yolov5/models/experimental.py
Jiacong Fang 808bcad3bb
Add TensorFlow and TFLite export (#1127)
* Add models/tf.py for TensorFlow and TFLite export

* Set auto=False for int8 calibration

* Update requirements.txt for TensorFlow and TFLite export

* Read anchors directly from PyTorch weights

* Add --tf-nms to append NMS in TensorFlow SavedModel and GraphDef export

* Remove check_anchor_order, check_file, set_logging from import

* Reformat code and optimize imports

* Autodownload model and check cfg

* update --source path, img-size to 320, single output

* Adjust representative_dataset

* Put representative dataset in tfl_int8 block

* detect.py TF inference

* weights to string

* weights to string

* cleanup tf.py

* Add --dynamic-batch-size

* Add xywh normalization to reduce calibration error

* Update requirements.txt

TensorFlow 2.3.1 -> 2.4.0 to avoid int8 quantization error

* Fix imports

Move C3 from models.experimental to models.common

* Add models/tf.py for TensorFlow and TFLite export

* Set auto=False for int8 calibration

* Update requirements.txt for TensorFlow and TFLite export

* Read anchors directly from PyTorch weights

* Add --tf-nms to append NMS in TensorFlow SavedModel and GraphDef export

* Remove check_anchor_order, check_file, set_logging from import

* Reformat code and optimize imports

* Autodownload model and check cfg

* update --source path, img-size to 320, single output

* Adjust representative_dataset

* detect.py TF inference

* Put representative dataset in tfl_int8 block

* weights to string

* weights to string

* cleanup tf.py

* Add --dynamic-batch-size

* Add xywh normalization to reduce calibration error

* Update requirements.txt

TensorFlow 2.3.1 -> 2.4.0 to avoid int8 quantization error

* Fix imports

Move C3 from models.experimental to models.common

* implement C3() and SiLU()

* Fix reshape dim to support dynamic batching

* Add epsilon argument in tf_BN, which is different between TF and PT

* Set stride to None if not using PyTorch, and do not warmup without PyTorch

* Add list support in check_img_size()

* Add list input support in detect.py

* sys.path.append('./') to run from yolov5/

* Add int8 quantization support for TensorFlow 2.5

* Add get_coco128.sh

* Remove --no-tfl-detect in models/tf.py (Use tf-android-tfl-detect branch for EdgeTPU)

* Update requirements.txt

* Replace torch.load() with attempt_load()

* Update requirements.txt

* Add --tf-raw-resize to set half_pixel_centers=False

* Add --agnostic-nms for TF class-agnostic NMS

* Cleanup after merge

* Cleanup2 after merge

* Cleanup3 after merge

* Add tf.py docstring with credit and usage

* pb saved_model and tflite use only one model in detect.py

* Add use cases in docstring of tf.py

* Remove redundant `stride` definition

* Remove keras direct import

* Fix `check_requirements(('tensorflow>=2.4.1',))`

Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>
2021-08-17 13:18:16 +02:00

116 lines
4.2 KiB
Python

# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
"""
Experimental modules
"""
import numpy as np
import torch
import torch.nn as nn
from models.common import Conv
from utils.downloads import attempt_download
class CrossConv(nn.Module):
# Cross Convolution Downsample
def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False):
# ch_in, ch_out, kernel, stride, groups, expansion, shortcut
super().__init__()
c_ = int(c2 * e) # hidden channels
self.cv1 = Conv(c1, c_, (1, k), (1, s))
self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g)
self.add = shortcut and c1 == c2
def forward(self, x):
return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
class Sum(nn.Module):
# Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070
def __init__(self, n, weight=False): # n: number of inputs
super().__init__()
self.weight = weight # apply weights boolean
self.iter = range(n - 1) # iter object
if weight:
self.w = nn.Parameter(-torch.arange(1., n) / 2, requires_grad=True) # layer weights
def forward(self, x):
y = x[0] # no weight
if self.weight:
w = torch.sigmoid(self.w) * 2
for i in self.iter:
y = y + x[i + 1] * w[i]
else:
for i in self.iter:
y = y + x[i + 1]
return y
class MixConv2d(nn.Module):
# Mixed Depth-wise Conv https://arxiv.org/abs/1907.09595
def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True):
super().__init__()
groups = len(k)
if equal_ch: # equal c_ per group
i = torch.linspace(0, groups - 1E-6, c2).floor() # c2 indices
c_ = [(i == g).sum() for g in range(groups)] # intermediate channels
else: # equal weight.numel() per group
b = [c2] + [0] * groups
a = np.eye(groups + 1, groups, k=-1)
a -= np.roll(a, 1, axis=1)
a *= np.array(k) ** 2
a[0] = 1
c_ = np.linalg.lstsq(a, b, rcond=None)[0].round() # solve for equal weight indices, ax = b
self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)])
self.bn = nn.BatchNorm2d(c2)
self.act = nn.LeakyReLU(0.1, inplace=True)
def forward(self, x):
return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1)))
class Ensemble(nn.ModuleList):
# Ensemble of models
def __init__(self):
super().__init__()
def forward(self, x, augment=False, profile=False, visualize=False):
y = []
for module in self:
y.append(module(x, augment, profile, visualize)[0])
# y = torch.stack(y).max(0)[0] # max ensemble
# y = torch.stack(y).mean(0) # mean ensemble
y = torch.cat(y, 1) # nms ensemble
return y, None # inference, train output
def attempt_load(weights, map_location=None, inplace=True, fuse=True):
from models.yolo import Detect, Model
# Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a
model = Ensemble()
for w in weights if isinstance(weights, list) else [weights]:
ckpt = torch.load(attempt_download(w), map_location=map_location) # load
if fuse:
model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().fuse().eval()) # FP32 model
else:
model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().eval()) # without layer fuse
# Compatibility updates
for m in model.modules():
if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU, Detect, Model]:
m.inplace = inplace # pytorch 1.7.0 compatibility
elif type(m) is Conv:
m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatibility
if len(model) == 1:
return model[-1] # return model
else:
print(f'Ensemble created with {weights}\n')
for k in ['names']:
setattr(model, k, getattr(model[-1], k))
model.stride = model[torch.argmax(torch.tensor([m.stride.max() for m in model])).int()].stride # max stride
return model # return ensemble