Tensorrt inference for a batch size = 16 fails. Works when batch size = 1

[BipinJG]

I am trying to extract feature vectors from my resnet50 based CNN optimized with TensorRT 7.0.

I am getting correct output when single input is given to the trt model. But when I am giving batch input to the model, then I get correct output only for the first sample of the batch. The remaining outputs are just zeros.

I have also built my trt engine with "builder.max_batch_size = 16" and "EXPLICIT_BATCH = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))"

How do I get the correct outputs for all the samples in the batch?

imgs = np.ones([16,3,256,128])  # batch_size = 16
output shape = [(16,3072)]

trt_logger = trt.Logger(trt.Logger.INFO)
def load_engine(trt_logger):
    TRTbin = 'resnet50_onnx_trt/resnet50mid.model.tar-60.trt'
    with open(TRTbin, 'rb') as f, trt.Runtime(trt_logger) as runtime:
        return runtime.deserialize_cuda_engine(f.read())

engine = load_engine(trt_logger)
context = engine.create_execution_context()

class HostDeviceMem(object):
    #Simple helper data class that's a little nicer to use than a 2-tuple.

    def __init__(self, host_mem, device_mem):
        self.host = host_mem
        self.device = device_mem

    def __str__(self):
        return "Host:\n" + str(self.host) + "\nDevice:\n" + str(self.device)

    def __repr__(self):
        return self.__str__()

def alloc_buf_N(engine):
    """Allocates all host/device in/out buffers required for an engine."""
    inputs = []
    outputs = []
    bindings = []

    stream = cuda.Stream()

    for binding in engine:
        size = trt.volume(engine.get_binding_shape(binding)) * engine.max_batch_size
        # size = 1572864 = 16*3*256*128 for inputs
        # size = 49152 = 16*3072 for outputs

        dtype = trt.nptype(engine.get_binding_dtype(binding))
        # dtype = # <class 'numpy.float32'> for both input and output

        # Allocate host and device buffers
        host_mem = cuda.pagelocked_empty(size, dtype)
        # host_mem = [0. 0. 0. ... 0. 0. 0.], 
        # host_mem.shape) = (1572864,) and (49152,) for inputs and outputs respectively

        device_mem = cuda.mem_alloc(host_mem.nbytes)

        # Append the device buffer to device bindings.
        bindings.append(int(device_mem))
        
        # Append to the appropriate list.
        if engine.binding_is_input(binding):
            inputs.append(HostDeviceMem(host_mem, device_mem))
            # print("inputs alloc_buf_N", inputs)
        else:
            outputs.append(HostDeviceMem(host_mem, device_mem))
            # print("outputs alloc_buf_N", outputs)

    return inputs, outputs, bindings, stream
 
def do_inference_v2(engine, context, inputs, bindings, outputs, stream):
    """
    Inputs and outputs are expected to be lists of HostDeviceMem objects.
    """
    # Transfer input data to the GPU.
    cuda.memcpy_htod_async(inp.device, inp.host, stream) for inp in inputs]

    # Run inference.
    context.execute_async(batch_size=16, bindings=bindings, stream_handle=stream.handle)

    # Transfer predictions back from the GPU.
    [cuda.memcpy_dtoh_async(out.host, out.device, stream) for out in outputs]

    # Synchronize the stream
    stream.synchronize()

    # Return only the host outputs.
    return [out.host for out in outputs]

inputs = imgs.astype(np.float32)
engine = load_engine(trt_logger)
context = engine.create_execution_context()

inputs_alloc_buf, outputs_alloc_buf, bindings_alloc_buf, stream_alloc_buf = alloc_buf_N(engine)

inputs_alloc_buf[0].host = np.ascontiguousarray(inputs)

trt_feature = do_inference_v2(engine, context, inputs_alloc_buf, bindings_alloc_buf, outputs_alloc_buf, stream_alloc_buf)
print("len(trt_feature)",len(trt_feature))
trt_feature = np.asarray(trt_feature)
trt_feature = trt_feature.reshape(16,3072)

print("trt_feature[0][0:15]",trt_feature[0][0:15])
print("trt_feature[1][0:15]",trt_feature[1][0:15])
print("trt_feature.shape",trt_feature.shape)

This gives me output

len(trt_feature) 1

trt_feature[0][0:15] [ 0.  0.  0.  23.91  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0. ]
trt_feature[1][0:15] [ 0.  0.  0.  0.     0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0. ]

The output "trt_feature[0][0:15] [ 0. 0. 0. 23.91 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. ]" is correct. However, it does not work for remaining samples in the batch.
The output was supposed to be a list of length 16 (each of dimension 3072), i.e. "len(trt_feature)" = 16. But I got the length of the output equal to 1, i.e. "len(trt_feature)" = 1.