512,) dtype=float32_ref>,
 <tf.Variable 'bn5c_branch2b/moving_variance:0' shape=(512,) dtype=float32_ref>,
 <tf.Variable 'bn5c_branch2c/moving_mean:0' shape=(2048,) dtype=float32_ref>,
 <tf.Variable 'bn5c_branch2c/moving_variance:0' shape=(2048,) dtype=float32_ref>]

The change in order is why you got an error about tensor shapes. The weight values saved in the hdf5 file are matched to the wrong weight tensors in Step 2 mentioned above. The reason that everything works fine when you freeze all layers is because your model checkpoint is saved also with all layers frozen and thus the order is correct.

---

Possibly better solution:

You can avoid a nested model by using the functional API. For example, the following code should work without error:

base_model = ResNet50(include_top=False, weights="imagenet", input_shape=(input_size, input_size, input_channels))
x = Flatten()(base_model.output)
x = Dense(80, activation="softmax")(x)
model = Model(base_model.input, x)

for layer in base_model.layers:
    layer.trainable = False
model.save_weights("all_nontrainable.h5")

base_model = ResNet50(include_top=False, weights="imagenet", input_shape=(input_size, input_size, input_channels))
x = Flatten()(base_model.output)
x = Dense(80, activation="softmax")(x)
model = Model(base_model.input, x)

for layer in base_model.layers[:-26]:
    layer.trainable = False
model.load_weights("all_nontrainable.h5")

Madelene