|
| 1 | +# Copyright (c) Qualcomm Innovation Center, Inc. |
| 2 | +# All rights reserved |
| 3 | +# |
| 4 | +# This source code is licensed under the BSD-style license found in the |
| 5 | +# LICENSE file in the root directory of this source tree. |
| 6 | + |
| 7 | +import json |
| 8 | +import logging |
| 9 | +import os |
| 10 | +import types |
| 11 | +from multiprocessing.connection import Client |
| 12 | + |
| 13 | +import numpy as np |
| 14 | + |
| 15 | +import torch |
| 16 | +from executorch.backends.qualcomm.quantizer.quantizer import QuantDtype |
| 17 | +from executorch.examples.qualcomm.utils import ( |
| 18 | + build_executorch_binary, |
| 19 | + get_imagenet_dataset, |
| 20 | + make_output_dir, |
| 21 | + parse_skip_delegation_node, |
| 22 | + setup_common_args_and_variables, |
| 23 | + SimpleADB, |
| 24 | + topk_accuracy, |
| 25 | +) |
| 26 | +from transformers import AutoModelForImageClassification |
| 27 | +from transformers.models.cvt.modeling_cvt import CvtSelfAttention |
| 28 | + |
| 29 | + |
| 30 | +# Copy from transformers/models/cvt/modeling_cvt.py in transformers 4.47.1 |
| 31 | +def attention_forward_without_einsum(self, hidden_state, height, width): |
| 32 | + if self.with_cls_token: |
| 33 | + cls_token, hidden_state = torch.split(hidden_state, [1, height * width], 1) |
| 34 | + batch_size, hidden_size, num_channels = hidden_state.shape |
| 35 | + # rearrange "b (h w) c -> b c h w" |
| 36 | + hidden_state = hidden_state.permute(0, 2, 1).view( |
| 37 | + batch_size, num_channels, height, width |
| 38 | + ) |
| 39 | + |
| 40 | + key = self.convolution_projection_key(hidden_state) |
| 41 | + query = self.convolution_projection_query(hidden_state) |
| 42 | + value = self.convolution_projection_value(hidden_state) |
| 43 | + |
| 44 | + if self.with_cls_token: |
| 45 | + query = torch.cat((cls_token, query), dim=1) |
| 46 | + key = torch.cat((cls_token, key), dim=1) |
| 47 | + value = torch.cat((cls_token, value), dim=1) |
| 48 | + |
| 49 | + head_dim = self.embed_dim // self.num_heads |
| 50 | + |
| 51 | + query = self.rearrange_for_multi_head_attention(self.projection_query(query)) |
| 52 | + key = self.rearrange_for_multi_head_attention(self.projection_key(key)) |
| 53 | + value = self.rearrange_for_multi_head_attention(self.projection_value(value)) |
| 54 | + # ====================Qualcomm Changed================================= |
| 55 | + attention_score = query @ key.transpose(-1, -2) |
| 56 | + attention_score = attention_score * self.scale |
| 57 | + # attention_score = torch.einsum("bhlk,bhtk->bhlt", [query, key]) * self.scale |
| 58 | + # ===================================================================== |
| 59 | + attention_probs = torch.nn.functional.softmax(attention_score, dim=-1) |
| 60 | + attention_probs = self.dropout(attention_probs) |
| 61 | + # ====================Qualcomm Changed================================= |
| 62 | + context = attention_probs @ value |
| 63 | + # context = torch.einsum("bhlt,bhtv->bhlv", [attention_probs, value]) |
| 64 | + # ===================================================================== |
| 65 | + # rearrange"b h t d -> b t (h d)" |
| 66 | + _, _, hidden_size, _ = context.shape |
| 67 | + context = ( |
| 68 | + context.permute(0, 2, 1, 3) |
| 69 | + .contiguous() |
| 70 | + .view(batch_size, hidden_size, self.num_heads * head_dim) |
| 71 | + ) |
| 72 | + return context |
| 73 | + |
| 74 | + |
| 75 | +def _replace_attention( |
| 76 | + module: torch.nn.Module, |
| 77 | +): |
| 78 | + for _, child in module.named_children(): |
| 79 | + if isinstance(child, CvtSelfAttention): |
| 80 | + child.forward = types.MethodType( # pyre-ignore |
| 81 | + attention_forward_without_einsum, child |
| 82 | + ) |
| 83 | + else: |
| 84 | + _replace_attention(child) |
| 85 | + return module |
| 86 | + |
| 87 | + |
| 88 | +def main(args): |
| 89 | + skip_node_id_set, skip_node_op_set = parse_skip_delegation_node(args) |
| 90 | + |
| 91 | + # ensure the working directory exist. |
| 92 | + os.makedirs(args.artifact, exist_ok=True) |
| 93 | + |
| 94 | + if not args.compile_only and args.device is None: |
| 95 | + raise RuntimeError( |
| 96 | + "device serial is required if not compile only. " |
| 97 | + "Please specify a device serial by -s/--device argument." |
| 98 | + ) |
| 99 | + |
| 100 | + data_num = 100 |
| 101 | + if args.ci: |
| 102 | + inputs = [(torch.rand(1, 3, 224, 224),)] |
| 103 | + logging.warning( |
| 104 | + "This option is for CI to verify the export flow. It uses random input and will result in poor accuracy." |
| 105 | + ) |
| 106 | + else: |
| 107 | + inputs, targets, input_list = get_imagenet_dataset( |
| 108 | + dataset_path=f"{args.dataset}", |
| 109 | + data_size=data_num, |
| 110 | + image_shape=(256, 256), |
| 111 | + crop_size=224, |
| 112 | + ) |
| 113 | + |
| 114 | + module = ( |
| 115 | + AutoModelForImageClassification.from_pretrained("microsoft/cvt-13") |
| 116 | + .eval() |
| 117 | + .to("cpu") |
| 118 | + ) |
| 119 | + # Fix prepare failed due to einsum |
| 120 | + module = _replace_attention(module) |
| 121 | + pte_filename = "cvt_qnn_q8" |
| 122 | + build_executorch_binary( |
| 123 | + module.eval(), |
| 124 | + inputs[0], |
| 125 | + args.model, |
| 126 | + f"{args.artifact}/{pte_filename}", |
| 127 | + inputs, |
| 128 | + skip_node_id_set=skip_node_id_set, |
| 129 | + skip_node_op_set=skip_node_op_set, |
| 130 | + quant_dtype=QuantDtype.use_8a8w, |
| 131 | + shared_buffer=args.shared_buffer, |
| 132 | + ) |
| 133 | + |
| 134 | + if args.compile_only: |
| 135 | + return |
| 136 | + |
| 137 | + adb = SimpleADB( |
| 138 | + qnn_sdk=os.getenv("QNN_SDK_ROOT"), |
| 139 | + build_path=f"{args.build_folder}", |
| 140 | + pte_path=f"{args.artifact}/{pte_filename}.pte", |
| 141 | + workspace=f"/data/local/tmp/executorch/{pte_filename}", |
| 142 | + device_id=args.device, |
| 143 | + host_id=args.host, |
| 144 | + soc_model=args.model, |
| 145 | + shared_buffer=args.shared_buffer, |
| 146 | + ) |
| 147 | + adb.push(inputs=inputs, input_list=input_list) |
| 148 | + adb.execute() |
| 149 | + |
| 150 | + # collect output data |
| 151 | + output_data_folder = f"{args.artifact}/outputs" |
| 152 | + make_output_dir(output_data_folder) |
| 153 | + |
| 154 | + adb.pull(output_path=args.artifact) |
| 155 | + |
| 156 | + # top-k analysis |
| 157 | + predictions = [] |
| 158 | + for i in range(data_num): |
| 159 | + predictions.append( |
| 160 | + np.fromfile( |
| 161 | + os.path.join(output_data_folder, f"output_{i}_0.raw"), dtype=np.float32 |
| 162 | + ) |
| 163 | + ) |
| 164 | + |
| 165 | + k_val = [1, 5] |
| 166 | + topk = [topk_accuracy(predictions, targets, k).item() for k in k_val] |
| 167 | + if args.ip and args.port != -1: |
| 168 | + with Client((args.ip, args.port)) as conn: |
| 169 | + conn.send(json.dumps({f"top_{k}": topk[i] for i, k in enumerate(k_val)})) |
| 170 | + else: |
| 171 | + for i, k in enumerate(k_val): |
| 172 | + print(f"top_{k}->{topk[i]}%") |
| 173 | + |
| 174 | + |
| 175 | +if __name__ == "__main__": |
| 176 | + parser = setup_common_args_and_variables() |
| 177 | + |
| 178 | + parser.add_argument( |
| 179 | + "-d", |
| 180 | + "--dataset", |
| 181 | + help=( |
| 182 | + "path to the validation folder of ImageNet dataset. " |
| 183 | + "e.g. --dataset imagenet-mini/val " |
| 184 | + "for https://www.kaggle.com/datasets/ifigotin/imagenetmini-1000)" |
| 185 | + ), |
| 186 | + type=str, |
| 187 | + required=False, |
| 188 | + ) |
| 189 | + |
| 190 | + parser.add_argument( |
| 191 | + "-a", |
| 192 | + "--artifact", |
| 193 | + help="path for storing generated artifacts by this example. " "Default ./cvt", |
| 194 | + default="./cvt", |
| 195 | + type=str, |
| 196 | + ) |
| 197 | + |
| 198 | + args = parser.parse_args() |
| 199 | + try: |
| 200 | + main(args) |
| 201 | + except Exception as e: |
| 202 | + if args.ip and args.port != -1: |
| 203 | + with Client((args.ip, args.port)) as conn: |
| 204 | + conn.send(json.dumps({"Error": str(e)})) |
| 205 | + else: |
| 206 | + raise Exception(e) |
0 commit comments