TypeError: 'NoneType' object is not subscriptable

[zsrabbani]

I have QCNN model( I used qkeras library) and when I setup the below configuration and I got an error.
I updated the last version of hls4ml.

Model:

Layer (type) Output Shape Param

rf_input (InputLayer) [(None, 1024, 2)] 0

q_conv1d (QConv1D) (None, 1024, 64) 640

q_batch_normalization (QBa (None, 1024, 64) 256
tchNormalization)

q_activation (QActivation) (None, 1024, 64) 0

max_pooling1d (MaxPooling1 (None, 512, 64) 0
D)

q_conv1d_1 (QConv1D) (None, 512, 32) 10240

q_batch_normalization_1 (Q (None, 512, 32) 128
BatchNormalization)

q_activation_1 (QActivatio (None, 512, 32) 0
n)

max_pooling1d_1 (MaxPoolin (None, 256, 32) 0
g1D)

q_conv1d_2 (QConv1D) (None, 256, 16) 2560

q_batch_normalization_2 (Q (None, 256, 16) 64
BatchNormalization)

q_activation_2 (QActivatio (None, 256, 16) 0
n)

max_pooling1d_2 (MaxPoolin (None, 128, 16) 0
g1D)

flatten (Flatten) (None, 2048) 0

q_dense (QDense) (None, 128) 262144

dropout (Dropout) (None, 128) 0

q_dense_1 (QDense) (None, 128) 16384

dropout_1 (Dropout) (None, 128) 0

q_dense_2 (QDense) (None, 7) 896

activation (Activation) (None, 7) 0

=================================================================
Total params: 293312 (1.12 MB)
Trainable params: 293088 (1.12 MB)
Non-trainable params: 224 (896.00 Byte)

Here is my hls4ml setup:
hls_config = hls4ml.utils.config_from_keras_model(model, granularity='name')
hls_config['Model']['ReuseFactor']=16
hls_config['Model']['Strategy']='Resources'

for Layer in hls_config['LayerName'].keys():
hls_config['LayerName'][Layer]['Strategy'] = 'Resources'
hls_config['LayerName'][Layer]['ReuseFactor'] = 16

hls_config['LayerName']['softmax']['exp_table_t'] = 'ap_fixed<16,6>'
hls_config['LayerName']['softmax']['inv_table_t'] = 'ap_fixed<16,6>'
hls_config['LayerName']['output_softmax']['Strategy'] = 'Stable'

cfg = hls4ml.converters.create_config(backend='Vivado')
cfg['IOType'] = 'io_stream'
cfg['HLSConfig'] = hls_config
cfg['KerasModel'] = model
cfg['OutputDir'] = 'CNN_16_6'
hls_model = hls4ml.converters.convert_from_keras_model(
model, hls_config=hls_config, output_dir='CNN_16_6', backend='VivadoAccelerator', board='zcu102')

hls_model.compile()

The Result:

[jmitrevs]

It it possible to put the model somewhere or paste a script to create a simple untrained model that has the same issue?

[zsrabbani]

rf_in = Input(shape=(1024, 2), name = 'rf_input')

x = QConv1D(64, 5, kernel_quantizer="quantized_bits(16,6)", padding='same', use_bias=False)(rf_in)
x = QBatchNormalization()(x)
x = QActivation("quantized_relu(16,6)")(x)
x = MaxPooling1D(2, strides = 2, padding='same') (x)

x = QConv1D(32, 5, kernel_quantizer="quantized_bits(16,6)", padding='same', use_bias=False)(x)
x = QBatchNormalization()(x)
x = QActivation("quantized_relu(16,6)")(x)
x = MaxPooling1D(2, strides = 2, padding='same') (x)

x = QConv1D(16, 5, kernel_quantizer="quantized_bits(16,6)", padding='same', use_bias=False)(x)
x = QBatchNormalization()(x)
x = QActivation("quantized_relu(16,6)")(x)
x = MaxPooling1D(2, strides=2, padding='same') (x)

x = Flatten()(x)

dense_1 = QDense(128, activation="quantized_relu(16,6)", use_bias=False)(x)
dropout_1 = Dropout(0.25)(dense_1)
dense_2 = QDense(128, activation="quantized_relu(16,6)", use_bias=False)(dropout_1)
dropout_2 = Dropout(0.5)(dense_2)
softmax = QDense(7, kernel_quantizer="quantized_bits(16,6)", use_bias=False)(dropout_2)
softmax = Activation('softmax')(softmax)

opt = keras.optimizers.Adam(learning_rate=0.0001)
model= keras.Model(rf_in, softmax)
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=["accuracy"])
model.summary()

[jmitrevs]

I think I understand the problem. The hls4ml software assumes that QDense will always have kernel_quantizer defined, but that is not the case here. I will add a check for it, but in the meantime, here is a workaround. Replace:

dense_1 = QDense(128, activation="quantized_relu(16,6)", use_bias=False)(x)

by

dense_1_noact = Dense(128, use_bias=False)(x)
dense_1 = QActivation(activation="quantized_relu(16,6)")(dense_1_noact)

[jmitrevs]

Try using https://github.com/fastmachinelearning/hls4ml/tree/weight_quantizer_none

[zsrabbani]

I think I understand the problem. The hls4ml software assumes that QDense will always have kernel_quantizer defined, but that is not the case here. I will add a check for it, but in the meantime, here is a workaround. Replace:

dense_1 = QDense(128, activation="quantized_relu(16,6)", use_bias=False)(x)

by

dense_1_noact = Dense(128, use_bias=False)(x)
dense_1 = QActivation(activation="quantized_relu(16,6)")(dense_1_noact)

I added the two code line to my code, but still the same error. nothing change.

dense_1 = Dense(128, use_bias=False)(x)
dense_1 = QActivation("quantized_relu(16,6)")(dense_1)
dropout_1 = Dropout(0.25)(dense_1)
dense_2 = Dense(128, use_bias=False)(dropout_1)
dense_2 = QActivation("quantized_relu(16,6)")(dense_2)
dropout_2 = Dropout(0.5)(dense_2)
softmax = Dense(7, use_bias=False)(dropout_2)
softmax = QActivation("quantized_relu(16,6)")(softmax)
output = Activation('softmax')(softmax)

[zsrabbani]

Even I installed hls4ml library again and still same error.

[johanjino]

+1 Same issue when using QKeras layers (QLSTM)