vai_q_tensorflow Usage - 3.0 English

Vitis AI User Guide (UG1414)
Document ID
UG1414
Release Date
2023-02-24
Version
3.0 English

The options supported by vai_q_tensorflow are shown in the following tables.

Table 1. vai_q_tensorflow Options
Name Type Description
Common Configuration
--input_frozen_graph String TensorFlow frozen inference GraphDef file for the floating-point model. It is used for quantize calibration.
--input_nodes String Specifies the name list of input nodes of the quantize graph, used together with –output_nodes, comma separated. Input nodes and output nodes are the start and end points of quantization. The subgraph between them is quantized if it is quantizable.
Recommended: Set –input_nodes as the last nodes for pre-processing and –output_nodes as the last nodes for post-processing because some of the operations required for pre- and post-processing are not quantizable and might cause errors when compiled by the Vitis AI compiler if you need to deploy the quantized model to the DPU. The input nodes might not be the same as the placeholder nodes of the graph.
--output_nodes String Specifies the name list of output nodes of the quantize graph, used together with –input_nodes, comma separated. Input nodes and output nodes are the start and end points of quantization. The subgraph between them is quantized if it is quantizable.
Recommended: Set –input_nodes as the last nodes for pre-processing and –output_nodes as the last nodes for post-processing because some of the operations required for pre- and post-processing are not quantizable and might cause errors when compiled by the Vitis AI compiler if you need to deploy the quantized model to the DPU.
--input_shapes String Specifies the shape list of input nodes. Must be a 4-dimension shape for each node, comma separated, for example 1,224,224,3; support unknown size for batch_size, for example ?,224,224,3. In case of multiple input nodes, assign the shape list of each node separated by :, for example, ?,224,224,3:?,300,300,1.
--input_fn String Provides the input data for the graph used with the calibration dataset. The function format is module_name.input_fn_name (for example, my_input_fn.input_fn). The -input_fn should take an int object as input which indicates the calibration step, and should return a dict`(placeholder_node_name, numpy.Array)` object for each call, which is then fed into the placeholder operations of the model.

For example, assign –input_fn to my_input_fn.calib_input, and write calib_input function in my_input_fn.py as:

def calib_input_fn:
# read image and do some preprocessing
    return {“placeholder_1”: input_1_nparray, “placeholder_2”: input_2_nparray}
Note: You do not need to do in-graph pre-processing again in input_fn because the subgraph before –input_nodes remains during quantization. Remove the pre-defined input functions (including default and random) because they are not commonly used. The pre-processing part which is not in the graph file should be handled in the input_fn.
Quantize Configuration
--weight_bit Int32 Specifies the bit width for quantized weight and bias.

Default: 8
--activation_bit Int32 Specifies the bit width for quantized activation.

Default: 8
--nodes_bit String Specifies the bit width of nodes. Node names and bit widths form a pair of parameters joined by a colon; the parameters are comma separated. When specifying the conv op name, only vai_q_tensorflow will quantize the weights of conv op using the specified bit width. For example, 'conv1/Relu:16,conv1/weights:8,conv1:16'.
--method Int32 Specifies the method for quantization.
  • 0: Non-overflow method in which no values are saturated during quantization. Sensitive to outliers.
  • 1: Min-diffs method that allows saturation for quantization to get a lower quantization difference. Higher tolerance to outliers. Usually ends with narrower ranges than the non-overflow method.
  • 2: Min-diffs method with strategy for depthwise. It allows saturation for large values during quantization to get smaller quantization errors. A special strategy is applied for depthwise weights. It is slower than method 0 but has higher endurance to outliers.

Default value: 1
--nodes_method String Specifies the method of nodes. Node names and method form a pair of parameters joined by a colon; the parameter pairs are comma separated. When specifying the conv op name, only vai_q_tensorflow will quantize weights of conv op using the specified method, for example, 'conv1/Relu:1,depthwise_conv1/weights:2,conv1:1'.
--calib_iter Int32 Specifies the iterations of calibration. Total number of images for calibration = calib_iter * batch_size.

Default value: 100
--ignore_nodes String Specifies the name list of nodes to be ignored during quantization. Ignored nodes are left unquantized during quantization.
--skip_check Int32 If set to 1, the check for float model is skipped. Useful when only part of the input model is quantized.

Range: [0, 1]

Default value: 0
--align_concat Int32 Specifies the strategy for the alignment of the input quantizeposition for concat nodes.
  • 0: Aligns all the concat nodes
  • 1: Aligns the output concat nodes
  • 2: Disables alignment

Default value: 0
--align_pool Int32 Specifies the strategy for the alignment of the input quantize position for maxpool/avgpool nodes.
  • 0: Aligns all the maxpool/avgpool nodes
  • 1: Aligns the output maxpool/avgpool nodes
  • 2: Disables alignment

Default value: 0
--simulate_dpu Int32 Set to 1 to enable the simulation of the DPU. The behavior of DPU for some operations is different from TensorFlow. For example, the dividing in LeakyRelu and AvgPooling are replaced by bit-shifting, so there might be a slight difference between DPU outputs and CPU/GPU outputs. The vai_q_tensorflow quantizer simulates the behavior of these operations if this flag is set to 1.

Range: [0, 1]

Default value: 1
--adjust_shift_bias Int32 Specifies the strategy for shift bias check and adjustment for DPU compiler.
  • 0: Disables shift bias check and adjustment
  • 1: Enables with static constraints
  • 2: Enables with dynamic constraints

Default value: 1
--adjust_shift_cut Int32 Specifies the strategy for shift cut check and adjustment for DPU compiler.
  • 0: Disables shift cut check and adjustment
  • 1: Enables with static constraints

Default value: 1
--arch_type String Specifies the arch type for fix neuron. 'DEFAULT' means quantization range of both weights and activations are [-128, 127]. 'DPUCADF8H' means weights quantization range is [-128, 127] while activation is [-127, 127]
--output_dir String Specifies the directory in which to save the quantization results.

Default value: “./quantize_results”
--max_dump_batches Int32 Specifies the maximum number of batches for dumping.

Default value: 1
--dump_float Int32 If set to 1, the float weights and activations are dumped.

Range: [0, 1]

Default value: 0
--dump_input_tensors String Specifies the input tensor name of Graph when graph entrance is not a placeholder. Add a placeholder to the dump_input_tensor, so that input_fn can feed data.
--scale_all_avgpool Int32 Set to 1 to enable scale output of AvgPooling op to simulate DPU. Only kernel_size <= 64 will be scaled. This operation does not affect the special case such as kernel_size=3,5,6,7,14

Default value: 1
--do_cle Int32
  • 1: Enables implement cross layer equalization to adjust the weights distribution
  • 0: Skips cross layer equalization operation

Default value: 0
--replace_relu6 Int32 Available only for do_cle=1
  • 1: Allows you to ReLU6 with ReLU
  • 0: Skips replacement.

Default value: 1
--replace_sigmoid Int32
  • 1: Enable replace sigmoid with hard-sigmoid
  • 0: Skips replacement.

Default value: 0
--replace_softmax Int32
  • 1: Enable replace softmax with hard-softmax
  • 0: Skips replacement.

Default value: 0
--convert_datatype Int32
  • 4: Do fold bn and convert to data type fp32
  • 3: Do fold bn and convert to data type bfloat16
  • 2: Do fold bn and convert to data type double
  • 1: Do fold bn and convert to data type fp16
  • 0: Skips conversion.

Default value: 0
--output_format String Indicates what format to save the quantized model, 'pb' for saving tensorflow frozen pb, 'onnx' for saving onnx model.

Default value: 'pb'
Session Configurations
--gpu String Specifies the IDs of the GPU device used for quantization separated by commas.
--gpu_memory_fraction Float Specifies the GPU memory fraction used for quantization, between 0-1.

Default value: 0.5
Others
--help Shows all available options of vai_q_tensorflow.
--version Shows the version information for vai_q_tensorflow .

Examples

show help: vai_q_tensorflow --help
quantize: 
vai_q_tensorflow quantize --input_frozen_graph frozen_graph.pb \
                          --input_nodes inputs \
                          --output_nodes predictions \
                          --input_shapes ?,224,224,3 \
                          --input_fn my_input_fn.calib_input
dump quantized model: 
vai_q_tensorflow dump --input_frozen_graph quantize_results/quantize_eval_model.pb \
                      --input_fn my_input_fn.dump_input

Refer to Xilinx Model Zoo for more TensorFlow model quantization examples.