Retinanet with mobilenet backend

[siriusmehta]

@srihari-humbarwadi

Can we adapt to different backend for RetinaNet such as mobilenet ?

[srihari-humbarwadi]

@siriusmehta Yeah you can added any backbone you wish, only thing that you need to take care is the scales of the feature map that go into the FPN

[siriusmehta]

Thanks @srihari-humbarwadi

I tried replacing Resent50 backbone with below:

default MobilenetV2 pertained downloaded is 224x224

def get_backbone(): """Builds ResNet50 with pre-trained imagenet weights""" backbone = keras.applications.MobileNetV2( include_top=False, input_shape=[None, None, 3] ) c3_output, c4_output, c5_output = [ backbone.get_layer(layer_name).output for layer_name in ["block_12_add", "block_14_add", "out_relu"] ] return keras.Model( inputs=[backbone.inputs], outputs=[c3_output, c4_output, c5_output] )

and getting below error:

---

InvalidArgumentError Traceback (most recent call last) in () 17 epochs=epochs, 18 callbacks=callbacks_list, ---> 19 verbose=1, 20 )

8 frames /usr/local/lib/python3.6/dist-packages/tensorflow/python/eager/execute.py in quick_execute(op_name, num_outputs, inputs, attrs, ctx, name) 58 ctx.ensure_initialized() 59 tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name, ---> 60 inputs, attrs, num_outputs) 61 except core._NotOkStatusException as e: 62 if name is not None:

InvalidArgumentError: Incompatible shapes: [2,24,32,256] vs. [2,48,64,256] [[node gradient_tape/RetinaNet/FeaturePyramid/add/BroadcastGradientArgs (defined at :19) ]] [Op:__inference_train_function_27062]

Function call stack: train_function

[srihari-humbarwadi]

I didn't run the code with your changes. But by looking at the error message, its likely that you are feeding in the wrong feature layers from the backbone. IIRC C3, C4 and C5 should have strides /8, /16 and /32. Just pick the layers accordingly

[siriusmehta]

Hi @srihari-humbarwadi

Thanks for your reply.

I am bit new and in exploring and learning stage .

Can you please point me to some useful links where I can read and understand more about

IIRC C3, C4 and C5 should have strides /8, /16 and /32

How the layers strides calculates to /8, /16 and /32 ?

Would appreciate your help!

Thanks

[vuducnghia]

Hi @siriusmehta, bellow is Mobilenetv2: base_model = MobileNetV2(include_top=False, input_shape=INPUT_SHAPE) base_model.trainable = training c3_output, c4_output, c5_output = [ base_model.get_layer(layer_name).output for layer_name in ["block_6_expand_relu", "block_13_expand_relu", "out_relu"] ]

[vitorbds]

Is it possible use imaginet weight ?