fastmachinelearning
Pytorch converter doesn't work.
I trained VGG16 model on CIFAR100 dataset on pytorch. When I run:
import hls4ml
import plotting
config = hls4ml.utils.config_from_pytorch_model(model, granularity='layer')
print("-----------------------------------")
print("Configuration")
plotting.print_dict(config)
print("-----------------------------------")
hls_model = hls4ml.converters.convert_from_pytorch_model(
model, hls_config=config, output_dir='model_3/hls4ml_prj', part='xcu250-figd2104-2L-e'
)
I get the error on the last line:
TypeError: cannot unpack non-iterable NoneType object
While I ran pre-trained VGG16 model of keras on hls4ml, it runs smoothly without any error. The cause of the error I found out is the config file generated from config = hls4ml.utils.config_from_pytorch_model(model, granularity='layer'). When I print this variable config, it shows:
{'Model': {'Precision': 'ap_fixed<16,6>', 'ReuseFactor': 1, 'Strategy': 'Latency'}}
which shows there is no information regarding the layers. In case of Keras i.e. config = hls4ml.utils.config_from_keras_model(model, granularity='layer') generates following output:
Interpreting Model
Topology:
Layer name: input_1, layer type: InputLayer, input shapes: [[None, 224, 224, 3]], output shape: [None, 224, 224, 3]
Layer name: block1_conv1, layer type: Conv2D, input shapes: [[None, 224, 224, 3]], output shape: [None, 224, 224, 64]
Layer name: block1_conv2, layer type: Conv2D, input shapes: [[None, 224, 224, 64]], output shape: [None, 224, 224, 64]
Layer name: block1_pool, layer type: MaxPooling2D, input shapes: [[None, 224, 224, 64]], output shape: [None, 112, 112, 64]
Layer name: block2_conv1, layer type: Conv2D, input shapes: [[None, 112, 112, 64]], output shape: [None, 112, 112, 128]
Layer name: block2_conv2, layer type: Conv2D, input shapes: [[None, 112, 112, 128]], output shape: [None, 112, 112, 128]
Layer name: block2_pool, layer type: MaxPooling2D, input shapes: [[None, 112, 112, 128]], output shape: [None, 56, 56, 128]
Layer name: block3_conv1, layer type: Conv2D, input shapes: [[None, 56, 56, 128]], output shape: [None, 56, 56, 256]
Layer name: block3_conv2, layer type: Conv2D, input shapes: [[None, 56, 56, 256]], output shape: [None, 56, 56, 256]
Layer name: block3_conv3, layer type: Conv2D, input shapes: [[None, 56, 56, 256]], output shape: [None, 56, 56, 256]
Layer name: block3_pool, layer type: MaxPooling2D, input shapes: [[None, 56, 56, 256]], output shape: [None, 28, 28, 256]
Layer name: block4_conv1, layer type: Conv2D, input shapes: [[None, 28, 28, 256]], output shape: [None, 28, 28, 512]
Layer name: block4_conv2, layer type: Conv2D, input shapes: [[None, 28, 28, 512]], output shape: [None, 28, 28, 512]
Layer name: block4_conv3, layer type: Conv2D, input shapes: [[None, 28, 28, 512]], output shape: [None, 28, 28, 512]
Layer name: block4_pool, layer type: MaxPooling2D, input shapes: [[None, 28, 28, 512]], output shape: [None, 14, 14, 512]
Layer name: block5_conv1, layer type: Conv2D, input shapes: [[None, 14, 14, 512]], output shape: [None, 14, 14, 512]
Layer name: block5_conv2, layer type: Conv2D, input shapes: [[None, 14, 14, 512]], output shape: [None, 14, 14, 512]
Layer name: block5_conv3, layer type: Conv2D, input shapes: [[None, 14, 14, 512]], output shape: [None, 14, 14, 512]
Layer name: block5_pool, layer type: MaxPooling2D, input shapes: [[None, 14, 14, 512]], output shape: [None, 7, 7, 512]
Layer name: flatten, layer type: Reshape, input shapes: [[None, 7, 7, 512]], output shape: [None, 25088]
Layer name: fc1, layer type: Dense, input shapes: [[None, 25088]], output shape: [None, 4096]
Layer name: fc2, layer type: Dense, input shapes: [[None, 4096]], output shape: [None, 4096]
Layer name: predictions, layer type: Dense, input shapes: [[None, 4096]], output shape: [None, 1000]
{'Model': {'Precision': 'fixed<16,6>', 'ReuseFactor': 1, 'Strategy': 'Latency', 'BramFactor': 1000000000, 'TraceOutput': False}}
Please resolve this issue.
No, I am still waiting for their reply. It seems their Pytorch converter has some bug. I temporarily switched to keras, which works fine.
No, I am still waiting for their reply. It seems their Pytorch converter has some bug. I temporarily switched to keras, which works fine.
Hi,did you use Vitis HLS or the Vivado HLS? Please reply.
Can you please help me how you run the code on Vitis HLS. The github repo is not working for Vitis hls