Skip to content
Snippets Groups Projects
Unverified Commit 60ac1c6c authored by Jan Buethe's avatar Jan Buethe
Browse files

prepared quantization implementation for DRED

parent 2192e85b
No related branches found
No related tags found
No related merge requests found
Pipeline #4148 failed
Stage: build
Stage: test
Hide whitespace changes
Inline Side-by-side
dnn/torch/rdovae/export_rdovae_weights.py
+ 34
− 34
View file @ 60ac1c6c
......@@ -115,75 +115,75 @@ f"""
# encoder
encoder_dense_layers = [
('core_encoder.module.dense_1' , 'enc_dense1', 'TANH'),
('core_encoder.module.z_dense' , 'enc_zdense', 'LINEAR'),
('core_encoder.module.state_dense_1' , 'gdense1' , 'TANH'),
('core_encoder.module.state_dense_2' , 'gdense2' , 'TANH')
('core_encoder.module.dense_1' , 'enc_dense1', 'TANH', False,),
('core_encoder.module.z_dense' , 'enc_zdense', 'LINEAR', False,),
('core_encoder.module.state_dense_1' , 'gdense1' , 'TANH', False,),
('core_encoder.module.state_dense_2' , 'gdense2' , 'TANH', False)
]
for name, export_name, _ in encoder_dense_layers:
for name, export_name, _, _ in encoder_dense_layers:
layer = model.get_submodule(name)
dump_torch_weights(enc_writer, layer, name=export_name, verbose=True)
encoder_gru_layers = [
('core_encoder.module.gru1' , 'enc_gru1', 'TANH'),
('core_encoder.module.gru2' , 'enc_gru2', 'TANH'),
('core_encoder.module.gru3' , 'enc_gru3', 'TANH'),
('core_encoder.module.gru4' , 'enc_gru4', 'TANH'),
('core_encoder.module.gru5' , 'enc_gru5', 'TANH'),
('core_encoder.module.gru1' , 'enc_gru1', 'TANH', False),
('core_encoder.module.gru2' , 'enc_gru2', 'TANH', False),
('core_encoder.module.gru3' , 'enc_gru3', 'TANH', False),
('core_encoder.module.gru4' , 'enc_gru4', 'TANH', False),
('core_encoder.module.gru5' , 'enc_gru5', 'TANH', False),
]
enc_max_rnn_units = max([dump_torch_weights(enc_writer, model.get_submodule(name), export_name, verbose=True, input_sparse=True, quantize=True)
for name, export_name, _ in encoder_gru_layers])
for name, export_name, _, _ in encoder_gru_layers])
encoder_conv_layers = [
('core_encoder.module.conv1.conv' , 'enc_conv1', 'TANH'),
('core_encoder.module.conv2.conv' , 'enc_conv2', 'TANH'),
('core_encoder.module.conv3.conv' , 'enc_conv3', 'TANH'),
('core_encoder.module.conv4.conv' , 'enc_conv4', 'TANH'),
('core_encoder.module.conv5.conv' , 'enc_conv5', 'TANH'),
('core_encoder.module.conv1.conv' , 'enc_conv1', 'TANH', False),
('core_encoder.module.conv2.conv' , 'enc_conv2', 'TANH', False),
('core_encoder.module.conv3.conv' , 'enc_conv3', 'TANH', False),
('core_encoder.module.conv4.conv' , 'enc_conv4', 'TANH', False),
('core_encoder.module.conv5.conv' , 'enc_conv5', 'TANH', False),
]
enc_max_conv_inputs = max([dump_torch_weights(enc_writer, model.get_submodule(name), export_name, verbose=True, quantize=False) for name, export_name, _ in encoder_conv_layers])
enc_max_conv_inputs = max([dump_torch_weights(enc_writer, model.get_submodule(name), export_name, verbose=True, quantize=False) for name, export_name, _, _ in encoder_conv_layers])
del enc_writer
# decoder
decoder_dense_layers = [
('core_decoder.module.dense_1' , 'dec_dense1', 'TANH'),
('core_decoder.module.output' , 'dec_output', 'LINEAR'),
('core_decoder.module.hidden_init' , 'dec_hidden_init', 'TANH'),
('core_decoder.module.gru_init' , 'dec_gru_init', 'TANH'),
('core_decoder.module.dense_1' , 'dec_dense1', 'TANH', False),
('core_decoder.module.output' , 'dec_output', 'LINEAR', False),
('core_decoder.module.hidden_init' , 'dec_hidden_init', 'TANH', False),
('core_decoder.module.gru_init' , 'dec_gru_init', 'TANH', False),
]
for name, export_name, _ in decoder_dense_layers:
for name, export_name, _, _ in decoder_dense_layers:
layer = model.get_submodule(name)
dump_torch_weights(dec_writer, layer, name=export_name, verbose=True)
decoder_gru_layers = [
('core_decoder.module.gru1' , 'dec_gru1', 'TANH'),
('core_decoder.module.gru2' , 'dec_gru2', 'TANH'),
('core_decoder.module.gru3' , 'dec_gru3', 'TANH'),
('core_decoder.module.gru4' , 'dec_gru4', 'TANH'),
('core_decoder.module.gru5' , 'dec_gru5', 'TANH'),
('core_decoder.module.gru1' , 'dec_gru1', 'TANH', False),
('core_decoder.module.gru2' , 'dec_gru2', 'TANH', False),
('core_decoder.module.gru3' , 'dec_gru3', 'TANH', False),
('core_decoder.module.gru4' , 'dec_gru4', 'TANH', False),
('core_decoder.module.gru5' , 'dec_gru5', 'TANH', False),
]
dec_max_rnn_units = max([dump_torch_weights(dec_writer, model.get_submodule(name), export_name, verbose=True, input_sparse=True, quantize=True)
for name, export_name, _ in decoder_gru_layers])
for name, export_name, _, _ in decoder_gru_layers])
decoder_conv_layers = [
('core_decoder.module.conv1.conv' , 'dec_conv1', 'TANH'),
('core_decoder.module.conv2.conv' , 'dec_conv2', 'TANH'),
('core_decoder.module.conv3.conv' , 'dec_conv3', 'TANH'),
('core_decoder.module.conv4.conv' , 'dec_conv4', 'TANH'),
('core_decoder.module.conv5.conv' , 'dec_conv5', 'TANH'),
('core_decoder.module.conv1.conv' , 'dec_conv1', 'TANH', False),
('core_decoder.module.conv2.conv' , 'dec_conv2', 'TANH', False),
('core_decoder.module.conv3.conv' , 'dec_conv3', 'TANH', False),
('core_decoder.module.conv4.conv' , 'dec_conv4', 'TANH', False),
('core_decoder.module.conv5.conv' , 'dec_conv5', 'TANH', False),
]
dec_max_conv_inputs = max([dump_torch_weights(dec_writer, model.get_submodule(name), export_name, verbose=True, quantize=False) for name, export_name, _ in decoder_conv_layers])
dec_max_conv_inputs = max([dump_torch_weights(dec_writer, model.get_submodule(name), export_name, verbose=True, quantize=False) for name, export_name, _, _ in decoder_conv_layers])
del dec_writer
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment